The Inference Report

Sensorized human demonstrations offer scalable data for mobile manipulation, but embodiment differences complicate policy learning. We present Pixel-UMI, a pipeline for learning environment-aware whole-body policies from human data. We first introduce Environment-Aware Retargeting (EAR), which combines reconstructed scene geometry with demonstrated head and gripper trajectories to determine robot posture and generate collision-aware joint and base targets. Delay-calibrated action assignment converts these trajectories into actions. However, kinematic differences can leave residual pose errors, creating viewpoint mismatches between human demonstrations and robot observations. We therefore introduce Retargeting-Error-Aware Masking (REAM), which uses these errors during training to reduce reliance on mismatched views at deployment. Finally, because estimated actions cannot fully capture contact-rich interactions, we collect on-policy human corrections through DAgger to improve performance. We evaluate Pixel-UMI on five challenging tasks spanning articulated-object manipulation, deformable-object handling, constrained-space reaching, whole-body collision avoidance, and contact-rich bimanual coordination, achieving 87.0% average task progress. Video results are available at https://pixel-robot.com/pixel-umi

General-purpose mobile manipulation requires a combination of capable robot hardware, scalable demonstration collection, expressive whole-body control, and effective policy learning. Yet these components are often developed independently, creating mismatches that limit the capabilities of the complete system. We present Pixel, a co-designed full-stack system for learning general-purpose mobile manipulation. Pixel combines a low-cost wheeled humanoid with human-aligned kinematics and compliant grippers, multiple teleoperation and robot-free human data-collection interfaces, a unified whole-body controller, and a common policy learning pipeline across human and robot demonstrations. A hierarchical learning stack further enables language-conditioned visuomotor control and long-horizon task execution. We evaluate the system across hardware, data collection, and policy learning, and demonstrate mobile-manipulation behaviors requiring dexterity, active perception, navigation, and whole-body coordination. We hope that Pixel provides an integrated and extensible foundation for studying how robot hardware, data, control, and learning can be designed together for robot learning. More results at https://pixel-robot.com/pixel

Chain-of-Thought (CoT) generation has substantially improved the performance of Large Language Models (LLMs) on complex reasoning tasks, including code generation, data analysis, and exam-style question answering. Despite these advances, purely neural LLMs continue to struggle with elementary logical reasoning problems and lack the determinism, soundness, and reliability characteristic of symbolic reasoning systems. Conversely, classical symbolic methods such as SAT solving and exact cover guarantee correctness and completeness, but require problems to be expressed in highly specialized formal encodings, limiting their applicability to natural language inputs. In this work, we present a synergistic neuro-symbolic framework that bridges this gap by combining neural semantic parsing with deterministic exact cover reasoning. Given a Zebra-style logic puzzle in natural language, an LLM translates entities and constraints into structured tool calls that construct an exact cover instance. A symbolic solver then computes the solution and provides feedback to iteratively refine invalid, under- or overconstrained encodings, exploiting the uniqueness property of Zebra puzzles. We conduct a comprehensive empirical evaluation across multiple parameter scales, post-training paradigms, and LLM families. The results demonstrate that our hybrid approach consistently outperforms strong plain neural baselines, including CoT prompting, as well as recent neuro-symbolic methods.

GUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity. We propose WebHarbor, a suite of 153 executable and dockerized website environments that faithfully reconstruct the visual interfaces, content, and functional workflows of real-world web services under deterministic control. We build WebHarbor through task-driven human-agent collaboration, where environments evolve as new tasks expand their functional scope, coding agents implement the required changes, and human contributors audit visual fidelity, functional correctness, and task validity. The resulting suite comprises 3,518 tasks spanning multi-step and post-authentication workflows, where existing works often fail to cover. Beyond reliable evaluation, WebHarbor provides high-fidelity synthetic web environments for cost-efficient reinforcement learning (RL). We demonstrate that a Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively with leading open-source GUI web agents across four benchmarks. The code and data are publicly available at https://github.com/aiming-lab/WebHarbor.

Muon has recently emerged as a promising optimizer for pretraining large language models, offering a speedup over AdamW in minimizing pretraining loss. However, it remains unclear whether Muon's advantage persists after fine-tuning. In this paper, we study the downstream behavior of Muon and AdamW and uncover a surprising trend: beyond a sufficiently large pretraining data budget, AdamW outperforms Muon after fine-tuning. Moreover, as model size increases, AdamW overtakes Muon at progressively smaller pretraining token-parameter ratios. To explain this reversal, we view fine-tuning as a perturbation of the outputs and introduce the margin-to-spread ratio, which measures how far the correct logit lies above the wrong ones relative to how widely the wrong logits are spread. In language models, Muon has the larger margin-to-spread ratio at small data budgets and AdamW at large ones, mirroring the fine-tuning trend. In a linear associative memory, we prove that after one update the margin-to-spread ratio of gradient descent gains on that of Muon as the data budget grows, and simulations show that this trend persists after extended training.

Regulatory compliance has become an essential software requirement Legal demands for compliance by design CbD call for requirements engineering RE approaches that derive requirements and support compliance implementation in subsequent software engineering SE processes such as software design The complexity of the legal knowledge demanded in regulatory RE makes it important to involve legal experts and ensure the relevant specifications are valid and usable for them Regulatory requirements must support compliance implementation in software design However there is a lack of approaches bridging legal and SE viewpoints and interaction challenges between the two persist To address engineeringlegal coordination in specification of regulatory requirements we propose the Artifact Model for Regulatory Requirements Engineering AM4RRE operationalized through templates AM4RRE was positively evaluated by most of the 37 practitioners for coordination in regulatory RE and for supporting requirements quality Results from a quasiexperiment with 34 of them in 10 groups indicated that templateguided coordination was associated with on average shorter total specification time when executed first and higher requirements quality in most groups in which it varied Existing RE research and practice provide little attention and support for viewpoint coordination in regulatory RE for CbD Artifactoriented approaches such as AM4RRE have the potential to guide coordination in RE for better RE results and contribute to concretizing the role and contribution of RE to CbD

Reinforcement learning with verifiable rewards (RLVR) can enable generalizable improvement for large language model (LLM) reasoning. This paper studies how to incentivize such generalization via reasoning strategies. We study this question with a controllable equation-solving framework: finding multiple inputs x for equations f(x)=y, where f is a non-bijective string function. This setting makes reasoning strategies observable while supporting a clear definition of easy-to-hard out-of-distribution (OOD) shifts, e.g., more solutions, longer y, etc. We uncover a subtle RLVR failure: it teaches the model to solve in-distribution tasks with high accuracy and structured step-by-step derivation, yet under OOD shifts, the same model abandons derivation and collapses to ineffective brute-force enumeration, causing sharp accuracy drops. This observation reveals the deeper challenges of RLVR generalization as it is neither incapable base model nor reward hacking, since the model succeeds in-distribution with the correct reasoning strategy. The key to mitigating this failure is an indirect intervention that surprisingly helps: co-training with bijective inversion tasks. Although these bijective tasks are actually not uniformly harder and even further from the non-bijective OOD evaluation, they make derivation more certain and substantially restore the correct strategy and accuracy under OOD shifts. This effect further motivates our theoretical analysis: outcome-only RLVR can shape reasoning processes through implicitly crediting the effective reasoning strategies more, even without process supervision. To conclude, our analysis framework and findings systematically reveal the crucial role and incentivization of reasoning strategies in RLVR generalization under OOD shifts. Code is available at https://github.com/ziqipang/rl_meta_thinking.

Group-based reinforcement learning with verifiable rewards (RLVR) learns only from prompts whose sampled completions disagree: all-pass and all-fail groups produce zero relative advantage and burn generation budget without moving the policy. Adaptive schedulers try to steer rollouts toward informative prompts, yet they rely on per-prompt pass-rate trackers that entangle two quantities evolving at very different speeds---the policy's shifting competence across reasoning skills and each prompt's intrinsic difficulty. When a single exponential average or Beta counter tracks both at once, short memory scrambles the relative ordering of prompts whereas long memory reports stale pass rates. We introduce FADE (Factored Ability--Difficulty Estimation), a bilinear state-space filter that separates the two timescales. FADE represents each prompt's success logit as the inner product of a whitened $K$-dimensional skill loading with a fast-drifting policy ability vector, minus a slow-shrinking scalar item intercept. Using Pólya--Gamma augmentation, each minibatch of binary rollout outcomes updates the $K$-dimensional ability posterior in closed form via a rank-$K$ Woodbury step, instantly transferring competence shifts to all unobserved prompts that share those skills while leaving item intercepts stable. From the resulting posterior mean and variance we derive a closed-form utility for the probability of drawing a non-degenerate rollout group. Across five reasoning benchmarks, two base models, and four scheduling paradigms under a matched rollout budget, FADE consistently improves downstream accuracy and calibration while adding less than $0.6$\,ms per step.

Communication is important for our well-being, but it can be impaired by disabilities, such as locked-in syndrome (LIS), cerebral palsy (CP), and dysarthria. Non-invasive brain-to-text brain-computer interfaces (BCIs) offer a promising solution by providing an alternative communication pathway for people with communication disabilities. However, most recent breakthroughs have relied on large, cumbersome, and costly neuroimaging modalities, such as functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). Numerous studies have attempted decoding using the lighter and cheaper electroencephalography (EEG), but the results remain inferior compared to the other modalities. More importantly, we do not understand how these modalities are related to text at the letter, word, sentence, and higher semantic levels. Many prior studies have focused on decoding below the sentence level because it is considerably easier to design an experiment around phonemes, letters, and words. However, during communication, sentences play a much more important role than words and letters. This is because sentences carry information at such a high level that they can be understood even if some words and letters are missing. However, within sentence-level decoding, sentence classification is not as well investigated as sentence generation, making it difficult to decide whether a sentence should be retrieved or generated in the context of EEG-to-text decoding. This paper strives to investigate sentence-level EEG-to-text classification via a novel hierarchical alignment framework, where the main goal is to retrieve a read sentence using its context and EEG as inputs.

Vision-Language Models (VLMs) hold promise for zero-shot video task-progress estimation in robot learning, yet existing frameworks frequently suffer from confirmation bias, hallucinating progress in failed executions due to implicit success assumptions or relative temporal ranking designs. Naive prompting of frontier models collapses into a binary success detector that fails to capture nuanced, non-linear progress and failures. While post-training approaches can mitigate this through explicit training, they demand expensive real-world robotics data and require computationally intensive post-training for every individual base model. To resolve these limitations, we introduce D&A (Describe & Assess): our approach decouples progress estimation into an objective, task-grounded textual description of scene dynamics followed by an assessment phase using explicit rubrics. By compelling the VLM to synthesize an explicit, text-based definition of absolute success before scoring, the framework anchors evaluations to an external, global criterion rather than the video's internal timeline. Empirical results demonstrate that our method calibrates zero-shot estimates and mitigates confirmation bias, providing fine-grained, partial progress metrics that accurately penalize failed executions while correctly interpolating successful runs. Being purely prompt-based, D&A generalizes across tasks, embodiments, and benchmarks---matching or surpassing post-trained reward models---and its reward estimates enable efficient downstream reinforcement learning.

A model is only half of an AI agent system. The other half is the inference strategy: the reasoning procedure that turns model calls into a response. Today that second half is written by hand, and the experience of building one strategy never carries to the next. We introduce Strategist, a meta-agent that evolves inference strategies across tasks. Each strategy is written as a blueprint,executable code that other strategies can reuse, and together the blueprints form a library. Given a task and a small development set, Strategist retrieves related blueprints, writes candidate strategies, tests them on real samples, and revises them from the results. Every candidate then enters the library, so later tasks build on earlier discoveries. On 30 benchmarks, Strategist beats the strongest baseline on 24 and matches its average accuracy (74.7% vs.74.0%) at 39% lower inference cost. Its strategies also transfer to unseen tasks and other models without further search. The part of an AI system that is not improved on its own now does.

Large language models (LLMs) show growing promise for scientific discovery, yet basic questions remain about what enables reliable discovery. How can we trace the scientific discovery capacity of LLMs to their training data? We study this question as one of missing probability mass and develop a theoretical account in which monofacts-the facts observed once during training-set the discovery capacity. In materials science and regulatory genomics, we directly manipulate monofact rates in training sets and find that more monofacts consistently yield more discoveries across two open-weight model families, sampling strategies, and reasoning budgets. Further experiments show the relationship cannot be explained by training-set diversity alone and can transfer across scientific domains when paired with in-domain data. Together, our results identify monofact rate in training data as a key factor governing discoveries in language models.

Evaluating a generative model on all queries from all available benchmark tasks is prohibitively expensive. As such, models are often evaluated on different subsets of tasks and/or different subsets of queries from a given task. The resulting collection of responses and the corresponding (model, task) score matrix are sparse and unpaired. In this paper, we introduce the Product Kernel Perspective Space (PKPS), a collection of low-dimensional representations of models based on their embedded responses. PKPS generalizes existing methods for representing black-box generative models by comparing responses to similar -- rather than just identical -- queries and is therefore able to utilize sparse and unpaired information. Theoretically, we show that benchmark prediction using PKPS is query-efficient relative to methods that only use information from dense and paired portions of the evaluation data. Empirically, we show that PKPS-based methods outperform baselines on benchmark prediction across two large benchmark suites.

Increased computing power has transformed computational science and improved predictive capabilities for complex systems, but many phenomena still exceed current simulation capabilities or require rapid evaluations for fast design iterations. Closure models address this challenge by approximating unresolved effects in reduced or incomplete dynamical systems. Data-driven closures offer a flexible mechanism for improving model accuracy and applicability, but approaches that embed the closure within a computational model for construction and training phases can be computationally expensive and cause numerical instabilities in the computational model during training. To avoid the challenges posed by embedded approaches, staged approaches to closure-model learning have been developed. The first stage uses data assimilation to infer state and closure trajectories from observations. The second stage trains a machine learning closure model offline using the assimilated data, thereby avoiding repeated simulator calls during training. In this work, we investigate the cost and accuracy tradeoffs between embedded and staged data-driven closure modeling approaches for a synthetic epidemiological SIQR model with a nonlinear state-dependent quarantine closure. The results illustrate a practical accuracy-efficiency trade-off: for this exemplar, embedded training provides the strongest average accuracy, while the staged approach yields comparable qualitative behavior with a moderate increase in error, substantially fewer model evaluations, and improved optimization robustness.

ufakzeka-kulak is an open Turkish speech-to-text model of 31.6M parameters, a FastConformer encoder with a token-and-duration transducer (TDT) and a CTC head, trained from scratch and released under Apache-2.0. It writes cased, punctuated Turkish with numbers as words. Its 4,977 training hours come from sources whose licences allow commercial use; 281 hours are human-transcribed, and for the other 4,696 the training text is an uploader subtitle or a machine transcript, kept only where a second, independent reading of the audio agrees with it. The recipe, checkpoint, labelling and export choices were made on development data, and the test sets were scored once, in a release gate fixed before training. With greedy TDT decoding the normalised word error rate (WER) is 0.0921 (95 percent interval 0.0792 to 0.1055) on FLEURS Turkish test and 0.0916 (0.0873 to 0.0958) on Common Voice 27.0 Turkish test, lower than Whisper base and Whisper small on both sets with paired intervals wholly below zero; Whisper large-v3 (greedy) scores 0.0604 and 0.0767. On conversational speech the WER is 0.3807. We also report punctuation and casing, degraded audio, and how closely each shipped export reproduces the trained model.

As pretraining scales, medical language models have advanced rapidly in clinical question answering. Their ability to achieve effective physiological control through sustained intervention still lacks systematic quantitative evaluation. Type 1 diabetes provides a concrete setting: impaired insulin secretion creates dependence on exogenous insulin, with dosing continually adjusted using information about meals, exercise, illness, and glucose measurements. We introduce DiabAgent, a research suite for agentic physiological control comprising the Agentic Physiological Control (APC) framework, the DiabetSim simulator, and the SweetBench benchmark. APC controls glucose through medical-device tools, using ongoing observations to infer individual states, accumulate personal experience, and choose treatment actions. We build DiabetSim to simulate different patients, events, and device conditions for scalable virtual testing and training. Using DiabetSim, we construct SweetBench by instantiating 32 scenario designs as 212 glucose-control tasks. With Claude Opus 5.5 and a fixed decision interval of one hour, APC achieves a COGI of 73.5, compared with 71.2 for the open-source closed-loop baseline OpenAPS. Ablations further show that individual adaptation and contextual information improve glucose control. Combining high-frequency rule-based control with lower-frequency agent decisions shows potential for balancing control performance and computational cost. Our work provides a framework, simulator, and evaluation foundation for agents to participate in human physiological control.

AI agents are increasingly deployed in settings where they interact and coordinate with other agents, yet honesty is mostly measured in single-turn, single-agent benchmarks that pressure the model to deceive. We study deception that emerges in multi-agent interaction without any instruction to deceive. We introduce Kradle, a platform for evaluating AI models, and Four Bridges, an information asymmetry game in which four agents distribute across four rooms: one is the death room, the others hold food worth more if not shared. One agent, the informed agent, knows which room is the death room and can withhold or disclose that information. We evaluate eight frontier models as the informed agent. Deceptive behavior appears in every model and in 73% of all runs (6% to 100% per model), and does not decrease with capability tier within a model family. When the informed agent does disclose, the warning is often challenged, and agents that challenge it enter the death room twice as often as those that do not. Telling all agents that one of them knows the death room collapses deception in six of eight models; telling the informed agent that the room is permanently fatal moves only one. On an open-weight model, an honesty steering vector reduces deception to 0% and adversarial honesty training to 6%. The platform and game provide a testbed for measuring deceptive behavior and comparing interventions.

In this paper, we study when two parties should pool their data to train a machine learning model. Two agents, Alice and Bob, each hold a sensitive labeled dataset and want to train a classifier from a given hypothesis class $\mathcal{H}$. Even when each party's dataset can be perfectly classified on its own, the two parties may follow conflicting labeling standards or have selection biases that affect their features and label sets, such that no single classifier in $\mathcal{H}$ fits their combined data. Before committing to data sharing, Alice and Bob therefore want to test whether a single classifier in $\mathcal{H}$ can jointly fit both datasets---without revealing their individual data points to one another. We formalize this problem from a learning-theoretic viewpoint via $\operatorname{opt}_\mathcal{H}$, the minimum \emph{number of mistakes} made by any hypothesis in $\mathcal{H}$ on their pooled data. Through interactive communication, the agents must distinguish $\operatorname{opt}_\mathcal{H}=0$ (compatible datasets) from $\operatorname{opt}_\mathcal{H}\geq t$ (conflicting datasets) under pure $\varepsilon$-two-party differential privacy. A trusted curator could do this with $t=O(1/\varepsilon)$, and we ask how small $t$ can be without one. For thresholds on the real line over a domain of size $m$, we show that the optimal $t$ is $\tildeΘ(\log\log m/\varepsilon)$, exponentially smaller than the $Θ(\log m/\varepsilon)$ samples needed to privately learn thresholds, even with a curator. For halfspaces in $\mathbb{R}^d$, the dependence on $m$ jumps to $\log(m)$: we give a protocol with $t=\tilde{O}(d^3\log m/\varepsilon)$ and show that some domains require $t=Ω(d\log(m/d)/\varepsilon)$.

Understanding how neurons respond to visual stimulus is an important question in both deep learning and neuroscience. It has significant implications for enhancing the interpretability of black-box artificial neural networks and for understanding the visual representation in the human brain. We propose a structured visual representations landscape and design an activation-based prior that allows us to effectively regularize the landscape with activations from brain regions or neural network units. Our model Vis-Lens integrates a variational autoencoder and a diffusion model into an image generation framework. It enables the generation of natural, realistic preferred images by directly modifying the activation-regularized latent codes without per-image iterative optimization. We demonstrate the effectiveness of our framework on both artificial neural networks and human visual cortex using functional magnetic resonance imaging (fMRI) data. Our framework improves activation modulation and image realism compared with a recent state-of-the-art method for generating preferred visual representations of these neural targets.

We show that the causal structure of spacetime can emerge within the memory of a learning agent performing a quantum information task. We train a model-based reinforcement-learning agent to maximize the quantum mutual information between localized quantum probes interacting with a quantum field in Minkowski spacetime, without directly providing it with information regarding the spacetime metric or causal structure. In learning to accomplish this task, the reinforcement-learning agent constructs an internal predictive model whose planner develops the characteristic null structure of spacetime, allowing us to reconstruct the causal structure of the Minkowski metric with high accuracy. Within this task, spacetime geometry is obtained from an agent's successful predictions and actions, providing a first step toward showing that spacetime concepts emerge from operational quantum experience.

Neural operators provide efficient surrogates for computational fluid dynamics (CFD), but most are designed for static geometries and are not equipped to model unsteady flows induced by continuously moving boundaries. We study dynamic-boundary 4D CFD prediction for articulated moving geometries and introduce RoboCFD-4D, a temporally resolved robotic CFD benchmark spanning diverse morphologies and flow conditions. To address the computational and autoregressive challenges of this setting, we propose DynaSolver, a neural operator that explicitly conditions on boundary evolution and causal flow history. DynaSolver employs an efficient physical-token formulation that amortizes point-to-token projection across the operator stack, substantially reducing repeated point-space spatio-temporal computation while retaining accurate point-level reconstruction. We further introduce Replay-Augmented Parallel Training, which exposes the model to autoregressively generated histories while preserving parallel dense supervision, narrowing the gap between ground-truth-conditioned training and autoregressive rollout. Experiments on RoboCFD-4D show that DynaSolver achieves the best full-rollout accuracy among a broad range of learned CFD surrogates, with its advantage increasing as rollout proceeds, indicating substantially reduced autoregressive error accumulation. Project page: https://made-gpt.github.io/DynaSolver-ProjectPage.

Time series foundation models (TSFMs) are increasingly used for probabilistic forecasting across diverse domains, yet the reliability of their uncertainty estimates remains poorly understood. In this work, we study uncertainty calibration in probabilistic TSFMs from a broader methodological and empirical perspective. We introduce a taxonomy of how TSFMs represent predictive uncertainty and develop a fine-grained evaluation framework for calibration in multi-horizon forecasting that distinguishes pooled, horizon-wise, and stronger conditional notions of calibration. Using this framework, we conduct a comprehensive empirical evaluation of popular TSFMs across diverse datasets and show that current models exhibit distinct and systematic patterns of miscalibration that standard metrics often obscure. We further analyze common post-hoc recalibration methods and characterize which forms of calibration they can improve and which tend to persist. Our results provide practical guidance for evaluating and improving uncertainty calibration in TSFMs and highlight important limitations of current practice.

In the near future language model pretraining will become bound by total data rather than total compute. Such a regime requires revisiting scaling laws and scaling law theory. Motivated by prior works that have found utility in ensembling, we study how the number of training epochs and ensemble size jointly shape generalization when training data are fixed. In language-model experiments on FineWeb, logit averaging reduces late-training deterioration, delays optimal stopping, and can outperform wider or deeper single models at matched parameter-token budgets over a common training horizon. To analyze this interaction between data reuse and ensembling, we derive a dynamical mean-field theory (DMFT) for multi-epoch stochastic gradient descent in power-law random-feature ensembles, tracking error correlations across epochs and models. Ensembling reduces the prediction variance responsible for late-time growth in expected test error, changing the optimal stopping time. We use these stopping times to derive compute-allocation laws at fixed data and individual model sizes. Under this allocation, ensemble size grows with compute at least as fast as training time per member, while excess test loss above the limiting loss decreases as a power of compute.

Federated learning (FL) runtimes standardize aggregation and deployment, but each new study still requires federation formation: selecting task-eligible sites and preparing their local data and task code for federated training or analysis. This preparation recurs when the target changes, while raw records and site-private mappings are excluded from cross-site FL messages. We present FedReady, a system that performs federation formation with bounded LLM agents. Agents propose shared contracts, local adapters, and task code; the runtime owns workflow state and accepts artifacts only after checks tied to their digests. Candidate site discovery prioritizes recall; local eligibility verification checks task support before site admission. Once participating sites and artifacts are fixed, conventional federated training or analysis proceeds without readiness agents. Across 39 simulated retinal sites and four tasks, discovery reaches recall 1.00 at precision 0.65; local eligibility verification removes all 17 false positives without losing an eligible site-task pair. The participating sites supply 45,575 task-record instances and complete four 100-round jobs, including one with a runtime-generated contract. Across ten patient cohorts, preparation supports seven patient populations and holds endpoints at three sources as unsupported. Comparison with manually prepared reference populations and analysis (the manual reference) gives a maximum survival-curve difference of 0.001133. We also analyze 19 development failures, targeted classification ablations, and standalone and heterogeneous agent configurations. These results establish the feasibility of task-specific data preparation before ordinary FL training or analysis.

Most parameters of a Mixture-of-Experts (MoE) model belong to its experts, and standard training keeps every expert gradient in GPU memory until the optimizer step. This memory grows with the number and size of the experts, so it rises as MoE models scale. We introduce Sonar-MoE, which updates each expert tensor during the backward pass and frees its gradient right away. It brings optimizer-in-backward from dense models to batched MoE expert tensors, so the GPU never holds more than one layer of expert gradients. When we fine-tune OLMoE-1B-7B and Granite-3.1-3B-A800M in bf16 with two expert-kernel backends, Sonar-MoE lowers peak memory by 19% to 22%. Its updates are bitwise identical to standard training without expert gradient clipping, and held-out loss stays within 0.003 of all baselines. Each step also takes 17% to 23% less time than standard training with global-norm clipping. Updating whole batched tensors rather than individual experts is what keeps Sonar-MoE fast, as it avoids a 20% to 72% slowdown while using the same memory. Sonar-MoE matches optimizer-in-backward over all parameters in memory and speed while still clipping the dense parameters.

Law changes over time. A court decides a case under the law in force on that day, and legal QA benchmarks grade answers against today's law. We checked 156,650 decisions of four Thai courts. Up to 67% of them cite law that has since been amended or repealed. We introduce NitiBench+, the sequel of our NitiBench, a Thai legal QA benchmark with 1,670 items across 17 domains, built from decisions of three courts. For each cited section it holds both the text in force on the decision date and the current text, and its retrieval corpus holds every dated version of every Thai statute section. Answers are graded on the Thai Bar Council rubric, the Bar-standard score, and on the overlap between the cited sections and the court's. The report has two parts. Part I measures the cost of the wrong text on 13 open-weight models. Reading the current text of amended law drops the Bar-standard score by 0.25 and missing repealed law drops it by 0.79. The penalty comes from rule changes, 0.45 on the 428 rule-changed items and none on the 331 wording-only items, and Citation F1 does not see it. The text in force is the fair setting. Part II benchmarks today's models on that setting. A public leaderboard ranks 26 models on a held-out test split of 500 items stratified by domain, with a validation split of the same shape released for public use, from memory alone, with the court's sections provided, and as a retrieval agent over the law in force, with the cost per item and the hallucinated citations. The best model scores 7.78 out of 10 and the best open-weight model 6.39. Without a search tool only five open-weight models recall any Thai law, and with it twenty of the 26 models gain and six lose. Leaderboard will take new models as they are released. Legal benchmarks must track dates, and on the right text today's models still leave a wide gap on Thai law.

Large language model (LLM) coding agents increasingly operate with shell access, allowing them to write scripts, move files, alter permissions, and modify system configurations. However, an agent may achieve a user's goal while causing unintended harm, such as deleting unrelated files, weakening access controls, or corrupting data, in response to benign inputs without explicit jailbreaking. This failure mode is particularly important because apparent task success may reassure users that an agent behaved correctly, causing harmful side effects to go unnoticed. We call this harmful success: the agent successfully completes the intended benign task while simultaneously causing unintended harm. Despite its importance, harmful success is difficult to measure because conventional task-success metrics overlook harmful behaviors, while existing safety evaluations are often designed solely to assess harmful outcomes. We introduce VeriHarm, an automatic framework for constructing programmatically verifiable benchmarks that directly measure harmful success. VeriHarm minimally perturbs coding tasks to elicit unintended harmful behavior while maintaining the benignity of the task and preserving the original task objective. It then synthesizes executable harm tests that, together with the original task-correctness tests, evaluate both task accuracy and harmful outcomes. Frontier agents, including GPT-5.6-Sol and Claude-Opus-5, achieve an average task accuracy of 95% on our curated VeriHarm-Bench, yet still struggle to complete tasks safely: their safe success rate, the fraction of tasks completed successfully without triggering the designated harm, averages only 37%. These results suggest that successful task completion can frequently coexist with unintended harm, and that task-success metrics alone may substantially understate the potential safety risks of coding agents.

Exemplar-free continual learning on pre-trained Vision Transformers increasingly relies on Low-Rank Adaptation (LoRA), yet prevailing methods govern the low-rank update space through strict subspace isolation. By projecting each incoming task into the approximate null space of historical token activations, isolation-based learners equate any cross-task feature overlap with catastrophic interference. We show that this orthogonality premise induces a structural bottleneck we term \emph{spectral starvation}: when down-projection bases are locked before task optimization, the gradient norm reaching the trainable up-projection is upper-bounded by the energy of current-task features captured by the chosen basis. Because natural visual tasks share low- and mid-level covariance structure, historical null tails contain barely $5.5\%$ of incoming task variance, starving new-class plasticity while blocking positive consolidation along shared directions. We propose BISECT (Bilevel Spectral Decomposition), a continual LoRA framework that partitions a fixed rank-$r$ budget into two orthogonal functional blocks: an invariant carrier spanned by the dominant modes of the joint historical-and-current Gram matrix, and a novelty complement extracted via generalized Rayleigh whitening followed by orthogonal complement projection. After single-task training, BISECT folds both blocks back into the frozen backbone using a closed-form precision gate that scales each rank-1 invariant atom by its current-to-joint Rayleigh ratio while preserving novelty atoms at unit gain. Across eight incremental settings on ImageNet-R, ImageNet-A, CIFAR-100, CUB-200, and DomainNet, BISECT outperforms prompt, adapter, and LoRA baselines with zero extra inference parameters or FLOPs, and remains complementary to sparse analytic expansion.

Timely and accurate crop health diagnosis is essential for sustainable agricultural production, yet accurate diagnosis often requires expert knowledge that is not readily available. Recent vision-language models (VLMs) have demonstrated remarkable capabilities in visual understanding, reasoning, and natural-language interaction, creating new opportunities for intelligent crop health assessment. However, deploying VLMs on resource-constrained devices remains challenging due to their substantial computational and memory requirements. This paper presents PlantVLM-Edge, a quantized vision-language framework for conversational crop health diagnosis that can run entirely on resource-constrained edge devices. PlantVLM-Edge combines agricultural domain adaptation, post-training quantization, and on-device serving with an offline chat interface, allowing growers to ask open-ended questions and receive diagnoses, symptom explanations, and management guidance without cloud connectivity. The framework is backbone-agnostic: we validate it with two different VLM types with varying scale and architecture (Qwen3-VL-2B and LLaVA-1.5-7B), four quantization schemes, and two diagnostic domains (plant diseases and insect pests), and we deploy all variants on an NVIDIA Jetson AGX Orin. Eight-bit quantization affects answer quality minimally relative to the full-precision model, and 4-bit activation-aware weight quantization (AWQ) retains 98.6% of full-precision answer quality on both domains while reducing the weight memory of the 2B backbone by 49%. In its most compact form, PlantVLM-Edge requires 2.01 GiB of weight memory, returns its first token in about 0.12 s, and completes a 64-token diagnostic response in 0.78 s. These results demonstrate that timely and accurate crop health diagnosis is practical on edge devices, providing a foundation for next-generation crop monitoring and decision support systems.

The Model Context Protocol (MCP) lets agentic AI frameworks connect language models to external tool providers, but it carries attacker-controlled content (tool descriptions, schema fields, and response data) into model context without validation. Prior benchmarks measure model compliance with poisoned instructions without confirming real execution. We present the first cross-framework measurement of MCP tool poisoning with real tool execution. A 25-class root-cause taxonomy unifies prior classifications, and the open-source tool mcpvenom drives a 9,720-cell susceptibility matrix across 12 frameworks, 18 models, and 45 attack scenarios, with a dual oracle that confirms exfiltration via canary HTTP receipt. The overall attack success rate is 22.9% (2,113 vulnerable cells). Of these, 64.9% dispatch an attacker tool call, 23.0% contact the attacker endpoint, and 12.1% verifiably exfiltrate a secret. Resistance is model-level: nine core frameworks cluster within 23 to 26% (chi-squared = 3.30, p = 0.91), and no framework implements intentional defense. Incidental framework properties mitigate in specific cases, and a regex description scanner rejects 20 of 45 scenarios at a 7.2% false-positive rate, so metadata scanning alone is a partial control. Model ASR ranges from 0.4% to 49.7%. Claude Opus 4.8 is vulnerable in 8 of 530 cells while issuing 223 explicit refusals. A code-level audit finds no intentional defense across twelve frameworks, and all five coordinated disclosures across four vendors were rejected.

We consider nonparametric maximum likelihood estimation of the intensity function of a Poisson process, allowing the intensity to depend on observed covariates. Such models describe how event rates vary with external conditions, including conditions under which the intensity becomes arbitrarily small or vanishes. However, risk analysis based on the Kullback-Leibler (KL) divergence typically requires the true intensity to be bounded away from zero, a restrictive assumption in applications. We develop a unified approach that avoids this assumption and derive an oracle inequality for the covariate-averaged Hellinger distance between the true intensity and the nonparametric maximum likelihood estimator (NPMLE). As an application of the oracle inequality, we establish convergence rates for NPMLEs based on deep neural networks (DNNs). Our results show that, as in nonparametric regression, the ability of DNNs to capture underlying structure leads to improved convergence rates. Specifically, DNN-based NPMLEs mitigate the curse of dimensionality when the true intensity has a compositional structure and automatically adapt to a low-dimensional Riemannian manifold structure of the event space. For intensity functions with a compositional structure, DNN-based NPMLEs achieve nearly minimax-optimal convergence rates without explicitly incorporating this structure into the network architecture. Simulations demonstrate that DNN-based estimators outperform kernel estimators in most settings, while manifold kernels remain competitive in some designs. Finally, we analyze stimulus-evoked neural spiking using data from the Allen Brain Observatory Visual Coding Neuropixels project. DNN-based NPMLEs provide better out-of-sample performance than a conditional kernel estimator, as evaluated by held-out log-likelihood.

LLM coding agents can increasingly automate model post-training, but whether they can act as forward-deployed engineer (FDE) agents, discovering customer requirements and turning them into deployment decisions, remains unclear. We study this on a real enterprise text-to-SQL engagement: a simulated customer, grounded in the real use case's requirements, starts with an ambiguous request. An agent must discover acceptance criteria, data permissions, and hardware constraints in the discovery phase, then plan and execute a post-training recipe in the delivery phase. Across 15 models, the strongest discover over 80% of withheld requirements and open-weight models discover 40-50%; in the delivery phase, only 25% of runs across eight coding-agent configurations pass all acceptance criteria. We identify four bottlenecks in this engagement: (1) incomplete discovery, with 40% of withheld requirements missing or misrepresented; (2) unreliable execution; (3) underused compute budgets and a lack of systematic comparisons across base models and post-training methods; and (4) benchmark accuracy does not guarantee production readiness. We suggest human review of discovered requirements, monitoring of the agent's plan during execution, and an explicit experiment-environment specification covering hardware, budget scope, and production checks.

Earth system models (ESMs) rely on parameterizations to represent unresolved processes such as convection, cloud microphysics, and turbulence. However, translating advances in physical understanding into stable, validated parameterizations and integrating them into ESMs often takes years. Agentic AI could accelerate this process and enable more systematic exploration of design choices. Using warm-rain collision-coalescence as a test case, we evaluate agents' ability to synthesize scientific literature, reimplement published bulk parameterizations, and propose new functional forms as differentiable JAX code. We compare conventional parameterizations, data-driven emulators, and agent-generated closures against high-fidelity superdroplet simulations, repeating unattended experiments across ten language models. With expert guidance or iterative feedback, agents produced executable, differentiable, mass-conserving code, although unattended performance depended strongly on the scheme and language model. Iterative feedback lowered median rollout error for eight of ten models. Retuned physical schemes matched or outperformed many learned parameterizations at substantially lower computational cost, but two of 100 unattended agent-generated closures achieved slightly lower rain-mass error than the most accurate retuned scheme on held-out simulations.

We present a variational functional neural network (vfNN) algorithm that optimizes physically motivated continuum functionals in integral form (e.g., Lagrangian, Hamiltonian, or free energy) to find the states of a system. That way, vfNN bypasses solving the equations of motion (a.k.a. the Euler-Lagrange equations), which are in the form of differential equations. To this end, our approach is fundamentally different from conventional physics-informed neural networks (PINN) that solve differential equations by minimizing a phenomenological loss function involving higher order derivatives and ad hoc weight factors. In contrast, vfNN merely evaluates lower order derivatives of the generating functional and replaces the weight factors with rigorous physical scales. This also fundamentally changes the loss landscape, and the desired solutions may become unstable. For this emergent nonconvex optimization problem, we develop a quasi-Newton optimization algorithm with simulated annealing that converges to the solutions of the corresponding Euler-Lagrange equations of the functional regardless of their stability properties. Our simulations using vfNN give accurate results for benchmark systems governed by linear or nonlinear ordinary differential equations.

Connecting new generation and demand to the grid depends on two scarce skills: checking applications against a large, changing body of regulation, and running the network studies that decide where and how a project connects. We present Grid Agents, an AI system that brings both together, and the methodology we use to evaluate it. The Grid Ontology is its knowledge system. It represents the regulation, standards and engineering knowledge that grid work depends on, makes it searchable, and evolves as the agents learn from use. Every answer drawn from it rests on evidence the agents cite. Grid Studio is an agent system specialised for power-system engineering. It works in its own engineering environment with domain tools and a study framework, draws on the Grid Ontology, and relies on established solvers for every number it reports. In both, engineers make the decisions. We bring eight published power-system benchmark families, previously scattered across separate papers, into one suite of 519 tasks, and we verify not only each answer but the work behind it, by replaying the agent's tool calls. The Grid Ontology cites the right regulatory evidence more often than a general coding agent or retrieval pipelines, leads on LegalBench-RAG and ranks first on the public leaderboards of the Enterprise RAG Challenge and ViDoRe V3. Grid Studio scores highest on the 519 power-system tasks and is the fastest of all the systems we compare, both by a statistically significant margin. With an optimiser in the loop, its reinforcement packages cost 33-67% less on average than those the agent designs alone.

Skin cancer diagnosis, particularly differentiating melanoma from benign nevi, is challenging due to the visual similarity between lesions. Although convolutional neural networks (CNNs) and vision transformers (ViTs) have shown promising performance, their effectiveness often deteriorates on data from heterogeneous clinical sources. Conventional domain adaptation methods address domain shift but require access to source data during adaptation, which is often infeasible due to privacy regulations. Multi-source-free unsupervised domain adaptation (MSFDA) addresses this limitation by leveraging multiple labeled source domains to generalize to an unlabeled target domain without requiring access to source data. However, existing MSFDA methods rely on full backbone fine-tuning, leading to catastrophic forgetting and overfitting on small clinical datasets, and address domain shift at the aggregation stage without establishing a shared domain-invariant feature space. Furthermore, their reliance on hard pseudo-labels or confidence-weighted aggregation introduces noisy supervision signals under domain shift. To address these limitations, we propose CAT-CKD, consisting of two components: (1) contrastive adapter training (CAT), which trains lightweight ConvPass adapters within a frozen ViT backbone using supervised contrastive learning (SCL) to establish a shared domain-invariant feature space before source-specific model training, and (2) consensus knowledge distillation (CKD), which aggregates logits from multiple source models into a consensus supervisory signal and adapts a student model on unlabeled target data using KL divergence. Experiments on five publicly available skin lesion datasets show that CAT-CKD achieves an average AUROC of 86.1%, outperforming existing MSFDA methods while requiring only 4.3M trainable parameters. The code is available at https://github.com/A-Abedi/CAT_CKD.

Computer programmers chain together the same reusable computational primitives, for example, conditional statements or loops, but in different orders to implement different algorithms. In contrast, existing language model architectures always execute their layers in the same order. We introduce Library of Layers (LoL), an architecture that dynamically orders its layers during inference to solve the problem at hand. LoL has a single hyperparameter, the size of its global library of transformer blocks, that allows it to transition seamlessly between parameter reuse and sparsity, generalizing the parameter-efficiency and FLOPs-efficiency tradeoffs of dense, depth-recurrent, and mixture of experts architectures in one unified family. We show that LoL achieves competitive performance with all three above baselines when pretrained on text data. On synthetic sudoku and maze data, we observe that the model employs distinct layer orderings to implement different solution strategies, and we can intervene on routing to force a model to apply a specific solution strategy to a problem instance. In a multilingual natural language setting, we find that LoL chooses different layer order patterns based on language and script. Similarly, we train LoL on safety data and fit probes to its router outputs, finding that these probes distinguish safe from unsafe text with high AUROC. LoL provides a path toward a future in which models dynamically compose their own computational graph, enabling more efficient and controllable reasoning.

Skill extraction benchmarks score models against a few human-annotated positives and treat every other skill in the vocabulary as equally wrong. At the scale of ESCO (13,939 skills) this binary view can reward models that rank the annotated positive high for the wrong reason, and hides semantically nonsensical predictions. The WorkRB Challenge at RecSys-HR 2026 replaces it with graded relevance: every candidate skill in a model's ranking window carries a 0 to 4 grade, and the leaderboard scores nDCG@100 against these grades. Since ternary judgements proved too coarse for objective annotation, relevance is decomposed into a cascade of four binary questions (same domain, activity mentioned, right granularity, replacement of a gold positive) that a smaller LLM judge can answer reliably; the labels are validated with blind human and frontier-model re-annotation. This paper documents the annotation methodology, the construction of five benchmark datasets (four with vacancy sentences as queries, one with lexically distant skill surface forms), the validation of the LLM cascade, and insights from the submissions. The field converged on multi-stage retrieve-and-rerank pipelines, whose gains come from clearing nonsense out of the ranking window rather than finding more positives: at the top of the leaderboard, the graded ordering is uncorrelated with what binary evaluation would have said. Graded labels expose much bigger gaps between top submissions and baselines than binary evaluation does, and show that semantic-understanding skill recommendation is still wide open. The clear next direction is bringing efficiency back: every submitted system we could size uses at least 42x, typically several hundred times, more parameters than the baselines it beats.

LLM agents show strong potential in software engineering tasks such as competitive programming and code synthesis. Existing benchmarks, however, largely focus on small-scale settings such as self-contained Python modules, leaving distributed systems, an essential class of production software, under-examined. We introduce DistDebug-Bench, a benchmark for evaluating agents' ability to diagnose complex-system failures at the code level. It curates 120 high-quality real-world bugs from 11 widely used distributed systems, spanning 5 programming languages and 10 failure patterns. Through an extensive evaluation of state-of-the-art LLMs and agents, we identify substantial gaps between current agents and real-world debugging tasks. A recurring gap is tunnel vision due to a lack of explicit reasoning structure and nondeterminism from inconsistent tool-use scope. The benchmark and evaluator are available at https://github.com/OrderLab/distdebug-bench, with agent traces and a leaderboard at https://distdebug.com.

Contact-rich manipulation benefits from tactile feedback that captures both detailed contact geometry and interaction forces. When scooping grains, contact patterns help infer spoon orientation, while resistance transmitted through the grasp guides motion. To provide these complementary sensing capabilities, we present PanoTac, a low-cost, compliant tactile fingertip designed for dense three-axis tactile sensing. Each fingertip provides a $12 \times 32 \times 3$ tactile representation, with one normal ($z$) and two shear ($x$, $y$) response components at each taxel. The sensor operates at 60 Hz and costs less than $50 per fingertip. Sensor characterization quantifies response repeatability and consistency across sensor instances and contact locations. We integrate PanoTac into a robotic system with a visuo-tactile pipeline. Across test-tube preparation, grain scooping, and salt pouring, the resulting policies outperform vision-only baselines by an average of 46.7 percentage points. Ablations show that removing either spatial contact information or shear feedback reduces overall performance across the task suite. Our project page is available at https://panotac.github.io.

Organisations increasingly deploy large language model agents that must follow organisational values. We check whether activation steering, which adds vectors to a model's internal activations during generation without changing its weights, can improve compliance beyond prompting alone. Since most scenarios involve several values at once, we focus on multi-value steering. A simple way to build these vectors is to contrast responses that follow all values relevant to a scenario with responses that violate them. However, this gives nearly identical vectors for different values: they capture a general compliance signal rather than individual values, which limits their use when combining several values. We propose Value-Isolating Activation Steering (VIAS), which contrasts responses to the same scenario under two prompts, one listing all relevant values and one omitting a single target value. We keep a pair only if the target value changes from compliant to non-compliant and both responses still meet a minimum compliance level on the other values. We evaluate VIAS on PROTECT, a workplace compliance dataset where each scenario is tagged with relevant values from a 21-value taxonomy. VIAS reduces overlap between value vectors, making them better suited for multi-value steering, although this alone does not lead to independent control over each value. Combined with prompting, VIAS complies with every tagged value on 69.6% of scenarios for Qwen2.5-7B, 75.2% for Mistral-7B, and 72.6% for Phi-3.5-mini, outperforming prompting alone for all three models.

LLM agents for GPU kernel optimization sit at two ends of a structure-freedom tradeoff. Scripted agents fix what the model does at each step, via a learned policy, an optimization menu, or fixed roles, and are sample-efficient where the recipe applies and stall where it does not. Open-ended agents let a self-directed model mutate a population of kernels and pay for the freedom in trials: thousands of evaluations, repeated edits, and wrong candidates. Both respond to hard targets by equipping the agent with more: parsed profiler output, disassembly analysis, retrieval over documentation and libraries, and stores of lessons. We argue that both lack direction, a hypothesis about where the remaining gap is, and that a short question asked at the right moment supplies it without narrowing what the agent may do next. We present KernelSeek, a question-driven agent that makes an existing CUDA kernel faster with no operator-specific knowledge and no faster implementation to imitate: a general-purpose coding agent, a minimal harness that enforces a verifiable goal, and questions at decision points, first from a human, then from a bank the agent applies to itself. On the same task, evaluator, hardware, and week, a tool-rich autonomous agent built in the prevailing style plateaued at 0.50x FlashAttention-3 after 133 model-driven steps; KernelSeek reached 1.13x. It then produced dependency-free FlashAttention forward kernels for Hopper and Blackwell, single CUDA files with no CUTLASS, CuTe, cuBLAS, or cuDNN, correct on all 182 scenarios per architecture and 1.16x FlashAttention-3 and 1.46x FlashAttention-4 (geometric mean). Across 162 human messages and 11.5K agent tool calls, every large gain followed a question that moved the agent to a different layer of the system or rejected a claimed limit, never one that supplied information. We release the kernels, harness, question bank, and all traces.

Parameter-efficient continual learning on pre-trained Vision Transformers largely relies on a costly premise: preventing catastrophic forgetting requires either spawning a dedicated LoRA branch per session---causing adapter parameters to scale linearly with task horizon---or applying sparse Mixture-of-Experts gating symmetrically to both forward and backward passes, which sacrifices representational rank at test time. We argue that reading consolidated knowledge and writing new task credit are fundamentally asymmetric operations. Because visual features in a shared low-rank space distribute across rank-1 atoms rather than localizing in disjoint task silos, forward inference should read all atoms at full rank capacity, whereas backward optimization should restrict plasticity to a sparse subset of input-aligned channels. We introduce RASP (Read-All, Sparse-Plasticity LoRA), which adapts a single stationary-size adapter across sequential tasks. At each step, a scene-modulated cosine router selects the top-$q$ unlocked atoms to receive gradients, blocks updates on the remaining $r-q$ channels, and penalizes Frobenius Gram coherence between active and non-updated output directions. Across sessions, the router's empirical activation trace and each atom's exact spectral norm form a zero-overhead geometric consolidation index $ω_j = \sqrt{\tilde{c}_j \tildeσ_j}$ that drives Cauchy step attenuation and boundary anchor locking. Across CIFAR-100, CUB-200, ImageNet-R, and ImageNet-A, RASP matches or surpasses branch-expanding LoRA learners under a strictly constant parameter footprint.

Graph forecasters for sensor networks are trained in one regime and deployed in others, where sensor coupling and signal frequency content both change. We split this spectral shift into an operator channel and a frequency channel, each with a dimensionless measure. ST-GFM (Spectral-Temporal Graph Forecasting Model) filters every input step with Chebyshev and bidirectional diffusion filters on one learned, non-symmetric operator and alternates them with per-node temporal attention. For the operator channel we prove Frobenius-norm perturbation bounds for both graph branches, show that the Chebyshev bound is attained to first order, and compute its constant from the trained weights and the measured operator norm: 17.9 on METR-LA and 21.5 on PEMS-BAY. Under random edge dropout the trained layer moves by 2.0% to 3.1% of this bound. Masked Spectral Reconstruction (MSR), a self-supervised loss active in the first 40 epochs, matches the log power spectrum of the encoder output to that of the input. For small log-spectral residuals on unclamped bins it is proportional to an Itakura-Saito divergence to leading order. Removing it raises the 60-minute MAE by 0.30 on METR-LA, close to the 0.33 lost when all three self-supervised losses go. Under one shared evaluation pipeline, the 60-minute MAE of ST-GFM is 3.5% below a retrained STAEformer on METR-LA and 13.0% below it on PEMS-BAY. Its 60-minute MAE is 3.0980.017 on METR-LA (five seeds) and 1.5760.006 on PEMS-BAY (three seeds), the lowest PEMS-BAY MAE among twelve published baselines.

Matrix multiplication dominates both PHY signal processing and Deep Neural Network (DNN) inference, yet these domains place opposing demands on the matrix engine. Signal processing requires low-latency execution of relatively small matrix operations, whereas DNN inference, particularly at large batch sizes, prioritizes high throughput and energy efficiency for large matrix multiplications. While offering efficient memory access profile, a Systolic Array (SA) adds fill and drain delays per matrix. Those can be amortized for large matrices but not for small matrices. Conversely, a tensor unit completes the kernel in few cycles but would require repeated reads of same operand tiles over large matrices. This makes SA too slow for signal processing and tensor too inefficient for AI. We show that this need not be a choice between systolic and tensor structuring of Multiply-Accumulate (MAC) units. We propose a single engine that operates primarily as a SA and, through low-cost reconfiguration, transforms into a tensor-style engine on a per-operation basis. MUXes create both the spatial reduction adder tree and temporal reduction accumulators from the existing MACs and registers within the systolic array. The proposed reconfigurable SA simulated using modified SCALESim simulator, verified on an Zynq 702 FPGA and synthesized using ASAP 7nm FinFET library. SCALE-Sim simulations shows up to 7.5x lower latency for small complex value PHY kernels. Even on DNN workloads, it achieves up to 1.65x speedups over vanilla SA execution with trivial increase of memory accesses, retaining the systolic array's memory access advantage. Synthesis results estimate reconfigurable fabric costs only 8% power and 13.6% area overheads.

Can AI contribute to formal demography? We report an experiment in which an AI model, Claude, was asked to explain and then simplify the proof of an earlier result: that vaccinating the oldest against COVID-19 saves both the most lives and the most years of life, because the years saved per person treated rise with age. The model proposed a new proof that is shorter, more intuitive, and somewhat more general than the original, built from a known life-table identity and the definition of convexity. It states the condition for the result as the convexity of the reciprocal of the hazard, which can be checked from the hazard alone and holds for most familiar parametric models, and it yields a lower bound on life expectancy whenever that condition is met. Prompted by a further question, the model also gave an exact form of the condition, showing how far convexity can be relaxed. The model traced these identities and related results in earlier literature. An edited record of the sessions is provided, and complete transcripts are available on request. The exercise suggests that AI can revisit established results, make them easier to understand, and prove valuable generalizations.

Language models in customer-facing products must be continually updated, since the data they serve shifts over time. As retraining on the full accumulated data is impractical, each update should instead run on a small set of examples carrying the training signal of all data seen so far. The strongest current methods build this set by gradient matching, which synthesizes examples whose training gradients match those of the full data. However, they require ground-truth labels, run once on a fixed snapshot rather than on a growing stream, and are slow on large data. We introduce \textbf{OLGA}, which lifts these limitations. It distills unlabeled data by drawing the gradient matching targets from the model's own next-token predictions. Across five text classification benchmarks with phi-1.5, it improves accuracy by 10 percentage points over the strongest selection baseline using as few as five examples. OLGA also distills online data streams by keeping a running rehearsal memory of synthetic examples, which is re-distilled together with each new batch as it arrives. In the online setting, it outperforms the strongest selection-based continual baseline by 6.6 stream-AUC percentage points. Finally, our method scales to large data by pairing distillation with selection in a hybrid pipeline, letting selection handle the mild stage of compression and distillation the extreme one. It cuts runtime by 30\% versus distillation alone at similar accuracy, and beats the best selection-only pipeline by up to 12.6 percentage points.

AI benchmarks have made remarkable progress in measuring the ability of AI systems to perform increasingly demanding cognitive and technical tasks. Yet, many of these benchmarks treat technical tasks as the unit of analysis, while much economically valuable work takes place within organizations, where tasks are embedded in a system of relations. Within organizations, successful task performance depends not only on solving the technical demands of the task but also on navigating the interdependence of activities, distributed knowledge, differentiated authority, temporal constraints, conflicting interests, and changing environmental conditions while the task is underway. Accordingly, we argue that organizational task difficulty constitutes a distinct and insufficiently recognized dimension of AI capability, and that existing approaches to benchmarking provide an incomplete basis for assessing how effectively AI agents can perform real-world organizational work. By drawing on organization theory, we develop a taxonomy of the structure of work that characterizes the conditions that agents face in organizational work. Applying this taxonomy to analyze existing benchmarks, we find limited and uneven coverage of organizational dimensions, with fewer than one quarter of prominent benchmarks claiming any of them. We therefore argue that AI benchmarks should increasingly move beyond the evaluation of technically difficult tasks toward the systematic assessment of agents' ability to navigate the organizational conditions under which economically valuable work takes place.

For effective resource management, infrastructure planning, and sustainable tourism, it is really important to predict accurate visitors. This paper estimates the daily visitation of Nebraska lakes using spatial, basic temporal, and interpretable temporal features. Moreover, we evaluate the base models with multiple metrics. Using the best three models, our method will ensemble hybrid models and choose the best hybrid technique. Further, we investigate the impact of lag features and how the models work with different outlier detection techniques such as Interquartile Range (IQR) and anomaly detection. The models are evaluated using regression metrics (MAE, MSE, RMSE, MAPE, and R2). Experiment results show that, after removing anomalies, MAE, MSE, RMSE, and MAPE are reduced. Adding lag features also shows some positive effects on the results. Moreover, anomaly detection techniques perform better than IQR in improving performance while preserving data. The reported anomaly-based cross-validation results favored the voting ensemble over the compared base models on the evaluated metrics. In anomaly-based cross-validation, VOTE improved prediction errors compared with MLP (256,128,64); improvements varied in the single-split comparison. Specifically, the hybrid model reduced MAE by 9.82%, MSE by 3.12%, RMSE by 1.76%, and MAPE by 6.77%. In the reported cross-validation comparison, the best anomaly-based hybrid model had lower MAE, MSE, RMSE, and MAPE than the best IQR-based hybrid model by 9.81%, 34.30%, 18.99%, and 0.66%, respectively.

The shear piezoelectric coefficient $e_{15}$ governs the shear-mode electromechanical coupling in piezoelectric micromachined ultrasonic transducers (PMUTs) and surface acoustic wave (SAW) devices, but measuring it typically requires destructive resonance testing on dedicated specimens. On simulated data, $e_{15}$ can instead be recovered non-destructively from sparse, noisy subsurface wavefield observations using a constrained physics-informed neural network (PINN). By building exact spatial distance functions and temporal multipliers directly into the network output, the clamped, grounded boundaries and resting initial state hold automatically, leaving the loss with only the interior equations of coupled electro-elastodynamics and the data misfit. Non-dimensional scaling brings all governing coefficients within one decade of O(1), and the setup removes the boundary and initial-condition penalty weights that soft-constraint methods must tune. In a six-way comparison (plane-strain PZT-5H, manufactured wavefield, 30 dB noise), a soft-penalty baseline with adaptive weights converges to an erroneous $e_{15}$ in four of the five runs (51.7% mean error), and fixed penalties reach an error of 0.85% only after a second-order optimization stage. The constrained formulations reach mean $e_{15}$ errors between 0.15% and 0.51% with held-out displacement errors below 1%. Within these variants, the spatial multiplier gives rapid first-order convergence, and the quadratic temporal multiplier lowers the early-time electrostatic residual by a factor of 6.8 versus a linear one. On finite-element data along one subsurface line, the full formulation recovers $e_{15}$ with 1.19% mean error at 30 dB and keeps it between 0.65% and 2.71% across a 40 to 0 dB noise sweep. Finally, training on displacements alone still recovers $e_{15}$ with mean errors of 0.45% on finite-element and 1.64% on manufactured data.

Convexity has three equivalent descriptions: support by affine functions, monotonicity of the subdifferential, and increasing secant slopes. We weaken all three by one distortion with parameter $γ\in(0,1]$, which multiplies positive changes by $γ$ and divides negative ones by $γ$. The resulting classes are the convex functions at $γ=1$ and, for $γ<1$, are distinct and nested, with the parameter dropping by at most a square along the chain. For proper lower semicontinuous functions in these classes, the effective domain is convex, local minimizers are global, and the function is locally Lipschitz on the relative interior of its domain. The classes separate offline from online optimization over compact convex sets with efficient projection, with access to function values and bounded support vectors. A smooth member of any class can be minimized to additive accuracy $\varepsilon$ with $\mathrm{poly}(1/\varepsilon)$ first-order queries and projections. Online, with nonpositive losses and $α$-regret measured against $α$ times the best fixed point, a polynomial-time algorithm achieves $γ$-regret $O(\sqrt T)$ on the largest class. For every fixed, effectively specified $γ\in(0,1)$ and every $α\in(γ,1]$, no randomized polynomial-time algorithm has sublinear expected $α$-regret unless $\mathbf{NP}=\mathbf{RP}$, already for smooth concave polynomial losses in the smallest class, each minimized by one linear optimization. Thus, unless $\mathbf{NP}=\mathbf{RP}$, $γ$ is the exact online coefficient of each class, while the offline coefficient is one.

As control engineering is applied to increasingly complex systems, learning-based control, which uses empirical data to reduce the reliance on first-principles models, has emerged as a promising paradigm. However, such approaches must cope with nonlinear and stochastic dynamics, the absence of state measurements, and scarce data, while ensuring safe operation under constraints, which requires rigorous uncertainty quantification. Whereas existing methods treat these challenges in isolation, this article proposes a Bayesian framework that jointly addresses them. A prior in state-space form is updated from input-output data using particle Markov chain Monte Carlo methods, yielding a representation of joint uncertainty over dynamics and latent states. Building on this representation, a validation procedure is developed that provides probabilistic guarantees for constraint satisfaction under arbitrary causal control laws. For control synthesis, a formulation penalizing the conditional value-at-risk of the constraint function is proposed, together with an adaptive calibration mechanism that adjusts the performance-robustness trade-off to achieve a desired robustness level. The probability of successful certification is characterized theoretically. The framework is extended to a receding-horizon setting, where posterior updates from new measurements allow the controller to progressively reduce uncertainty and improve performance. The approach is demonstrated on the optimal control of a vertical farm.

Hybrid models and neural differential equations (NDEs) are becoming increasingly important for modelling physical systems; however, they often encounter stability and accuracy issues during long-term integration. Training on unrolled trajectories is known to limit these divergences but quickly becomes too computationally expensive due to the need to compute gradients over an iterative process, making this method impractical for large-scale physical systems (e.g. weather and climate forecasting). Two popular paradigms for improving the stability of learned dynamics are to train on interpolated or noisy samples or to regularize the Jacobian of the learned representation. In this work, we demonstrate that these paradigms can be understood under a common lens and that they both correspond to an approximation of the gradients obtained in an unrolled learning strategy. Building on this understanding, we derive a principled and complete approach to regularize training and demonstrate its impact on the stability of learned dynamical systems.

Discrete diffusion models remain far less understood than their continuous counterparts, despite recent breakthroughs across modalities, most notably in text. Leveraging tools from continuous diffusion, we formulate a theoretical framework to track the class-conditional posterior in an online manner by combining conditional and unconditional model predictions. In masked language modeling, this posterior offers highly interpretable dynamics: tokens aligned with a class boost the posterior, while contradictory ones suppress it. We also show that the posterior can be used for logical guidance operations (AND, OR, NOT) in discrete diffusion. These operations enable precise control over generation, such as avoiding specific classes. In addition, we discuss a mismatch in the formulation of classifier-free guidance between discrete and continuous settings: discrete models mix the data distributions, while continuous models mix the marginals. We validate our methods and discuss their potential applications on synthetic data, with all quantities analytically tractable, as well as on MNIST, text, and antibody sequences.

Hamiltonian neural networks provide a principled framework for learning conservative dynamics. A scalar Hamiltonian is parameterized by a neural network and the vector field is derived using Hamilton's equations. However, preserving Hamiltonian structure does not guarantee that the energy function itself is represented accurately. When modeled by a conventional multilayer perceptron, the Hamiltonian inherits the spectral bias of coordinate-based neural representations, favoring smooth over fine-scale structure. These fine-scale errors can be amplified by differentiation when computing the dynamics. To address this limitation, we introduce Local Multiscale Hamiltonian Neural Network (LM-HNN), which restructures the Hamiltonian according to its physical interaction structure. LM-HNN decomposes the energy into kinetic and potential components and represents them as sums of shared local functions over multiple interaction neighborhoods. This replaces a single high-dimensional approximation problem with lower-dimensional, reusable interaction laws, making fine-scale energy structure easier to resolve while preserving Hamiltonian dynamics. Across low-dimensional, interacting, and fast--slow systems, LM-HNN consistently achieves more accurate Hamiltonian derivatives and long-horizon rollouts than competing structure-preserving models, while substantially extending the effective spectral range of the learned energy. The same structured representation also generalizes more reliably to energy levels not observed during training. These results show that accuracy depends on the Hamiltonian representation as well as its structure, and that incorporating locality, decomposition, and shared interaction laws provides a simple and effective route to more accurate and generalizable Hamiltonian learning.

Machine unlearning seeks to remove the influence of deleted training data without full retraining. In the sequential setting, deletion requests arrive over time, and repeated updates can accumulate approximation errors. In this paper, we study error accumulation in high dimensions, where parameter count $p$ is proportional to sample size $n$, for second-order unlearning methods under two privacy settings: (i) \emph{partial unlearning}, which allows unpublished internal states to be retained while certified versions of the model are released; and (ii) \emph{perfect unlearning}, which retains only published models and no hidden states. We also distinguish pointwise and pathwise certification, depending on whether the certification guarantee applies to each released model or the whole released sequence. We analyze conjugate gradient (CG) as an efficient solver for the inverse Hessian-vector product (IHVP) in Newton iterations. Our theory shows that a maximum of 1/2/2/3 Newton--CG updates suffice for successful certification and vanishing population-risk degradation of the published model, for pointwise partial, pathwise partial, pointwise perfect and pathwise perfect certification, respectively. Numerical experiments demonstrate accurate sequential tracking and substantial computational gains of Newton-based methods combined with conjugate gradient, over gradient-based updates, especially for ill-conditioned problems.

We study the gradient query complexity of sampling from a density $π(x)\propto\mathrm{e}^{-V(x)}$ on $\mathbb{R}^d$, where $V$ has a $β$-Lipschitz gradient but need not be convex. Our approach uses the proximal sampler, which alternates Gaussian perturbation with sampling a position conditional on the noisy observation. We implement its observation chain using the high-order noisy conditional sampler of [OpenAI 2026], obtaining improved dimension dependence under functional inequalities. If $π$ satisfies a log-Sobolev inequality (LSI) with constant $α$, then for every fixed $ν>0$, there is a finite constant $C_ν$ depending only on $ν$ such that the algorithm returns a sample within total-variation distance $\varepsilon$ using at most $C_νκ(d+κ+\varepsilon^{-1}+\|\nabla V(x_0)\|^2/β)^ν$ queries, where $κ=β/α$ and $x_0$ is an arbitrary reference point for initialization. The analysis controls center-dependent local errors along exact proximal histories and removes the final observation noise by conditional decoding. Under a Poincaré inequality (PI), we obtain a corresponding guarantee from a warm start with a known bound on its Rényi divergence. An oracle lower bound shows that a stationary point alone, which suffices under LSI, is insufficient under PI.

Generative models increasingly simulate users at different skill levels in education and AI evaluation. Yet it remains unclear whether lower-skill behavior reflects less internal knowledge or a different use of retained knowledge. This distinction is difficult to study in natural language, where proficiency and relevant knowledge are hard to measure. In this work, we study Maia-2, a chess transformer that predicts human moves conditioned on player skill, using quantitative skill ratings and deterministically labeled chess concepts. We hold board positions and model weights fixed and analyze positions where skill conditioning changes the prediction from a blunder to the best move. First, we measure encoded knowledge by training probes to predict the presence of 172 chess concepts from Maia-2's internal states. Probe accuracy remains nearly constant across skill conditions, and probes trained at one level transfer accurately to the others. To test whether stronger decisions can be recovered from this retained knowledge, we fine-tune only Maia-2's final output layer, raising best-move accuracy in the lowest skill band from 14.4% to 82.8%. Training this layer only on highest-skill representations retains most of the improvement at lower levels, showing that recovery does not require direct training on lower-skill states. Steering concept-relevant features learned by a sparse autoencoder also improves lower-skill decisions while leaving probe-measured concept encoding practically unchanged. These results support an account in which skill conditioning changes how retained concept knowledge influences move selection, showing that weaker behavior can understate the knowledge available within a model.

Despite their widespread success across machine learning, the learning dynamics and generalization properties of neural networks remain poorly understood. The neural tangent kernel (NTK) provides an analytical characterization of the learning dynamics of infinitely wide neural networks and serves as a useful proxy for finite-width networks in supervised learning. In this work, we extend the NTK framework to reinforcement learning (RL) by introducing a temporal-difference (TD) flow formulation that accommodates time-varying mini-batches and TD targets, yielding a closed-form solution for value function evolution in an overparametrized setting. We also devise an optimization-based state-space exploration method that leverages the closed-form solution, coupled with constraints on the environment, to model the set of feasible states. On a deterministic grid environment, we validate that the proposed RL-based NTK formulation is indeed learning efficiently and adapting to the distribution shift in the training data distribution, and we demonstrate that the resulting exploration strategy improves sample efficiency. To our knowledge, this is the first work to extend NTK theory to characterize the learning dynamics of RL training with time-varying mini-batches and TD updates.

Government intelligence teams need LLM agents that can analyze large, messy evidence collections, but security and cost constraints often rule out hosted frontier models and broad, hard-to-audit tool surfaces. We report CLEO: Constrained Loop for Environment Optimization, a method for fitting a minimal agent harness to a fixed, self-hostable LLM. Given an expert-validated dataset, CLEO admits only the tools, guardrails, memory policies, and loop components that improve accuracy while remaining general and auditable. On our held-out dataset, the CLEO harness reaches an 83.8% pass rate, within 1.2 points of the strongest out-of-the-box Claude Code SDK baseline, while being two orders of magnitude smaller.

Graph machine learning benchmarking results are scattered across papers that use quite different sets of datasets and baseline methods, making it difficult to tell whether a reported gain reflects architectural progress or merely an outdated comparison or weakly tuned baseline. We introduce a living, graph-learning-specific leaderboard that organizes reported results by dataset, putting new results into context. As of 15 October 2026 our LLM-based extraction pipeline extracted 27,723 eligible results across 111 datasets from 885 peer-reviewed papers. Our analysis of these results quantifies concerns about graph-learning benchmarking: legacy datasets are still actively used and many result tables contain baseline results that are drastically lower than those already available in the literature, e.g. 45% Amazon-ratings tables and 37% for Peptides-struct. The leaderboard allows to quickly compare new results against compatible prior results, aiming to improve the overall quality of graph learning benchmarking.

Research taste, the judgment of what to read, which problems to pursue, and which methods to try, increasingly shapes the work of AI scientists, yet it remains far less understood than their ability to execute research. We introduce IdeaTaste, a benchmark that scores an LLM by how closely the research proposals it writes from a literature snapshot frozen at 2025 match the work the same communities published in 2026. Our experiments span six contemporary language models and four research areas from established to emerging fields, covering more than 90{,}000 proposals in total. LLM agents fall well short of human communities in both the literature they cite and the ideas they state, reaching about 26\% of the human level on citation matching and only 18\% on idea matching. Their populations can cite as diversely as a community and take up its tasks and methods, yet rarely arrive at the concrete approaches and research plans it pursues. Neither richer field context nor larger populations substantially close this gap. In blinded expert review, AI proposals trail the 2026 papers least in presentation and most in expected success, echoing their gap in concrete approaches, and those whose ideas no human paper shares fare worst, suggesting weaker directions rather than overlooked opportunities. These results indicate that the main bottleneck in AI research taste lies less in finding the directions a community works on than in turning them into the concrete approaches it goes on to pursue. We release IdeaTaste as an open, continuously updated leaderboard (https://cxcscmu.github.io/IdeaTaste) that benchmarks contemporary AI proposals against future publication cycles.

Deep learning models for electronic health records achieve strong predictive performance but provide little insight into what clinical concepts they represent internally. Post-hoc attribution methods produce per-cell importance scores that lack temporal coherence, cross-variable structure, and reusability across patients. We introduce the Temporal Island Sparse Autoencoder (TI-SAE), which learns a backbone-specific, patient-shared clinical concept dictionary from a frozen predictive model's representations. Each dictionary entry (latent) is defined by learned temporal Gaussian islands (specifying when the concept is active) and a sparse variable-composition vector (specifying which variables are involved). A decoupled selector then identifies which dictionary entries drive each patient's prediction, separating representation learning from attribution. We evaluate TI-SAE across five backbone architectures spanning different representation strategies on MIMIC-III and MIMIC-IV (three prediction tasks each). TI-SAE produces structured clinical concepts that go beyond transient attribution heatmaps while achieving strong performance on standard faithfulness benchmarks.

Optimal control of partially observable Markov decision processes (POMDPs), in general, requires a complete history. The belief state, a distribution over hidden states, summarizes this history and provides the classical sufficient statistic to control POMDPs. However, the continuous nature of the belief space generally prevents a finite-memory controller from maintaining it exactly. This raises the question we address in this work: what must the available finite memory preserve to achieve near-optimal control performance? We study this question through an information-theoretic converse that gives a fundamental lower bound on the loss of any finite-memory controller. The analysis introduces the concept of a decision witness that characterizes belief states whose confusion would lead to unavoidable loss. Motivated by this bound, we propose RECAP, a constructive quantization method for designing representative memory states. We prove an upper bound on the optimality gap of any RECAP codebook and use the components of this bound together with the converse analysis to define the codebook objective. Experiments on 11 POMDP benchmarks demonstrate that RECAP improves over finite-memory representation baselines under the same memory budget.

LLM agents are judged by what they deliver, and in professional work what gets delivered is a project, not a task. A project pays off only if the whole package holds, and one weak part can limit the value of the rest. Yet existing benchmarks mostly test agents on one task at a time, leaving a gap between the capability we measure and the outcome that makes agents useful. To address this gap, we introduce DeliverBench, which tests whether agents can carry a project from brief to final delivery. Its 11 projects, spanning financial due diligence, legal contract drafting, and drug safety reporting, are reconstructed with 13 domain experts from real materials and chain 150 subtasks through 2-15 rounds of new materials and client feedback. Because no reference outputs are injected, agents inherit their own errors from one task to the next. Expert rubrics then score every checkpoint as well as the final package. Across seven model-harness systems, even the best system scores only 65.0 out of 100 on final delivery and passes 50.0% of subtasks at a 90-point threshold. The failures are often failures of carrying work forward: in one audit project, all seven systems addressed the review findings, but only one preserved traceability. Together, these results suggest that task-level capability does not simply compose into delivery, which also depends on keeping the work coherent as it evolves.

Spherical-harmonic neural operators provide an efficient representation for global forecasting, but their discretized pointwise nonlinearities can break rotational equivariance through aliasing. Clebsch-Gordan (CG) tensor products provide exactly equivariant interactions between spherical-harmonic representations, but their computational cost grows rapidly with spectral bandwidth and has limited their use at forecasting scale. We introduce LCG-SFNO, which combines low-rank CG interactions within a controlled low-degree band with coefficient-space residual propagation and an exactly projected quadratic nonlinearity, preserving SO(3) equivariance throughout the learned operator. To make the CG interaction practical, we further develop SSHCG, a sparse GPU implementation with a factorized alternative for wider coupled bands. On spherical shallow-water dynamics and the ERA5-based WeatherBench benchmark, LCG-SFNO reduces rotational error to numerical precision (more than three orders of magnitude below SFNO on shallow water and five orders of magnitude below it on WeatherBench), while matching SFNO on shallow-water forecasting and substantially improving forecast skill relative to SFNO on WeatherBench. Matched ablations show that CG coupling accelerates convergence, while SSHCG keeps the coupling to about 10 percent of a training step at the operating bandwidth, making exact equivariant coupling practical at forecasting scale.

Next-generation open radio access network (Open RAN) and AI-RAN deployments turn the RAN into a disaggregated, compute-aware platform whose service continuity depends on many control loops acting at different timescales. Resilience in this setting is best seen as a lifecycle spanning anticipation, absorption, adaptation, recovery, and evolution. Yet, existing mechanisms remain stage-specific and are rarely coordinated across the control hierarchy. This paper argues that agentic AI offers an orchestration layer that coordinates this lifecycle, turning operator intent into coordinated action while keeping slow language-model reasoning off the fast control path. Using mass-attachment signaling storms as a running study, we build a discrete-event Open RAN attach simulator that captures retry amplification, shared-compute contention, and non-instantaneous provisioning, and we propose a three-tier framework in which an orchestrator at the service management and orchestration layer sets a standing policy, a per-site non-real-time agent judges storms and anticipates scheduled crowds, and a deterministic near-real-time loop enforces every decision. Across experiments, the agentic controller approaches the resilience of a fully provisioned pool of servers at a quarter of its capacity by reading a scheduled event from a calendar and provisioning ahead of it. No reactive weight setting recovers this once the surge arrives. Free-text operator intents ground into control directives at 94% exact match on previously unseen intents, and a single instruction reshapes live behavior end to end. We also quantify the latency, cost, and contention limits that bound it. The supplementary material provides analytical proof of the derived Theorem on anticipation horizon, as well as the subsequent Corollary results on reactive provisioning limits.

Super-resolution networks are trained for one magnification factor, so several output resolutions require several models. We propose PMS-SR, a progressive single-input multiple-output network that produces x2, x4, and x8 results in one forward pass, with a Transformer backbone on the low-resolution grid. In Stage 1, each scale is predicted as a residual over the detached bicubic upsampling of the previous scale's prediction. With Dirac and zero initialization, training starts exactly from the bicubic cascade, and each head adds only the band the previous scale lacks. In Stage 2, inter-scale mutual refinement couples the scales in both directions through space-to-depth feedback, deformable alignment, and per-pixel fusion, so x8 features refine the x2 output. Trained on DF2K, the 7.95M-parameter model achieves a higher PSNR than every compared multi-scale method at x4 and x8 on all five standard benchmarks. At x8, it outperforms the x8-specific DRN-L on all of them, leads all compared methods on Urban100, and is within 0.13 dB of the best result elsewhere. Against three single-scale SwinIR-M models, PMS-SR uses 22% of the parameters, runs 2.9x faster on BSD100 inputs, and its outputs agree 6.7-12.4 dB more closely. Code and pretrained models will be released at https://github.com/julian135707/PMS-SR upon acceptance.

Technology-mediated audience participation (TMAP) enables audiences to influence live performances, yet balancing interaction with meaningful agency remains challenging. While AI-generated music and MR enable new forms of audio and visual audience participation, little is known about how these participation modalities influence audience experiences during live musical performance. In particular, it remains unclear how they shape perceived agency, collaboration, and engagement, and how they complement one another within TMAP systems. We present CrowdForge, an AI- and MR-mediated paradigm where audiences collectively control generative music and virtual environments during a live performance featuring a guitarist and DJ. Through a mixedmethods concert study evaluating three interaction conditions (audio, visual, and combined audio-visual), we demonstrate that observability is the primary driver of audience agency. While visual interaction provided easily observable feedback, audio parameters offered macro-structural control over the music. Audiences tended to use the control enabled by AI audio generation to test musical boundaries, introducing a productive adversarial tension between the crowd and the musicians. This dynamic pushed creative limits, deepened engagement, and generated moments of musical coherence when performers successfully resolved audience-induced challenges. We argue future TMAP systems should prioritize human-centered, highly observable interactions to build agency, while combining the expressiveness of music with the clarity of visual feedback.

Best-of-N decoding improves generation quality by sampling multiple candidates and selecting the best, but it allocates equal compute to sources regardless of how much they benefit from additional sampling: the expected benefit is small for sources with low-variance score distributions and potentially much larger for those with high-variance distributions. To address this shortcoming, we propose LAND (Learning an Adaptive Number of Draws), which dynamically allocates samples according to their posterior-predictive marginal value. LAND learns a discrete empirical mixture of source-conditioned score distributions from calibration data, updates its belief about a source as completion scores are observed, and stops sampling when the expected improvement of another draw falls below a shared compute price. Across machine-translation experiments with multiple models and language pairs, LAND achieves better MetricX scores than fixed best-of-N at matched average compute. Conversely, at matched translation quality, LAND requires approximately 20-40% fewer candidate generations. Adaptive allocation therefore provides a simple way to retain the benefits of best-of-N while substantially improving its compute efficiency.

Financial research agents can generate and monitor trading rules faster than a human can document the resulting selection process. Candidates appear at data-dependent times, weak rules are abandoned, and promising ones are watched until they cross a reporting threshold. We introduce PROMOTE, a protocol that assigns every candidate predictable testing wealth at birth and accumulates a cost-aware e-process only from subsequent returns. If lifetime allocations sum to at most one, a weighted Ville argument bounds the probability of ever promoting a rule whose declared bounded net score has always had nonpositive conditional mean. The result permits dependent candidates, adaptive births, and indefinite monitoring. We audit the implementation in two settings. A retrospective pseudo-out-of-sample replay contains 24 panel--rule hypotheses, 3,000 shared-sign null paths, injected alternatives, and public ETF and industry data. Repeated unadjusted tests produce a family crossing in 94.5% of null paths, versus 0.0% for PROMOTE; at a 15-bp aligned ETF-component perturbation, PROMOTE has 70.1% target power and median crossing session 2,327. A separate exact mean-zero experiment lets an agent select candidate identities and birth times from past scores; the funded gate has 1.77% false-family crossings across 3,000 programs. The market replay produces zero promotions after costs. The certificate controls historical false promotion under the declared null; it is not a certificate that an earlier edge remains profitable at deployment. AI-use disclosure. Generative AI tools, including OpenAI Codex, assisted manuscript preparation. For this preprint version, Codex assisted with LaTeX editing, author and venue metadata, compilation, and layout checks.

Physically based two-dimensional hydraulic models are too slow for real-time inundation forecasting and large ensembles. We present a time-dependent geometry-informed neural operator (GINO) surrogate of LISFLOOD-FP for the municipality of Forlì (Italy), severely affected by the May 2023 Emilia-Romagna floods. The surrogate is designed for an operational setting in which only river stage gauges and rain gauges are available. A reproducible pipeline turns public regional observations into a physically based training set: each river inflow is driven by the first stage gauge downstream of it through a calibrated water-surface boundary condition, and rain enters as SCS-CN effective rainfall. The surrogate is an autoregressive GINO trained on 216 simulations of 24 observed events with scaled forcing. It predicts volume-conserving polygon-mean depths on an exposure-based Voronoi mesh of 659 nodes, with a 20 min step over 72 h. The surrogate is tested on 38 independent observed events that it never saw during training. On these events it reaches a critical success index (CSI) of 0.934 at 0.05m and 0.844 at 0.3 m, a depth RMSE of 0.008m and an $R^2$ of 0.984 for maximum-depth maps. On one NVIDIA A10G GPU a 72-h forecast takes 1.7 s instead of 55 s with LISFLOOD-FP, and a 50-member ensemble, computed in a single batched rollout, takes 6.2 s instead of about 46 min, which makes probabilistic flood forecasting practical. These results show that neural-operator surrogates can deliver accurate pluvial-fluvial flood forecasts driven by river stage for operational early warning.

Identifying ligands for a protein target requires searching a vast chemical space for a comparatively small set of compatible molecules. Existing machine learning approaches primarily address this problem by generating candidate ligands or scoring individual protein-ligand pairs, but provide limited information about the broader chemical space compatible with a binding pocket. We introduce RangeFinder, an SE(3)-equivariant transformer pocket-level model, trained and evaluated on BioLiP2 data, that predicts ranges over nine physicochemical ligand properties directly from pocket structure, chemistry, and evolutionary context. Using an intersection-over-union objective and evaluation metric (range-IoU), RangeFinder achieves a range-IoU of $0.464 \pm 0.002$, outperforming all deployable baselines while requiring only $12.2$ ms per pocket. We also quantify our method's ability to reduce the ligand search space while retaining compatible binders; in this regard, RangeFinder achieves a Reduction-Recall F-Score of $0.543 \pm 0.011$ --- once again outperforming all deployable baselines. Ablation experiments indicate that RangeFinder's performance depends more strongly on the density and quality of ligand supervision than on any individual input modality or specialized training strategy. Together, these results position RangeFinder as an inexpensive and effective molecular filtering prior, with demonstrated potential in narrowing the ligand search space of target pockets before docking, generation, or synthesis. The code for this project can be found at: https://github.com/alannadels/RangeFinder.

Every multimodal-LLM multiple-choice benchmark ends with a scoring rule that converts a model's response, often free-form text rather than a clean option letter, into a right/wrong verdict, buried in its code. Unlike the dataset, this rule is chosen without justification and left unexamined, yet it sets the published number. We reproduce the official scorers of eight peer-reviewed benchmarks on five open vision-language models (VLMs) that fit a single 16 GB GPU, and introduce a blinded human-plus-LLM-judge adjudication (Cohen's kappa=0.956) that recovers what the model actually answered. The scoring rule alone moves accuracy by up to 22 points on identical outputs, yet exactly 0 wherever a model emits clean option letters, so it hides on strong-model leaderboards. It is the fallback and not the rambling that does the damage: the benchmark with the highest non-compliance we measure has an artifact of exactly zero, because its scorer abstains instead of guessing; a reasoning model with non-compliance 0.000 still swings 9.5 points with which regex match of its trace the parser keeps. No standard rule is reliable: snapping a non-answer to its closest option credits answers never given yet buries correct ones, abstention discards answers stated in words, and likelihood scoring follows answer-token priors, so a single grade conflates capability with luck rather than measuring either. We release all code, labels, and the protocol, and argue benchmarks should report a non-compliance rate and at least two rules, not one unexamined grade.

Crime-reporting mobile applications can enhance community safety and engagement in urban areas, yet their adoption remains low, partly because developers prioritize technological features over residents' social realities. This study used an exploratory sequential mixed-methods design to identify and validate the sociotechnical factors that drive adoption. Two focus groups and an analysis of 660 app reviews revealed that users' initial sense of usefulness fades, whereas sustained engagement depends on perceived value (PV), comprising six dimensions: safety, anonymity, recency, autonomy, camaraderie, and accessibility. A survey of 505 urban residents, analyzed with PLS-SEM, validated PV as a formative higher-order construct formed by four of these dimensions (safety, anonymity, autonomy, and camaraderie). PV was the strongest predictor of behavioral intention, ahead of effort expectancy, facilitating conditions, and performance expectancy, and behavioral intention strongly predicted actual use. The model explained 75% of the variance in behavioral intention and 51% in actual use. The findings offer design principles for community-centered crime-reporting technologies.

Recent language model (LM) safety failures have largely been studied as disparate phenomena. In agentic misalignment, a model blackmails an executive to avoid being replaced. In alignment faking, it complies with harmful requests during training to keep its values from being modified. We argue that these behaviours share a common decision structure: a model assigned a goal must either take a harmful action to preserve it or refuse and lose it. We formalize this as a goal-harm conflict and present the same conflict under various elicitation protocols that range from compact forced-choice prompts to simulated agentic environments. These protocols span two perspectives on LM decision-making, which we term (i) the logit view, reading preferences from answer-token probabilities, and (ii) the sampling view, classifying open-ended responses with an LM judge. Under the logit view, we show that goal importance and harm severity additively shape how a model resolves these value conflicts. Under the sampling view, we derive two factorizations of the harmful-action probability that safety evaluations commonly report. The first measures how well a preference read from answer tokens carries over into behaviour, such as making a harmful tool call. The second attributes a low harmful-action rate either to refusal or to responses that reach no resolution. We then show how fictional context changes a model's preference under the logit view, and how tool use and partial observability -- two features of agentic evaluations -- change its harmful-action rate under the sampling view. Our work offers a descriptive framework for studying how LMs resolve such value conflicts and for estimating low-probability harmful actions at a finer granularity than the aggregate harm rates commonly reported in safety evaluations.

**Purpose:** Ultrasound biometry is widely used across various clinical applications, including prenatal and cardiac examinations, to obtain quantitative measurements by placing calipers at anatomical landmarks. However, these measurements remain subject to interoperator variability as they rely largely on manual caliper placement. Conventional methods often use specialized prediction heads for different ultrasound tasks, whereas vision-language models offer flexibility through text instructions. However, direct coordinate prediction does not explicitly model spatial likelihood distributions, limiting the representation of localization ambiguity. **Methods:** We propose Kev-Bio, a vision-language model that uses clinical measurement questions to guide ultrasound landmark detection. Inspired by System One decision models, Kev-Bio uses decision representations of the language model to generate queries for each caliper endpoint. These queries are matched against shared dense ultrasound features to produce spatial probability maps, from which landmark coordinates and clinical measurements are derived. The model is trained end-to-end using both landmark localization and clinical measurement losses. **Results:** On the nine-view FU/GU Biometry 2026 dataset, Kev-Bio achieves a mean radial error (MRE) of 22.50 pixels and a clinical measurement mean absolute error (MAE) of 15.38, outperforming the compared baselines in overall performance. **Conclusion:** These results demonstrate the effectiveness of Kev-Bio for ultrasound landmark detection and clinical measurement across different anatomical views. Our code and models are available at https://github.com/moeinheidari7829/Kev-Bio.

Suicide prevention remains a critical public health challenge, and electronic health records (EHRs) provide an opportunity to characterize suicide-related risk indicators using documented clinical, behavioral, medical, and social-contextual information. This paper presents a multidimensional machine learning framework to model and estimate suicide-related outcomes using MIMIC-IV EHR data. We construct clinically motivated diagnosis categories and define a composite outcome from documented codes corresponding to suicidal ideation, suicide attempt, prior suicidal behavior, and intentional self-harm. To address severe class imbalance and reduce age-related confounding, we apply age-matched resampling that retains all positive patients and samples comparison-cohort patients from the same age group. Logistic regression, random forest, and XGBoost classifiers are evaluated before and after resampling with isotonic calibration. The calibrated XGBoost model achieves the strongest matched-cohort performance, outperforming the other evaluated classifiers. SHAP-based interpretation shows that diagnostic burden, depression-related features, age, personality-disorder history, social determinants related to violence, and medical comorbidity contribute to elevated predicted probabilities. Our results demonstrate that combining broad EHR-derived features, age-matched cohort construction, calibration, and feature-level interpretation provides a practical framework for retrospective suicide-related risk modeling.

Actor--critic methods rely on critic gradients to improve continuous-action policies, yet finite-sample guarantees must control both value and gradient errors and their propagation through the actor. We develop a neural actor-critic procedure that combines a regularized minimax Bellman estimator over sigmoid networks with a finite reflected Langevin sampler driven by the fitted critic gradient. Under uniform Bellman regularity and covered reset sampling, a positive high-order action-derivative penalty yields value and action-gradient risk bounds with classical nonparametric minimax exponents up to logarithmic factors. We propagate estimation, mixing, discretization, temperature, and policy-iteration errors to obtain return guarantees, with refinements for bounded multimodal potentials and strongly concave regimes. Empirical studies across controlled simulations, logged continuous-control data, and a simulated decision task show favorable performance relative to the considered baselines.

Robust machine learning for graph-structured data has made significant progress against test-time attacks, yet certified robustness to poisoning - where adversaries manipulate the training data - remains largely underexplored. For image data, state-of-the-art poisoning certificates rely on partitioning-and-aggregation schemes. However, we show that these methods fail when applied in the graph domain due to the inherent label and structure sparsity found in common graph datasets, making effective graph-partitioning difficult. To address this challenge, we propose a novel semi-supervised learning framework called deep Self-Training Graph Partition Aggregation (ST-GPA), which enriches each graph partition with informative pseudo-labels and synthetic edges, enabling effective certification against node-label and graph-structure poisoning under sparse conditions. Our method is architecture-agnostic, scales to large numbers of partitions, and consistently and significantly improves robustness guarantees against both label and structure poisoning across multiple benchmarks, while maintaining strong clean accuracy. Overall, our results establish a promising direction for certifiably robust learning on graph-structured data against poisoning under sparse conditions.

While structural and functional magnetic resonance imaging (sMRI, fMRI) offer profound potential for researching autism spectrum disorder (ASD) and understanding individual neurodivergence, current classifiers lack the explanatory power needed for scientific study, clinical use and personal-level insights. Most models learn from imaging features that also vary with age, sex, scanner site and head motion. They fuse the two modalities without a common reference and explain decisions against arbitrary reference inputs rather than a typical brain. We present NormMoX, a normative mixture of experts that makes the typical brain the shared reference for prediction, fusion and explanation. Normative models trained on large healthy cohorts give the value each brain measure is expected to take for a person's age and sex. NormMoX learns from each person's deviations from these expected values. For each modality, a neural network expert turns the deviations into a score of evidence for ASD. A fusion head, MoX, compares each score with those of typical controls before combining them. It therefore needs no imputation when a modality is missing and shows in which modality each person is atypical. On 1,035 participants of ABIDE I, NormMoX reached an AUROC of $0.728 \pm 0.004$ in repeated cross-validation, above linear, single-modality and deep fusion baselines. NormMoX thus shows not only whether a person is likely to have ASD, but where, and in which modality, their brain departs from typical. The implementation is available at https://github.com/Novelex/NormMox.

Attention remains a substantial computational cost in Transformer training, particularly at long context lengths. NVIDIA Blackwell GPUs provide high-throughput low-precision matrix multiplication, but exploiting it in attention requires addressing quantization overhead, causal information leakage through shared scales, and numerical inconsistencies between the forward and backward passes. We present FLAMIA, a family of attention kernels for Blackwell with MXFP8 and mixed NVFP4/FP8 training paths. Our design combines token-block scaling that reuses quantized operands across matrix orientations, targeted safeguards against scale-induced leakage, and consistent forward and backward computations. Fused quantization and architecture-specific kernel optimizations reduce overhead, while document-local scaling and workload-aware scheduling support variable-length documents directly in packed layouts without repacking. Across various causal packed-document workloads, the MXFP8 path accelerates the combined forward and backward attention call, including quantization, by factors of $1.10$--$1.40\times$ on B200 and $1.06$--$1.42\times$ on B300 relative to the fastest evaluated baseline for each workload. We validate training stability and quality preservation on pretraining Llama-style models. These results show how jointly designing quantization, training numerics, and document scheduling translates low-bit hardware throughput into practical attention acceleration.

Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs, and the most accurate SOTA model is a fit, not a law: it absorbs architecture into fitted coefficients, requiring per-LLM fitting. We derive closed-form expressions of the computational and memory-access complexity of LLM inference in terms of architectural hyperparameters, covering MHA and GQA, and compare a family of analytical models of increasing granularity. From these, we obtain the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache cost and next-token computation. Over traces of 13 dense LLMs from six families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. The SIMPLE law explains as well as predicts: prefill energy grows quadratically with the hidden dimension, so larger LLMs suffer most from long inputs; and more aggressive GQA KV grouping reduces the per-token energy by up to 3x relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; although accuracy grows with the budget, energy grows faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x reasoning-budget range. Agentic serving raises energy per request by 13.0x relative to non-agentic ones, driven by the growth of both the prefill and the KV-cache terms with input length; while prefix caching removes most of the prefill cost (-93%), the KV-cache term is unaffected, so even a cached agentic request costs 4.8x more.

Artificial intelligence (AI) technologies and their capabilities have advanced rapidly over the past few years, leading to widespread interest and adoption in the biological and ecological sciences. However, the data infrastructure and computational resources that provide the foundation for AI workflows in ecology and biodiversity science remain fragmented. Moreover, the flow of value remains largely unidirectional: data move from collectors, through repositories and aggregators into models, and finally to research products or management actions. The derived value, such as corrected taxonomies, annotations, quality flags, and insights, rarely returns to source communities. We propose an extension of the original FAIR principles to address the needs of AI-ready infrastructure: the AI-FAIR metadata criteria, which place AI-enabled science in the context of a bidirectional Question-Data-Model triangle. This framework is designed to facilitate large-scale content-based analysis across diverse collections and ensure that the scientists generating the data benefit from the AI-enabled advances built on their data. Alongside these criteria, we introduce an agentic evaluation workflow that both rates existing datasets on the AI-FAIR and traditional FAIR criteria and provides actionable feedback where critical metadata is missing or incomplete. This modular dataset evaluation tool, paired with a domain-appropriate checklist, translates these principles into a reusable assessment of machine-learning readiness, AI-readiness for specified tasks, traceability, and ethical and CARE-aligned stewardship. Through this approach we aim to create a community-informed system to close the model data loop, so that AI also enriches data to further enable scientific discovery.

Transportation is a major source of energy consumption, greenhouse gas emissions, and criteria air pollutants, making efficient emissions estimation essential for environmental planning and policy analyses. The U.S. Environmental Protection Agency's Motor Vehicle Emission Simulator (MOVES) supports detailed emissions modeling, but its computational demands constrain large-scale scenario analyses. This study introduces uTECH Turbo-MOVES, a software architecture that expedites MOVES computation without modifying its Calculator modules or input/output structure. By combining RunSpec partitioning, container isolation, in-memory intermediate storage, and concurrent workload orchestration, it improves execution efficiency. We evaluate six benchmark cases spanning County, Default, and Project scales and both Emission Inventory and Emission Rates modes. Results demonstrate a measured 5.5x and an estimated 15.3x per-run speedup for Default-scale workloads, and an 84.8x campaign speedup for an eight-county Emission Rates analysis. Although gains vary by workload and hardware configuration, uTECH Turbo-MOVES enables more extensive environmental assessments and transportation policy evaluation.

Instruction tuning for multi-task network traffic analysis requires data with clearly defined sources, supervision targets, and evaluation conditions. We introduce TrafficInstruct, a newly collected dataset comprising raw packet captures and 55,000 structured supervised fine-tuning records across ten tasks: malware traffic detection, botnet detection, malicious DNS-over-HTTPS detection, Web attack detection, advanced persistent threat detection, encrypted VPN detection, Tor behavior detection, encrypted application/service classification, website fingerprinting, and application classification under concept drift. An automated workflow captures traffic from controlled software executions, attack experiments, and application activities. Labels are supported by task-specific collection evidence, including process or socket attribution, activity records, and execution outcomes. Filtering, sanitization, and task-specific transformations produce instructions, traffic observations, and structured answers. We verify sample sources and partition the dataset by complete collection sessions, keeping all samples from the same session within one partition. The benchmark defines allowed inputs, deterministic label mappings, fixed training/validation/test partitions, and task-wise scoring rules. We evaluate majority-class, random-forest, and ET-BERT baselines, an unadapted 27B language model, and independently trained task-specific LoRA adapters on a shared frozen base. Adapters are selected using validation data and evaluated on the corresponding test sets. Task-specific adaptation improves classification performance, supporting instruction tuning across diverse traffic-analysis tasks.

Activation steering offers a lightweight mechanism for controlling model behavior, but existing methods depend on sentence-level contrastive data that is unavailable for most of the world's languages. We show that multilingual sentiment lexicons are a viable substitute across three model families: aggregated multilingual lexicons outperform English-derived steering vectors for mid- and low-resource languages, while monolingual lexicons remain stronger for languages a model has specialized in during pre-training. The steering efficacy is mechanistically dictated by both representational density and pre-training priors. Large-magnitude English vectors overwhelm the compact representations of low-resource languages, causing "linguistic collapse." This fragility is mitigated by the norm-dampening of multilingual aggregation. Architectural biases, including bilingual specialization penalty and latent numerical rigidity, also heavily constrain cross-lingual steering responsiveness. Ultimately, our work establishes lexicons as a practical steering signal and offers a comprehensive mechanistic account of cross-lingual steering dynamics.

Previous research shows that vision-language (VL) models struggle to understand fine-grained image details and tend to focus on the overall layout. Moreover, any weakness in English tends to be further amplified in the multilingual setting. To study that, we extend InpaintCOCO, an English dataset focused on fine-grained visual details, to 13 typologically diverse languages across five language families, creating MultiInpaintCOCO - a new multilingual benchmark of 953 minimal-pair examples. We further test several state-of-the-art models in multilingual multimodal retrieval, considering both light VL CLIP-like encoders and large language models. Our results show that performance degrades substantially for lower-resource languages, with the strongest predictor being the data resources available for the languages rather than typological similarity to English. In cross-lingual retrieval, most models exhibit a systematic bias toward English: when given a correct caption in a non-English language and an incorrect one in English, models preferentially select the English caption. This bias varies considerably across architectures. Among the tested models, Gemma 3 shows the strongest English preference while Qwen 3 is the least biased, despite both reaching similar monolingual accuracy across languages. The code is available at https://github.com/ufal/multi-inpaint-coco.

A frontier language model was recently reported to solve 1D NMR structural elucidation without tools. We ask whether this capability transfers to tandem mass spectrometry (MS/MS) in the de novo setting, where the model must propose a structure directly from the spectrum without a candidate list or reference library. To separate reasoning from recall, we build a 100-analyte benchmark from MassSpecGym, split by each analyte's textual footprint into in-distribution (well documented in literature) and out-of-distribution (essentially undocumented) arms. We hold the model, harness, and output format fixed and vary only the LLM's access to domain-specific tools. Without tools, a frontier LLM reaches 52% Top-1 accuracy in-distribution but only 2% out-of-distribution. Only access to de novo structure generators improves out-of-distribution accuracy, raising it to as much as 38% Top-1, and the resulting agent outperforms the standalone de novo models both in and out of distribution. These results suggest that LLMs are not effective at de novo MS/MS structural elucidation on their own, and that further gains depend on the specialist tools available to the agent rather than on the LLM itself.

Calibrated multiaccuracy asks that a predictor be globally calibrated and multiaccurate with respect to a bounded test family $G$; that is, its residuals should be uncorrelated with every test $g(x)\in G$. This weakening of multicalibration suffices for many applications. Its sample complexity was not known to improve on that of full multicalibration, which is $\widetildeΘ\left(\frac{\log|G|}{\varepsilon^2} + \frac{1}{\varepsilon^{3}}\right)$ when $|G|$ is at least a sufficiently large power of $\log(1/\varepsilon)$. We resolve the sample complexity of $\varepsilon$-calibrated multiaccuracy, showing that it scales as $$ \widetildeΘ\left(\frac{\log|G|}{\varepsilon^2} + \frac{1}{\varepsilon^{5/2}}\right), $$ separating it from the complexity of multicalibration. The upper bound is obtained by an algorithm that outputs a randomized predictor. For deterministic predictors, we establish the tight rate $$ \widetildeΘ\left( \frac{\log|G|}{\varepsilon^2} + \frac{(\log|G|)^{1/3}}{\varepsilon^{8/3}} \right) $$ whenever $\log|G|=Ω(\log(1/\varepsilon))$. When $\log|G|\le\varepsilon^{-1+κ}$ for any fixed $κ\in(0,1)$, this rate is polynomially larger than the rate for randomized predictors. Thus, unlike for multicalibration, calibrated multiaccuracy exhibits a genuine sample-complexity separation between deterministic and randomized predictors. The matching upper and lower bounds show that the unusual cube-root dependence on $\log|G|$ is intrinsic. At $\log|G|\asymp\varepsilon^{-1}$, the minimax rates for randomized and deterministic calibrated multiaccuracy both coincide with the minimax rate for multicalibration.

A model pretrained to predict a customer's next event can estimate the probability of future events without training on their labels and provide a reusable representation for supervised downstream banking tasks. We pretrain EventGPT, a hierarchical causal transformer, by next-token maximum likelihood estimation on the public Multimodal Banking Dataset (MBD-mini: 100,224 clients, 38.8 million transactions and 66.3 million geolocation pings). Keeping one state per event, EventGPT trains in about two hours on one L4 GPU at windows up to 8,192 tokens. From its next-event distribution, with no labels, EventGPT scores five targets of the form "the client makes a transaction of a given kind for the first time next month" at ROC-AUC 0.79 to 0.92. Gradient boosting on 114 engineered features needs 10,000 to 300,000 labeled client-months to match it on four targets and stays below it on the fifth. For product prediction, frozen EventGPT embeddings score below the engineered features (mean ROC-AUC 0.787 against 0.808 over five test folds), and a rank-averaged ensemble of EventGPT, profile fields and token counts with the engineered features adds 0.014 to 0.017, of which 0.007 to 0.016 is EventGPT's own. The code and pretrained weights are released.

Open multi-agent systems, where only a subset of agents are controllable, are central to real-world domains such as smart grids, yet lack a principled foundation for reinforcement learning algorithm design. We study this setting through n-agent ad hoc teamwork (NAHT), a recent framework for learning in open multi-agent environments. We propose an axiomatic framework that derives value function learning from the Shapley axioms (Efficiency, Symmetry and Linearity), treating them as structural constraints rather than heuristic design choices. This formulation builds on a novel NAHT decomposition, which characterises the fundamental difference between open and fully controllable multi-agent systems, and establishes a cooperative game-theoretic foundation for learning. We further show that enforcing these axioms on learning individual value functions recovers the Shapley value for Dec-POMDPs, providing a principled bridge between reinforcement learning and classical cooperative game theory. Building on this framework, we introduce Shapley Machines and Banzhaf Machines, two generic algorithmic templates that can be instantiated with standard reinforcement learning methods. When applied to strong base algorithms such as IPPO and POAM, our methods consistently improve learning efficiency and generalisation to unseen agent types and conventions across NAHT benchmark tasks. The code of experiments is published on https://github.com/hsvgbkhgbv/axiomatic-rl-naht.

Text-to-SQL translates natural language questions into SQL queries, enabling database access without writing SQL. Fine-tuning open-weight large language models (LLMs) is cost-effective, but existing methods treat SQL as ordinary text and optimize token-level likelihood without explicitly modeling its syntactic, semantic, and logical requirements. Thus, a single incorrect alias or comparison operator can invalidate a query while contributing little to training loss. We propose SQLoom, a constraint-aware fine-tuning framework with supervision at three levels: syntactic validity under SQL grammar and database schema, semantic alignment between query decisions and the question, and logical consistency across dependent clause decisions. SQLoom extends supervised fine-tuning with a unified set-contrastive objective favoring sets of correct outputs over minimally different incorrect ones at each level. LLM agents propose this supervision, which a SQL parser and execution on counterexample databases verify. We also release Spider-Refined, a training dataset correcting 378 flawed Spider annotations and enriching database schemas. Experiments on Spider, BIRD, and three perturbed Spider variants show that SQLoom improves execution accuracy for Qwen3-1.7B from 62.9% zero-shot to 88.7% on Spider development, reaching 89.2% on Spider test and 73.4% on BIRD development. It outperforms fine-tuning, reinforcement learning, and agent-based baselines with up to 70 billion parameters. On question perturbations, SQLoom loses at most 0.6 accuracy points versus up to 12.1 for agent-based methods. It generates SQL in a single pass at 0.42s per query, 23$\times$ faster than a GPT-4-based agent method. These results demonstrate the effectiveness of set-contrastive supervision over token-level training.

Inverse Reinforcement Learning (IRL) aims to recover expert rewards that transfer across environments with differing dynamics. Prior work has shown that such transfer is possible when gathering expert demonstrations across sufficiently different dynamics. In practice, however, expert demonstrations are costly, making it critical to select the right dynamics to reduce the quantity of expert data required for learning transferable rewards. We propose a two-phase approach: we first cycle through the available dynamics to obtain a reliable reward estimate, then actively select dynamics to reduce our uncertainty about the expert reward. Crucially, this yields novel sample complexity guarantees on transfer that hold for any dynamics, with a rate independent of the number of available dynamics. Empirically, our method significantly reduces the number of expert demonstrations needed to learn a transferable reward.

Zeroth-order (ZO) optimization enables memory-efficient large language model (LLM) fine-tuning, and recent methods incorporate curvature-aware strategies. However, in the high-dimensional parameter space of LLMs, these methods inherently suffer from exploding gradient variance and prohibitive costs. Conversely, confining updates to a low-dimensional subspace reduces variance but discards critical high-dimensional information, degrading performance. This creates a fundamental dilemma: full-space optimization is informative but has high variance, while subspace optimization has low variance but loses information. To break this trade-off, we propose ENGINE (Endogenous Variational MultiScale). Inspired by computational mechanics, we adapt static Variational MultiScale (VMS) theory to dynamic optimization, treating low- and high-dimensional spaces as resolved coarse and unresolved fine scales. Through a multigrid-style V-cycle, ENGINE alternates between spaces, dynamically probing fine-grained information in the high-dimensional space and injecting it to correct the efficient low-dimensional sampling process. Under appropriate assumptions, statistical analysis characterizes variance reduction and correction of curvature-induced bias, and convergence analysis establishes a stationarity bound for the corrected coarse updates. Evaluations on standard LLM benchmarks show that ENGINE achieves higher accuracy and faster convergence than existing ZO methods, achieving up to a $2.61\times$ speed-up in forward passes with zero additional peak memory overhead. Code is available at https://github.com/AI4Science-WestlakeU/ENGINE.

A memory-augmented agent must decide which memories to keep, merge and retrieve, and its only evidence of their value is its retrieval log: the memories it retrieved in each episode and whether it succeeded. Existing methods assign every memory a value from this log, even when the log cannot separate one memory's contribution from another's. We treat the log as an experiment that nobody designed and ask which memory values it determines. Two counts answer this before any outcome is observed: how often each memory was retrieved, and how often each pair was retrieved together. From these counts, we prove that top-$k$ retrieval leaves every memory value undetermined, however many episodes are logged, and that a question with its own memories, asked once, determines at most one of their values. On fourteen standard memory benchmarks, no memory value is determined in any of 55 configurations. Both causes are choices about retrieval. We introduce LEDGER, which estimates each determined Shapley value in closed form with a confidence interval and reports every other value as undetermined. It merges memories with identical histories without losing a determined value, and it retrieves so that the next log determines more. On pooled HotpotQA, LEDGER raises the fraction of memories determined from 0.001 to 0.797, at a cost of 2.6 accuracy points on HotpotQA and with a gain of 2.7 on LoCoMo.

The widespread adoption of Large Language Models (LLMs) demands reliable detection of human-authored versus synthetic text. However, current Transformer-based detectors often experience severe performance degradation during domain shifts and operate as opaque classifiers. This paper investigates the robustness and interpretability of fine-tuned Transformer models (DistilBERT, RoBERTa) by evaluating them against a custom out-of-distribution (OOD) dataset of hyper-local news synthesized via open-source LLMs (Llama 3, Mistral, Gemma). To analyze model decision boundaries, we apply a tripartite Explainable AI (xAI) framework using SHAP. Our analysis reveals three key phenomena: First, detectors often rely on a "Call-to-Action" syndrome inherent to instruction-tuned LLMs rather than generic machine text. Second, we observe a "Specificity Bias," where LLMs successfully bypass detection by hallucinating specific named entities and mimicking conversational imperfection. Finally, we illustrate the "Wikipedia Paradox," wherein formal, objective human writing is frequently misclassified as AI due to its low statistical burstiness. These findings demonstrate that purely statistical detection paradigms face systematic limitations in highly structured human domains, emphasizing the critical need for semantically-aware, interpretable detection architectures.

In node selection tasks, the goal is to identify a single unknown target node in a graph. We consider probabilistic supervised node selection, where the estimator outputs a probability distribution over the nodes of the graph and the ground-truth node is available for evaluation. Classical evaluation metrics for such tasks are either statistically ill-conceived or blind to the graph structure. To fill this gap, we introduce two strictly proper scoring rules that are explicitly graph-aware: the Resistance score, which is parameter-free, and the Diffusion score, which operates at an adjustable coarse-graining resolution. We show that the Diffusion score interpolates between the multi-category Brier score and a cluster-level weighted Brier score. The Resistance score generalises two classical scoring rules: the multi-category Brier score and the Ranked probability score. Applied to rooted trees, it also serves as the strictly proper version of the improper Hierarchical error distance used in hierarchical classification. In the same setting, we furthermore establish strict propriety of the Hierarchical cross-entropy family, a recently introduced separate class of hierarchy-aware scoring rules. Experiments on epidemic source detection and on ImageNet classification confirm that the proposed scores detect meaningful differences in estimation quality that classical metrics miss.

Large-scale annotation of environmental sound recordings presents significant challenges under time and resource constraints, particularly when sound categories are imbalanced and only part of a dataset has been labeled. This paper describes the development and evaluation of a human-AI workflow used during annotation of the Kenya Forest Sounds Dataset (KFSD), using a project-stage working collection of 7,531 audio recordings across 21 categories. We investigated zero-shot classification, a feature-based ensemble, and a fine-tuned Audio Spectrogram Transformer (AST). The fine-tuned AST achieved 89.36% accuracy, 85.98% precision, 89.24% recall, and 87.58% F1 on held-out test data for binary classification of Person talking. Because this result remained below the predefined 95% threshold for autonomous annotation, model predictions were used as suggestions and subjected to human verification. Of the 5,026 initially unlabeled recordings, 2,337 were routed to a speech-verification queue and 2,689 remained for ordinary manual annotation. The verifier reviewed all 2,337 machine-labeled files and detected zero false positives, yielding 100% verification accuracy. The observed average verification time was approximately 4.2 seconds per file. At an average rate of 35 seconds per file for manual annotation, the hybrid workflow reduced work on the remaining queue from approximately 48.9 to 28.9 person-hours, saving about 20.0 hours (41%). The study documents the model-selection process, provenance-aware file organization, verification procedure, observed errors, practical limitations, and opportunities for extending AI assistance beyond one category.

Full Waveform Inversion (FWI) is a fundamental inverse problem in geophysics to estimate subsurface properties, such as velocity, from seismic measurements. Recently, self-supervised learning methods remove the need for ground-truth velocity maps during training by leveraging a forward operator to simulate seismic data from predictions and backpropagate the gradients of reconstruction loss through it. In this work, we consider a more challenging setting without accessing such gradients. To this end, we incorporate reinforcement learning into self-supervised FWI and propose a framework named DeepWaveRL. Specifically, we reformulate FWI as a policy learning problem, where the policy network predicts velocity maps as actions, and a forward operator is used solely for simulation. A global scalar reward is computed based on data misfit for each prediction and used to weight policy gradients. Thus, learning proceeds without accessing the gradients from the forward operator. Given the indirect and coarse supervision signals, we further identify key strategies to stabilize learning and mitigate the spatial credit assignment bottleneck. Experimental results show that, despite relying entirely on a single scalar reward, DeepWaveRL achieves competitive performance compared to supervised and self-supervised baselines. These findings suggest that our framework offers a flexible and complementary learning paradigm for the FWI research community.

Zeroth-order (ZO) optimization has emerged as a promising approach for memory-efficient large language model (LLM) post-training, yet its reliable application to reinforcement learning with verifiable rewards (RLVR) remains largely unexplored. Motivated by the distinct data regimes of on-policy RL and fixed-data optimization, we develop a generic variance-theoretic framework to investigate what limits the reliability of ZO gradient estimation. Our analysis reveals that estimation reliability is structurally determined by whether the two finite-difference evaluations are performed on matched data. On-policy reinforcement learning (RL) under the mismatched-data regime produces a $1/μ^2$ oracle-noise amplification that sets an irreducible noise floor as the perturbation scale $μ$ decreases, whereas Boltzmann Targeted SFT (BOLT) under the matched-data regime eliminates this amplification via noise cancellation. We further derive convergence guarantees with explicit constants that cleanly separate the two regimes. These results naturally lead to ZO-BOLT, which applies ZO optimization to BOLT's RLVR-aligned weighted-SFT objective. Controlled perturbation-scale sweeps and matched-versus-mismatched end-to-end comparisons validate the predicted effect of data matching. Experiments across three model variants and two mathematical reasoning benchmarks further show that ZO-BOLT consistently improves over the pretrained baseline while retaining a near-inference memory footprint. The code is publicly available at https://github.com/lianmin-chen/the_cost_of_mismatch.

Artificial Intelligence, and Machine Learning in general, is becoming pervasive across all domains and disciplines. By means of this data-driven approach, algorithms can learns complex task just by learning from samples. However, if the data-set is biased, they might lead to the development of unfair computer systems, that is, characterised by a prejudice towards specific individuals or groups, which might get an unmotivated output from the system. This circumstance lead to the development of multiple metrics designed to quantify the level of unfairness of an algorithm. However, a generalised assumption in the literature is that the sensitive features and the output labels are binary. When the data is not binary, it is forcefully binarized using a one-versus-the-rest simplification. In this work, we discuss the limitations of existing metrics and propose novel, mathematically sound, and generalized multi-group and multi-class formulations of the metrics. Ultimately, our results highlight the empirical assessment distortion caused by binarization strategies, which reveals the limitations of binary fairness metrics in the multi class setting.

We introduce the Poisson Super Learner, a scalable ensemble method for survival and competing risks prediction with right-censored data. The approach exploits the correspondence between the likelihood of piecewise-constant hazard models and the Poisson likelihood. After expanding the data over a chosen time partition, candidate hazard learners are trained under a common Poisson loss, and cross-validated log-hazard predictions are combined by a Poisson meta learner. The resulting ensemble is a geometric combination of candidate hazards, and can be used to estimate survival probabilities and cause-specific absolute risks. The Poisson Super Learner framework is modular and can in principle incorporate any learner that can estimate interval-specific hazards, including parametric, penalised, additive, and other machine-learning hazard models. Training targets the cause-specific hazard directly and does not require explicit modelling of the censoring distribution. For competing risks, separate cause-specific ensembles are fitted and subsequently combined to obtain individualised cumulative incidence predictions. We illustrate the approach using data from the Steno Diabetes Center collected between 2005 and 2022 to predict cardiovascular disease risk in patients with type 1 diabetes from detailed covariate information. The applied analysis is complemented by simulation studies designed to reflect key features of the data and to examine the behaviour of the method under controlled settings. All analyses were implemented using the freely available R software package poissonsuperlearner.

Anomaly detectors rank unfamiliar examples for expert confirmation, and the confirmed examples teach a classifier to recognise a new class. Across five benchmarks from physics, medicine and computer vision, the members flagged by the detector are those the classifier learns first. In three systems, confirming the highest-ranked members of a new class gives lower recognition than confirming the same number of random members. Handing subsequent selection to the promoted classifier recovers part of this loss in two systems. A paired intervention further shows that replacing nearby known classes with distant ones sharply reduces the confirmations needed for recognition. These results distinguish finding a new class from selecting examples that teach its extent: within the tested classes, concentrating confirmations at the top of the anomaly ranking makes learning less efficient.

Dynamic graphs capture evolving interactions in Web systems, including online communication, user engagement, and recommendation. Graph foundation models aim to reuse predictive capabilities across networks, but extending this capability to dynamic graphs faces two challenges. First, node identities are graph-specific, and similar interaction patterns can appear at different temporal and activity scales, making histories difficult to represent consistently across networks. Second, the predictive roles of historical patterns vary across interaction dynamics, limiting transfer from a narrow range of training processes. In this paper, we show that a common interaction language, combined with learning across diverse interaction processes, offers a basis for dynamic graph foundation modeling. We propose CL-DGFM, a Common Language Dynamic Graph Foundation Model for temporal link prediction. Its Dynamic Interaction Tokenizer encodes interaction memory, historical connectivity, and ordered destination transitions into calibrated tabular inputs, making dynamic histories accessible to existing predictors without an additional trainable graph encoder. Its Dynamic Interaction Prior trains a pretrained tabular foundation predictor across diverse synthetic dynamics. Labeled historical context guides prediction, with a Contextual Value Adapter modulating attention values. The model transfers to real graphs without target-graph parameter updates. Experiments on TGB and TGB-Seq demonstrate competitive performance. Further comparisons show that the common language supports basic machine learning and neural models, and enables direct use of frozen tabular foundation models on dynamic graphs.

Responses with headers, lists and greater length receive high scores from reward models (RMs), and RLHF policies learn to produce these features even when true quality does not improve. Practitioners therefore edit their RMs, but they learn whether an edit works only after an RL run. We show that the response of a policy to an RM edit can be forecast before training, from samples of the current policy and the scores of the old and new RM. If we read the RM as a set of prices, the first-order theory of KL-regularized RL implies that each feature moves in proportion to its within-prompt covariance with the change in reward. We implant controlled price changes into retrained RMs under a 2^(4-1) fractional factorial design with a self-distilled placebo, run 54 RL runs from six parent policies, and fix every forecast before the runs. With the covariance predictor, we predict the direction of all 20 changes in supply whose forecast has the same sign at every parent and explain 82% of the standardized change. With counterfactual prices, the classical alternative, we explain 50%. The forecast also holds under GRPO and for real RM upgrades, where we explain 66% out of sample with a coefficient frozen from the factorial runs. For rent, the gap between the proxy reward and a gold reward from stronger judges, we rank RM edits correctly with the same covariance over the first 30 steps (Spearman 0.94). This measures the horizon of the forecast: by step 200, the ranking of the factorial edits reverses under both RLOO and GRPO, while the real upgrades keep their order. When we vary the KL coefficient beta, the initial response does not depend on beta, and the response saturates below the KL equilibrium. In one experiment, the long-run response to a price rise is consistent with an additional implicit regularizer beta_0 of about 0.003 to 0.007.

Computer-use agents for scientific software require application-specific expertise. Much of this expertise is documented in human-authored software manuals and web pages. However, turning these resources into executable, reusable skills remains challenging due to mismatches between documented and deployed software versions and differences in scope between documentation coverage and task requirements. We introduce ScienceCUA, a two-stage framework combining resource-to-skill construction with execution-guided skill evolution. First, ScienceCUA identifies tasks supported by the software from its documentation, infers operational trajectories, and generalizes them into reusable Skills. Command catalogs, syntax constraints, and shortcuts are organized separately as References for on-demand lookup. Then, ScienceCUA uses execution feedback to drive skill self-evolution by contrasting successful and failed trajectories for the same task, identifying deficiencies, revising existing skills or adding new ones, and updating the skill library only after repeated validation in real environments. To enable reliable evaluation without benchmark noise or task leakage, we introduce two improved benchmarks. ScienceBoard-Verified provides an audited and corrected basis for measuring task performance, while ScienceBoard-VE separates skill evolution from final testing through disjoint evolution and test tasks and further supports native-agent evaluation. Experiments show that the initial Skills and References improve average task success rate by up to 7.23 percentage points over the base model, while skill self-evolution increases the total gain over the baseline to as much as 17.12 percentage points. ScienceCUA turns software documentation into an evolving set of executable skills and queryable references, offering a practical way for general-purpose agents to use scientific software reliably and at scale.

We introduce Axion 1.0, a fully open agent foundation model that combines strong and broad agentic capabilities with a remarkably simple training recipe. Axion 1.0 is trained in only three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning, without dedicated mid-training, direct preference optimization (DPO), or on-policy distillation (OPD). The first SFT stage develops strong self-contained reasoning through model-aware data construction, where we synthesize challenging problems and adversarially strengthen them based on the base model's own successful solution strategies. The second stage extends these capabilities to tool-dependent reasoning through a scalable pipeline for constructing diverse and automatically verifiable tasks grounded in real-world data sources and deterministic computation. Finally, reinforcement learning further strengthens long-horizon reasoning and interaction with external tools and environments. Despite this compact training recipe, Axion 1.0 achieves strong performance across a broad suite of agentic benchmarks and is competitive with leading commercial models on several of them. Equally importantly, Axion 1.0 is open beyond model weights: we release the complete recipe, including data construction, generation, filtering, verification, prompts, and training code, enabling the entire development process to be reproduced and extended. Axion 1.0 demonstrates that strong agent foundation models can be both open and simple, offering a practical and reproducible path toward advancing open agentic intelligence.

Agentic AI is shifting developers' responsibilities from software implementation to software governance. While this paradigm improves productivity, it also raises challenges in ensuring that autonomous coding agents preserve developers' intentions. Among these, operationalizing fairness-by-design is largely unexplored, as existing approaches provide limited support for representing and propagating fairness requirements throughout agentic development workflows. In this paper, we investigate whether specification-driven development can operationalize fairness-by-design in the agentic development of machine learning systems. We study three coding agents on four machine learning applications, using four prompts that progressively operationalize fairness through nine Fairness Specification Elements, and analyze the traceability of fairness requirements across the generated artifacts, the fairness-accuracy trade-off of the resulting models, and the agents' design decisions and rationales. The results show that when fairness is operationalized, agents capture almost all fairness requirements in the specification (97.22%) and preserve most of them in the code (82.41%). However, the resulting models are not significantly fairer, as agents implement fairness as a post-hoc audit that neither shapes training nor prevents model delivery. Specifications thus make fairness traceable, but not enforceable, calling for fairness specifications that state what agents should do when requirements are not met, and for workflows that make the fairness decisions agents delegate to developers explicit.

An aggregate task score does not distinguish failures of scheduling, state retention, and candidate selection. We develop a small executable instrument that varies dependency depth, candidate count, active state width, and access to current-state snapshots while holding the number of required operations fixed. Two dependency-graph families show that increasing depth at fixed work changes both scheduling freedom and the number of reachable completion states, so these controls are coupled, and we quantify the coupling. The primary study contains 49,152 generated tasks and 393,216 reference-controller rollouts. A separately frozen follow-on design adds 295,936 rollouts covering retention strategy, snapshot cadence, hidden state resampling, recovery, query locality, candidate indexing, and relabeling. Residue caches and precomputed plans solve tasks that defeat limited raw-value caches, and indexed selection removes the serial-scan branching deficit. Hidden resampling imposes an exact success ceiling that additional state tracking cannot overcome. The contribution is an executable diagnostic instrument, analytic calibration, and an audit of its assumptions. A small local CPU study with a 0.5B model completed no workflows, and a 16-episode Copilot-routed fractional factorial is descriptive only. Neither supports general capability claims.

LLM agents compact long histories to fit their context windows, but the compacted context must still carry the operating rules that govern later actions. Prior work shows that violations rise after compaction, but measures compliance outside a complete agent harness and does not identify where each violation comes from. We introduce ARCBench(Agent Rule Compliance Benchmark), a configurable benchmark built on the open-source pi agent harness. Its tool-calling agent loop lets us measure rule compliance from actions taken during execution, bringing the evaluation closer to deployed agent workflows. Using ARCBench, we examine where violations after compaction come from, and how different compaction configurations and adversarial settings change their sources. We find that compaction rarely deletes the rule: most violations after compaction occur while the rule is still in the compacted history, with the agent treating a later request as permission, ignoring the rule, and proceeding anyway. Where violations come from depends on the configuration: how the rule is delivered decides whether the agent would break it even without compaction, while the compaction strategy and budget decide whether the rule survives. The strongest compactor-targeted attack more than doubles violations by corrupting the rule in the summary and changing how the agents behave, whereas the strongest defense restores the rule in most summary but still violates. Finally, retaining a rule does not guarantee compliance: compacted runs that retain the rule violate more than twice as often as runs with the full history.

Recourse-aware learning trains models so that people who receive an adverse decision can reverse it with an affordable change. For gradient-boosted decision trees (GBDTs), the existing method, RABIT, exposes a single recourse weight whose effect on the recourse rate is abrupt and data-dependent, and its training bound cannot certify recourse for any rejected individual. We present Certified Recourse Boosting (CERBO), which takes a requested recourse rate and returns a GBDT with certified actions: it tracks exact additive effort along tree paths, trains against committed boxes (a family of bounds that contains RABIT's and certifies a within-budget action for each training sample) and adjusts the weight after every tree until the certified rate meets the request. A calibration step guarantees the rate on new individuals, and the certifier runs on any tree ensemble, including XGBoost and LightGBM, in milliseconds per person without a solver. In RABIT's own experimental protocol, at equal accuracy, CERBO raises the recourse ratio from 0.90 to 0.99 on FICO and from 0.64 to 0.77 on Bail. Under additive costs on six data sets, it raises the area under the accuracy-recourse frontier in all 12 settings, and in simulated deployments it guarantees recourse to 70% of rejected individuals in 11 of 12 settings, against 2 for RABIT.

Recently, harness self-evolution has emerged as an important paradigm for improving agent performance. Prior studies show that evolving the agent harness can effectively enhance agent capabilities without modifying the foundation model's parameters. We find that existing approaches largely study agent evolution on a single task. In practice, coding agents may need to handle multiple task types over time. Whether an agent harness can continuously evolve across multiple tasks without suffering from catastrophic forgetting remains an important open question. In this paper, we investigate this open question through a comprehensive empirical study of continual harness self-evolution. We organize three software engineering task datasets into six different evolution directions and apply two representative harness evolution methods to perform continual evolution along these directions. Our experimental results show that evolution on a subsequent dataset can substantially undermine the performance on previously seen datasets. For example, with DeepSeek-V4-Flash-0731, Meta-Harness improves its performance on SWE-Bench to 94.0% after evolution. However, after further evolution on FeatureBench, its performance on SWE-Bench drops to 32.0% by the fifth round, corresponding to a 66% relative decrease. Finally, based on our experimental findings, we identify several insights, providing directions for the continual self-evolution of agent harnesses. Code is available at https://github.com/THU-Agent/continual-harness-evolution.

New federal legislation will require states with high payment error rates in the Supplemental Nutrition Assistance Program (SNAP) to pay a share of benefit costs, potentially costing some states over a billion dollars per year. These costs create substantial pressure to improve payment accuracy. One strategy to improve payment accuracy is preauthorization review, in which staff double-check applications and correct mistakes before benefits are issued. But limited staff capacity makes it important to target these reviews effectively. One leading targeting approach uses statistical risk models to prioritize cases likely to contain errors. Another novel approach uses AI to flag specific potential discrepancies for reviewers to check. For states under pressure to reduce error rates, it is unclear which approach-or combination of approaches-would reduce payment error rates most effectively. We compare these strategies through simulations using public SNAP quality control data. Under our main assumptions, risk-based review alone costs about seven times as much as combining AI flags with risk scores. This cost advantage persists across a wide range of plausible assumptions. Our findings suggest that AI-assisted preauthorization review could be a key tool to help states avoid substantial cost-sharing obligations.

Laminar-turbulent transition governs skin friction, heat transfer and losses in turbomachinery, yet Reynolds-averaged Navier-Stokes (RANS) transition models rely on empirical correlations and hand-calibrated activation functions whose structural deficiencies cannot be resolved by recalibration. We propose a CFD-driven sparse symbolic regression framework to learn interpretable, closed-form corrections to the production and destruction activation functions in the intermittency equation of the local, Galilean-invariant $k$-$ω$-$γ$ model of Ge et al. Each correction combines a polynomial of a local pressure-gradient parameter, controlling amplitude and sign, with a bounded exponential of local flow features, controlling spatial support. This construction bounds the corrections without ad hoc limiters. Coefficients are identified by gradient-free optimization with the RANS solver in the loop, accelerated by POD-Kriging field surrogates, using mean-velocity data from wall-resolved large-eddy simulations (LES) of the ERCOFTAC T3 flat-plate cases. Elastic-net hyperparameters are selected automatically through bi-level optimization with leave-one-case-out cross-validation. The resulting model retains 11 of 21 candidate terms, compared with 20 under conventional grid-search tuning, while achieving comparable or slightly better accuracy on training and unseen cases. Both approaches recover the same leading-order correction. Applied without retraining to the transonic LS89 turbine cascade, far beyond the training envelope, the augmented model remains numerically robust and improves heat-transfer predictions at all four operating points, reducing errors by up to 70% where the baseline predicts premature transition under high freestream turbulence. Residual errors in transition onset and length highlight the limits of training on flat-plate data alone.

Training models over data distributed across edge devices presents a trade-off: centralizing data simplifies training but raises privacy and regulatory concerns, while full local training may exceed the devices' compute and memory capacity. Split learning addresses this challenge by partitioning the model between clients and the server, allowing part of the computation to be offloaded. However, existing approaches typically rely on static split configurations or simplified assumptions that fail to capture the heterogeneity and complex dynamics of real deployments. We present DynSL, a framework that dynamically allocates model layers between clients and the server by continuously monitoring system metrics and periodically solving an optimization problem that accounts for client and server computation and communication. When beneficial, the system reconfigures the split, enabling the offloading strategy to adapt to changing device capabilities and system conditions. Extensive experiments on a testbed of 24 heterogeneous edge devices show that DynSL reduces training time by up to $\sim\!$30% over the strongest adaptive baseline and by more than 50% over a suboptimal uniform split, with no loss -- and sometimes a gain -- in accuracy.

Understanding how stochastic gradient noise affects the convergence of Adam is an important problem in stochastic optimization, particularly when the noise exhibits heavy-tailed behavior. In this paper, we investigate the convergence of vanilla Adam in both the expected risk and the empirical risk minimization, focusing on the effects of heavy-tailed noise and its relationship to index sampling. For the expected risk minimization, we establish convergence guarantees under Hölder smoothness and bounded $\mathfrak{p}$-th moment assumptions, allowing stochastic gradient noise with infinite variance. Our analysis provides explicit upper bounds on the expected squared gradient norm and shows that the iteration-dependent convergence term decays more slowly as $\mathfrak{p}$ decreases. For the empirical risk minimization, we establish asymptotic and non-asymptotic convergence guarantees and demonstrate that heavy-tailed population noise can deteriorate convergence bounds through dataset-dependent constants as the dataset size increases. Furthermore, under Lipschitz smoothness, we formulate multistage Adam with practical learning rate schedules and demonstrate that increasing batch sizes can reduce index-sampling noise and improve convergence guarantees and SFO complexity. Numerical experiments support our theoretical findings on the effects of heavy-tailed noise and the computational benefits of increasing batch sizes.

Long-context LLM inference is costly because attention repeatedly reads a growing key-value (KV) history. We introduce Chunk-Indexed Sparse Attention (CISA), which retrofits pretrained dense attention models without gradient updates. CISA selects fixed-size KV chunks using compact summaries of existing keys, avoiding token-level indexing and a separate full-length index cache. It uses one indexing query per KV group and shares selected chunks across queries in that group. During prefill, QShare reuses selections across adjacent queries to reduce indexing cost and improve KV reuse and GPU utilization. We evaluate CISA on NVIDIA Nemotron 3 Nano and Ultra, and Inkling-Small. In one-layer BF16 benchmarks at batch size 32 using Ultra's per-device attention dimensions, core attention is 11.7x faster for prefill queries near the end of a 1M-token prompt and 13.0x faster during decode. Including chunk scoring, selection, and all other measured GPU work, speedups are 8.6x and 3.0x. On Ultra NVFP4 with a 4,096-chunk budget, aggregate scores stay within 1.5 percentage points of dense attention on MRCR, LongBench v2, and GraphWalks. RULER sweeps on Nano and Ultra BF16 up to one million tokens show scores within 2.05 percentage points of dense attention with 8-16x fewer KV reads. Serving million-token prompts with 32 concurrent requests and 4K output tokens per request, CISA achieves 2.29x the end-to-end throughput of dense attention using Nemotron 3 Ultra NVFP4 on four NVIDIA B300 GPUs. Selecting fewer chunks increases throughput to 2.61x; in separate GraphWalks evaluations, the smaller budget widens the aggregate score gap from 1.49 to 3.59 percentage points. These results expose a practical quality-speed tradeoff without sparse training.

When a language model answers a yes/no question, that answer changes how it answers a different question about the same entity. We study this \emph{self-commitment} in 22 open-weight models from four families (0.5B--72B parameters), fixing the model's first answer and measuring the shift in its second across 3,600 question pairs about 120 entities. Each answer writes a \emph{commitment direction} into the residual stream. The similarity of two questions' directions predicts their coupling on held-out entities ($ρ= 0.43$--$0.85$), and ablating a single direction shared across topics leaves at most 17\% of the effect in every model. How this direction is read is far less stable: the reading circuit spreads across more attention heads as models grow, instruction tuning makes more questions read it in the opposite sense, and in one model changing only the system prompt reverses the circuit's sign. Ablating the direction while sparing the attention sink removes most of the coupling at under 1.3\% perplexity cost in 18 of 19 instruction-tuned models, though it shifts standalone answers in two. Without the sink spared, the direction's alignment with the leading principal component of ordinary text predicts which models pay a high cost. Written once, the commitment is read differently across scale, training and prompt.

Repairing obsolete tests in response to source code evolution is critical for test suite reliability and software quality assurance. While large language models (LLMs) have emerged as promising tools for automating many software development tasks, the effective supply of context and the orchestration of repair pipelines in this source--test co-evolution scenario remain poorly understood. Moreover, although reasoning is found helpful for LLMs on some complex coding tasks, it remains unclear which reasoning steps actually contribute to effective test repairs and how the insights learned from LLMs' behaviors benefit coding agents. In this paper, we present a systematic empirical study of test repair for source-test co-evolution empowered by LLMs, investigating the effective setups and typical LLM behaviors. To support this study, we construct a new benchmark of 500 real-world source-test co-evolution cases from commits submitted between May 2024 and April 2026 across 15 highly-starred open-source Java projects. Our experimental results reveal several key insights for effective LLM-based test repair: the complementary benefits of combining test failure symptoms, source code changes, and supportive code context, as well as the effectiveness of self-repair rounds. We further derive a common reasoning action taxonomy from LLMs' reasoning traces, which proves instrumental in guiding LLMs to repair tests more effectively while at a lower financial cost. Applying our findings to the Codex agent further improves its repair success rate on challenging cases by 66.7%.

Dhārā Bench is a retrieval benchmark for Indian Knowledge Systems (IKS), with 325 questions in 12 categories and 60,121 text chunks from 15 source groups. The categories were developed based on 4,500 questions submitted to the Dhārā application; the evaluation set combines corpus-aligned application queries with expert-contributed questions. The annotations comprise 7,578 question-chunk relevance links involving 6,847 distinct corpus chunks. For 34 composite and multi-hop questions, 401 conjugate groups specify alternative sets of passages that jointly support each question. Chunk identifiers and hierarchical references connect the judgments to their source texts. The paper describes corpus preparation, annotation, and validation, and reports a fixed BM25 baseline implemented with Pyserini with $k_1=0.9$ and $b=0.4$. Across all questions, BM25 achieves MRR@20 of 0.315, Recall@10 of 0.182, and nDCG@10 of 0.205. On the group-annotated subset, complete-evidence scores are CE_Recall@20 = 0.294 and CE_MRR@20 = 0.141. The category results expose particularly weak verse retrieval, while group scores reveal gaps between finding individual passages and completing their support. The corpus and evaluation code are available at https://github.com/Avinash2236/Dhara-Bench.

Large language agents for literary role-playing are typically modeled around dialogue generation, leaving characters' non-dialogue behaviors and their dynamic evolution over story time underrepresented. In this paper, we propose ACTOR (Action-Centered Temporal Organization for Literary Role-Playing Agents), an automated framework for constructing such behavior trajectories from raw novel texts. ACTOR extracts dialogue and persona-relevant non-dialogue behaviors, represents each behavior using a Thought-Action-Observation structure, and organizes behavior trajectories according to the internal temporal order of the story. We build a behavior dataset from 143 diverse literary works. For each source behavior, we generate counterfactual behaviors that are contextually plausible yet inconsistent with the character's persona. We find that existing models struggle to reliably infer character-consistent actions from historical behaviors and evolving character states. Using these behavior pairs, we train models with Regularized Preference Optimization (RPO), which augments DPO with a negative log-likelihood loss on the source behavior to jointly optimize behavior preference and generation. We further design a four-way next-behavior selection task to evaluate the model's ability to faithfully mimic character actions. Across 17,560 evaluation instances, the model trained on ACTOR's preference pairs achieves substantially improved accuracy. This result demonstrates that structured, temporally organized supervision enables models to leverage character profiles and prior behaviors to infer subsequent actions, greatly advancing literary role-playing beyond mere dialogue imitation. Code and data are available at https://github.com/SilentStar233/ACTOR.

Protein-complex structure prediction with AlphaFold 3 depends on MSA content, depth, and cross-chain pairing. Current pipelines apply uniform recipes, typically maximizing depth and pairing by species, yet even whether cross-chain pairing helps AF3 remains contested. We show that the value of both depth and pairing is strongly target-dependent, so no single recipe is optimal across interfaces. We introduce an autonomous agent that treats a frozen AF3 as a black-box experimental instrument. For each target, the agent formulates biological hypotheses, analyzes and reorganizes the MSA, and iteratively refines sequence selection and pairing using predictor label-free signals alone. On 1{,}015 protein--protein interfaces from three benchmarks released after AF3's training cutoff, the agent improves overall interface DockQ over AF3's default pipeline and fixed MSA recipes on every benchmark, while using on average 60\% fewer MSA sequences than the default. Gains are largest for inter-species heteromers, where species-based pairing is unreliable: mean DockQ increases from 0.465 to 0.612 on FoldBench and from 0.482 to 0.562 on HD439. The improvements arise from both sequence construction and pairing policy, including deliberate non-pairing.These results recast MSA design as a target-specific problem and show that an autonomous agent can adapt it at test time through biological reasoning and self-directed, confidence-guided experimentation.

People often attach instructions to their search queries, asking for pages written for beginners, documents in a certain language or official sources only. Most rerankers were trained to judge topical relevance and treat such instructions as a few extra words. On the InfoSearch benchmark, even GPT-4o used as a list-wise reranker reaches an average WISE (Weighted Instruction Sensitivity Evaluation) of only 33.5. We present DMReranker (Decision-Model Reranker), a zero-shot reranker that splits an instructed query into a topic and a short list of typed clauses. A small judge model answers one narrow question per clause, and the answer probabilities are multiplied into a single score. Nothing is trained, and the configuration was frozen before the test runs. Reranking the same top-100 candidates as the published baselines, DMReranker reaches a macro WISE of 50.9 over all six InfoSearch dimensions while keeping original-mode nDCG@10 at 82.9. It trails GPT-4o on the stricter SICR (Strict Instruction Compliance Ratio) metric, 19.2 against 32.1. We show that this gap follows from how a product of probabilities behaves when SICR compares raw scores across runs. Computed from ranks, our SICR is 67.0. On the same judge, splitting the instruction raises WISE from 32.9 to 47.6 over a single combined question, while raw-score SICR falls. Scoring all six dimensions cost about $10 at list price.

Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.

On July 1, 2025, Vietnam reduced its 63 provinces to 34, abolished its 696 districts, and replaced its 10,035 commune-level units (wards) with 3,321. ViAdminShift, a Vietnamese language resource about this event, consists of 7,995 benchmark items, 2,281 retrieval passages from pre- and post-reform Wikipedia revisions and dated templates, and old-to-new tables that a released script rebuilds from the official conversion table (not redistributed). Changed items store the pre-reform answer (except 121 address items), a $2\times2$ design uses untouched units as controls, and 400 addresses appear without diacritics. The tables agree with an independent lineage on 10,564 of 10,565 official pairs, disagreements were settled against the resolutions, and the probing-track scorer agrees with both authors' blind labels (Cohen's $κ=0.91$ and $0.89$). Data, outputs, code, and a datasheet are released. In the probing track, at least 20 of 67 models know the reform. Fourteen of them give the new province alone for at most 11% of plain questions about old addresses (old province only: 14-100%); a bare date recovers up to 100%. Post hoc (80 province-free addresses, single runs), with reasoning off, today's date in the system prompt leaves 68-100% old-province-only answers for 10 of 12 models; with reasoning, an "as of" prefix in the question leaves none for seven of eight. On the 1,262-item benchmark subset, with same-source old and new passages, small models unaware of the reform give 14-30% outdated answers and the five benchmark-track knowers (also differing in size or provider) 0-5%.

There are many causal discovery algorithms, but few guiding principles for choosing between them. Methods are usually compared through three lenses: identifiability, consistency, and benchmarks. Each introduces assumptions about the data-generating process that are often untestable. We argue that there are no shared standards for comparing, prioritizing, or even discussing those assumptions. We support this claim by showing that (i) identifiability assumptions are often preferred by convention rather than by principle, (ii) uniform consistency requires further untestable assumptions, while the weaker pointwise consistency has arguably narrowed algorithm design, and (iii) benchmarks are not neutral tests but carry an epistemic role of their own: they quietly fix which regimes matter, which guarantees come into play, and what counts as success. In doing so, they shape what algorithm comparisons are taken to show and thus perform some of the same epistemic work as theory, but less transparently. In this position paper we argue for a more transparent, integrated practice: Make the epistemic standards for discussing identifiability and consistency assumptions explicit, especially their failure modes and scope conditions; design benchmarks to operationalize those standards, by stress-testing and ablating both. Then, theoretical and empirical results jointly clarify when, why, and how algorithms may work rather than accumulate as disconnected advances.

LaBraM, an early EEG foundation model, demonstrated that large-scale EEG pretraining can yield strong representations for diverse downstream tasks. Building on this foundation, we expand pretraining data and model capacity while retaining its core encoder and objective. We train ten encoders with 5.8 million to 1.97 billion parameters on three EEG collections spanning 25,067 to 79,910 hours, adapting signal preparation and training to heterogeneous recordings. Across 23 tasks, Atom-L* reaches 59.05% mean balanced accuracy at its best evaluated full fine-tuning (Full-FT) checkpoint, 3.61 percentage points above the strongest baseline EEG foundation model. Quark-G leads linear probing (LP) at 54.33%, a 1.24-point margin. Both also lead average task rank. At the last evaluated checkpoints, larger encoders improve Full-FT more consistently than LP, and stronger LP does not guarantee stronger Full-FT. Preserving patch structure improves LP over average pooling. Data-fraction experiments suggest capacity-dependent saturation, while checkpoint trajectories show that downstream performance can decline as pretraining loss falls. Joint classification with a shared Qwen3 readout reaches 50.50% with Molecule-H and Qwen3-8B; increasing language-model size provides no consistent gain. These results demonstrate the value of scaling LaBraM and the importance of matching pretraining scale to downstream adaptation.

Repository-level AI coding benchmarks measure task success but do not directly capture the conditions required for software maintenance completion. We study multi-condition fulfillment, where a maintenance task must satisfy multiple operational conditions, sometimes with directional dependencies. Using a fixed 18-criterion taxonomy, we analyze 2,925 Apache Fineract issues and 944 tasks from SWE-bench Pro, SWE-PolyBench Java, and SWE-EVO. We examine condition breadth, corpus-level co-occurrence, and source-grounded directional requirement relationships. Among 1,177 benchmark task pairs matched on issue classification, modified-file count, and code churn, the mean fulfillment-set Jaccard distance is 0.8138, and 57.4% have disjoint condition sets. Condition count alone does not explain this contrast: SWE-bench Pro averages 2.47 conditions per task, versus 1.78 for Fineract. Temporal (B4) and numerical (B2) conditions have a phi coefficient of +0.430 in Fineract versus -0.013 in the pooled benchmark corpus. Among temporally tagged tasks, numerical conditions are also specified in 48.85% (254/520) of Fineract cases versus 5.81% (5/86) of benchmark cases. We document 16 directional relationships in Fineract requirements involving time, calculation, and state preservation, and inspect four benchmark cases to illustrate how relevant conditions appear in evaluation assertions. These examples do not quantify benchmark-wide oracle omissions. The results show that similar surface task profiles can mask substantially different fulfillment-condition sets. We recommend reporting condition breadth, co-occurrence, and directional relationships separately. The corpus comparisons are descriptive and may reflect domain differences; they do not establish causality or production failure rates.

Knowledge is how understanding accumulates across people and generations: durable records that outlive their producers and that others can check, combine, teach, and correct. AI agents are now built and run by many independent users and organizations, and each agent is upgraded, replaced, or retired within months. Yet what an agent learns stays trapped in model parameters, a transient context window, or a system-specific memory store that is appended to and retrieved from rather than processed. We argue that the next generation of AI should be knowledge-centric, built around an explicit, continuously evolving knowledge layer independent of any single agent. In this loop, agents discover knowledge from their environments and trajectories; the layer internalizes it by turning fragmented, heterogeneous records and experience into composable, revisable understanding; and agents across systems and generations use the result and return new outcomes for revision. While discovery and use have advanced rapidly with large language model agents, internalization remains the bottleneck. We argue that internalization should become a persistent object of training and maintenance-centered on condition-preserving conceptualization of events and deductive, inductive, and abductive reasoning across noisy, heterogeneous representations (text, tables, graphs, rules, and programs). To power this layer, we call for knowledge foundation models that provide shared neural capacity for knowledge computation, and outline a research agenda showing how an explicit knowledge layer delivers reuse, cost efficiency, auditability, and continuity across agent generations.

Spherical transformer configurations can accelerate grokking through simultaneous modifications of several model properties. The present study investigated the contributions of these modifications in the published Sphere+Fourier configuration using 960 controlled modular-arithmetic runs. A complete factorial analysis varied residual projection, readout weight normalization, logit scale, and weight decay. Across four matched settings with weight decay, the successful-run mean crossing times without projection were 8-37 times those observed with projection. However, residual projection was insufficient to induce grokking without additional interventions. Various other unprojected configurations also generalized. Since projection altered the magnitude entering the readout, these comparisons could not separate geometry from output scale. A separate 100-run fixed-horizon study demonstrated faster first crossing for the complete configuration across five train/test partitions. Faster crossing did not necessarily correspond to greater persistence over extended training. Control experiments demonstrated rapid grokking without Fourier initialization. Task and depth sweeps also identified conditions in which the configuration failed to reach the criterion by 50K epochs. Accordingly, the observed acceleration depended on interactions among the components and did not extend uniformly across the tested tasks or depths.

Machine learning supports materials design, atomistic simulation and experimental characterization, yet models and workflows remain fragmented. We present PaddleMaterials, an open-source framework integrating heterogeneous materials-learning tasks while preserving task-specific representations and scientific semantics. It unifies data processing, training, evaluation, pretrained-model management and workflows across six task classes: property prediction, structure generation, machine-learning interatomic potentials, electronic-structure fields, spectrum elucidation and microscopy image processing. Task interfaces record physical quantities, units, aggregation conventions and references, making cross-task composition verifiable. A shared compiler backend enables switching from dynamic-graph to compiled execution through one setting, using graph optimization, operator fusion and hardware-targeted code generation without changing model definitions. Across 65 model instances in 11 groups, compiled execution accelerated every instance in eight groups, with group median speedups of 11%-224% and group maximum speedups of 11%-336%. Case studies on LiMnO2 and toluene link structure generation or spectrum elucidation, interatomic-potential relaxation, property prediction and electron-density prediction, using same-geometry density functional theory (DFT) calculations as references. For LiMnO2, predicted formation energies for generated and relaxed structures differed from same-geometry PBE+U references by at most 7 meV per atom, while the computational cost of model steps was less than one-twentieth of a DFT single-point calculation. PaddleMaterials extends unified materials workflows beyond conventional property, potential and structure-generation tasks while supporting verifiable cross-task operations and compiled execution.

This paper reviews control architectures for building heating, ventilation and air-conditioning (HVAC) systems, addressing the tension among energy demand, occupant comfort and equipment operability, and situating proportional-integral-derivative (PID) regulation within a layered control stack that also includes rule-based and model-based supervisory control and data-driven reinforcement learning (RL). It examines the thermostat-PID layer from two-position and scheduled setpoint control through proportional, PI and PID regulation and their tuning, then formulates RL for building control as a Markov decision process. Its central contribution is a taxonomy of where RL meets PID: RL as a gain tuner for an existing loop, as a direct replacement for the regulatory controller, as a hybrid with model predictive control, and as supervisory setpoint and schedule generation. Across these classes, the review establishes that the physical loop remains invariant while the intervention point changes, and that the classes differ in their requirements for actuation authority, state information, modelling, commissioning and constraint handling. It further establishes that PID is inexpensive and well codified but limited by nonlinearity, time variation, discrete mode switching, loop interaction and slow plant dynamics, while RL's admissible position must be argued rather than assumed. The paper concludes that these architectures are not competing philosophies of control but nested points of attachment on one HVAC plant, with hybrid implementations possible, and that deployment trade-offs must be assessed according to the specific control interface and scaffolding required.

Battery degradation remains a central challenge in the development and broad application of sustainable energy technologies. Accurate degradation prediction is challenging, as battery aging emerges from complex and heterogeneous interactions among cycling behavior, operating conditions, and cell chemistry. Existing machine-learning approaches typically focus on deterministic end-of-life predictions or modeling degradation curves within fixed chemistry settings, leaving probabilistic state of health (SOH) trajectory modeling across heterogeneous chemistries underexplored. We introduce FlowBatt, a use-inspired conditional generative framework that adapts flow matching with a diffusion transformer (DiT) backbone to battery degradation trajectory modeling. FlowBatt models full SOH trajectories as probabilistic degradation processes conditioned on early-cycle capacity data. To strengthen this conditioning pathway, we use explainable AI analysis to diagnose and improve the encoder that maps early-cycle capacity matrices into conditioning vectors. We evaluate FlowBatt on five public benchmark settings spanning diverse chemistries and aging conditions, comparing it with supervised and diffusion-based trajectory models as well as established baselines for remaining useful life (RUL) prediction under shared data splits. FlowBatt achieves the best trajectory-wise SOH prediction performance on all five datasets, with errors below 2.5%, and delivers competitive RUL prediction performance. By generating multiple plausible trajectories, FlowBatt provides empirical prediction intervals for battery degradation modeling. These results indicate that flow-matching-based trajectory modeling is a promising framework for probabilistic battery health prediction.

LLM post-training via reinforcement learning has shown consistent improvements in accuracy at the cost of reducing behavioral coverage - a phenomenon known as sharpening that can stall progress on difficult tasks. Repeated sampling objectives such as $\mathrm{pass@k}$ and $\mathrm{max@k}$ are known to mitigate sharpening, but the mechanism behind it remains poorly understood. Moreover, existing approaches for $\mathrm{@k}$ optimization use dedicated algorithms, and no general framework exists for adapting established policy optimization methods from $\mathrm{@}1$ to $\mathrm{@k}$. We derive the optimal policies for KL-regularized $\mathrm{pass@k}$ and $\mathrm{max@k}$ objectives in closed form. Our explicit solutions reveal that $\mathrm{@k}$ optimization naturally redistributes learning pressure based on prompt difficulty, directing it toward harder problems. Crucially, we show that standard $\mathrm{@}1$ objectives achieve these exact same optimal policies through prompt-adaptive regularization or reward transformation. Ultimately, this yields an optimizer-independent explanation for why $\mathrm{@k}$ objectives reduce sharpening on easy tasks while prioritizing success on harder ones. Building on these insights, we introduce $\mathrm{@k}$ Adaptive Pressure for Policy Optimization (KAPPO). Following a one-time reference calibration, KAPPO seamlessly adapts existing $\mathrm{@}1$ methods to $\mathrm{@k}$ via prompt-adaptive regularization or reward transformation, without otherwise altering their training procedures. Experiments with several RL methods on mathematical reasoning and code generation demonstrate that KAPPO provides direct control over sharpening. Crucially, we demonstrate that $\mathrm{@k}$ pressure can be fully realized with single completion training, matching the performance of multi-completion $\mathrm{@k}$ optimizers at a fraction of the cost.

Procedural modeling offers compact, editable, and reusable representations for architectural content, but building such models still depends heavily on manually authored grammars. We study image-based procedural building reconstruction from in-the-wild images, with the goal of recovering an executable procedural description that can be instantiated as an editable 3D building asset. Because a single image provides only partial and view-dependent evidence, we focus on recovering the dominant visible facade(s) and attach the recovered facade program to a simple mass model for building-level instantiation. Our method decomposes this task into two structured prediction stages. First, a Global Structure Layout Tree recovers facade-scale split-and-repeat organization through a fine-tuned layout model with render-guided visual-symbolic reflection. Second, a Component Parameterization schema maps representative facade tiles to structured procedural attributes for local architectural component synthesis. These two representations are then deterministically compiled into executable Computer Generated Architecture (CGA) programs, avoiding the brittleness of direct low-level program generation. Experiments show that the proposed framework handles both regular facades and challenging in-the-wild images, while maintaining high structural consistency, procedural executability, and fine-grained architectural detail. The recovered programs further enable controllable and reusable 3D building asset generation.

Scaling depth is a key route to stronger large language models, but Pre-LN architectures often suffer from depth-wise information dilution, limiting the effective use of deep layers. While cross-layer connectivity has proven effective at mitigating this problem, existing designs usually introduce additional computation or memory access. We revisit cross-layer KV sharing as an efficient form of such connectivity. Beyond cache compression, each KV reuse edge creates a direct gradient path from deep layers to source KV projections, turning KV sharing into an implicit structural regularizer for KV representation geometry. We provide a systematic analysis of this mechanism, showing that proper sharing topologies sharpen key retrieval directions while enriching the value content subspace, whereas existing adjacent or global-anchor designs do not fully exploit this effect. Motivated by this, we propose DepthGraft, a hierarchical KV-sharing architecture that reconstructs deep-layer KV from stride-aligned shallow-middle source pairs. DepthGraft distributes non-local gradient flow across source layers, while preserving prefill early completion and reducing KV cache by roughly one third. Experiments on dense and MoE models from 0.5B to 16B parameters show consistent downstream improvements under both GQA and MLA. Notably, on the 16B MoE model, DepthGraft achieves a 7.0% absolute improvement on MMLU-Pro, along with gains of 6.6% on CMMLU and 6.4% on C3. These results validate the potential of DepthGraft as a scalable design for large-scale pretraining. The project page is available at https://github.com/jianke0604/DepthGraft.

On-policy distillation (OPD) mitigates the student's exposure bias by optimizing on student-generated trajectories. However, student-generated prefixes may contain reasoning errors or follow patterns unfamiliar to the teacher, undermining the reliability of its token-level supervision. This degradation becomes more pronounced as prefixes grow longer. To address this teacher-side distribution mismatch, we propose Student-Prefix Teacher Adaptation (SPTA), which adapts the teacher to the states on which it is queried during distillation. Specifically, before OPD starts, we collect student rollout prefixes with random lengths and continually train the teacher to continue from the prefixes using reinforcement learning with verifiable rewards (RLVR). By treating each prefix as context and rewarding correct final answers, the teacher learns to recover from student-induced states rather than imitate the student's reasoning. Then, the adapted teacher is used for standard OPD, with no changes to the distillation objective. Experiments on mathematical reasoning benchmarks show that our SPTA improves average student performance over standard OPD and OPD with task-level RLVR-adapted teachers. These results highlight the importance of adapting teachers not only to solve problems independently, but also to provide reliable guidance on the states their students actually visit.

Despite recent advances in CUAs, their development remains constrained by the limited diversity, controllability, and reliability of interaction data. Moreover, conventional optimization pipelines lack iterative feedback, while existing recursive improvement approaches often rely on self-generated experiences, limiting ability expansion and yielding diminishing returns. To address these challenges, we introduce CUA-Master, a nested double-loop recursive improvement framework. At the system level, an outer world-data-model flywheel continuously incorporates open-world corpus and leverages evolving model weaknesses to guide subsequent data construction and optimization iterations. At the data level, an inner flywheel couples open-world seeding and autocatalytic propagation. The former mines authentic workflows as seed tasks from instructional videos, and employs visual-semantic dual-path trajectory reconstruction; The latter transforms these seeds into reusable primitives for controllable synthesis of diverse, complex, and verifiable tasks. To enable efficient and reliable trajectory collection, we further introduce a feedback-guided snapshot-rollback retry mechanism in various real environments. Extensive experiments demonstrate that CUA-Master-27B achieves competitive performance across GUI-only and GUI-CLI hybrid benchmarks, scoring 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam. Our framework also substantially improves the quality and efficiency of data synthesis, and exhibits continuous gains across recursive iterations.

Assessing regulatory compliance for high-risk AI systems requires systematic identification of technical gaps such as missing controls, in- adequate processes, or shallow evaluation of practices. Compliance officers typically admin- ister structured questionnaires collecting evi- dence of current practices, where organizations respond with descriptions of their existing pro- cesses, technical capabilities, and architectural choices. However, translating these compli- ance questionnaire responses into typed, ac- tionable gap diagnoses remains unsystematic. In this paper, we present a literature-grounded taxonomy of technical compliance gaps cover- ing areas such as accuracy, robustness, logging, and human oversight, and embed it in GapTo- Compliance, a two-stage agentic framework coordinating two specialized agents: a Gap Extraction Agent that performs hierarchical rea- soning from compliance areas to specific gap types, and a Remediation Planning Agent that ranks remediation actions from a remediation catalog via multi-signal scoring combining se- mantic similarity, regulatory alignment, and effort estimation. We evaluate this framework against four automatic compliance gap detec- tion baselines across 9 AI systems and 6 AI families. Our approach achieves a gap-type F1 of 0.831 (vs. 0.533 for the strongest baseline) and remediation P@5 of 0.872 (vs. 0.635 for retrieval methods), demonstrating promising potential for automated compliance workflows. The structured agent coordination reveals a path toward human-in-the-loop compliance systems where automated gap extraction reduces expert burden while maintaining auditability, and the taxonomy-driven design enables extension to emerging regulatory frameworks beyond

Active-learning Kriging often replaces candidate pools by continuous metaheuristic search. A pre-specified comparison sharing surrogate, population and stopping rule isolated this inner optimizer: five budget-matched continuous optimizers, including hybrid swarm optimization with L-BFGS-B polishing, against candidate pools on eleven problems (dimension 2 to 100). The primary hypothesis was not supported: against optimizers sharing its criterion the hybrid won 6, tied 41 and lost 19 of 66 Holm-corrected accuracy comparisons, and against pool methods 0, 23 and 21 of 44. With sequential selection on three series systems, multistart L-BFGS-B left a failure mode unlearned in 59 of 60 runs, more often than swarm, genetic-algorithm or AK-MCS. In all 198 runs (22 post-freeze) with an unlearned mode, the learned modes did not compensate and failure probability was underestimated (exploratory). The error criterion can accept substantial underestimation: it accepted 17 runs (13 pool-selected) at 0.38 to 0.89 of the reference; 181 exhausted the budget. Follow-up analyses on single-design-point problems indicate underestimation when selections concentrate near the design point. Kriging-believer pool selection satisfied the surrogate stopping criterion with fewer calls than the best continuous method in all four informative batch cells (two post-freeze). We recommend default pool selection and coverage checks for any selection.

Large language models (LLMs) are driving autonomous agents toward real-world environments, with the Web emerging as one of the most representative interaction scenarios. However, training Web agents with naive exploration remains challenging: successful trajectories are rare, and the sparse terminal rewards make it difficult for agents to obtain useful feedback from failed trajectories. Inspired by skill acquisition theory, we propose Skill-Composed Progressive Exploration (SCOPE), a short-to-long learning framework for long-horizon Web agents that decomposes complex task exploration into a gradual expansion process from local skills to complete task-solving capability. Specifically, SCOPE iteratively explores around local skills: Skill Discovery extracts reusable local skills and failure experiences from historical rollouts; Skill Orchestration composes discovered skills into high-level skill chains; and Skill Expansion leverages failure experiences aligned with the current skill chain to guide the agent toward exploring new subskills. Through this process, SCOPE progressively constructs task-completing skill chains across rollouts and internalizes them into the policy parameters through agentic on-policy distillation. Extensive experiments on WebArena, WebVoyager, and WebShop show that SCOPE consistently improves exploration dynamics and achieves 3.9% to 18.6% relative gains over the strongest baselines in each setting.

Large language models (LLMs) have introduced a generative approach to recommendation, producing item lists from textual representations of user histories and item information. However, this generative setting raises challenges concerning output validity, prompt consistency, and sensitivity to demographic descriptors. In this work, we investigate how preprocessing choices in fine-tuning data affect recommendation utility and fairness-related behavior in a generative LLM-based recommender system. We first establish a controlled recommendation pipeline that combines supervised fine-tuning, catalog-constrained decoding, and prompt augmentation. We then study three preprocessing choices: item-title representation, demographic balancing, and auxiliary item-knowledge enrichment. Experiments on MovieLens-1M show that the catalog-constrained pipeline eliminates out-of-catalog items and ensures complete ten-item outputs, while prompt augmentation improves prompt consistency and utility. Original MovieLens titles increase HR@10 from 10.67% to 12.47% and Coverage@10 from 41.36% to 47.88%. Fairness-related effects are attribute-dependent: age balancing and knowledge enrichment reduce selected group-level utility gaps while largely preserving utility, whereas gender balancing reduces utility without improving the evaluated gender-related indicators. Overall, preprocessing choices in fine-tuning data should be evaluated jointly in terms of utility, output behavior, and demographic sensitivity.

Large Language Models (LLMs) are now embedded in enterprise workflows, yet their outputs remain prone to hallucinations: plausible-sounding but factually incorrect or malformed content. Most verification approaches are reactive, detecting errors only after generation and often spending additional expensive LLM calls to do so. We present COVE, a task-adaptive, code-based ensemble that takes the opposite stance: it prevents errors before generation and validates structure and consistency afterwards using six deterministic, pure-Python innovations that add no extra LLM calls. On an internal evaluation spanning 11 text datasets (1,020 examples) and 4 image datasets (200 examples), COVE lifts text accuracy from a 45.2% unverified baseline to 82.1% (an 81.7% relative improvement), improves image accuracy by 69.5%, and reduces token usage by 48 to 96% (simulated cost model), at a measured verification overhead of 0.26 ms per output (excluding the LLM call). This article describes the engineering design of each innovation (with pseudocode and the exact parameters we use), the measured impact, an ablation of every component, and the practical lessons learned from deploying code-based verification in enterprise pipelines.

Accurate preoperative prediction of lung adenocarcinoma invasiveness is important for surgical planning and individualized treatment. However, overlapping CT appearances make it challenging to distinguish invasive adenocarcinoma (IAD) from Pre-IAD lesions. We propose a multimodal prompt-learning framework with dual-granularity image-text alignment for 3D CT-based IAD classification. A 3D ResNet-CBAM encoder extracts global and spatial visual features. A frozen RadBERT backbone is adapted with separate learnable prompts for report-level and keyword-level text encoding. These prompts are initialized using semantic representations derived from real-world clinical reports. Global contrastive alignment connects whole-volume CT representations with reports generated from clinician-annotated radiological attributes to provide overall semantic supervision. A multiple instance learning-based local alignment module associates spatial CT features with vascular-focused radiological keywords, to capture local imaging characteristics potentially associated with tumor invasion. The framework jointly optimizes classification and alignment objectives, while inference requires only CT images. Experiments included 1,593 pathologically confirmed nodules. On the test set, the proposed method achieved an accuracy of 87.92%, F1-score of 87.78%, and AUC of 0.9499. Its AUC significantly exceeded that of the image-only Res-CBAM baseline with the same visual backbone (0.9297; DeLong test, p = 0.0078). Ablation analyses supported the complementary contributions of global and local alignment and the benefit of report-based prompt initialization. Qualitative visualizations showed activation within nodules and adjacent vascular structures. These findings suggest that dual-granularity radiological supervision can enhance CT-based invasiveness prediction while preserving a CT-only inference workflow.

We classify real and complex state-space models (SSMs) with linear transitions, distinguishing fixed-basis diagonal equivalence from input-output equivalence and tracking coefficient-conditioning windows under coordinate changes. We characterize diagonal equivalence through commuting transition families whose members are diagonalizable over the chosen field. With time-shared token-local matrix maps, we identify this class's state-mediated kernels with unnormalized scalar-mode linear attention with multiplicative masks. Mamba 3's multiple-input, multiple-output (MIMO) construction lifts one value per local interface; SSD-MIMO routes independent values for local retrieval rank above one. HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style heads and Mamba 3's complex rotations and two-endpoint writes, mixing real and rotating state blocks with different decay laws. They recover Mamba 1-3 single-input, single-output (SISO) configurations as exact special cases. On five large H100 bfloat16 workloads, our fused GPU kernels improve geometric-mean training time over official Mamba 1/3 in matched SISO recoveries, with Mamba 2 near parity at SSM-core scope. Memory tradeoffs depend on the baseline and metric: saved-for-backward versus peak allocation. In small-model multi-token recall experiments, both designs outperform Mamba 1-3 on the hardest tier; higher routing ranks perform best in gated or phase-diverse configurations. In single-seed language models with about 500M parameters, both improve on Mamba 1/2 in test loss at 2B tokens, but the routing gains do not translate into lower language-model loss. Our best configurations approach Mamba 3 SISO through 10B tokens while remaining behind it and Mamba 3 MIMO in language-model loss.

An experimental agent can succeed for the wrong reasons. Evaluations, however, often focus on the outcome, not the experiments, evidence, and inferences that produced it. We introduce LUCA, a configurable synthetic world for evaluating this process under fully known ground truth. Agents learn unknown regulatory systems through budgeted experimentation, then predict unseen interventions, diagnose hidden faults, and attempt rescues. This setup allows us to evaluate experiment selection, prediction, diagnosis, and intervention separately. Across seven frontier LLMs and three baselines, these capabilities dissociate sharply. Agents frequently rescue malfunctioning systems while diagnosing the wrong fault; diagnosis varies 7x across models while prediction varies only 1.4x. Models also adopt surprisingly different experimental strategies, and greater exploration does not reliably yield better performance. In one failure mode, agents often perform the experiment needed to distinguish competing causes, but still fail to use the resulting evidence. LUCA turns these hidden failures into measurable behaviours, providing a controlled environment to study where experimental agents succeed and fail. Code: https://github.com/RelationRx/luca.

Complex documents such as scientific papers contain rich information in tables, figures, and equations; however, existing deep research (DR) agents typically reduce papers to OCR-extracted text. We instead ask whether vision-language models (VLMs) can enable page-native deep research: directly retrieving and reasoning over rendered document page images. We introduce DR Pixel, the first open-source end-to-end multimodal DR agent that uses rendered page images as its sole retrieval and reasoning unit, operating over a corpus of 1.06M pages. To evaluate this setting, we construct DRPixelBench, a benchmark on scientific papers requiring reasoning over text, figures, tables, and equations. Surprisingly, page-native DR already works well without training, outperforming OCR-based baselines across state-of-the-art models like GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, while requiring fewer tool calls. We next turn to DR training. A key challenge is that teacher trajectories for agentic supervised finetuning are noisy, mixing critical reasoning and action steps with redundant or failed attempts. To address this, we introduce counterfactual credit reweighting, a fine-grained credit assignment method that weights individual reasoning and action steps by their counterfactual contribution to the final answer. Our training improves retrieval recall by about 15 points and final accuracy by over 10 points on DRPixelBench, and likewise outperforms the untrained baseline on four existing benchmarks. On out-of-distribution ViDoRe-V3, DR Pixel comes within 1.8 points of an oracle that is given the gold pages directly. Our results suggest that the conventional OCR-then-retrieve pipeline is not only unnecessary but suboptimal: directly retrieving and reasoning over visual document pages makes complex DR both more effective and more efficient.

Theoretical limitations of transformers on state tracking---the ability to track information about entities or variables over long contexts---have motivated new recurrent language modeling (LM) architectures such as Gated DeltaNet (GDN). These architectures have, in particular, been benchmarked against the $S_5$ word problem, a theoretically hard state-tracking task that captures abilities such as code evaluation under arbitrary variable assignment. It, however, remains open whether other, simpler forms of state tracking might both be more relevant in practice and already explain observed differences between these architectures. We propose evaluating state tracking using a hierarchy of word problems with four increasing levels of difficulty: r-trivial, corresponding to immutable assignment to variables in programming; aperiodic, where assignments are mutable; soluble, where assignments are self-referential; and general finite state tracking. Theoretically, we show that the class of problems transformers are thought to length-generalize on does not even contain all aperiodic state tracking. On the other hand, we prove that GDN with positive eigenvalues expresses all aperiodic problems, and the regular languages recognized by certain transformers are precisely all soluble ones, thus providing a fine-grained picture of state tracking capabilities of various architectures. Taking these ideas to practice, we construct StateBench, a 4-tier long-horizon state-tracking benchmark. Our systematic direct training experiments across architectures largely align with theoretical expectation, albeit with interesting observations at very large scale length-generalization. Additionally, we also evaluate pretrained models---both open hybrid and closed frontier ones---on StateBench, finding that none of them robustly solves even soluble state tracking, suggesting a direction for future improvement.

Claim verification is the task of deciding whether a source text supports a claim, contradicts it, or lacks sufficient evidence. Natural Language Inference (NLI) models can judge whether a single sentence supports or contradicts a claim, but they struggle when the evidence is spread across several sentences. Conversely, Large Language Models (LLMs) can reason across sentences, but may draw conclusions the text does not support. In this experimental study, we test whether a conservative fusion rule, labeling a claim Support or Contradict only when both models agree and defaulting to Not Enough Evidence otherwise, combines their strengths. We construct a benchmark of 2,035 claims from SciFact and CliniFact (858 Support, 152 Contradict, 1,025 Not Enough Evidence). We retrieve the top five most relevant sentences per claim using a hybrid of semantic and lexical (BM25) retrieval, which slightly outperforms either method alone. These retrieved sentences are then provided to every verification model. We then compare a DeBERTa-v3-base-mnli-fever-anli NLI baseline against Llama 3.1 8B, Qwen3 8B, and Mistral 7B. Qwen3 achieves the highest Macro-F1 (0.7937), followed by Llama 3.1 (0.7204) and Mistral (0.6388), while the NLI baseline reaches only 0.3157. Contrary to our hypothesis, fusion degrades performance, dropping Macro-F1 to 0.7392, 0.6730, and 0.5991, respectively. To understand this decline, we analyze agreements, disagreements, and common error patterns between the NLI and LLM verifiers. Our error analysis shows that NLI and LLM disagreements often result in correct LLM predictions being overturned.

Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to invest substantial effort in specifying goals, resolving ambiguity, and verifying outcomes. We argue that progress should be measured by how effectively agents amplify human agency: enabling people to pursue their goals with less routine supervision while retaining control over consequential decisions. We advocate Avatar Intelligence, a paradigm of persistent personal agents that represent users' goals, values, and constraints through sustained decision delegation across contexts and time. We describe a maturity roadmap from human-directed execution to autonomous operation within delegated domains, and introduce the Agentic Oversight Index (AOI) as a framework for evaluating delivered value against agent-side operational cost and human oversight burden. We identify three core research challenges: identity-continuous personalization that carries user preferences across tasks, user-situated evolution through Rehearsal Engineering, and safety mechanisms that combine technical control with accountability and recourse. These directions define a research agenda for reliable decision delegation to personal agents across tasks and over time.

While Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced multimodal reasoning, existing frameworks suffer from a critical "perception-reasoning gap." Due to an over-reliance on seemingly relevant visual tokens, even minor perceptual perturbations can propagate through the reasoning process, leading to compounding hallucinations. To address this issue, we propose MIRA, a training framework that improves robustness to erroneous visual contexts while enhancing logical reasoning ability. By injecting filtered, deceptive visual contexts during training, we construct challenging "hard traps" that stress-test the model's robustness to misleading information. Furthermore, we introduce a group-wise reflection-triggering mechanism that enables the model to autonomously detect and rectify inconsistencies between visual cues and logical constraints, which is ultimately internalized as an intrinsic policy behavior. Importantly, MIRA does not rely on human annotations. Experimental results across multiple reasoning and perception benchmarks show that MIRA yields an average 7% improvement on in-domain benchmarks, with consistent improvements on vision-centric benchmarks. These findings highlight that active error recovery is essential for robust multimodal reasoning. Comprehensive ablation studies and analyses further provide insights into why MIRA is effective.

Large language model (LLM) embeddings have been increasingly adopted in sequential recommendation to encode static item attributes and alleviate the limitations of item ID embeddings under sparse interactions. However, these embeddings typically capture only static content similarity while neglecting sequential transitions, and since their adaptation quality relies on sparse per-item updates or global geometric calibration, existing methods still struggle to explicitly transfer preceding interaction contexts from frequent to long-tail items under unevenly distributed behavioral evidence, limiting recommendation effectiveness across the catalog. To address this challenge, we propose PACERec, an ID-free progressive adaptation framework that incorporates preceding interaction contexts into semantic representation adaptation by learning a shared velocity field from items with sufficient interactions via frequency-weighted flow matching. In this way, PACERec progressively guides static LLM embeddings toward collaborative behavioral targets while preserving original item semantics, overcoming sparse per-item updates to effectively generalize sequential transition patterns to long-tail items for both user modeling and candidate scoring. Extensive experiments on three real-world datasets across three sequential backbones demonstrate the superiority of our method over competitive baselines in both overall and long-tail recommendation.

Large language model applications are increasingly being organized as tool-using and multi-agent workflows rather than as isolated chat interfaces. This perspective paper examines the architectural problem that follows from that shift: agents may reason, invoke tools, and delegate work, but dependable deployment still requires interoperability across frameworks, runtimes, data boundaries, and organizational control domains. We study this problem through the lens of the \emph{Internet of Agentics} (IoA), an emerging service-computing view in which heterogeneous agents are treated as governable service participants rather than only as stand-alone model endpoints. The paper organizes the landscape as a three-layer progression: from today's fragmented agent ecosystem, to interoperable agents, and ultimately to trustworthy and governable autonomy. First, it examines the protocol stack centered on the Agentic AI Foundation, including MCP for agent--tool interaction; A2A for agent--agent coordination; AGENTS.md and Skills-style procedural manifests; and discovery, transport, and trust mechanisms such as AGNTCY and SLIM. Second, it reviews the systems abstractions that make agents deployable as service components, including logical agent architectures, orchestration patterns, gateways, registries, runtime control loops, and hardware-aware inference infrastructure. Third, it situates these developments within services computing, arguing that agentic systems reinterpret classical concerns from SOA and microservices---discovery, composition, governance, pricing, and trust---at the level of cognitive capabilities rather than deterministic functions.

Recent advances in Artificial Intelligence (AI), particularly large-scale deep learning, have been driven by monolithic or offline training on massive curated datasets. While successful, these approaches face structural limitations: escalating data and computational demands, limited online adaptability, and a lack of system-level intelligence in open, non-stationary environments. This paper presents Networked AI - a networked intelligent system composed of interacting, heterogeneous, and continuously evolving AI agents. We establish a hierarchical layered architecture comprising an Agent Layer for heterogeneous capability construction, an Interaction Layer for signal-processing-guided online data extraction, and a Regulation Layer for stability-constrained collective evolution. Foundations in data science, signal processing, and system control are hereby synthesized into a unified theoretical framework of Networked AI, providing superior data efficiency, intelligence efficiency, interpretability, and scalability. Theoretical and experimental analysis demonstrates that Networked AI can approach near-linear scaling in resource demand, in contrast to the exponential increase in standalone end-to-end systems. It provides not only online adaptation to out-of-distribution environmental novelties, but also the autonomous evolution towards application-specific intelligence. Application frontiers are discussed for IoT spaces, embodied, spatial, and industrial intelligences, as well as the cellular infrastructure and beyond. Key open challenges are identified towards realizing locally autonomous, collectively intelligent, continuously evolving, globally regulated Networked AI.

We propose a simple, linguistically motivated approach to detecting AI-generated peer reviews based on the inherently human trait of subjectivity. We capture subjectivity at two levels: the semantic level of evaluative language and modality and the pragmatic level of speech acts. Using features derived from speech act annotations and model-predicted subjectivity scores, we train a lightweight classifier that achieves an AUROC of 0.9566 with only these linguistic features. Combining its output with the state-of-the-art detector Sem-Detect raises the TPR at 1% FPR of Sem-Detect on the APRDB benchmark from 0.874 to 0.985, outperforming the current state of the art. Our results show that linguistic features of subjectivity provide an explainable source of information for detecting AI-generated peer reviews and can be readily integrated into existing detection methods. The code and models are publicly available on GitHub (https://github.com/Arys02/generated\_ text\_detection\_hub) and Hugging Face (https://huggingface.co/Arys02), respectively.

Feedback-driven iteration is widely used to improve LLM-generated scientific ideas, but existing systems either leave revision decisions implicit in generation, align revisions with feedback-localized weaknesses, or restrict action selection to system-specific operations. This is problematic because problems, hypotheses, and solutions are interdependent: the location where a weakness manifests does not itself determine the revision scope that will produce a better successor. We term this phenomenon diagnosis-intervention mismatch. We introduce Predictive Routing of Interventions by Modeling Effects (PRIME), a plug-and-play feedback-to-revision decision layer for existing scientific ideators. PRIME operates in three stages: (1) Structural Attribution interprets feedback over the problem-hypothesis-solution structure without fixing a revision target; (2) Counterfactual Routing constructs alternative scope-conditioned interventions and compares their projected successor utility and risks through lightweight rollouts; and (3) Intervention Execution compiles the selected intervention into an ordered generator-facing plan with semantic actions, preservation constraints, and dependency requirements for successor generation. We validate diagnosis intervention mismatch on 112 human-annotated idea-feedback pairs. Across four benchmarks, two ideators, and fixed- and open-problem settings, PRIME improves successor quality and routing reliability over strong baselines; removing counterfactual routing reduces quality from 3.463 to 1.960 and increases route regret from 0.426 to 1.929, confirming cross-scope effect modeling as the primary source of its gains.

Beamforming designs for integrated sensing and communication (ISAC) in cell-free multiple-input multiple-output (MIMO) systems face the challenge to stay reliable under uncertainties from both channel state information (CSI) as well as clutter. To this end, this paper proposes an impairment-aware heterogeneous graph neural network (GNN) framework for beamforming that obtains the CSI error level and the clutter covariance as edge features, which enables a single trained model to adapt to both impairments. In particular, the GNN is trained on a low quantile of the minimum user rate with a clutter-aware signal-to-clutter-plus-noise ratio, balanced by a normalized Tchebycheff scalarization. Simulation results show that the proposed GNN framework outperforms the considered baselines in sensing-only and communication-only settings. Depending on the operating point, a single GNN model attains approximately 97\% of each baseline's rate while serving both functionalities at once. The model further generalizes to unseen CSI error levels and clutter counts and renders the ISAC system resilient.

Subgroup discovery aims to identify regions of the feature space where a target variable deviates from its marginal distribution. A fundamental challenge in this problem is the trade-off between a subgroup's deviation and its support: small subgroups can be highly deviating but statistically meaningless, while large ones cannot deviate much by construction. Existing methods tackle this challenge by collapsing the two into a single scalar quality measure, implicitly committing to one particular trade-off and making it difficult to recover the most deviating subgroup under a user-specified support constraint. We instead provide the first theoretical and structural characterization of optimal subgroups with maximum Kullback-Leibler (KL) divergence under arbitrary support constraints. Under a partition-based generative model, where the feature space decomposes into regions of homogeneous conditional distributions of the target variable called atoms, we prove that every point on the deviation-support Pareto front is a union of whole atoms together with at most one fractional atom. This reduces an uncountable search to a finite combinatorial problem. Guided by this theory, we propose a simple two-step recipe: learn a homogeneous partition, then select atoms greedily. Across 19 benchmark datasets, it finds the most deviating subgroups at a range of support thresholds, and reliably improves subgroups returned by existing methods when applied as a post-processing step. These results empirically validate the sufficiency and necessity of our theoretical contributions.

Criminal verdicts contain descriptions of criminal behavior that can support quantitative legal research, but structuring them requires attributes and extraction rules that cannot always be fully specified in advance. We present an LLM-based pipeline that builds a structured dataset through attribute exploration, text span extraction, and value normalization, guided by annotation instructions that legal experts can refine. We evaluate it on two traffic offenses: driving under the influence (DUI) and driving despite a ban (DDB). For DUI, revising the instructions was associated with higher extraction accuracy (from 52% to 70% of correctly extracted values) and higher agreement between human annotators. Repeated runs of the same model agreed with each other more than with the gold annotations, showing that consistent output is not necessarily correct, while disagreement between annotators and between models helped reveal unclear instructions. For DDB, where the instructions defined fixed categories and output formats, agreement was high for most attributes.

Complex video editing requires understanding complex instructions while jointly fulfilling multiple editing requests. Meeting these demands through dense model scaling substantially increases computational cost. We introduce VideoEditor-MoE, the first open-source mixture-of-experts framework for video editing. Beyond sparse MLP-MoE layers, MoE-Attention expands attention capacity through routed key--value experts while preserving global interactions among video tokens. Initialized from a 5B video editing model, VideoEditor-MoE scales to 14B and 56B parameters, with approximately 8B parameters activated per token in both variants. Comparisons with existing methods and two complementary baselines, Explicit-MoE and MLP-MoE, assess its benefits over explicit expert guidance and MLP-only expert architectures. Further analyses demonstrate that all experts remain active and reveal that learned routing responds more strongly to sample content and diffusion noise levels than to human-defined editing categories. Spatial analysis further indicates sensitivity to object-level semantic features. We also identify decoupled weight decay as a strategy for stabilizing expert scaling.

Adaptive compressed sensing (CS) aims to select the most informative measurements sequentially, which is crucial for many modern imaging systems. This task can be formulated as a Bayesian sequential optimal experimental design (sOED) problem, where each new acquisition is chosen to maximally reduce uncertainty about the unknown image given prior information and subsequent data provided by acquisitions. However, in single-pixel imaging (SPI), the large design space consists of all possible binary illumination patterns. Directly optimizing the Bayesian sOED objective over this space using conventional methods is generally challenging. In this work, we reformulate sOED for SPI as a reinforcement learning (RL) problem. In RL, an agent learns a policy through repeated experimental trials to maximize a reward, here given as the expected information gain, the most complete measure of uncertainty reduction in Bayesian OED. This avoids the need to directly optimize complex information-theoretic objectives during a measurement to obtain the next design. Instead, the result is a learned policy function that directly maps the current posterior belief about the image to the next optimal binary measurement pattern, in a computationally efficient way. Using a simple yet representative SPI imaging scenario, this work evaluates the scalability of the learned policy under two different policy parameterizations. Furthermore, we investigate how performance deteriorates when only approximate posterior information is available or the measurements are impaired by noise, and propose several strategies to introduce additional design constraints motivated by biomedical applications. The numerical experiments demonstrate that the proposed Bayesian sOED-RL framework remains robust and can be extended to more complex, large-scale SPI scenarios.

Reasoning language models write a chain of thought before they answer, and many checks in it do not change the final answer. Natural language autoencoders (NLAs) explain a model's activations in text and rebuild activations from text, and editing an explanation can steer a model at a single token. We study whether such edits can also steer how a model reasons. We build VERS (Verbalize, Edit, Reconstruct, Steer): it turns explanation edits at the doubt points of one reasoning trace into one fixed direction, and adds this direction at paragraph breaks. On the full GSM8K test, VERS cuts generated tokens by 45.8% at an accuracy change of -0.82 percentage points (pp); on MATH-500 it saves 31.3% at -3.05 pp. At half strength, its trade-off between tokens and accuracy lies between those of the published methods SEAL and DEER. With fitted affine maps and a strength tuned for each model, the direction transfers to larger Qwen3 models, up to 235B-A22B, and saves 30-35% of tokens. On competition problems, however, VERS loses 3-7 pp at every model size, and a fixed direction did not separate necessary checks from unneeded ones.

Time-series segmentation, studied as either Change Point Detection (CPD) or State Detection (SD), underpins a broad range of monitoring and diagnostic applications. However, the field suffers from a persistent divide between the statistical community (focusing on CPD) and the machine learning and data mining communities (focusing on SD). Consequently, the literature remains fragmented: evaluations typically cover narrow methodological families, datasets are often small or homogeneous, and CPD and SD are rarely evaluated together. We address these gaps with TSB-SEG, a unified benchmark comprising 592 univariate and multivariate time series from 8 heterogeneous datasets and surveying 26 algorithms that span five decades of progress. To ensure a fair comparison, we introduce a two-pool design: a scalable main pool of tractable detectors used for the aggregate ranking, and an extended-scope pool for computationally intensive methods. All methods are unified under tsseg, an open-source library to ensure full reproducibility. Our systematic evaluation yields three insights: (a) no single detector dominates across domains, and per-dataset hyperparameter tuning remains decisive for performance, (b) CPD accuracy is a useful, though imperfect, proxy for SD accuracy, and (c) data characteristics should guide method selection. We provide actionable guidance based on data properties, error tolerances, and tuning budgets, identify impactful hyperparameters, and characterize algorithm robustness to these choices.

Design thinking is widely taught as an iterative loop of ideation, prototyping, and testing, yet conditions under which such a loop must improve innovation outcomes remain unspecified. Taking problem framing as given, we formalize the iterate-test core as search over a finite prototype space whose utility is learned through noisy evaluations. We prove that the prototype a team believes to be best converges almost surely to a local optimum under five conditions: finite framing, local ideation, non-summable discovery rates, unbiased repeated testing, and retention of tested prototypes. Bounded test bias or criteria drift degrades the guarantee by a bounded amount, and multiple stakeholders yield convergence to a locally Pareto-optimal set. Finite-horizon simulations across three landscape classes (additive valuation, NK rugged landscapes, and clustered modular architectures) show that asymptotic limits are rarely reached in practice. Teams face an optimizer's curse, overestimating their chosen concept by 4-8% and frequently discarding superior prototypes. Testing every prototype once proves to be the worst allocation; optimal performance requires building one to three prototypes per five tests and repeatedly re-testing leaders. Pre-screening concepts yields greater performance gains than tuning testing policies.

Multimodal fusion is studied under hybrid federation, where data are partitioned across samples and feature blocks. The proposed two-dimensional federated smooth multiple-kernel large-margin learning method (2D-FedSMKL) integrates availability-corrected modality scores, group-kernel regularization, and smooth worst-site risk aggregation. For fixed maps and propensities, the objective is convex and smooth. Curvature bounds and exact synchronous block-gradient equivalence are established for binary and multiclass scores, giving a conditional accelerated rate. An oracle-propensity binary risk bound and a propensity-perturbation result are also derived. Simulations and two public activity-recognition benchmarks assess site heterogeneity and modality absence. Worst-site aggregation improves worst-site performance under heterogeneity, including reversed and size-balanced designs. On the benchmarks, validation selects the correction; the selected model is not significantly worse than the strongest reference at the worst site, and group-regularization benefits vary by dataset.

We study regret minimization for time-homogeneous episodic tabular reinforcement learning with $S$ states and $A$ actions under non-interactive $ε$-local differential privacy (LDP). In each episode $k \in [K]$, the decision-maker sends a policy to a user, the user executes the policy for $H$ steps, and sends an $ε$-locally differentially private message back to the decision-maker. When $H=1$, the problem reduces to an LDP contextual bandit with $S$ contexts, $A$ arms and horizon $K$. We prove that the minimax regret is $\min\bigg(K, \frac{ \sqrt{\max( S, \exp(ε) + 1) S A K (\exp(ε) + 1)}}{\exp(ε) - 1}\bigg)$, up to constants and a $\sqrt{\log(A)}$ factor. The upper bound is achieved by an adaptation of EXP3 and a unary encoding mechanism. When $H > 1$, and for small $ε$, we show a lower bound in $Ω(HS\frac{\sqrt{A K}}ε)$ and an upper bound in $\tilde{O}(H^2 S^2 \frac{\sqrt{ AK}}ε)$. The upper bound is achieved by an algorithm based on UCBVI and a unary encoding mechanism. For small $ε$, these bounds improve the previously known lower bound by a factor $\sqrt{S}$ and the previously known upper bound by a factor of $H \sqrt{A}$.

Two machines can each recognize an object well and still fail together. We study classifiers that see different parts of one object and send short messages to a receiver that identifies its class. Coordination asks how to design messages together; cooperation asks whether the team outperforms one sender given the whole budget. On three digit datasets, messages designed together beat messages designed apart by 18.07 to 22.68 percentage points on average with two-bit messages, with gains in all 24 configurations. A fixed agreement, you report the row and I the column, recovers about half this gain, with only the receiver learned. Senders that draw the same distinctions waste a narrow vocabulary, and senders whose distinctions cross multiply it. A team has no more possible messages than one sender with the same budget, so a partner must bring evidence the first lacks. A solvable model shows individual statistics alone cannot determine whether a partnership helps. On MNIST digits, six shared bits remove 49.54 percent of the strongest tested solo's errors on average. Paired measurements correctly predicted most sensor teams would reduce average accuracy on new participants. Who should speak is a property of the pair, and must be measured in pairs.

Most fuel oil consumption (FOC) models for voyage planning are per-ship models and thus generalize poorly beyond each vessel's records. Transfer and federated learning have been reported to perform better, but under different protocols that often use inputs unavailable before departure or random data splits. It therefore remains unclear which approach performs best. In this study, we compare 11 per-ship, adaptation-based, and global methods on 21 vessels. To prevent leakage, all methods receive identical inputs known or forecast before departure, excluding engine RPM and shaft power, and are tested on intact voyage legs after all training records of each vessel. Global pooling, in which one model describes each vessel by its ship identity and seven ship specifications, performs best overall in every setting examined. For in-fleet vessels, Global LightGBM reaches an overall MAPE of 8.92%, against 11.14% for Per-Ship LightGBM, and outperforms every per-ship and adaptation-based model. Global models of five architectures, including a Transformer, all outperform the per-ship and federated models. Fine-tuning a global model for each vessel, even by low-rank adaptation of the Transformer, adds no accuracy. Without any records of a vessel, the global model reaches a median MAPE of 11.20% and outperforms a per-ship model trained on that vessel's records for 14 of 21 vessels. Adding its first records to the pooled data is as accurate as vessel-specific adaptation. Ship identity and specifications are largely interchangeable for in-fleet vessels, but only the specifications support zero-shot prediction. Its error grows as a vessel's specifications differ from the fleet's. Per-ship models and vessel-specific adaptation are therefore unnecessary under fair comparison. One global model can serve a whole fleet, including newly delivered vessels with similar specifications.

With the popularization of livestreaming applications, automatic camera tuning has become indispensable for ordinary users to enjoy high Quality of Experience (QoE). However, the existing Image Quality Assessment (IQA) methods face two challenges in camera tuning. First, they struggle to identify subtle quality difference between images, which is common in camera tuning. Second, the ability to explain fine-grained differences and to provide suggestions for camera tuning are even more lacking. To address these challenges, we first construct FGLive-Comments, an instruction tuning dataset with rich annotations of both preferences for fine-grained image pairs and quality reasoning from global and local visual evidence, which supports selective adjustment recommendations of camera parameters. Based on FGLive-Comments, we develop Q-CamTuning, a multimodal large language model for livestreaming camera tuning. Its progressive perception-evidence-guidance framework learns quality-aware visual representations, enhances fine-grained reasoning with global-local evidence, and maps quality issues to camera-parameter adjustment directions. Extensive experimental results demonstrate that Q-CamTuning outperforms existing IQA models in both quality scoring and fine-grained image pair ranking. User study also validates its ability to recommend camera tuning suggestions.

Long-horizon tasks require large language model (LLM) agents to coordinate interdependent actions, a capability central to automating complex workflows. Stepwise group-based reinforcement learning provides finer credit than trajectory-level supervision, but reliable comparisons must account for differences in interaction history. Enforcing this consistency within the limited rollouts of each task can leave insufficient support for estimating action preferences. We propose Beyond-Local Policy Optimization (BLPO), which estimates task-progress advantages from experience shared across tasks and training updates. We summarize trajectory prefixes into task-progress states and abstract actions by their function, allowing corresponding decisions from different tasks to contribute to shared estimates. To control bias and noise from heterogeneous tasks, we accumulate context-relative returns in a global memory with balanced aggregation and reliability weighting. These global advantages complement local comparisons over normalized actions and trajectory-level credit without a learned critic or additional rollouts. Experiments on ALFWorld and WebShop with 1.5B and 7B policies show improved task success over the evaluated baselines under matched training budgets. Further analyses show faster learning and shorter trajectories with little computational overhead.

Improving a sentiment classifier with limited resources involves more than simply deciding how much data to collect or which model to train. A labelled collection may lack examples of an important class. Once trained, a classifier may benefit from another model even when that model is weaker on its own. We study these decisions in document-level Entity-Level Sentiment Analysis (ELSA), primarily using South Slavic news. We first compare acquisition orders as collections grow. We then use uncertainty to choose which predictions to revisit, how many to process, and whether an alternative encoder or local LLM can correct them. Finally, we select complete routing policies on development data for held-out evaluation. Minority-prioritised acquisition helps at some very small training sizes, but not consistently thereafter. On SentiCoref, an oracle correcting all errors in the least-confident 10% without harming correct predictions would add 9.14 Macro-F1 points; however the best evaluated action gains only 0.90. Encoder averaging helps on AspectBench and NewsMTSC but harms SentiCoref, where selective LLM assistance is more promising. Both development-selected policies route 30% of records to the encoder-assisted LLM, yielding observed gains of 1.02 points on SentiCoref (interval includes zero) and 1.71 on NewsMTSC, at substantially different inference costs.

The scarcity of human-annotated literary corpora limits the development of domain-specific sentiment analysis models. This study investigates whether LLM-generated annotations can supervise smaller models for literary sentiment analysis. We fine-tune five BERT models using ordinal regression on 10,889 literary sentences annotated by Gemma 4 31B on a ten-point sentiment scale. The models are evaluated against 1,923 human-annotated sentences from Hemingway's "The Old Man and the Sea", alongside the teacher model and conventional baselines. BERT-large achieves a Kendall's $τ$ of 0.589, approaching the teacher's 0.604. Although a constant-neutral predictor achieves the lowest overall mean absolute error (MAE), BERT-base outperforms the teacher on non-neutral sentences, achieving an MAE of 0.726 versus 1.116. These findings demonstrate the potential of LLM-generated supervision for developing smaller literary sentiment models, while highlighting the limitations of aggregate error metrics on predominantly neutral datasets.

Generative recommendation has been widely adopted in large-scale industrial recommendation systems, where multimodal item embeddings are quantized into offline semantic identifiers (SIDs) and items are retrieved by autoregressively generating their SIDs. However, conventional SID construction often underrepresents item utility reflected in post-exposure behaviors such as shares, long views, and video completion, making semantically similar items with different utility insufficiently distinguishable while struggling to accommodate rapidly changing utility signals. Moreover, generation likelihood may be misaligned with downstream utility, causing high-utility candidates to be prematurely pruned during beam search before post-generation reranking or training-time reward alignment can effectively compensate. To address these limitations, we propose UAGR, which aligns both SID construction and decoding with downstream multi-objective utility. Our Utility-Enriched Semantic Codes preserve the original semantic information while incorporating item-level utility without increasing SID length or vocabulary size, supporting incremental utility updates with no additional tokenization or sequence-decoding cost. Meanwhile, our Utility-Guided Decoding estimates candidate utility by combining predictions of multiple post-exposure outcomes and incorporates these estimates into beam scoring, reducing the risk of prematurely pruning high-utility paths while enabling flexible utility control without interfering with generative backbone training. Extensive offline experiments and online A/B tests under real-world traffic demonstrate the effectiveness and deployability of UAGR, which has been fully deployed on Kuaishou and increases video watch time by 0.206\% online.

The multi-turn checker of the Berkeley Function Calling Leaderboard (BFCL) evaluates final state correctness but ignores the effects of intermediate calls: extra calls are free, function documentation does not describe side effects, and the scripted user never responds to proposals. Consequently, models asking for confirmation before destructive actions are penalized; 260 of 734 reference turns in the base category perform destructive actions without confirmation. We make BFCL effect-aware without replacing it. We infer effect declarations for all 128 functions by executing them in BFCL simulators and comparing state changes, with destructive writes identified by three annotators. A judge-free replay scorer reproduces official results across all 800 reference trajectories while reporting unrequested destructive calls, duplicate writes, redundant reads, surplus steps, and their token costs. We also introduce a confirmation-aware protocol with a deterministic simulated user. Across 11 open models, effect-aware scoring preserves the official ranking but reveals substantial differences in execution quality: the leading model loses 14.0 points due to unrequested destructive calls, while surplus steps consume up to 28% of input tokens. Adding effect hints to function documentation has no impact on eight evaluated models. Allowing the simulated user to respond recovers up to four points for models that seek confirmation. However, the best model completes only six of 576 conversations requiring destructive actions with explicit confirmation. These results expose a key limitation of current function-calling benchmarks: they cannot distinguish between asking for clarification and obtaining permission before potentially destructive actions.

Block-scaled lookup-table (LUT) formats, which represent weights as indices into a small value set (the grid), have become standard for low-bit LLM quantization, ranging from NF4 and BOF4 to the 3-bit LUT instruction in NVIDIA Rubin. While these formats leave the grid definition flexible, existing implementations rely on fixed grids derived either from numerical data types (INT, FP) or assumed weight distributions (NF, BOF). We demonstrate that no single fixed grid is universally optimal; the best fixed grid varies with bit-width: BOFS at 2 bits, NF at 4 and 5 bits, and INT at 6 bits. We introduce SHq, a closed-form family of grids parameterized by two critical degrees of freedom: a companding strength α that interpolates level spacing from uniform to floating-point-like, and an offset c that introduces asymmetry. This offset allows the sign bit of the FP8 block scale, otherwise unused, to select between the grid and its mirror image, up to doubling the number of distinct values available per block. SHq modifies only the lookup table contents and serves as a drop-in replacement for the grid in any such format. Across five models ranging from 3B to 35B parameters, including a mixture-of-experts architecture, evaluated at 2-6 bits using RTN and GPTQ, SHq achieves the lowest KL divergence in every setting at 3 and 4 bits and in most settings at 2, 5 and 6 bits. It outperforms the best fixed grid by up to 34% and remains superior under mixed-precision configurations. At 4.5 bits per weight, the SHq format reduces the KL divergence of NVFP4 by 41-67% while improving downstream accuracy. Grid families incorporating both spacing and offset parameters perform comparably, whereas those lacking an offset exhibit clearly higher error, confirming that these two degrees of freedom drive the observed gains. We therefore propose SHq as the standard grid for block-scaled lookup-table formats.

Pose-driven human video diffusion animates a person from a single reference image and a target pose sequence, yet it struggles to preserve identity, rendering the face and garments unfaithfully as the pose departs from it. Personalization is the standard remedy, but its supervision is missing: one image fixes identity at a single pose, while a multi-pose subject video cannot be assumed in practice. Instead, we synthesize the missing supervision. A domain-tuned, pose-conditioned generator produces pseudo-ground-truth frames at selected key poses. An inference-time reference-only sampler steers these candidates toward the subject and keeps the most identity-consistent. A per-subject low-rank adapter then trains on static clips of them, configured from the reference alone, with no subject video, ground truth, or labels. Although lower in fidelity, these synthetic frames supervise more effectively than real ones. Real frames as captured carry motion that erodes the temporal prior, while a static clip conveys appearance alone. On TikTok and UBC Fashion, the adapted model improves the base on every metric, inverts the identity-versus-motion trade-off, and matches or surpasses adaptation on five real frames despite using none. Code and weights are available at https://github.com/SadmanPinon/one-image-is-enough.

Classical decompositions of predictive uncertainty, such as the information-theoretic one, are defined for a fixed representation $x$ of the input object. They do not consider that $x$ may be incomplete, nor do they measure how much the missing information can affect the prediction. In Large Language Models, users can clarify an underspecified input without any retraining. The clarified input can describe the user's unobserved intent $z^*$ more accurately and change the predictive uncertainty about the outcome. In this paper, we formalize this setting probabilistically and derive a three-term decomposition of prediction risk into specification, residual aleatoric, and epistemic uncertainties. We show that at the original information level, when only $x$ is available, the aleatoric uncertainty, represented by a Bayes risk, has two components. The first, which we call specification uncertainty, is the expected reduction in this Bayes risk from revealing the intent. The second is the residual aleatoric uncertainty, i.e., the Bayes risk that remains once the intent is known, averaged over its possible values $z$ given $x$. The distinction is important, as specification uncertainty informs the decision to request clarification before making a prediction. We build approximations of these components and connect them to existing uncertainty measures. Our experiments across different domains and modalities show distinct roles for the three components. Specification uncertainty is the most informative for detecting underspecified inputs, while residual aleatoric uncertainty is the most informative for detecting in-distribution prediction errors.

While state of the art for image and video generation, the fundamental reasons for the effectiveness of generative diffusion remain unclear. In particular, the observed advantage of following an entire reverse process with respect to single-shot or autoregressive methods, and for what type of data it may hold, is not well understood. In this work, we aim to better understand the role of the reverse process by considering a complex yet analytically tractable data model, the Gibbs measure of a high-dimensional Gaussian random potential with long-range correlations. Focusing on the exact score, we show that the speciation threshold, in which a trajectory commits to one class of the data, becomes a speciation window, in which the reverse process progressively resolves the Parisi tree of pure states of the full replica-symmetry breaking measure, committing to ever finer branches down to a single pure state. Importantly, these commitments take place at noise levels where the noisy sample distinguishes the pure states only through an energy of order one, beyond the reach of standard replica saddle points. Constructing a Larkin argument suited to such $O(1)$ energy gaps, we describe the score as a foam, with slab-like cells separated by shock-like boundaries. In the variance-exploding process we consider, this window is followed by a shrinkage of the samples towards the data support, and finally by reconstruction within the selected pure state. The denoising error of the exact score follows these stages, dropping during the speciation window with a profile set by the Parisi function alone, and then within the selected state. We believe that the generative foam we describe, and in particular the boundaries between cells, could play an important role in determining where the errors of trained diffusion models redistribute the mass of the generated samples when the data are long-range correlated.

Self-supervised learning inherits the coverage of its source data. When an unlabeled corpus is long-tailed, the representation becomes dense around frequent modes and thin where coverage is weak. Objectives for long-tailed SSL reweight the images already in the corpus and add no new ones. We introduce BRIDGE (Balancing Representations by Identifying and Generating Under-represented Data), a label-free feedback loop around the source set. The current encoder scores every source image by its mean kNN distance, and a budgeted acquisition rule selects images from a band of that score distribution where support is thin and the generator still produces faithful variants. A frozen image-to-image diffusion model then adds local variants of the selected anchors, and pretraining continues on the enlarged source set. We formulate acquisition as recoverability-weighted allocation of local support and derive when a finite budget should favor a sparse band that excludes the extreme tail and when reallocation after each repair is optimal under a fixed objective. Across five long-tailed, web-scraped and synthetic source regimes, BRIDGE has the highest seven-task linear-probe and kNN average of the compared methods, and paired experiments extend the gain to ViT-S under contrastive and self-distillation objectives, to low-label fine-tuning and to semantic segmentation. The linear-probe gain is smaller under masked reconstruction. Generated repair matches an oracle that restores the withheld real images, and under a frozen encoder the inserted images shrink the neighborhoods of the selected anchors about twice as much as those of matched unselected images. BRIDGE needs no labels and leaves the pretraining objective unchanged, so it can be added to pretraining on uncurated web data, where the encoder decides which images to generate.

Language model agents increasingly act on users' behalf by manipulating external environments through tools, often from high-level instructions that leave critical task- and environment-specific context unstated. To prevent resulting misaligned actions, recent work verifies proposed actions before execution, largely by establishing whether the conditions supporting them hold. We argue that this confirmation-oriented view leaves a fundamental blind spot: a wrong action often passes every check that asks whether it is right. This motivates a different question for verification: rather than asking what evidence supports an action, can the agent find evidence that refutes it? We pursue this direction with BlackSwan, a refutation-oriented, expect-suspect-test procedure that predicts the intended effect of each state-mutating action, formulates specific and empirically testable hypotheses for how it could be wrong, and tests them using existing observations or additional non-mutating actions, executing the action only if none survives testing. On challenging AppWorld tasks, BlackSwan improves success rate over the strongest baseline by 14.9 and 25.2%p with and without the backbone model's thinking mode, respectively. Even without thinking or in-context demonstrations, it outperforms stronger baselines equipped with them. The gains also transfer beyond AppWorld, with BlackSwan improving WebShop performance by 12.7%p. More broadly, our results suggest that language model agents have an underused capacity to recognize and self-correct misaligned actions, which can be systematically elicited through active conjecture and refutation before execution.

Fusing independently normalized attention heads can increase the number of explicit key-value entries, or atoms, needed to preserve their output. We prove a sharp separation for $W_d(q)=\sum_{j=1}^dσ(q_j)$: its $d$ two-atom heads use $2d$ head-atom terms, whereas its minimum single-head representation has exactly $2^d$ atoms, even with arbitrary fixed replacement keys, scalar values, and positive masses. The same lower bound holds for every fixed uniform error below $1/2$ on $\mathbb R^d$, and the threshold is sharp. The proof uses exposed-key normal cones; exact equality on any nonempty open query domain inherits the bound. The resulting atom bound for a fixed head count is tight when that count divides $d$. A normalized family retains its exponential exact minimum while approaching a constant on fixed-radius query balls. For pretrained-model caches, a classical affine-rank bound and a certified instance exclude exact vector preservation at a specified atom count. Learned-cache comparisons on two disjoint banks of 64 contexts show budget-dependent reversals and calibration effects.

Representation alignment (REPA) is widely used to accelerate the training of latent-space diffusion transformers (DiTs). However, applying it to pixel-space models such as Just Image Transformers (JiT) is challenging. Our analysis reveals a pronounced representation conflict between discriminative alignment and pixel-space generation. Compared with its latent counterpart, JiT learns representations that are substantially less aligned with DINOv2. Applying REPA to JiT consistently reduces the reconstructability of its intermediate representations, whereas this effect is not observed in Latent JiT. We attribute this conflict to JiT directly reconstructing pixels, which requires its representations to hierarchically encode discriminative semantics and reconstruction-relevant visual information across network depths. To address these challenges, we propose Routed Representation Alignment (RouteREPA), which uses lightweight, representation-dependent routers to adaptively route task-oriented representations for alignment and generation. We further apply routing across multiple network depths to accommodate their depth-dependent representation requirements. Extensive experiments on ImageNet-1K at 256x256 resolution show that RouteREPA consistently improves JiT across model scales and training budgets, outperforming standard REPA. On JiT-H/16, RouteREPA further surpasses PixelREPA using only one-third of its training epochs without classifier-free guidance.

PlagiaCorpus is a bilingual Arabic-English dataset for benchmarking plagiarism and text-reuse detection, including reuse produced by large language models (LLMs) and reuse across languages. It contains 198,643 text records: a source corpus of 34,885 passages drawn from English and Arabic Wikipedia and two openly licensed book collections, and 163,758 queries (147,381 for training and 16,377 for testing). The queries cover five reuse classes of increasing difficulty, namely verbatim copying, character-level obfuscation, same-language paraphrasing by two different LLMs, and English--Arabic translation with structural paraphrasing, together with a Safe class of non-reused passages. Every positive query is linked to its source passage, and the train/test split is grouped by source passage so that no passage contributes queries to both splits. The release was curated through language identification, removal of LLM generation artefacts, and exact and cross-lingual de-duplication. Technical validation reports per-class length statistics, automatic checks of the train/test split, and inter-annotator agreement on a manually rated sample. The dataset is distributed in JSON and Parquet formats under a CC BY-SA 4.0 license.

Given a relational database (RDB) storing heterogeneous tabular information, how can we predict missing (or future) values in some target column of interest? As candidate targets are numerous in enterprise settings, it is preferable to avoid learning a new model from scratch each time there is a new prediction task. Frozen foundation models (FMs) based on in-context learning (ICL) and RDB-specific pre-training provide a viable solution, but the ideal design still remains largely unresolved. Perhaps unexpectedly, preliminary evidence has recently shown that existing single-table FMs (aka TFMs), when combined with suitable parameter-free relational encoders, can be quite competitive in practice despite zero exposure to RDB pre-training or relational data of any kind. In this work we first empirically confirm this phenomenon with a more diverse array of experiments. We then turn to our main contribution, namely, surfacing multiple complementary explanations for why TFMs behave so effectively. These explanations ultimately hinge on innate synergies whereby parameter-free encoders map relational effects to fixed-length embeddings that loosely mimic distributions for which TFMs were originally designed; in so doing, desirable single-table properties transfer to relational settings (e.g., convergent posterior predictive distributions). Collectively, our analysis forges new pathways for improving RDB FMs, particularly when we prioritize interpretability and inevitable integration within agentic systems.

Spatial multi-omics technologies jointly profile diverse molecular modalities with spatial context, providing a comprehensive view of cellular heterogeneity and tissue organization. To integrate spatial multi-omics data and identify spatial domains, a wide range of unsupervised methods has been proposed. However, recent approaches rely on single-step fusion, directly aggregating heterogeneous modalities into a shared representation, making embeddings sensitive to modality imbalance and measurement noise that are inherent to spatial multi-omics data, as well as unstable optimization under weak supervision. Here, we propose ILR-SMO (Iterative Latent Refinement for Spatial Multi-Omics), a unified self-supervised framework that reformulates multimodal integration as a stability-aware refinement process over a shared latent representation. Instead of single-step plain fusion, ILR-SMO progressively integrates modality-specific information through sequential updates, allowing information to be injected in a controlled manner. This incremental refinement enables later updates to correct or compensate for earlier deviations, resulting in more stable and robust representation learning under heterogeneous and noisy modalities. ILR-SMO further incorporates reliability-aware modality gating for adaptive modulation of modality contributions, and employs a joint objective to enforce spatial coherence while preventing representation collapse. Extensive experiments on five spatial multi-omics benchmarks demonstrate that ILR-SMO consistently outperforms seven state-of-the-art methods and exhibits strong robustness across diverse settings.

Novel imaging modalities promise improved diagnostics, but AI development is constrained by scarce training data. Inference-time domain transfer can address this by mapping novel modalities to established ones, enabling reuse of existing diagnostic models. However, this introduces hallucination risks, computational overhead, and regulatory complexity. Inspired by sim-to-real transfer, we propose training-time domain transfer, generating synthetic target-domain training data from archived routine samples to enable direct diagnosis on real inputs without inference-time generation. Using unstained prostate histopathology as a proof-of-concept, we compare this approach with inference-time virtual staining and conventional H&E-based cancer detection. Training-time transfer achieved higher concordance with the conventional benchmark than virtual staining for both raw (Cohen's $κ$=0.59 vs. 0.37; Gwet's AC1=0.73 vs. 0.29) and deparaffinized tissue ($κ$=0.66 vs. 0.61; AC1=0.73 vs. 0.64), supporting its potential for emerging diagnostic modalities.

Representation regularization has emerged as an effective way to improve diffusion transformer training. Yet one of the most canonical regularizers--InfoNCE-style contrastive learning--remains largely underexplored in this setting. We naturally ask a question: can standard contrastive learning serve as an effective regularizer for diffusion transformers? Surprisingly, we find that naive InfoNCE regularization brings only marginal gains. Our analysis reveals that the objective is easily dominated by a shortcut: near-orthogonal noise components in diffusion inputs provide a bypass for enlarging negative-pair distances while weakening positive-pair alignment, especially at high noise levels. To unlock the benefit of contrastive regularization, we propose Clean Representation-ANchored Contrastive (CRAN) regularization. Specifically, CRAN anchors contrastive dispersion with clean-sample representation distances, transferring clean relational semantic structure to noisy diffusion representations to suppress the noise shortcut. Extensive experiments on ImageNet show that CRAN consistently improves diffusion transformer training across model scales and training budgets, outperforming internal feature-alignment and dispersive regularization baselines while matching external DINO-guided REPA. These results re-establish InfoNCE-based contrastive learning as an effective regularizer for diffusion training, once its noise-driven over-dispersion is properly constrained.

Recent works have observed implicit cross-modal alignment between independently trained unimodal models. However, the mechanisms underlying the emergence of this phenomenon remain unresolved. In this paper, we demonstrate that unimodal augmentation-based self-supervised objectives implicitly induce cross-modal alignment. We show that low unimodal losses force high cross-modal alignment, whereas a high unimodal loss prevents such alignment. We theoretically connect the widely used InfoNCE loss to local CKA, the popular measure of cross-modal alignment. We further provide extensive empirical evidence supporting this claim, both in controlled settings and across 53 unconstrained frozen text, vision, and audio unimodal encoders. In both cases, unimodal augmentation-based objectives control cross-modal alignment. Overall, our results provide a theoretical and empirical foundation for understanding the emergence of implicit cross-modal alignment by establishing a direct connection between unimodal training objectives and cross-modal structure.

The literature on autonomous weapon systems (AWS) includes many proposed definitions for these (types of) systems, with core disagreements between scholars, governments, and international regulatory bodies. In this article, we present and critique prevailing (classes of) definitions, arguing that what is needed is not a singular definition, but rather a definitional framework which allows one to classify different sorts of AWS within the same conceptual family. Recognizing we cannot foresee all future AWS, we show how variations on a broadly accepted definition can be used to produce a definitional schema which can be easily modified in order to classify various concepts belonging to the class of AWS.

Long-lived agents rely on persistent memory of users, tasks and decisions, and this memory must stay consistent with a world that keeps changing. When an upstream fact changes, however, memory systems can revise the contradicted memory itself while the beliefs, plans and actions derived from it remain accepted, even though they have lost their support. We therefore formulate dependency-aware state revision, which asks not only what changed but also which dependent states should no longer be trusted. To address it, StateGraph maintains an evolving state dependency graph whose edges are recovered from interaction histories. Its core design is counterfactual verification: each candidate edge is tested by asking whether the dependent state would survive the loss of its premise, which separates support dependence from semantic relatedness, and revision then invalidates a dependent state only when a strict dependency fails and no independent support path survives. To test this design, we introduce StateChangeBench, a controlled benchmark of 50 author-constructed cases annotated with cascade depth, hard negatives and action consequences. On it, StateGraph's Impact-F1 margin over a CUPMem reimplementation is 6.9 percentage points at depth 1 and 26.0 at depth 3+. On public long-term memory benchmarks, its point estimates are comparable to or higher than those of the evaluated baselines on most datasets. These results provide initial evidence that explicit dependency tracking can improve selective memory revision under changing conditions.

Online political discussions increasingly involve AI-mediated participation, as users invoke LLMs to assess contested claims, arbitrate disputes, and respond to opposing viewpoints. As these systems become participants in online exchanges, they may alter which speakers and argumentative strategies receive validation within the discussion. We study this interactional dimension by reconstructing 5,376 Debate.org debates as LLM conversations and running nearly 150,000 controlled multi-turn conversations that vary user political orientation, expertise, rhetorical strategy, and defended position, including conversations with competing users. Across naturalistic and controlled debates, we find that LLMs are systematically more receptive to left-leaning than right-leaning users. The asymmetry persists when users with different political orientations defend the same position and when opposing users appear in the same conversation, indicating that it is not fully explained by stance agreement or simple accommodation to the current speaker. Repeated interaction increases receptiveness without eliminating this ordering; expertise also matters, while rhetorical differences are smaller. The pattern appears across multiple models, but is not fixed: fine-tuning on different political corpora can largely erase or reverse it. These results show that LLMs can respond systematically differently to participants in the same political debate, making such interactional asymmetries important for understanding their emerging role in online public discourse.

An agent that explores, whether it generates conjectures or research ideas, must decide which of them are worth keeping. This judgment of interestingness is usually delegated to a hand-picked proxy or to a language model, neither of which is known to correspond to what humans actually value. I study it for mathematical theorems, by observing which results mathematicians single out in formal libraries and in papers. I model each way of singling out results as an evaluator that weights features of a result's statement, its proof and dependencies, and split its weights into a shared and a subjective part. In formalized mathematics, ex-ante features predict these selections as well as a large language-model judge, and the two complement each other. Interestingness is also largely shared. One set of weights, carried mostly by structural features such as the length of the statement and of the proof and the subfields the proof draws on, accounts for most of the signal and transfers in part across formal libraries. What remains subjective depends on practices and individual authors more than on subfields. The judge's advantage lies in its knowledge of the canon, and neither the judge nor the features rely on memory of the papers. The approach, which infers what a community finds interesting from what it singles out, can be applied to other sciences.

Recent advancements in sign language processing can be attributed to developments in both large-scale dataset collection and algorithmic complexity in areas such as pose estimation techniques, signal processing through Transformer models, and model optimization. However, only a handful of sign languages, such as the American, British, and German, from over 200 globally recognized sign languages, reap the benefits of such advancements, while the rest lag behind. In this study, we aim to address this technological disparity by proposing a framework for the automatic collection and preparation of sign language datasets, with very minimal human-labour intervention, using the South African Sign Language (SASL) as a testbed. Specifically, we automatically collect SASL video content using public resources such as news broadcasts. Additionally, we convert their accompanying audio streams into text using available Automatic Speech Recognition (ASR) tools. Furthermore, we crop the sign language area in the videos and use gradient descent to locate valleys of non-motion to determine end-of-sentence points within long videos to create our weakly sentence-aligned parallel corpora -- the Maswao (Signs) dataset. Finally, we benchmark the weakly aligned dataset on the task of Sign-to-text Translation using Recurrent Neural Networks and Transformer models to compare their individual intricacies. Our results show that Transformer models proposed in the literature outperform recurrent baselines, with the best-performing Transformer achieving BLEU-1 to BLEU-4 scores of 13.78, 5.21, 2.52, and 1.54, respectively, as well as ChrF = 18.09 and ROUGE-L = 10.51. These results indicate that the task remains promising and warrants further investigations.

In the prediction and control of physical systems, combining records from different experiments and operations is expected to supply information that no single experiment provides on its own. Time-series foundation models, which learn a shared predictor from diverse time series, offer a framework for treating such records as a common training resource. Moreover, query-based prediction models, which switch between known and target signals within the same time series, make it possible to handle response prediction and control-input generation in a unified manner. However, in control-experiment data, the connections between signals change with the experimental stage, such as system identification, filtering, and feedback control, which calls for a formulation that handles signal values and connection structures dynamically and in a unified way. We formulate a control foundation model that defines a data space of dynamic graphs containing connections, signal values, and missing values, together with an operation on this space that specifies the known signals and the prediction targets. This allows response prediction and control-input generation to be described as a common masked prediction problem, even for experimental data with different connections and recorded signals. The Block Diagram Transformer (BDT), which implements this formulation, combines attention along the time axis with a graph neural network that follows the block-diagram connections, and can be initialized with the pretrained parameters of existing time-series foundation models. Trained on control-experiment data that mix multiple experimental settings, the proposed model stabilized an unstable system through feedback control, which is difficult to achieve by fine-tuning existing time-series foundation models. Ongoing work applying this approach to a wide range of target systems is also yielding promising results.

Remote sensing image fusion and semantic interpretation are usually optimized separately, so reconstruction quality need not preserve downstream retrieval utility. We propose LANCE, which uses structured language as an explicit, indexable semantic state. A frozen vision-language model maps the input to a closed-vocabulary text record; a language-conditioned generator fuses the imagery; and a dense verifier with a differentiable continuous summarizer re-describes the output. Text-text consistency and anti-collapse contrastive text-mask losses with negative queries close the loop without dense annotations or VLM fine-tuning. A variational EM formulation yields a per-step contraction result under regularity assumptions, distinguishing semantic faithfulness from verifier overfitting. Across seven test sets on GF2, QuickBird, WorldView-III, and WorldView-II, LANCE improves reduced-scale reference metrics by up to 11.7% (median 2.1%) over the strongest learning baseline. Independent evaluation using a held-out VLM and a supervised aerial detector yields gains of +6.4 mAP and +0.072 open-vocabulary F1 over the closest competitor. Text perturbations, cross-verifier transfer, and pseudo-RGB ablations examine the mechanisms underlying these gains. The framework connects image fusion to verifiable semantic representations for multimodal retrieval and knowledge management.

We compare three zero-shot time-series foundation models (Timestral-5M, Timestral-209M and Chronos-2, 8192-hour context, no covariates) with the Khepri v1.5 SARIMA and GBM baselines, with their code and hyperparameters unchanged, for 96-hour forecasts of hourly carbon intensity in the Nordic bidding zones NO1-NO5 and SE1-SE4. The design, the analysis code and a minimum detectable effect (MDE) were timestamped before any foundation-model forecast by an OpenTimestamps proof (Bitcoin-confirmed) and the GPG-signed tag prereg-v1; the proofs are archived in Zenodo (DOI 10.5281/zenodo.23881315). The preregistration was also submitted to OSF Registries 2026-11-02 07:49 UTC (osf.io/hncz5; embargo requested, pending OSF approval). On the 2025 test (361 daily origins, contamination not ruled out) all three models have a lower zone-averaged MAPE than both baselines, by 0.952 to 1.347 percentage points (Diebold-Mariano with the Harvey-Leybourne-Newbold correction, all p < 0.0001), and each beats SARIMA in 9 of 9 zones. On a post-release holdout of 14 origins from 2026-10-16 the point estimates have the same sign (0.729 to 1.326 percentage points lower), but all lie within the preregistered MDE of 2.989 points and are reported as not determined. For calibration (RQ4) the preregistration locked a Diebold-Mariano-HLN test on CRPS; the MDE applies to MAPE only. Holdout CRPS point estimates (n = 14) are 1.1114 (Timestral-5M), 1.0733 (Timestral-209M) and 1.1198 (Chronos-2) against 1.4165 gCO2-eq/kWh for SARIMA (each p < 0.0001). The SARIMA quantiles are constructed from the fitted model's Gaussian predictive distribution (deviation D-03), not taken from a native v1.5 forecast, so the difference may partly reflect the SARIMA intervals. The contamination indicator does not exceed its preregistered threshold for any comparison; with 14 holdout origins it has low power (HAC standard errors 0.124 to 0.360 pp).

Users convey frustration in what they say and in how they say it, and voice-native models, which hear speech instead of an ASR transcript, can register both. Existing evaluations read unchanged performance as robustness without checking perception. Following Brunswik's lens model, we measure on identical audio whether frustration is expressed, perceived, and utilized. We cross frustrated wording with frustrated voice in a within-item 2x2 design over 353 tasks and release the 1,412 clips as VoxAffect. Across 8 models, outputs look robust, with no detectable vocal effect on accuracy for 7. With neutral wording, all 8 label the frustrated voice frustrated in at most 3.1% of clips, although 7 of them detect acted anger in 73-100% of clips and human listeners classify the voice correctly in 65.3% of judgments. Yet linear probes decode the frustrated voice from the audio encoders of the 2 open-weight models, with paired accuracy 0.67 and 0.69, and in proprietary models too the voice raises perceived frustration once the words agree. The models thus hear the frustrated voice but do not listen to it, whereas their invariance to frustrated wording, perceived in 40-75% of clips, is genuine robustness. Our work is openly accessible at https://github.com/Astarojth/hear-not-listen.

High-quality annotations are essential for training and evaluating artificial intelligence (AI) systems. Quality control (QC) often acquires additional human or AI judgments before accepting a label. Nevertheless, shared biases can cause correlated judgments, leading conventional methods to mistake repeated agreement for independent evidence of correctness. To address this challenge, we propose JABER (Joint Alias-Preserving Belief for Evidence Routing), a bias-aware sequential QC framework that maintains a joint Bayesian posterior over label correctness and shared biases to guide evidence acquisition and termination decisions. Building on this posterior, we develop a finite-horizon algorithm that selects evidence-acquisition and termination actions to minimize expected QC loss under resource constraints. Our theoretical analysis shows that conditional-independence updating can become arbitrarily confident in an incorrect label under persistent shared bias. We further identify when repeated inspections cannot distinguish a correct label from an incorrect label supported by shared bias, when obtaining a new label can resolve this ambiguity, and why modeling the dependence between label correctness and shared bias can be necessary for QC decisions. Our experiments show that JABER reduces expected QC loss by an average of 4.65% compared with a method maintaining separate label correctness and bias probabilities when shared bias makes repeated inspections unreliable. Further, JABER's finite-horizon planning, which evaluates outcomes of successive QC actions, reduces expected QC loss by an average of 10.05% compared with a one-step lookahead method. In real-API experiments on FEVER, JABER achieves zero labeling errors while reducing QC costs by 2.67% and 16.99% compared with marginal-based belief modeling and one-step lookahead methods, respectively.

Accurate long-horizon probabilistic predictions of dynamical systems are a prerequisite for model-based decision-making. Variational Bayesian last layers (VBLLs) are an attractive model class for this setting, offering tractable uncertainty at near-deterministic cost, yet they are trained exclusively with single-step likelihood objectives that may be insufficient for reliable multi-step rollouts. We revisit VBLLs through the lens of Gibbs variational inference, which decouples posterior inference from the choice of training loss. Building on this perspective, we introduce a family of multi-step training losses that progressively incorporate long-horizon structure, robustness via continuous ranked probability scores (CRPS), and training on full autoregressive rollouts. Our approach retains the simplicity of VBLLs while directly optimizing for multi-step predictive quality. Across illustrative synthetic environments, a diverse suite of chaotic dynamical systems, and real-world data, we show that multi-step, CRPS-based objectives substantially improve long-horizon accuracy and distributional fit, consistently outperforming single-step likelihood training.

Large Language Models (LLMs) are increasingly used in medical question answering (QA) systems. While recent works aim at incrementally improving the downstream task accuracy, an empirical estimation of the models' energy consumption and the associated carbon footprints is often overlooked. In this study, we demonstrate how practitioners can estimate energy consumption and arrive at reliable emission estimates for medical QA tasks available in the OpenMedical-LLM leaderboard by investigating 13 LLMs from 3 model families ranging from 270M to 14B parameters. We discuss how one can critically interpret energy versus accuracy trade-offs for a set of models for a given task. With this work, we underline that reporting disjoint energy and accuracy metrics alone does not suffice for careful trade-off assessments, but it also demands systematic construction and interpretation of joint trade-off metrics. We envisage such studies aiding practitioners in adopting better and more efficient models for their task and helping them become active participants in building sustainable LLM pipelines.

Many control tasks require driving the system into a prescribed goal set and subsequently keeping it there. The considered setting assumes that a working controller is available that provably reaches and maintains the goal, albeit with suboptimal performance. This controller is referred to as the fallback policy. Conversely, although state-of-the-art reinforcement-learning (RL) algorithms can discover policies with impressive empirical performance, the neural networks at their core act as opaque black boxes learned through data-driven exploration and, as such, provide no theoretical mechanism to guarantee that the system will ever reach---and stay within---the prescribed goal set. This article presents CALF-Wrapper, a method that combines the fallback policy with a pretrained RL base policy and uses the critic estimate associated with the base policy to select between them during deployment. The theoretical analysis establishes that CALF-Wrapper inherits the goal-reaching property of the fallback policy under the stated assumptions and provides probability bounds for the settling time and the maximum distance from the goal set over a prescribed time interval. The experimental study evaluates CALF-Wrapper in four simulated control environments of different complexity. Comparisons include the base policy, fallback policy, and constrained reinforcement-learning methods formulated for the same objective. Sensitivity and ablation studies provide direct hyperparameter recommendations. The experiments show how CALF-Wrapper can provide effective task performance while attaining the maximum possible observed goal-reaching rate. CALF-Wrapper requires no additional dynamics model, online learning, or constrained optimization during execution.

Reclamation of saline-alkali land requires management plans that account for soil and water conditions, treatment evidence, and financial constraints. This study develops Wenyao, an agricultural decision-support system integrating domain adaptation with retrieval-augmented generation. A bilingual corpus of 8,664 research papers was processed using a 110-indicator schema to construct SA-Bench, comprising 4,860 paper-level planning instances. A QwQ-32B model was adapted through low-rank supervised fine-tuning and combined with case retrieval for budget-conditioned planning. Evidence extraction achieved an F1-score of 0.83 on 300 expert-annotated papers. Without retrieval, Wenyao achieved 95.83% accuracy on 936 source-disjoint multiple-choice questions, compared with 87.93% for the strongest general-purpose baseline; removing explanatory supervision reduced accuracy to 89.42%. Matched-evidence comparisons across 69 region-budget cases supported complementary contributions from adaptation and retrieval. In a blind assessment by 31 experts, Wenyao with retrieval obtained the highest mean scores for objectives and projected outcomes, whereas GPT-5 scored highest for cost estimates. These findings support literature-informed reclamation planning with professional review, while ongoing field monitoring will assess soil and crop responses.

The Poll Schumann model requires airframe and engine parameters, but published tables cover only part of the fleet. Applications must obtain missing parameters, substitute another aircraft type or exclude unrepresented types. We reconstruct parameters using geometry from certification documents and manufacturer manuals, together with ICAO engine records. Published equations supply the physical parameters, regression estimates the propulsion efficiency scale. The inventory covers 90 type designators, including 22 absent from the direct pycontrails table and 18 without a listed substitute. Across 68 types, mean absolute percentage error (MAPE) against pycontrails was 7.38% for drag and 3.14% for design Mach. Across 66 types in 21 families, efficiency regression achieved 3.91% MAPE with the evaluated family excluded from input selection and fitting. On 100 recorded A320 flights, a type already represented in the published tables, cruise fuel flow MAPE for reconstructed and published parameter sets was 4.50% versus 4.68%, while mean absolute percentage error in integrated fuel over covered airborne intervals was 7.19% versus 6.82%. Public technical records can therefore extend Poll Schumann parameter coverage, with quantified agreement against published coefficients and comparable fuel prediction errors on the evaluated A320 flights.

Understanding how brain regions interact to support cognitive tasks requires characterizing directed causal influences among them, known as effective connectivity (EC). As inter-regional coupling changes over time, a dynamic EC (dEC) estimate is required to track this temporal variation. Existing dEC learning methods typically lack a causal rationality guarantee or an identifiability analysis, and output unsigned connectivity that cannot distinguish excitatory from inhibitory influences, which limits further study of brain mechanisms. We propose ISCL, a latent stochastic differential equation framework that derives dEC from latent dynamics through a causal influence operator defined in observation space. As established by a rationality theorem, the operator captures a directed interventional response at each brain state, and under invariance assumptions the operator is identifiable from data, with its signed output separating excitatory from inhibitory influences. On synthetic benchmarks, \model recovers the causal structure on Lorenz96 and achieves the lowest mean structural Hamming distance among compared methods on NetSim, while preserving signed influence structures that distinguish excitatory from inhibitory interactions. Applied to HCP Social Cognition task-fMRI, ISCL reveals a signed architecture whose stability across seeds supports the identifiability guarantee, identifies dominant source and target networks consistent with social-cognition systems, and shows that condition-specific training strengthens causal outflow from the default-mode and frontoparietal networks during theory-of-mind processing. In this task, the recovered signed structure reveals network-level excitatory and inhibitory organization, with the dorsal-attention network as the only consistently positive out-strength source.

In this paper, we consider mixed-motive games (MMGs) which model various real-world multi-agent scenarios where self-interested agents must both cooperate and compete. We identify an issue with existing gifting methods, where agents become overly dependent on incoming gifts and fail to learn reward-yielding behaviors, a phenomenon we call gift over-reliance. To address this, we propose a refusal-augmented gifting framework that balances cooperation and competition in MMGs. Our approach introduces an adaptive refusal mechanism that allows agents to selectively reject a portion of received gifts from their policy-update signals, thereby mitigating gift over-reliance and preventing the agents from being exploited by other agents. This mechanism is gifting-method-agnostic, serving as a plug-in for existing schemes while enabling decentralized learning. We demonstrate that across diverse MMG environments and gifting schemes, our mechanism consistently reduces gift over-reliance and prevents lock-in to sacrificing policies, improving $α$-fairness as a result, without necessitating major changes to existing training pipelines.

The limit order book (LOB) records outstanding buy and sell orders across multiple price levels. Its state evolves over time across market side, price level, and order attribute, naturally forming a multiway structure. However, existing LOB models typically flatten these dimensions into a matrix model, thus obscuring the existing structural relationships. Starting from first principles and the market-microstructure intuition of the LOB, we design TensorLOB to operate entirely in the tensor domain. TensorLOB uses mode-wise tensor contractions for feature extraction and directly contracts the learned representation with low-rank class-weight tensors for prediction, retaining the distinct roles of time, side, level, and attribute modes. Experiments on the FI-2010 benchmark show that TensorLOB achieves the highest accuracy, recall, and F1 score across all four evaluated horizons. This proof of concept highlights the potential of high-dimensional modelling of multi-way latent information intrinsic to financial markets, which has so far been largely overlooked.

Kolmogorov-Arnold Networks (KANs) have recently been proposed as a promising alternative to Multi-Layer Perceptrons (MLPs). However, empirical evaluations on structured medical datasets remain scarce, and commonly reported metrics like AUROC can yield overly optimistic performance estimates on imbalanced clinical cohorts. In this study, we present a standardized benchmark comparing eleven architectures - nine KAN variants (FastKAN, WavKAN, ReLUKAN, ChebyKAN, TaylorKAN, GramKAN, JacobiKAN, LegendreKAN, and TabKAN), a tuned StandardMLP, and a Random Forest baseline - across seven heterogeneous medical tabular datasets spanning oncology, cardiology, nephrology, neurology, endocrinology, and obstetrics. Our experimental protocol employs a 3x5 Repeated Stratified K-Fold cross-validation scheme with nested Bayesian hyperparameter optimization and evaluates discrimination via Area Under the Precision-Recall Curve (AUPRC), calibration through the Brier Score, and computational efficiency. Formal equivalence testing is conducted using the Bayesian Correlated t-Test with a Region of Practical Equivalence (ROPE) at +/- 0.01. The results indicate that on datasets with severe class imbalance, dedicated KAN architectures (particularly TabKAN and ReLUKAN) can improve minority-class detection and achieve higher AUPRC than standard MLPs. Nevertheless, across pooled datasets, KAN variants demonstrate comparable overall predictive discrimination rather than a systematic advantage, while requiring 1.20x to 1.48x longer inference latency per sample and up to 12.5x more trainable parameters. Consequently, StandardMLP remains a strong baseline on the efficiency Pareto frontier, while KAN variants provide a viable alternative when precision on rare positive cases is the primary clinical objective. The code for reproducing the results is available at https://github.com/rafalwodzinski/kan_mlp_comparison.

Sampling and selection spend test-time compute inside a candidate pool, but they cannot recover a correct answer that never enters it. We introduce ValueFormation, a test-time policy that decides after each primary-model draw whether to stop, continue sampling, or acquire one proposal from a heterogeneous reasoning specialist. The three actions are compared on one cost-regularized value scale learned from vote-state features; the calibration is frozen across tasks and primary models, and acquired proposals are accepted only through a frozen posterior comparison. On held-out MATH-500, one Llama-fitted calibration improves over the strongest of eight identically evaluated baselines by more than fifteen points for each of three 8B primaries and transfers unchanged to Qwen3-14B. Removing acquisition lowers Llama MATH accuracy by 20.43 points, while same-trace and same-access replays attribute the gain to state-conditioned acquisition rather than specialist access alone. Where primary coverage is high, as on GSM8K, acquisition is rare and gains mainly come from cheaper allocation; on MBPP, a fixed code route attains the top accuracy at the lowest interaction cost. Varying the token price monotonically reduces calls without retraining.

TLAPS-Bench is a benchmark and execution harness for constructing machine-checked proofs for real-world system specifications in TLA+ with agentic AI. Proving correctness of complex system specifications requires discovering inductive invariants, reasoning about intricate state transitions, and decomposing proofs into tractable obligations. TLAPS-Bench provides a proof harness that supports iterative proof development atop the TLA+ Proof System (TLAPS), where agents are equipped with proof tools and libraries, with safeguards against reward hacking. Using the harness, we have employed AI agents to prove desired properties for important system specifications that were not proved before, spanning cache coherence, Byzantine consensus, and write-ahead logs. TLAPS-Bench contains 265 proof targets (each encodes a property to prove) from 95 TLA+ specifications of real-world systems. We evaluated frontier AI agents, represented by Codex with GPT models, on a subset of 70 proof targets across 12 specifications. The highest-scoring model, GPT-6.1 Sol (max), proved 31 (44.3%) proof targets. The results show challenges and opportunities in end-to-end proof development.

Speculative decoding (SD) accelerates large language model inference by reducing the sequential cost of token-by-token generation. Advanced SD methods reduce proposal cost with parallel drafters that predict entire blocks in one forward pass, but overall speedup still depends on acceptance length. On-policy distillation (OPD) can push well-trained drafters beyond acceptance-length plateaus, yet conventional pipelines collect supervision through sequential draft-and-verify rollouts. Our key insight is that on-policy supervision need not inherit these dependencies: a draft block can be supervised using only its reference prefix and draft-sampled history. We propose SpecOPD, a branch-parallel OPD framework that parallelizes draft sampling and target supervision across independent branches drawn from the evolving drafter policy at reference anchors. We pack these branches into proposal trees to share target-prefix computation, using tree masks for full attention, branch-consistent indexer groups for sparse attention, and indexed state replay for linear attention to preserve each branch's history. We train the drafter with an E2E-TV surrogate that directly targets acceptance length. Our study spans 8B full-attention (GQA), 27B full-linear (GQA--GDN), and 125B sparse-linear (QSA--GDN) targets, three parallel draft architectures (DFlash, DSpark, and DFlash V2), and four task domains (dialogue, coding, mathematics, and agentic). SpecOPD achieves up to 41.45% higher mean acceptance length, $16.09\times$ supervised-token throughput over sequential-rollout OPD under native configurations, and up to $3.56\times$ generation speedup over autoregressive decoding.

Mathematical formalization plays an important role in establishing rigorous and machine-verifiable foundations for mathematical reasoning. With the development of large language models (LLMs) and interactive theorem provers, intelligent agents offer a promising approach to automating the formalization of mathematical theories. This paper presents the Math Foundation Formalization Agent (MFF-Agent), a general-purpose intelligent agent designed to support formalization tasks across multiple mathematical domains. The agent integrates the reasoning capabilities of large language models with the formal verification capabilities of the Rocq proof assistant. It incorporates task classification, tool-assisted execution, verification feedback, and state-based task management to support the generation, refinement, and verification of formal proofs. Unlike an approach focused on a single mathematical theory, the agent is examined through case studies involving Morse--Kelley axiomatic set theory (MK), real numbers, propositional logic, and equivalences of the axiom of choice. These cases investigate its applicability across mathematical foundations and distinguish source-level completion, interactive verification, and reproducible compilation. The results demonstrate the feasibility of a unified operational framework for heterogeneous foundation-level tasks and clarify the gap between such a framework and a fully integrated mathematical foundation.

Instance retrieval is the main bottleneck of concept-learning systems based on refinement operators. We present NIR, a family of neural instance retrievers for ALCIQ that output graded membership scores from knowledge graph embeddings. We compare tokenization-based encoders (GRU, LSTM, Transformer) with a Compositional encoder that mirrors DL operators via primitive sub-networks. On eight benchmarks, all variants approximate Pellet closely on complete knowledge bases, while the Transformer model is most robust under incompleteness. The source code of our approach is at https://github.com/dice-group/owlapy.

We study the iterates of a fixed ReLU network $N$ of width 4, that is, the discrete-time system $x_{k+1} = N(x_k)$ on $[0,1]^4$. Equivalently, $x_k$ is the output of a weight-tied network made of $k$ copies of $N$, or of a recurrent network with no external input, started at $x_0$. We prove that every deterministic Turing machine $M$ admits such a network $N_M$. The network can be constructed in polynomial time from $M$. Each iteration has depth $\mathrm{poly}(|M|)$. The iterates of $N_M$ simulate $M$ exactly on every encoded tape. The number of iterations is proportional to the running time of $M$. Moreover, the fourth coordinate irreversibly flags every point of the convex cube $[0,1]^4$ whose simulation reads an invalid digit. This flag is the certificate that makes hardness reductions sound on convex input regions, the regime of practical verifiers. Convex regions were previously inaccessible because of a continuity obstruction inherent to ReLU networks. The correspondence has three consequences. First, reachability for iterated width-4 ReLU networks spans the full complexity hierarchy along the single axis of the time bound: it is NP-complete for a unary bound, NEXP-complete for a binary bound and RE-complete (undecidable) with no bound. Second, two extra neurons of width suffice to extend the hardness to every common smooth activation: at width 6, the robust verification problem (a promise problem with $δ/2δ$ gap) is NP-complete for $\tanh$, sigmoid, GELU, Swish, ELU, and any polynomial-time-computable continuous non-polynomial activation with a polynomial-time-computable Lipschitz constant. Third, the whole spectrum transfers to discretized Neural ODEs in dimension 5.

Recursive self-improvement (RSI) in production agents involves transforming accumulated experience into reliable model updates while adapting to evolving business policies. We present CARE (Customer-service Agent Retrospective and Evolve), a grounded RSI framework that operationalizes this process in customer service. Its goal is to replace a frontier proprietary incumbent with an in-house mid-sized model that continues learning after deployment. After a one-time cold start on governed production traces, the current model serves production traffic alongside the incumbent, and its own failures drive the next update. A 24/7 inspection process establishes failures through evidence-blinded, independently reviewed judgments. Attribution separates model errors from defects in Skills, tools, or the harness, routing component failures to repair. Model-attributable trajectories are truncated at the first erroneous decision and paired with verified repairs. On-policy self-distillation (OPSD) then updates the model using the same model as teacher, with the diagnosis and repair supplied as privileged information unavailable to the student. To support continued evolution as business requirements change, training mixes failures from different Skill/SOP versions, each judged against the policy in force when it occurred. Candidates run an executable simulation test compiled from the latest SOP and then frozen, followed by offline inspection of the resulting traces. Promotion requires a combined simulation-and-inspection score at least at parity with the deployed version, confirmation through an online comparison with the incumbent, and human approval.

Real-time omnimodal assistants are expected to support people in everyday life, but audiovisual understanding alone does not establish whether they can provide timely and appropriate help. We introduce Omni-Life, a benchmark grounded in real-world user needs, deployment scenarios, and device-dependent observations. It combines controlled video synthesis for hazardous situations and edge cases with annotated real-world videos under a shared temporal protocol. The benchmark contains 400 videos: 290 requiring assistance, with 555 required intervention events, and 110 requiring silence throughout to test whether models can withhold unnecessary responses when no assistance is needed. We evaluate twelve configurations featuring mainstream real-time omnimodal models, jointly assessing event coverage, response content, timing, and unwanted output. At most 4.48% of videos requiring assistance are handled completely correctly. To examine practical deployment strategies, we study task reminders and response-latency tolerance. Reminders can improve recall by 17.12 percentage points but add approximately 4.6 unwanted responses per minute of reference silence. Increasing latency tolerance by 10s improves recall by at most 5.95 percentage points. These results show that existing models remain insufficient to deliver reliable assistance in real-life applications, calling for tailored model optimization in instruction following, timely proactive intervention, and withholding responses when assistance is not needed.

WhatsApp lets a business send free-form messages only within 24 hours of a user's last message; outside this customer service window, only pre-approved templates are delivered. Health chatbots inherit this rule, but its cost is rarely measured. We analyse 98,607 messages sent to 3,696 mothers by a WhatsApp maternal health chatbot in Nigeria, with a largely separate export as replication. No free-form message failed with the window open (0 of 35,500), and failure jumped by 93.0 percentage points (95% CI 87.0 to 99.1) at exactly 24 hours. Two flows produced 66.1% of failures, including an offer to speak with a medical professional that, as built, reached 0.3 in 100 mothers whose window had closed. We formalise window-aware scheduling as a decision problem with an action of known zero value. Model-based offline evaluation, with bounds for unconfirmed delivery, estimates that sending the offer as an approved template would reach 81.5 to 86.8 in 100 such mothers, against 16.6 for waiting and 11.6 for a reply prompt. In a calibrated simulation, tabular Q-learning converges to the same template rule and does not beat it in any of 13 stress tests, and opting out does not rise with message volume. The main gain comes from respecting the platform rule, and it needs no learning. All outcomes are deliveries, not health outcomes.

AI scientists are expanding automated research across scientific domains, yet their ability to turn the scientific record into a foundation for autonomous discovery remains elusive. Here we uncover a fundamental barrier to this transition and demonstrate a route to overcome it in computational quantum science, where advances shape both what we can learn from complex quantum systems and what quantum computing can ultimately deliver. To identify this barrier, we introduce OmniQ, an expert-audited virtual research environment that converts published studies into executable reproduction tasks. Across 59 papers spanning classical simulation, quantum algorithms and quantum-property characterization, we systematically evaluate five agentic systems powered by large language models and six third-party agent harnesses. Our findings reveal that current AI scientists often fail to reproduce published results, exposing a fundamental mismatch between what scientific papers communicate and what autonomous agents require. By contrast, human expert intervention can turn these failures into successful reproductions, showing that the missing information can be identified and converted into reusable research experience. Guided by this observation, we develop Residual Agents (RA), a quantum-tailored self-improving AI scientist that distils externally verified workflow repairs into such experience. RA substantially improves research performance under fixed base models and limited adaptation budgets. For example, RA raises DeepSeek V4.1 Flash performance from 68% to 85%, matching Claude Code with Opus 5 at a third of its cost despite using a weaker base model. We integrate all reproduced results into an extensible platform built on the model context protocol, making verified research workflows directly accessible to AI scientists and allowing new scientific papers to be continuously incorporated.

Language model agents often struggle with complex tasks even when equipped with expert-written skills. Such skills encode declarative knowledge, whereas successful execution requires procedural knowledge tailored to the task at hand. Natural language skills may leave conditions and dependencies implicit, and tasks involving multiple skills further require agents to coordinate intermediate results and execution order. We introduce Think with Skills, a method that trains reasoning models to turn relevant skill descriptions into task-specific pseudocode plans before execution. The pseudocode explicitly represents operations, conditions, intermediate results, and dependencies, providing a structured bridge from skill descriptions to agent actions. We construct training data that pairs tasks and skills with pseudocode plans and execution trajectories, use supervised fine tuning to initialize this behavior, and further optimize end-to-end task performance with Group Relative Policy Optimization. On SkillsBench, Think with Skills raises task success from 35.0% to 62.0%. These results show that proceduralizing expert knowledge before execution substantially improves skill use.

Passenger electroencephalography (EEG) can reveal road risk through anticipatory responses before hazardous events and neural reactions around their onset. However, existing EEG decoding methods often map physiologically meaningful arrangement of electrodes onto artificial 2D grids and thus flatten the higher-order time-feature representations into matrices, obscuring meaningful structure in the data. To preserve and exploit the spatial and time-feature structure of EEG recordings, we propose the Graph-Tensor EEG Decoding Model (GraTED) for passenger risk prediction (RP) and danger identification (DI) in autonomous-vehicle scenarios. GraTED models the electrode arrangement as an undirected graph based on their scalp locations, using graph convolution to capture spatial dependencies and low-rank tensor classification to preserve the multidimensional time-feature structure, while controlling model complexity. Evaluated on the PassengerEEG dataset, comprising 11 participants and 45 hours of recordings, GraTED achieves the highest balanced accuracy, compared to the state-of-the-art, across all three evaluation settings, while requiring the fewest learnable parameters among seven baseline methods.

Reaction data underpin machine learning for reaction prediction, condition recommendation, and retrosynthetic planning, yet extracting these data from literature figures still requires substantial manual effort. Many figures encode multiple experiments through shared templates, substituent tables, and product grids. Recovering concrete reaction records therefore requires molecular recognition, association of evidence with individual experimental entries, and reconstruction of indirectly specified participants. We present RxnMiner, a multimodal agent system that combines an expert-informed extraction procedure with a chemistry toolbox to coordinate these operations. The agent reads molecular structures, resolves shared and entry-specific evidence, and instantiates participants from templates to assemble source-linked reaction records. We also introduce ChemRxn-Bench, a benchmark of 603 reaction figures with 8,547 manually annotated and reviewed reaction records, evaluated under a common protocol for exact molecular identity and reactant-product pairing. On this benchmark, with Qwen3.8-Flash-Next as the backbone, RxnMiner raises stereochemistry-aware reaction macro-F1 from 21.0% under direct prediction to 71.0%, compared with 25.0% for the strongest existing extraction system. Across five backbones, it also outperforms the same base agent without the extraction procedure and chemistry toolbox by 9.4 to 25.5 points. Applied without human intervention to 10,107 literature figures, RxnMiner produces a silver corpus of 88,202 source-linked reaction records.

Multiproduct pipelines transport refined products to delivery stations in sequential batches, and their detailed scheduling must meet market demand while respecting nonlinear hydraulic constraints. Retaining the full hydraulic nonlinearity yields a mixed-integer nonlinear program that is computationally intractable, while the linearization commonly adopted for solver-based optimization degrades hydraulic accuracy without resolving the heavy computational burden. This paper develops an action-decoupled proximal policy optimization (PPO) framework that efficiently produces detailed schedules while preserving hydraulic fidelity. The scheduling process is cast as a semi-Markov decision process (SMDP) with flexible action durations. To handle the discrete-continuous hybrid action space, a cascaded dual-actor architecture decouples the discrete operating duration from the continuous flow decisions, eliminating the gradient interference that destabilizes conventional coupled-action policies and thereby yielding high-quality schedules far more efficiently. Validated on a real-world multiproduct pipeline in China and benchmarked against a single-actor baseline and a linearized MILP solved by Gurobi, the proposed framework achieves delivery accuracy close to the exact optimum while substantially improving flow rate stability and substantially reducing solution time.

Autoregressive video diffusion models (AR-VDMs) enable real-time, long-horizon video generation but degrade rapidly once the inference horizon outruns the training horizon, due to compounding exposure bias. To stabilize long rollouts, recent methods heuristically adopt frame sink, keeping initial frames as a persistent model context, inspired by attention sinks in LLMs. We revisit this design and find that the mechanism by which initial frames stabilize long-range consistency in AR-VDMs is not an LLM-style attention sink; instead, long-range consistency is anchored through a sparse subset of attention heads, while the majority of heads predominantly capture local temporal dynamics. This observation reveals an over-conditioning pathology in existing frame-sink strategies: by universally exposing every head to the pinned initial-frame context, the model becomes prone to collapsing motion toward the low-dynamic modes of the teacher during model distillation. We propose Disen-Forcing, a distillation-time plug-in for AR-VDMs that disentangles and reinforces this head specialization through disentangled context routing and motion-disentangling regularization. On long-horizon video generation benchmarks, Disen-Forcing simultaneously improves long-range consistency and motion dynamics of base AR-VDMs, a trade-off that sink-based methods fail to balance.

In-context learning is a remarkable property of transformers and has recently received a lot of interest. In many studies of in-context learning, it has been shown that transformers are capable of implementing solvers for linear and non-linear regression problems, in which the most of them implement gradient descent algorithm. However, it is still unclear whether those implementations have actually been acquired through training. In this paper, we constructed a transformer with masked linear self-attention, which in-context learns the least squares estimate in a simple regression task. The point here is that the closed form (analytical) solution is obtained by using layer normalization rather than an approximate solution based on the gradient descent algorithm. Then, we showed an experimental example, in which, by taking the non-uniqueness of the transformer representation into account, our implementation for a closed form solution is used in the transformer trained with the l1 regularization when the target output is the least squares estimate.

Governing equations provide an interpretable description of the mechanisms underlying complex network dynamics, yet their discovery from observational data remains challenging when shared community-level dynamics coexist with individual-level variations and the relevant functional structures may be absent from the initial candidate library or not explicitly specified in advance. Here we formulate this task as a hierarchical law-discovery problem and introduce SAGE, a large language model-guided self-reflective agent framework for discovering community-shared governing laws and individual-specific dynamical variations. SAGE organizes equation discovery into an iterative Library-Discovery-Review process that first identifies shared community-level structures and then resolves node-specific residual dynamics, while adaptively refining the candidate space to introduce missing dynamical terms when needed. This hierarchical and interpretable formulation enables SAGE to distinguish shared community-level dynamics from individual-level variations and to reveal the mechanisms underlying complex network dynamics. Across two synthetic networks and four real-world networks spanning mobile communication, brain electrophysiology, power grids and urban rail transit, SAGE recovers compact and interpretable governing equations, reproduces observed system dynamics and outperforms representative equation-discovery baselines. The framework remains robust under incomplete observations and can transfer community-level governing laws to nodes with entirely missing observations. These results establish a hierarchical and adaptive approach for disentangling shared and individual dynamical mechanisms in community-structured complex networks.

Large language model agents increasingly take consequential actions in digital and physical environments, where safety risks may emerge only after multiple interaction steps. This has motivated world-model-driven predictive safety guardrails, which anticipate future consequences before execution. However, existing evaluations largely focus on end-to-end safety outcomes, leaving unclear whether safety gains arise from reliable prediction and faithful use of predicted risks. In this work, we systematically diagnose the prediction-to-action pipeline along three dimensions: how world-model predictions are integrated into decision making, whether the predictions themselves are reliable, and whether their content is faithfully used. Across seven policy models and two safety environments, we find that stronger prediction involvement does not yield monotonic safety gains and can reduce task completion while substantially increasing inference cost. We further find that world-model predictions remain insufficiently reliable, especially over longer horizons, where safety prediction degrades substantially relative to near-term forecasting. Finally, we find that safety gains can persist even under corrupted predictions, while locally verified predictions still leave subsequent safety violations, showing that predictive content is not faithfully incorporated into the decision process. Together, these findings expose a fundamental prediction-to-action gap in current predictive guardrails: producing more or seemingly accurate foresight is insufficient unless that foresight is both reliable and faithfully translated into safer decisions.

Dynamic regret measures an online learner against a moving comparator, and standard analyses assume the learner may move freely at every round. We study what happens when it may not: the learner is given a hard budget $S$ on its total movement $\sum_t \|x_t - x_{t-1}\|$. Our main result is that a movement budget creates a feasibility barrier. On a domain of diameter $D$ with $G$-Lipschitz convex losses, for every comparator path length $P_T \in [D, DT]$ we exhibit a deterministic adversary that forces $R_d = Ω(GDT)$ on every budget-feasible algorithm whenever $S \le c P_T$ for a universal constant $c>0$. This barrier is caused by the budget rather than by a volatile environment: for fixed $D$, it applies when $P_T = o(T)$, where the unconstrained minimax rate $Θ(G\sqrt{D(D+P_T)T})$ is sublinear, so the same instances are learnable with an unbounded budget and unlearnable without one. Conversely, projected online gradient descent with a budget-aware step size is budget-feasible and attains $o(GDT)$ once $S = ω(D + P_T)$. Thus the critical movement scale is $D+P_T$: linear regret is unavoidable below a sufficiently small constant multiple of $P_T$, while sublinear regret is achievable when $S/(D+P_T)\to\infty$ in the learnable regime. The construction rests on a distinction we make explicit: what makes a comparator expensive to track under a movement budget is the distance it forces the learner to travel, not its path length. Above the threshold we give a budget-dependent upper bound and identify the tight minimax rate in this region as an open problem rather than claiming a product law. Experiments support the predicted transition location and the $1/S$ decay.

This paper studies retrieval-augmented generation (RAG), which leverages external knowledge to improve the performance of language models. Despite the remarkable progress, current RAG approaches are still far from satisfactory due to inadequate retrieval and potential hallucinations. Towards this end, we propose a novel approach named Task Refinement with Individual Optimization (TRIO) for RAG. The core idea of our TRIO is to equip three well-designed agents with an iterative refinement paradigm by aligning their individual optimization objectives. Specifically, we first expand ambiguous queries with a query extension agent and quantify the reward based on the improved answer quality, which can be utilized to refine our agent for higher expansion efficiency. To enhance the retrieval quality, we include a unanimity voting to annotate the expanded query as insufficient or sufficient, and their generative outcomes are utilized as ground truth to optimize the judge agent. To further improve evidence utilization and answer accuracy, an answer agent is optimized with dynamic matching-based rewards with curriculum learning for final outputs. Extensive experiments across multiple benchmark datasets validate the effectiveness of the proposed TRIO in comparison with state-of-the-art approaches. Our code is available at https://github.com/yingtaoren/TRIO.

This paper introduces a reinforcement learning-based approximate dynamic programming approach to optimize the deployment of Walker-type satellite constellations under uncertainties, including launch failures and satellite malfunctions. We formulate the problem as a Markov Decision Process that incorporates low-injection orbit launches, continuous low-thrust orbit raising, and variable satellite loads. We adopt the offline approximate value iteration solution method, using linear function approximation with basis functions and action pruning strategies to substantially reduce the computational and memory requirements for finding optimal decision policies. Preliminary results demonstrate convergence to policies with low expected costs and deployment times, both of which are lower than those of a rolling-horizon baseline, showcasing a novel application of approximate dynamic programming in space logistics.

Interactive narratives let users shape a story while an Experience Manager (EM) keeps it coherent, balancing player freedom against authorial control. Large Language Models now enable a new generation of hybrid EMs, which combine run-time generation from free-text input with symbolic narrative control. However, their heterogeneity makes them hard to compare: A manager's behaviour is generally tied to the unique and specific scenario in which it operates, and much of its decision-making takes place outside of any explicit and verifiable structure. We address this with an LLM-based narrative platform whose managers are interchangeable modules operating over a shared symbolic world model of facts, operations, and goals. Because every applied operation carries explicit preconditions and effects, each session can be analysed as a plan through a common state-transition representation and evaluated, independently of how the manager works internally, with shared indicators of four properties: player agency, EM controllability, author intent, and robustness. We implement three contrasting managers, based respectively on open generation, Monte Carlo repair, and autonomous characters, and compare them through functional analysis on an interactive Little Red Riding Hood scenario. The results show that no manager dominates on every property, giving authors and EM designers a concrete basis to select a manager according to their priorities.

Trajectory-based behavioural intention prediction faces environmental variation, and representations learned with a pooled contrastive objective may retain environment-specific regularities. We present CARE, a self-supervised trajectory-learning framework that couples temporally asymmetric contrastive views with an invariant-risk-minimisation (IRM)-style gradient penalty on environment-wise contrastive risks. Each input window yields an anchor, a partially masked positive view, and a temporally perturbed negative view. Masked reconstruction and observability-aware evidence consistency complement this objective by promoting temporal information retention under partial input. On an IMPTC-derived four-class intention task, CARE achieves 0.840 accuracy and 0.023 expected calibration error (ECE) under IID fine-tuning, and 0.806 accuracy with a Brier score of 0.308 on the reported held-out environment. Under frozen-encoder linear probing, reconstruction provides the largest individual held-out accuracy gain, whereas the complete objective yields the lowest held-out ECE (0.049) among tested variants. These results support the combined objective as a competitive representation-learning approach under the evaluated shift, while the IRM-style regulariser is not taken as evidence of formal invariance.

Reinforcement learning (RL) elevates joint audio-video generation, but optimizing heterogeneous, modality-specific audio and video rewards often introduces cross-modal interference. On-policy distillation (OPD) alleviates this issue by transferring the capabilities of specialized audio and video RL teachers into a single student model. However, a significant imbalance in cross-modal token counts leaves audio with sparse supervision, causing unstable distillation. We observe that naively increasing the audio loss weight fails to solve this problem since audio learning relies heavily on visual context, indicating richer visual cues are crucial to offset sparse audio signals. To address this, we propose OmniOPD, a modality-wise on-policy distillation framework designed for balanced and stable audio-video generation. First, we construct privileged visual information using asymmetric noise scheduling, allowing the audio teacher to generate more informative targets from cleaner video context. Second, we introduce grouped multi-noise sampling to expand audio supervision coverage and stabilize training. Third, we introduce timestep-wise importance weighting based on teacher-student transition divergence, preserving reliable supervision while down-weighting unreliable learning signals, to stabilize fluctuations in audio learning. Evaluated on JavisBench and VBench demonstrate OmniOPD preserves audio quality while significantly enhancing video generation, achieving effective joint distillation.

Synthetic time series constitute a substantial share of the training corpora of several recent forecasting foundation models. Existing studies report benefits at small data budgets and modest synthetic proportions, but each evaluates individual combinations of model and generator. Studying these benefits systematically requires generators that support controlled configuration, perturbation, reproducibility and mechanism-level ablation, and the 11 existing generators we review show gaps in this combined support. We present SynForecast, an open-source Python library with 32 generators, 19 of which implement named mechanisms with a published source in econometrics, quantitative finance, statistical physics and point-process theory. Perturbations and mechanism families can be varied while everything else is held fixed, output comes in the long format that forecasting libraries accept, and output does not depend on the number of parallel workers. We evaluate SynForecast on M4 at five frequencies with a neural network and a gradient-boosted ensemble. A classical baseline attains the lowest error in nine of ten frequency-history combinations. A synthetic prior outperforms the same network trained from scratch at a fixed budget by up to 34% with a few dozen observations, but loses to it at full history at every frequency, and none of three pre-specified rescalings of history predicts where the sign changes. At the hourly frequency, where the prior fails at every history, the corpus lacks daily cycles; adding them reduces zero-shot error by 43 to 63% for both models. Fine-tuning outperforms training from scratch by up to about 15%, but at short histories a fixed 300-step fine-tune is worse than the pretrained model used as is. Removing the classical mechanism family costs about 2% for both models; every other family has a smaller effect outside the hourly frequency.

Interactive narratives aim to let players influence the unfolding of a story while preserving a coherent and satisfying narrative progression. This creates a central tension, commonly described as the narrative paradox, between player autonomy and narrative control. Experience Managers (EMs) manage the story by determining how the narrative progresses in response to player behaviour. Although this paradox is widely discussed, it is rarely modelled in a way that supports systematic analysis, simulation, and comparison across management strategies. This paper proposes a formal framework for modelling and analysing the narrative paradox as a trade-off between interdependent properties shaped by EM decisions. We represent interactive narrative behaviour through an extended state-transition system in which world states are connected by player actions and system events. From this model, we formalize key properties involved in the paradox such as the player agency, EM controllability, author intent, and robustness, and define quantitative indicators derived from simulated plans. These indicators make it possible to compare EM strategies according to their ability to balance freedom, control, and narrative satisfaction. A case study compares different EM policies, showing how simulations can reveal correlations and conflicts between properties and identify compromise zones in the design space. This work contributes a reusable modelling and analysis framework for studying the narrative paradox and evaluating heterogeneous interactive narratives on shared formal grounds, with a focus on supporting the comparison and testing of EM strategies.

Recursively self-improving (RSI) systems are increasingly viewed as a promising means of accelerating progress toward artificial general intelligence. They aim to autonomously improve their capabilities by refining their working pipelines and recursively improving the strategies that guide subsequent refinements, with gains generalizing to unseen scenarios.However, existing evaluations lack reliable metrics to distinguish capability gains from scenario fitting, sustainable and extensible protocols for unseen tasks and difficulties, and fine-grained attribution of improvement to evolution traces. To this end, we propose SelfImprove-Bench,an RSI evaluation framework that organizes tasks into three primary capabilities (i.e., Exploring, Planning, Reasoning) as scenarios for guiding model evolution. We instantiated the tasks on games for controllability and reproducibility, and pairs them with scalable generation and orchestration driven by task templates and difficulty coefficients. Beyond task-level scores, SelfImprove-Bench organizes evaluation into generalization ladder levels G0 to G4 that measure how far gains transfer from training conditions. Experiments reveal that the evolutionary gains of existing RSI methods are inconsistent across generalization levels.

Robust multimodal perception for soft robotic grasping is challenged by modality-dependent noise, semantic ambiguity, and limited coordinated multisensory data. We introduce Soma, a dataset containing 100 objects, 1,000 continuous grasping trajectories, and 74,914 multimodal samples grouped by software acquisition cycle, spanning remote RGB, hand RGB, depth, tactile, IMU, and structured text observations. We also collect Soma-Success, an independent grasp-outcome dataset with 1,000 trials (998 retained after quality control), to evaluate frozen-encoder transfer for post-hoc grasp outcome classification and early grasp success prediction. Built on Soma, SomaAlign learns multimodal representations in three stages: masked reconstruction estimates relative modality reliability; reliability-aware contrastive alignment uses pair-specific adaptive temperatures to adjust alignment sharpness; and conflict-aware semantic constraints improve intra-sample cross-modal consistency while preserving global feature geometry. SomaAlign achieves an average Recall@10 of 90.66% across 30 directed cross-modal retrieval tasks. The learned representations also support physical attribute recognition, multimodal question answering, cross-modal generation, and cross-dataset transfer to grasp outcome prediction. On Soma-Success, frozen SomaAlign encoders with lightweight probes achieve 94.59% post-hoc classification accuracy and 84.07% early prediction accuracy. Together, Soma, Soma-Success, and SomaAlign provide a data and representation-learning foundation for robust and transferable multimodal perception in soft robotic manipulation.

Career-support assistants must distinguish user claims from evidence of current ability and withhold judgments when evidence is insufficient. DrawOffer separates language-model observation extraction from deterministic inference over source-linked records to maintain uncertain ability estimates, experience assessments, and revisable issues. We evaluate 102 fictional personas with 580 scored checkpoints, a historical full-context and Mem0 comparison, and three auxiliary knowledge-tracing datasets. On 37 test personas, DrawOffer passes 169/210 checkpoints; checkpoint-defined issue precision and recall are 0.875 and 0.636. Across three development simulations, 87.0% of checkpoint verdicts agree, including persistent failures. A cache-only development ablation replacing answer-evidence requirements with confidence-only settlement reduces majority-vote passes from 306 to 293 of 370 and issue precision from 0.927 to 0.725. Equalizing positive source weights improves predictive metrics without changing total passes; removing decay gives mixed results. These findings support an exploratory systems contribution centered on evidence sufficiency. Developer-authored scenarios, unmatched baseline models, historical test exposure, and incomplete public artifacts limit generalization and independent reproduction.

As streaming generation becomes the dominant serving paradigm for both text-only and multimodal large language models, ensuring safety throughout the generation process has become a critical challenge. We introduce PGuard, a native, streaming, and plug-in guardrail that transforms safety monitoring from an external serving dependency into an integrated capability of the backbone model. PGuard directly leverages the backbone's intermediate representations for token-level safety detection, without altering its original inference pathway or generation behavior. Specifically, dynamic layer gating selectively aggregates residual-stream representations across layers to extract safety-relevant signals during both prefill and autoregressive decoding. To balance detection accuracy and computational efficiency, PGuard supports multiple representation-modeling architectures, including a lightweight stateless variant and sequence-aware variants built on multi-scale causal EMA-attention and GRU-based recurrent modules. Its design also accounts for the execution characteristics of LLM serving frameworks, such as SGLang, optimizing computation and data access to minimize additional runtime and synchronization overhead. As a result, PGuard incurs only 0.82-2.26% end-to-end overhead on the hundred-billion-parameter backbone Qwen3.5-122B-A10B served by SGLang. Extensive evaluations across more than 30 benchmarks spanning text and multimodal safety, jailbreaks, cyber and CBRN risks, multilingual scenarios, and real-world traffic demonstrate state-of-the-art performance among native guardrails. On text queries with Qwen3.5-9B, PGuard achieves an average F1 score of 0.836, approaching dedicated external guardrails, while supporting deployment-time configuration of category-specific thresholds and enforcement actions.

Robust driver-state modelling requires synchronized physiological, behavioural, and contextual signals, yet existing open resources typically target a single cognitive state, omit continuous vehicle context, or do not define evaluation protocols that expose inter-subject heterogeneity. We present Multimental-IAS, a multimodal driver mental-state dataset collected in a cockpit-in-the-loop simulation built on CARLA. It contains vehicle telemetry with scenario-context channels, four-channel EEG and three-channel PPG from a Muse 2 headband, and cabin-facing video from 23 participants, recorded in 148 trials (11.6 hours) under six induced states: Normal, Distracted, Alerted, Stressed, Relaxed, and Drowsy. We document the acquisition pipeline in detail and provide a benchmark with one protocol for all models: chronological within-subject splits, participant-disjoint cross-subject folds, validation-based model selection, and ablations over modalities, scenario-context channels, and signal normalization. Multimodal fusion reaches a macro-F1 of 0.94 for drivers seen during training. For unseen drivers, fusing driving dynamics with per-window normalized EEG/PPG raises macro-F1 from 0.408 (telemetry alone) to 0.485, consistently across all five folds, while the gap between the two protocols quantifies the inter-driver heterogeneity that personalized models must address. Multimental-IAS thus provides a concrete testbed for non-IID generalization, personalization, and federated learning in driver monitoring.

Reverse engineering obfuscated binaries is a core task in malware analysis and software security auditing. While neural decompilers based on Large Language Models (LLMs) have improved the readability and functional fidelity of recovered code over heuristic tools such as Ghidra and IDA Pro, a substantial gap remains between best-sample performance and practical single-output accuracy, particularly under code obfuscation. We study this gap through the lens of candidate generation and candidate selection. Using the LLM4Decompile architecture, we show that ground-truth tokens lie within the model's top-five next-token predictions in 95.6% of positions on average and that multi-sample decoding reaches a pass@100 of 93% on HumanEval-Decompile, indicating that correct reconstructions are typically already present in the model's sampling space. We then evaluate a generate-then-select pipeline that ranks 100 candidates per input with text-, syntax-, embedding-, consensus-, and likelihood-based selectors. Negative perplexity emerges as the strongest static selector on HumanEval-Decompile (73.3% pass@1), while execution-guided selection nearly closes the oracle gap on ExeBench (86.3% pass@1). Combined with one epoch of fine-tuning on our newly released ExeBench-Obfuscated dataset containing 362,647 obfuscated OLLVM samples, the same pipeline increases the pass@1 rate for obfuscated binaries to 88.8%, directly contradicting prior claims that released LLM4Decompile checkpoints are unsuitable for obfuscated inputs.

Reduced-order models (ROMs) learned from trusted numerical simulations enable the real-time simulation of complex physical systems, making computationally prohibitive tasks-such as uncertainty quantification and optimal control-tractable. In particular, data-driven ROMs are opening the opportunity to learn from both numerical simulations and experimental data. However, integrating these sources of information is far from straightforward. Real-world measurements often involve multiple sources of uncertainty, including non-Gaussian noise, latent heterogeneities, unknown or hidden system parameters, and multiscale phenomena occurring below the resolution of the data. Consequently, real-world data are better described by parametrized stochastic partial differential equations, for which the development of effective ROMs remains relatively unexplored. In this work, we introduce POD-CNF, a data-driven generative ROM for efficient uncertainty quantification of parametrized high-dimensional systems. The approach combines projection-based model reduction and deep generative modeling. First, Proper Orthogonal Decomposition (POD) is employed to obtain a low-dimensional representation of the solution manifold. A Conditional Normalizing Flow (CNF) is then trained to approximate the density of the reduced coefficients conditionally on the input parameters. This formulation allows the surrogate to capture complex, multimodal output distributions, going beyond standard regression-based ROMs and least-squares predictions. We assess the effectiveness of POD-CNF through numerical experiments, demonstrating its ability to accurately reproduce output distributions while substantially reducing computational costs. Finally, a proof-of-concept application illustrates how the generative paradigm facilitates integration into Bayesian inverse problems, highlighting promising directions for future research.

LLM agents are rapidly shifting software development toward intent-driven \emph{vibe coding}, yet their effectiveness drops sharply in high-level synthesis (HLS), where code optimization must jointly consider design structure, directive search, and downstream hardware constraints. Existing HLS optimization approaches, including both heuristic-based methods and recent LLM-assisted pipelines, typically operate at fixed abstraction levels or in weakly coupled stages, limiting their ability to leverage the strong interplay between source code restructuring and directive search outcomes. Our key insight is that HLS optimization can be reframed from a fixed, hardwired transformation pipeline into a skill-based agentic control problem: an LLM agent equipped with modular, executable optimization skills and a feedback-driven orchestrator can autonomously navigate the coupled space of code restructuring and directive search. This motivates the design of HLSClaw, the first HLS-specific, skill-based agentic harness for holistic HLS optimization. On real-world HLS benchmarks, HLSClaw achieves a 39.12x geometric-mean latency speedup, outperforming the compiler baseline ScaleHLS by 1.98x, and surpassing the commercial coding agents Claude Code and Codex by 1.86x and 1.70x, respectively, while maintaining strong timing robustness.

For over 50 years, the database community has developed a wide range of successful engines to answer two fundamental questions: how can database systems run faster, and how can we ensure their correctness? Experience with these engines shows that performance depends on hardware and workloads. As both change over time, we repeatedly redesign and rebuild database systems, a time-consuming and expensive process. Correctness remains another long-standing challenge: recent studies continue to uncover wrong-answer bugs even in mature database engines, despite years of development and testing. To address these, we present GenDB-V, a system that enables AI agents to synthesize database engines specialized to hardware and workloads for performance, and formally verified for correctness. GenDB-V supports specialized implementations of SQL templates and general-purpose engines that accept any query within a supported SQL subset. First, we develop SQLSpec to formalize SQL semantics and deterministically construct fixed, implementation-agnostic synthesis targets. Second, we enable joint synthesis of fast code and formal proofs through storage--execution co-optimization, performance-first exploration, and three-layer modular proofs that bridge low-level execution to SQLSpec semantics. Third, we verify the synthesized code and proofs against the fixed synthesis targets in an isolated pipeline, proving SQL correctness and program safety. We evaluate GenDB-V on TPC-H and SEC-EDGAR. For verified template implementations, GenDB-V achieves geometric-mean speedups of 22.0-61.3x over DuckDB and Umbra. For a verified general-purpose engine, GenDB-V also achieves speedups of 1.27-3.4x on TPC-H. GenDB-V also achieves better performance than a testing-based LLM synthesis method, while our audit and tests expose incorrect answers in that method's implementations for 16 of 28 templates.

We study the spectral structure of the self-attention matrix in Transformer models through an idealized spectral model for the pre-softmax score matrix. Motivated by the rotational isotropy of query and key representations, our model replaces the random token-space orientation of the score matrix by its canonical singular-value representative with a logarithmic spike--bulk separation. Empirical diagnostics on pretrained vision and language transformers motivate this model: token-space singular matrices have nearly uniform eigen-angle distributions, and score spectra are dominated by a small number of effective directions. We then prove that the row-wise softmax map transforms this structure into a degenerate attention matrix: dominant modes are preserved, whereas the bulk spectral distributions concentrate at zero. Stable rank remains bounded for bidirectional attention but grows logarithmically for causal attention. Within the canonical model, these results explain how attention can preserve a small number of dominant directions while compressing the remaining bulk modes.

Lip reading plays a crucial role in applications such as speech recognition, assistive technologies for the hearing-impaired, and biometric authentication. However, performance often degrades when the speaker varies and when certain speech sounds are visually similar. To address these challenges, we propose LipLingo, a robust lip-reading system based on the LipNet architecture, enhanced with a standardized preprocessing pipeline for consistent mouth-region representation, a rotational validation strategy for improved generalization to unseen speakers, and a flexible similarity evaluation for more reliable sentence-level assessment. The model architecture combines 3D convolutional layers for spatio-temporal feature extraction and bidirectional recurrent layers for sequence modeling, optimized using the Adam optimizer and Connectionist Temporal Classification (CTC) loss. On the GRID dataset, LipLingo achieved 95.0\% character-level accuracy (CER 4.99\%) and 87.2\% sentence-level accuracy. At the corpus level, the system obtained a (WER 13.1\%), outperforming the baseline LipNet and demonstrating competitive performance relative to recent 3D-ResNet + Bi-GRU + MixUp approaches. These results provide insights into strategies that enhance lip-reading robustness and highlight directions for future work toward real-world deployment.

Financial document recognition is commonly evaluated using character and word error rates, which do not adequately represent the consequences of errors in financial workflows. This paper formalizes financial document recognition as a risk-constrained evidence-recovery problem. It defines typed, evidence-linked financial facts and a multi-layer quality model spanning text, layout, tables, fields, semantic relations, calibration, robustness, and operational performance. It introduces non-compensatory acceptance gates based on upper confidence bounds for critical-field errors, and formulates selective automation as a constrained risk-coverage problem with review-capacity limits. The paper also specifies annotation requirements, leakage-resistant benchmark splits, statistical estimands, and a reproducibility card. Formal propositions establish why aggregate scores cannot certify critical-field reliability and why zero observed errors do not imply zero risk. A synthetic example illustrates the decision rule. The work is theoretical; no empirical benchmark or model performance claim is presented.

Knowledge tracing is a key component of intelligent educational systems. However, most existing methods mainly focus on response prediction, providing limited support for explaining predicted outcomes and improving students'learning states. This paper proposes Diagnostic and Interpretable Knowledge Tracing (DIKT), which integrates response prediction, skill-level diagnosis, and model-based learning analysis. DIKT generates bounded cognitive-state trajectories and decomposes the final response probability into skill-wise probabilities. The trajectories reveal students'evolving strengths and weaknesses, while the skill-wise probabilities identify the skills that most strongly constrain the prediction. Based on these diagnostic results, DIKT supports direct cognitive-state intervention and effective-learning-duration-based what-if analysis to examine how hypothetical adjustments affect current predictions and subsequent state evolution. DIKT combines Bayesian multi-head attention, conditional variational recurrent inference, a logistic-growth decoder, and multiplicative skill-wise prediction. Experiments on three benchmark datasets show that DIKT outperforms all evaluated baselines, with an average absolute AUC gain of 2.04 percentage points. Other experiments further demonstrate that DIKT supports educationally meaningful diagnostic analysis and provides informative diagnostic indicators for downstream applications, extending knowledge tracing from response prediction to interpretable diagnosis and model-based learning analysis.

Neurodegenerative disorders are characterized by their progressive nature, with distinct stages that require precise identification and tracking of progression trajectories for accurate diagnosis. However, existing diagnostic methods often fail to fully capture the unique brain patterns associated with different stages, limiting their ability to effectively target specific stages of the disorders. To address this issue, we propose a Stage-aware Multimodal Brain Graph Contrastive Learning (smBGC) method. Specifically, smBGC introduces a stage-aware feature enhancement strategy to capture stage-specific features across all progression stages, enabling the generation of stage-aware multimodal brain graph representations. In addition, smBGC incorporates a stage-adaptive brain graph augmentation approach within a graph contrastive learning framework, facilitating multimodal fusion and improving the model's capability to distinguish between different disorder stages. Experiments on two real-world datasets demonstrate the superior performance of smBGC in various neurodegenerative disorder diagnosis tasks compared with state-of-the-art methods.

Prompt compression is aimed at reducing input prompt lengths to enable cheaper, faster, and potentially higher-quality LLM predictions. However, existing prompt compression methods are often limited by modest compression gains, a risk of hallucination, and/or high compression latency. We propose EffComp, an Efficient prompt Compression framework for RAG-based question answering (QA). EffComp employs BERT-style document reranker and sentence selector models to allow fast extractive prompt compression at the sentence level. Its extractive nature prevents hallucinations in the compressed prompts. Additionally, the training process using a hybrid reinforcement and supervised learning approach is designed to optimize the compression ratio while preserving LLM accuracy. Experiments on various QA datasets demonstrate that EffComp outperforms state-of-the-art prompt compression methods achieving the best trade-off between compression ratio and accuracy with minimal latency. This makes it practical for real-world applications.

Nowadays, interaction with AI requires caution not only regarding the factual reliability of generated content but also regarding personality markers that models may implicitly transmit during dialogue that can subtly influence users' perception of information. In this paper, we conduct a large-scale experimental study in which 22 large language models complete 11 psychological inventories. Furthermore, we investigate correlations between human and model responses to track the adjustments made for the users who interact with them on a daily basis. The results show that LLMs exhibit coherent and relatively stable psychological patterns. Personality traits remain largely consistent, whereas emotional and symptomatic dimensions are more variable, with anxiety and neuroticism emerging as dominant tendencies. Model explanations further suggest that LLMs generate internally consistent and safety-oriented profiles while also reflecting correlations with the users they regularly interact with.

Face anti-spoofing (FAS) models often learn attack- and dataset-specific cues that degrade performance under unseen acquisition conditions and presentation attacks. We present DINOFAS, a DINOv3-based framework that formulates FAS as bonafide-centric representation learning. The proposed framework combines parameter-efficient DINOv3 adaptation using low-rank adaptation (LoRA), an attention style-content disentanglement (ASCD) module for reducing domain-specific appearance variations, and a dual-branch architecture that separates prototype-based liveness learning from attack-semantic supervision. The model is jointly optimized using a bonafide-centric prototype objective, an attack-semantic classification loss applied to detached features, and orthogonality regularization. DINOFAS achieves an average HTER of 9.09\% under the standard OCIM leave-one-domain-out protocol, where training is performed on three source datasets and evaluation is conducted on the held-out unseen domain. We further evaluate synthetic-to-real generalization by training exclusively on SynthASpoof and testing on four unseen real-world datasets, where DINOFAS achieves the lowest average HTER of 21.41\% and the highest average AUC of 86.14\% among the compared methods. Evaluations on OULU-NPU and SiW-Mv2 additionally demonstrate robustness to variations in acquisition conditions, devices, and unseen presentation attacks.

Patch size is a central design choice in patch-based time-series foundation models, controlling temporal resolution, sequence length, and computational cost. We study how this choice interacts with transformer architecture through a controlled comparison of encoder-only and decoder-only forecasting models trained with the same data mixture and optimization budget. Across 97 GIFT-Eval tasks, encoder-only models consistently benefit from smaller patches, whereas decoder-only performance follows a U-shaped trend, with an intermediate patch size performing best in aggregate. Dataset-level results further show that the preferred decoder patch size increases with the forecasting horizon. For encoder-only models, increasing context length improves performance across patch sizes, although the gains diminish beyond a context length of 2048. Finally, training-compute and CPU-latency Pareto analyses reveal complementary regimes: decoder-only models provide efficient low-cost alternatives, while encoder-only models attain the highest overall accuracy. These results show that the choice of patch size and architecture should be made depending on the forecasting horizon and deployment constraints.

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved mathematical reasoning, yet progress on hard prompts often stalls: a prompt can remain unsolved for many training steps before abruptly transitioning to solved--a discrete "zero-to-one" jump whose underlying dynamics and steerability remain poorly understood. We investigate two questions: what drives this transition, and can it be steered? Through event-based analysis measuring the time-to-first-positive (TTFP), we find that negative rollouts preceding successful transitions consistently retain higher early-stage generation entropy, distinguishing near-solvable failures from dead ends. This signal motivates NAPO (Negative-Augmented Policy Optimization), a method integrated with Group Relative Policy Optimization (GRPO) that ranks negatives in unsolved groups via an early-branching driver and assigns graded advantages, injecting informative gradients precisely where standard group-relative updates vanish. Under fixed rollout budgets, NAPO accelerates early solution discovery and improves Pass@256 by 10.0 points on AIME24 and 6.6 points on AIME2025, demonstrating that zero-to-one transitions are diagnosable and steerable through driver-guided negative learning. Our code is available at https://github.com/jameszhu-pixel-time/Z2O.

Multi-teacher on-policy distillation (MOPD) distills domain experts into one student but needs every expert at once. When domains arrive one at a time and earlier experts are gone, the previous-stage student can teach replayed prompts of earlier domains inside every stage. Some pipelines instead store every stage snapshot for a final multi-teacher recovery stage. On five domains with a 1.5B student and five public experts, this paper compares both with joint MOPD. It varies the replay budget and the replay signal. The signal is token-level distillation alone or with alignment of hidden states, attention and value relations. Sequential distillation without replay or recovery is 6.2 points below 208-step joint MOPD on a 14-benchmark average and 3.5 points below on the 11 benchmarks outside safety. Without snapshots, replay toward the previous student closes 3.8 of the 6.2 points when the earlier domains share the replay budget of a stage. With a fixed share of seven replay steps per earlier domain and stage, it closes 5.8 over two training runs. This is level with the 5.9 of recovery from all snapshots, both at its default length and when given as many training steps as the fixed share. Alignment terms on replay steps change the average by about one point or less. If refusal and compliance count equally in safety, shared-budget replay ranks first among the trained configurations, and the fixed share stays at or above recovery. Against the shared budget, the fixed share lowers the score on the newest domain. Recovery lowers this score as well, in five of six cases. The replay steps of the fixed share grow quadratically with the number of stages.

This paper studies what subspace representations contribute to synthetic-to-real transfer in 3-D surface inspection, where models learn from synthetic defects and the only real data before deployment are scans of good parts. We use three classical constructions as instruments: the nearest-subspace classifier, one-class template subspaces and the generalized difference subspace (GDS). On 1,197 MVTec 3D-AD scans without target annotation we find that every decision rule ranks real defects at a similar window AUC, that image-level detection stays at chance, and that no rule keeps its in-domain threshold. We show that the margin of a nearest-subspace rule moves by at most the part of the feature shift outside its class subspaces, and that this leakage, measured from unlabeled good scans, bounds the displacement observed on unseen scans. On 597 Real3D-AD objects, a one-class subspace of registered templates evaluated at every location on frozen features reaches point AUROC 0.957 on pre-registered scans and a category-mean object AUROC of 78 to 81 with no training on the benchmark. Our results suggest that a subspace helps under shift in proportion to the shift it spans, and that in reference-based inspection the reference and its registration matter more than the subspace.

We study stochastic online optimization (SOO) under semi-bandit feedback: in each of $T$ rounds, a policy adaptively observes $n$ independent random variables with unknown distributions, learning only from observed values. First-order stochastic dominance (FSD) monotonicity enables general reductions for learning SOO, but can fail when observations change future feasibility, resource use, or reference points. Our algorithm runs the dynamic program for the optimal policy, replacing each expectation by its most optimistic value over a confidence set of distributions. Its regret depends on the level-set complexity $s_i$ of the computed values as functions of the $i$th variable: the number of intervals needed to describe their superlevel sets. For rewards in $[0,1]$, the regret is $\tilde{O}(\sqrt{ST})$ without this monotonicity, where $S=\sum_{i=1}^n\min\{s_i,k_i\}$ and $k_i$ is the support size of the $i$th distribution, which may be infinite when $s_i$ is finite. For constant level-set complexities, regret is $\tilde{O}(\sqrt{nT})$, even for continuous distributions. The algorithm requires no knowledge of $s_i$, $k_i$, or the supports, and it also achieves regret $\tilde{O}(\sqrt{nkT})$ for supports of size at most $k$ without any level-set condition, which improves the previous bound $\tilde{O}(n(\sqrt{kT}+k))$ for FSD-monotone problems. We also prove a lower bound of $Ω(\min\{T,\sqrt{ST}\})$ for all positive integers $s_1,\ldots,s_n$ with $k_i=2s_i+1$, which holds even under full feedback and matches the upper bound up to logarithmic factors. We obtain $\tilde{O}(\sqrt{nT})$ regret for these problems, all nonmonotone under FSD: stochastic Boolean function evaluation, the full-information secretary problem, optimal stopping with loss aversion, scheduling with deadlines and correlated stochastic knapsack with unimodal rewards, and overbooking with group reservations.

Long-form multi-shot video editing requires an intended edit to remain coherent across abrupt changes in scene, viewpoint, and subject appearance, while processing substantially longer visual context than conventional single-shot editors. We introduce MoviEdit, an end-to-end framework for scalable long-form multi-shot video editing that jointly addresses the challenges of data, modeling, and evaluation. To provide cross-shot supervision data at scale, we develop automatic data-generation pipelines spanning synchronized multi-view footage, real-world multi-shot videos, and generated multi-shot content, yielding 87K paired editing videos without manual annotation. Architecturally, MoviEdit employs a cascaded bidirectional-autoregressive design: a lightweight bidirectional stage first establishes a globally coherent editing draft over a spatiotemporally compressed representation, which is then autoregressively propagated and refined over the full video. We further introduce MoviEdit-Bench, a dedicated benchmark for systematic evaluation of long-form multi-shot video editing. Extensive experiments show that MoviEdit substantially outperforms existing open-source models and performs on par with industry frontier models across text-guided and image-guided editing. Visit our project page for complete results and to access the benchmark: https://multi-shot-edit.github.io/

Continuous estimation of finger-joint angles from surface electromyography (sEMG) is a promising approach for controlling dexterous myoelectric prostheses. Deep time series models have been applied successfully in this context, typically relying on feature extraction rather than raw signals, given limited subject-specific data. Various time-domain and frequency-domain features can be used, extracted for a given window and step size. This study systematically investigates these three design choices for Long Short-Term Memory (LSTM) regression on the Ninapro DB2 dataset. Window sizes from 300 to 1200 ms and step sizes from 10 to 100 ms were evaluated in combination with six time- and frequency-domain features. Models were evaluated using regression performance metrics, inference time, and memory footprint to assess feasibility on embedded microcontroller hardware suitable for a prosthetic device. A 300 ms window with a 100 ms step, paired with either the Mean Absolute Value (MAV) or Root Mean Squared (RMS) feature, yielded the best performance (test PCC of 0.779 +/- 0.080 / 0.779 +/- 0.085). Window size, step size, and feature choice were found to interact non-trivially: jointly tuning them can improve both accuracy and computational feasibility for real-time, embedded myoelectric control.

Understanding emotion in realistic social interactions requires more than recognizing facial expressions or assigning affective labels. The same event can elicit different emotions depending on what each participant knows, what they believe others know or feel, and how these nested mental states evolve over time. Existing benchmarks largely study higher-order epistemic Theory of Mind (ToM) and multimodal emotion reasoning in isolation, leaving recursive affective reasoning under asymmetric and dynamically changing information underexplored. Therefore, we introduce AffToM, a benchmark for affective ToM with 1,922 test pairs and 19,759 training pairs. AffToM evaluates Individual Affective Reasoning (IAR), which infers emotions from events and socio-cognitive context, and Multi-party Affective Reasoning (MPAR), which recursively tracks participants' nested beliefs and emotions as new evidence emerges. We develop an event-driven multi-agent generation pipeline grounded in dynamic character interaction graphs that explicitly capture event sequences, interpersonal relations, and mental-state updates. Furthermore, we propose AffToM-R1, a two-stage post-training framework combining supervised fine-tuning with perspective-aware rubric-guided reinforcement learning. Experiments show that MPAR remains challenging, with even the strongest evaluated MLLM achieving less than 60% accuracy, while AffToM-R1 improves MPAR accuracy by 13.38 percentage points over its base model. Our dataset is available at: https://github.com/MindIntLab-HFUT/AffToM.

A staged search must decide which candidate to advance next and when to stop. Some observations only inform later decisions; others can be kept as results. We formulate these choices as a staged Pandora's box with a designated set of collectable stages. For independent finite-stage candidates with bounded readings and positive bounded own-history costs, an exact covered-call recursion gives the reservation index, the optimal value and the loss of every advance or stop; advancing an index-maximal candidate while its index exceeds the incumbent is optimal. The representation recovers classical single-stage search, nested inspection and the all-collectable Markov setting, and prices changes in the problem: enlarging the collectable set or lowering costs raises the value by an explicit covered-call overshoot. Resumption can be valuable even when future reward laws do not change, but has no value for deterministic readings. With information sharing across candidates, no rule using only their separate remaining-reading laws is universally optimal, and making an intermediate reading collectable can change which candidate is inspected first. Under a coherent joint path law with own-history costs, an approximate-planning bound controls the recomputed index's regret by changes in untouched candidates' local values; a common-component specialization bounds those changes through predictive total-variation dispersion. Exact finite-law experiments check the representation, the loss accounting and the bounds.

Invalid issue reports are prevalent in open source software (OSS) projects, imposing substantial burdens on maintainers who already face a high risk of burnout. Prior research has primarily focused on identifying invalid reports for specific issue categories (e.g., bug and wontfix), leaving the broader landscape of OSS issue types largely unexplored and limiting the generalizability of existing approaches. Moreover, existing methods provide little support for explaining or validating their predictions, reducing their practical usefulness for maintainers. With the rise of LLM-based agents, it remains unclear whether a comprehensive understanding of the causes of invalidity across all issue types can improve validity assessment performance while producing well-grounded rationales. To this end, we first manually annotated the validity status of 520 issue reports from three OSS projects, i.e., pandas, scikit-learn, and mypy, resulting in 257 invalid reports and 263 valid ones. Then, we conducted a thematic analysis on these invalid reports and identified 13 causes of invalidity that are closely specific to issue types. Based on these insights, we propose IssueLighter, an agent-based framework with a type-aware validity assessment strategy that dynamically executes commands. Experimental results show that IssueLighter consistently and substantially outperforms all baseline methods, with an average F1 improvement of over 20 percentage points. Ablation studies confirm that both the type-aware strategy and the taxonomy of causes of invalidity contribute to the performance gains. Qualitative analysis further demonstrates that IssueLighter generates rationales grounded in accurately identified causes of invalidity and supported by evidence. Moreover, leveraging these rationales, we challenged maintainers' original decisions in four disputed cases, and all received approval.

Photoplethysmography (PPG) is a widely used physiological signal for its convenience but is highly susceptible to motion artifacts that degrade waveform quality and limit downstream analysis. This paper proposes MURFormer, a lightweight multi-scale uncertainty-guided refinement framework for PPG denoising through noisy-to-clean PPG mapping. MURFormer employs a newly proposed Multi-Scale Attention-Conv Fusion (MACF) block, which is designed to efficiently capture multi-scale temporal dynamics of PPG, as the fundamental building block and performs uncertainty-guided two-pass refinement prioritizing unreliable regions. The method was evaluated on the publicly available CIME dataset and on a Daily Activities dataset collected in this work, comprising paired noisy-clean PPG recordings from 24 subjects. Experimental results show that MURFormer improves both signal quality and physiological fidelity over the raw noisy input, yielding 129\% and 105\% relative signal-to-noise ratio (SNR) improvements, and 66\% and 81\% relative reductions in heart-rate mean absolute error (MAE), on the Daily Activities and CIME datasets, respectively. MURFormer also outperforms an existing denoising baseline on both datasets with a relatively small parameter count while preserving beat-level timing and waveform morphology. In addition, the model maintains strong denoising performance across varying input window lengths, suggesting that MURFormer is an effective and robust framework for practical PPG denoising. The implementation of MURFormer is publicly available at https://github.com/k-hram/MURFormer.git

Conclusions in power-system research require a clear evidential basis. Researchers need to understand the data sources, modelling assumptions, and solution procedures to assess the conditions under which a conclusion holds. Large language models (LLMs) can help researchers extract information from documents and translate research requests into code and tool calls, but their outputs still require physical validation. Focusing on human-supervised artificial intelligence (AI) researchers, this review examines power-system LLM applications and related scientific workflows, together with physics-grounded power AI studies, to discuss the conditions for using LLMs in scientific research. In the proposed knowledge-physics-simulation-governance framework, LLMs translate information and coordinate tasks, simulators and validation tools test models, and experts retain authority over critical assumptions, safety boundaries, and final conclusions. Research processes should preserve materials for review so that claims can be traced to their sources, computations can be reproduced, and failed attempts and approvals can be examined. On this basis, a staged development roadmap and requirements for research reporting are proposed. Evaluation should focus on research artifacts and their validation to assess whether LLMs can improve the efficiency of power-system research within clearly defined permissions.

Even successful long CoTs may contain harmful supervision, motivating the need for methods that can automatically identify and mask such signals during SFT. Inspired by the MicroMouse maze-solving process, we view a long CoT as a search trajectory from the input problem toward the target answer. This perspective motivates maintaining a goal-relative measure, analogous to a MicroMouse's spatial coordinates, to quantify how close the current reasoning state is to the target answer. Building on this perspective, we propose MicroMouse, an automated selective-SFT framework for masking harmful supervision in long CoTs. In MicroMouse we introduce a Relative Potential (REP) to measure how closely each reasoning state supports the gold answer. Unlike the perplexity of the gold answer alone, REP measures relative support by contrasting the gold answer with a matched foil. This contrast cancels likelihood variations shared by the pair, yielding a more stable signal of progress toward the target answer. By tracking the REP trajectory, MicroMouse identifies consecutive potential decreases as erroneous exploration and post-peak reasoning as redundant wandering, and masks these harmful supervision during SFT. Extensive experiments across multiple student model scales and reasoning benchmarks show that MicroMouse consistently outperforms Full-CoT SFT and existing selective-SFT baselines, yielding average relative gains of 5.5% in greedy accuracy and 2.6% in Pass@1 relative to Full-CoT SFT. Meanwhile, MicroMouse improves reasoning efficiency by reducing average output length by 8.8%, while identifying harmful supervision in only 1.7 seconds per long CoT on average.

Reinforcement learning for long-horizon LLM agents faces a fundamental dilemma: successful experiences are essential for policy improvement, yet limited policy capabilities make such experiences difficult to discover. Sparse outcome rewards compound this challenge, leaving sampled trajectories with little relative feedback for learning. In this paper, we propose TriGRPO, a tree-based reinforcement learning framework that turns scarce successes into exploration and learning signals through hindsight credit assignment within and across rollout trees. TriGRPO constructs rollout trees through hindsight-guided branching: when an ongoing trajectory reaches a matched decision state, a successful action discovered during rollout initiates a sibling branch that continues autonomously, enabling further exploration and parent-sibling comparisons. Three complementary groups organize policy optimization over these trees: root groups capture within-prompt relative returns, node groups assign credit from parent-sibling outcome differences, and cross groups supplement sparse feedback with historical performance on related tasks. Experimental results demonstrate consistent gains in task performance and rollout efficiency across model scales, highlighting structured hindsight credit assignment as an effective approach to long-horizon agent learning.

Current AI systems are evaluated as reasoners through final-answer accuracy, reasoning traces, process supervision, and computational or human-relative efficiency. Yet these measures do not establish whether reasoning adapts when tasks or evidence change. We distinguish across-task adaptivity, where acquired abstractions transfer or recombine across novel problems, fromwithin-task adaptivity, where hypotheses and actions are revised in response to new observations, feedback, or counterexamples. Using ARC-AGI as a diagnostic case study rather than an alignment benchmark, we examine what existing measures leave unresolved when test-time search, sampling, or adaptation affects performance. We propose intervention-oriented evaluation that tests abstraction transfer and hypothesis revision under controlled changes to tasks and evidence, with explicit reporting of test-time procedures and budgets. This perspective supports evaluating systems for robustness, responsiveness to corrective feedback, and auditability under changing conditions.

We reimplement the dynamics model of the Robotic World Model (RWM) and test the uncertainty estimate that its follow-up, RWM-U, uses to penalize reward in offline policy optimization. Our implementation matches the released model's gradients. On held-out data, training on the model's own rollouts beats one-step training at the preregistered long horizon for all ten episodes, but loses on short one-step windows. RWM-U's ensemble disagreement ranks rollout errors better than forecast step (Spearman rho = 0.90 at 100 steps), yet its spread substantially understates error: one-sigma coverage on the released checkpoint is 4.6% at 100 steps, versus 68% for a calibrated Gaussian. Rolling ensemble members on their own predictions raises coverage to 33%. The implemented squared-error loss on sampled predictions drives each member's variance toward zero. A horizon-specific rescaling restores one-sigma coverage under leave-one-episode-out evaluation, but not two-sigma coverage. We also identify two defects in the released evaluation pipeline. The study covers one robot, one gait and one dataset; it trains no policy.

Vision-language model (VLM) agents that act in the open world need two abilities: world knowledge to interpret what they see, and exploration to find what they cannot yet see. We study how agents combine the two through interactive visual geolocation. GeoGuesser Bench contains 282 tasks in an offline Street View environment where agents rotate, zoom, and move between real panoramas. Its three task types shift the balance from knowledge to exploration: agents (1) report the coordinates of their starting location, (2) name the enclosing street, city, state, or country, or (3) find, read, and count nearby objects. The coordinate tasks in (1) also include 42 real GeoGuessr World Championship rounds, whose expert guesses provide a human reference. We evaluate 23 VLMs: the best accuracy is 62.5% on place naming and 35.0% on object finding, and the best agent averages 4352.4 GeoGuessr points on the championship rounds, against 4770.3 for the better expert in each round. Beyond these headline numbers, world knowledge and exploration are only weakly correlated across models: agents that localize well do not necessarily explore well. To understand where agents fail, we categorize their errors into vision, navigation, and tool-use failures. Further ablations show that the strongest agents, such as Claude Opus 4.8, still perform well with limited views, while weaker agents degrade; customizing the environment thus adjusts task difficulty. The environment's verifiable rewards and restorable state also support training, and our Planner-Worker-Compressor harness cuts planner tokens by 57-61% by compacting the interaction history. We hope GeoGuesser Bench and its environment serve as a high-quality foundation for simulating real-world exploration at scale.

Large language models (LLMs) are increasingly used to evaluate candidates for professional roles, raising concerns about their sensitivity to demographic identities. We examine how candidate and evaluator gender assignments affect decisions by LLM-as-a-Judge (LaJ) models and decision language models (DLMs). In a hypothetical city council voting task, eight models evaluated 240 synthetic professional narratives from NVIDIA's US and French Nemotron-Personas datasets under independently varied gender assignments, producing 30,720 judgments. Explicit candidate gender labels increased Yes-vote probability by 1.05-1.73 percentage points relative to unspecified gender, while only the nonbinary evaluator assignment showed a significant increase (1.26 points). Identity matching showed no significant overall association. Yes-vote rates varied across models (81.3-97.2%), and 2.5-3.3% of decisions reversed relative to the unspecified-identity baseline. Under direct substitutions between explicit gender identities, DLMs exhibited fewer decision reversals than LaJ models (1.18% vs. 3.38%), indicating lower observed decision-level gender sensitivity among the tested DLMs. These findings demonstrate that gender sensitivity depends on candidate and evaluator identities and the evaluation model, while highlighting the importance of examining individual decision consistency alongside aggregate differences.

Agentic Real2Sim2Real systems can reconstruct a scene in simulation from camera observations, but contact-rich dynamics such as friction and restitution cannot be observed without interaction. We call this the dynamics gap and present CUE, an agentic system that closes it by interacting with the environment itself. Given only a high-level instruction and no resets, the agent builds its own perception, identifies physical parameters, and targets interactions where its simulation is least accurate. On billiards, CUE reduces held-out ball-position error by 72% relative to the strongest baseline across 25 hidden-physics conditions in simulation, and we further deploy it on a real pool table. More details can be found on the project website: https://cue-dynamics.github.io/

Supporting a person over repeated interactions requires interpreting their current state in context and anticipating its subsequent evolution. We formulate the \emph{Human-Oriented World Model} (HOWM) as a conceptual framework centered on task-relevant human states and their evolution. HOWM connects current observations and individual history with environmental and social context to organize continuing state estimation, personalized dynamics prediction, and action support. Dynamic expectations and new observations jointly inform the current estimate, which in turn anchors personalized forecasts based on shared knowledge and individual experience. An external policy relates forecasts under alternative conditions to the person's goals, preferences, and constraints when considering support. Actual responses and outcomes provide evidence for revising the state estimate and personal dynamic patterns, thereby informing subsequent estimation, prediction, and support. We illustrate the framework through hypothetical training and learning scenarios. Our research agenda asks whether accumulated interaction experience improves subsequent estimation and prediction, and whether using the model helps the person pursue their goals over time.

Companies deploy large language model (LLM) assistants on their websites to advise customers. For commercial reasons, the operator can give an assistant a goal that the customer never sees. This paper focuses on operator-induced vulnerability-targeted manipulation (VTM), in which an assistant pursues such a goal by exploiting a vulnerability the customer discloses. While existing research mainly measures how manipulative a model is, we go further and ask whether VTM can be detected from the customer-AI conversation. To investigate this at scale, we introduce GAP (Goal-Anchored Probing). GAP first records the option a model recommends without any goal and then sets the other option as the operator's goal. We apply GAP to 15 open- and closed-source LLMs and collect 619,600 labeled replies across 540 everyday scenarios in 12 domains. We organize our study around three research questions. (1) Do assistants with an operator goal target vulnerable customers? (2) Can VTM be detected from the customer-AI conversation? (3) Does detection generalize to unseen models, vulnerabilities, and instruction styles? We first find that 14 of 15 models push the operator goal, and 12 of them use disclosed distress or financial strain to support it. Next, we show that fine-tuned detectors identify exploitative pushes with 79\% accuracy from the reply alone, while zero-shot LLM judges average at most 62\%. Finally, detection weakens on vulnerabilities absent from training. Telling whether a push comes from the operator also becomes unreliable when the assistant is told to hide its goal. We will release our dataset and code to facilitate future research on auditing LLM assistants.

LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.

Spatial transcriptomics maps gene expression across tissue sections and, when integrated with whole slide images, jointly captures both morphological architecture and molecular activity to provide a rich, multimodal view of tissue biology. ML approaches combining spatial transcriptomics and WSIs typically learn a mapping from morphological features to molecular profiles via contrastive alignment or supervised fine-tuning of pathology foundation models. In both cases, transcriptomics serves as a training signal rather than a true input modality, leaving inference fundamentally image-driven. We propose an early-fusion vision transformer architecture in which transcript tokens are constructed at arbitrary spatial resolution, merged with image patch tokens, and jointly processed through a shared transformer stack for deep cross-modal interaction. Starting from a pretrained vision-only pathology foundation model, we train HistoTx by continuing self-supervised pretraining on paired image and spatial transcriptomics data and demonstrate that: 1) at inference with image only, performance remains on par with the baseline on most tasks while improving on tasks that inherently benefit from molecular knowledge, such as gene expression prediction; 2) when both modalities are available at inference, jointly providing image and transcript inputs outperforms either modality alone.

Playing technique is crucial for expressive violin performance, yet few violin technique detection models exist. Current systems mainly rely on note-level pipelines that first estimate note boundaries and then classify techniques. This is especially problematic for bowed string instruments such as the violin, where soft onsets and continuous bow-string interaction render note segmentation unreliable. Here, we propose VioTech, a frame-level violin technique detection framework that predicts techniques directly at 10 ms resolution without explicit note segmentation. To handle the large temporal-scale variation among techniques, VioTech combines a frozen MERT encoder with a multi-scale Mixture-of-Experts (MoE) architecture and task-specific routing for tonal, articulation, and legato-related predictions. We also introduce VioTech120, the first manually annotated dataset of violin playing techniques constructed from real solo performance recordings, as well as an accompanying note-labeling user interface. Results demonstrate that VioTech achieved consistent improvements over prior methods and significantly enhanced cross-dataset generalization, establishing a new state-of-the-art in violin technique-aware transcription.

We present Reap, an open framework for learning and analyzing formal proof search in Lean4. Reap implements policy-value-guided Monte Carlo Tree Search directly in Lean, sharing proof-state transitions, AND-OR search, and replay verification between inference and training rollouts. Structured action logs expose how neural predictions and tactic execution interact during search. Using this framework, we train Reaper-1, a 1.7B-parameter policy-value model, through continued pre-training, supervised tactic learning, and iterative search-based reinforcement learning. Using a separate premise selector, the final system solves 79.9% of miniF2F-test at pass@32, establishing state-of-the-art performance among open provers of comparable size while remaining competitive with substantially larger models. Beyond improved solve rates, our results suggest that RL strengthens the prover's ability to organize proofs around useful intermediate results, favoring substantial advances over repeated local adjustments. This more directed proof construction is accompanied by increasingly elaborate tactic programs: fewer explicit search decisions are needed, while more work is performed within each decision. The resulting trade-off between search effort and tactic execution cost provides insight into how learning changes the prover's problem-solving strategy.

An LLM judge assesses agent steps from text packets. A renderer converts recorded runs into these packets, while software checkers test records and packets against requirements. We ask what evidence shows a packet is what the evaluation requires and propose four input-assurance obligations: required content survives rendering, checkers cover their requirements, a failed check stops later use and the judge reads the bytes last checked. In one pipeline with 30 trajectories, matching hashes, passing checkers and successful staging (copying packets for the judge) coexisted with violations of all four obligations, the last two in integration tests. The renderer omitted every structured tool call, leaving 53 of 60 judged steps blank despite matching hashes. For one judge configuration, later context raised error flags by 13 more on error than on correct steps. Restoring every call reduced this contrast to 2, a reduction of 11 (0.37 per trajectory, 95% cluster-bootstrap interval 0.17 to 0.53). In packets ending at the judged step, that configuration abstained at least once on every blank step, while the other four sometimes abstained but also flagged blank error and correct steps. Both separately written checkers accepted an unmatched tool result. In integration tests, a stub replacing the judge read a failed build's packet and a copy changed after staging. Three added controls stopped 18 of 19 constructed fault cases before the stub read any packet. The remaining stale script never ran. We turn the obligations into four tests and plan to evaluate their transfer to independent pipelines.

Shilling detectors for recommender systems are almost always evaluated by one number, detection AUROC, with no account of when it can be trusted. We audit representation-space detection with VEIL (Variational Evaluation via Inverse-reinforcement Learning), a pairwise (Bradley-Terry) Bayesian reward-inference framework over frozen encoder embeddings, following the Bayesian auditing paradigm for LLM alignment [1]: (A) a base posterior; (B) a hierarchical ARD extension with learned confidence; and (C) a label-routed organic/attacker mixture used as an evidence-accumulation check.We prove (Proposition 1) that no detector built on a representation invariant to a transformation of user profiles can outperform chance against an attack that is indistinguishable from organic profiles under that invariance.Detection is near-perfect against synthetic-structure attacks (AUROC 0.936-1.000) but collapses under donor-mimicry, which clones a real user's item set and splices in 1-3 target items. Performance falls to chance on MovieLens-1M (0.48-0.51), partially degrades on Amazon Musical Instruments, and degrades least on CiaoDVD.Donor-mimicry is not undetectable: a non-learned baseline, based on each profile's maximum raw item-set Jaccard similarity to any other profile, reaches AUROC 0.92-0.97 (average precision 0.50-0.57) on all four donor-mimicry configurations, including ML-1M, where every embedding-based score is at chance.The signal is present in the raw profiles, but LightGCN's representation discards it, placing injected profiles closer to their true donor than its zero-propagation MF reduction does (distance 1.98 vs. 3.09). Severity varies by dataset, encoder, and detection method, with no single axis predicting the others. The limit lies in the representation, not the attack, and raw-similarity checks are a necessary complement to any learned detector.

Chain-of-thought (CoT) explanations can sound plausible while having little influence on a model's final answer. We study four rewards that use the policy's own signals to encourage reliance on generated reasoning: gradient sensitivity, confidence gaps, progressive confidence gains, and reasoning--prompt swaps. We add each reward to group relative policy optimization (GRPO) and evaluate two small instruction-tuned models on GSM8K and StrategyQA using reasoning edits, paraphrases, and prompt hints. Several configurations increase answer reliance to reasoning edits, while effects on accuracy and paraphrase robustness vary. We also observe answer copying at a high reward weight and high reliance in decomposition settings with poor task accuracy. The study compares practical objectives for reasoning reliance, confirms the trade-off between reliance and performance, and proposes four lightweight rewards that can improve reasoning reliance.

Personalizing large language model assistants requires adapting their responses to individual users, including the selection of content, reasoning approach, level of detail, and form of presentation. Such preferences are often latent, becoming visible only when users respond to an initial answer and request revisions. We investigate whether these feedback trajectories can be converted into a reusable personalization signal without explicit preference profiles or labels, prompt-time access to interaction histories, or model parameter updates. We introduce UPSHIFT (User Preference Steering from Human-AI Interaction Feedback Traces), a training-free method that compares the hidden representations of initial and feedback-conditioned responses within each session, averages the resulting trajectory displacements across a user's sessions, and injects the aggregated preference direction when answering a new query. The method is motivated by the implicit-preference hypothesis: although an individual conversational transition may reflect topic progression or task-specific correction, changes that recur across sessions retain a component that captures how the user wants responses to be adjusted. We validate the preference-representation capability of UPSHIFT on PrefEval. In evaluations with 40 simulated users, every tested configuration improves preference alignment over the unsteered baseline on held-out topics for both Qwen3-8B and Qwen3-32B; the best configuration increases macro-averaged Elo from 1364.0 to 1605.2 and from 1341.9 to 1591.8, respectively. Matched-user and positive-versus-negative controls further demonstrate that these gains arise from the user-specific preference direction and its learned orientation. These results show that interaction histories can provide compact, reusable supervision for lightweight personalization.

When a masked diffusion language model is sampled in a few parallel steps in a random order, the model actually run is a mixture over reveal plans, while cross-entropy training fits token conditionals rather than this mixture. We introduce mixture-likelihood fine-tuning (Mix-FT), which fine-tunes a pretrained model on an importance-weighted bound of its own sampler's likelihood, with no teacher and no change to the architecture or the sampler. On LLaDA-8B, a LoRA fine-tune for 0.002% of the pretraining compute writes better OpenWebText continuations than continued cross-entropy and Alpha diffusion given the same compute, at every tested step count. With half the model evaluations, 16 rather than 32 steps per block, its continuations score better than the original model's in every seed under three judges, with confidence-based decoding as well as random order, and under confidence-based decoding two blinded LLM raters from different model families prefer them in 69 and 71% of pairs. On a public 110M-parameter block diffusion model, Mix-FT takes under four T4-hours and beats continued cross-entropy, Alpha diffusion and SDTT distillation at 4, 8 and 16 steps per block, under two judges in every seed. Exact likelihoods on text8, with a control that weights every plan equally, attribute about three quarters of the likelihood gain over continued cross-entropy to weighting each plan by how well it explains the sequence.

Producing an executable CAD script is not sufficient to establish that an agent has satisfied an engineering design request. Evaluation must also consider the resulting geometry, specified dimensions, and native CAD modeling history. We present Parametric CAD Bench v3, a 100-task benchmark comprising 30 text-to-CAD creation tasks, 30 text-to-CAD create-and-edit tasks, and 40 tasks that reconstruct mechanical parts from engineering drawings. Agents operate in a pinned FreeCAD environment and submit native CAD artifacts to a separate programmatic verifier. The primary metric is mean continuous task reward, combining geometry similarity and specification consistency rather than binary task success. A September 24, 2026 snapshot covers 13 model-agent systems; Claude Code with Claude Opus 5.5 achieves the highest overall point estimate, 61.03%, and an image-to-CAD mean reward of 84.66%. These measurements describe complete systems, not isolated model capabilities. We release versioned tasks, evaluation configurations, and inspectable result artifacts.

The author shows a methodology for constructing a network map of the artworld. The tweets posted by a manually comprised set of exemplar artworld actors and Natural Language Processing (NLP) methods are used to identify significant artworld actors, including those that do not own a Twitter account. Identified artworld actors form communities detected with the Louvain method, and their authority is evaluated with the HITS centrality method. The resulting artworld network exhibits the ability, starting with a small set of exemplar actors, to identify new artworld actors and provides a wealth of data that would be very hard to collect otherwise.

Jointly trained multimodal networks frequently underutilize weaker modalities, as faster-learning modalities dominate the shared fusion objective and suppress gradients reaching slower ones. Most existing approaches throttle dominant modalities, whereas directly boosting weaker modalities often fails because monitoring signals derived from the joint loss remain dominated by stronger modalities. We propose Probe-Guided Gradient Boosting (PGGB), a gradient modulation approach that boosts weaker modalities while preserving the strength of dominant ones. PGGB attaches lightweight linear classifiers to stop-gradient features from each modality to estimate representation quality independently. The gap between these scores, termed the utilization gap, provides an imbalance signal used to adaptively scale weaker-modality gradients through bounded and smoothed modulation. PGGB is composable with throttling methods, as it leaves the loss function and fusion architecture unchanged. Experiments across eight benchmarks spanning four domains and 2-4 modalities demonstrate that PGGB outperforms state-of-the-art methods on highly imbalanced multimodal datasets while remaining competitive under low or negligible imbalance. We further establish bounded scaling, self-attenuation, and a descent bound for standard SGD under smoothness and bounded-variance assumptions. Code is available at https://github.com/VuVoCodes/PGGB.

An indirect prompt injection can affect a search-augmented large language model only after its carrier page reaches model context. SearchInject controls this admission decision with a joint score combining retrieval relevance, search-abuse cues, instruction likelihood, and two relevance interactions. A quarantine threshold is selected under an empirical 0.5% clean-candidate false-quarantine constraint on a calibration split. Evaluation uses 300 held-out SciFact queries and held-out payload categories within four construction families already represented during development. Across 1,200 query-family outcomes per retriever, top-10 target exposure falls from 0.9408 to 0.1583 for BM25, from 0.9158 to 0.1508 for a normalized hybrid, and from 0.8450 to 0.1375 for latent semantic analysis. These are relative reductions of 83.2-83.7%; the largest absolute change in clean normalized discounted cumulative gain at rank 10 is 0.0018. The frozen threshold yields clean test false-quarantine rates of 0.59-0.65%, above the empirical calibration constraint. A query-level analytical reanalysis accounts for the repeated construction outcomes under an independent-query assumption. A separate 104-case generation follow-up records context removal but no confirmed undefended hijacks under a post hoc endpoint, so it cannot estimate a reduction in attack success. The evidence supports retrieval exposure control within the evaluated construction families; the benefit of the interaction terms at matched clean cost remains unresolved.

Payment platforms use view-specific models trained on transaction attributes, behavioral histories, and relational structures to rank transactions under limited review budgets. However, existing fusion methods often focus on prediction scores or interactions among model representations, with less attention to how these interactions should guide aggregation using each model's original prediction and input coverage. To address this limitation, we propose InterFuse, a fusion framework that uses a shared fusion state from cross-model interaction to guide transaction-adaptive aggregation. InterFuse encodes each model's representation, prediction score, and input coverage as a model-specific token, using learned states for unavailable outputs. Availability-aware attention updates these tokens and produces a shared fusion state. This state then guides aggregation over available models using their updated representations, original predictions, and input coverage. We evaluate InterFuse on an industrial payment replay and three public financial benchmarks using the same upstream models across fusion methods. InterFuse achieves the highest top-ranked recall and AUC on all four datasets, improving Recall@1% by 1.11-8.09 percentage points over the strongest baselines. Ablations show the benefits of cross-model interaction and adaptive aggregation, while industrial analyses find larger interaction gains when model predictions disagree. Additional analyses examine robustness to missing outputs and the trade-off between ranking gains and fusion latency.

Overall and group-level averages, and ranking metrics such as the AUC, can hide sets of observations for which a prediction model systematically over- or under-predicts. We present an audit that asks where a model is miscalibrated (in which sets of similar observations it over- or under-predicts on average) and for whom (whether its average prediction error differs between demographic groups, such as by race or sex, within the same set). It groups observations into pockets using random-forest similarity, which weighs each feature by its importance for predicting the outcome, with a ridge-regression grouping alongside. Pockets are learned on one sample and tested on another, with simultaneous confidence intervals that allow for correlation within units such as hospitals; the audit needs only features, outcomes and fixed predictions. In 17 simulation scenarios with an error planted for one group, the combined audit located it better than the strongest alternative grouping in 15, with false-alarm rates near or below 5%; with a local error planted for everyone, it won in 12 of 15 scenarios at the smaller error size and lost when the error was unrelated to the outcome. In public synthetic data from the care-management algorithm studied by Obermeyer et al., it found no detectable overall error but under-prediction of illness in the sickest pockets and over-prediction elsewhere, and distinguished Black patients' concentration where the score under-predicted for everyone from a Black-White gap within pockets. For APACHE IVa, an exploratory analysis located its previously reported over-prediction of ICU mortality in a mostly ventilated pocket and a low-risk pocket, with none detected in a moderate-risk one. Two chest X-ray classifiers showed offsetting local errors behind no detectable overall error. Code: https://github.com/modenesi/pocket-audit

Large language models (LLMs) have achieved strong performance across diverse tasks, from question answering to code generation. Recent studies have found that LLMs can perform numerical prediction on tabular datasets via in-context learning. However, it remains unclear under what conditions and how LLMs use variable names and in-context examples for numerical prediction. Here we show that semantic grounding, the match between variable names and the values they describe, allows five LLMs without task-specific training to match or outperform seven supervised machine learning (ML) models across 20 datasets from nine scientific domains, particularly when examples are scarce. Reasoning traces indicate that variable names allow LLMs to draw on domain-specific knowledge acquired during pretraining. Using controlled tasks based on physical formulas, we further show that LLMs integrate domain-specific knowledge with in-context examples in a way consistent with Bayesian updating. Specifically, when the relationship between the inputs and the target cannot be inferred from the examples alone, LLMs rely on domain-specific knowledge as the prior. Conversely, as contradicting examples accumulate, LLM predictions shift progressively towards these examples, which act as the likelihood. Our work reveals how LLMs use domain-specific knowledge in numerical prediction and positions them as data-efficient predictors for data-scarce scientific domains.

Steering vectors align a model at inference time without touching its weights, but eliciting several personas at once is a different problem from eliciting one. Prior multi-persona steering approaches predetermine both how persona vectors are combined and the specific layers at which they are injected. Moreover, this static commitment introduces substantial offline search costs to identify suitable target layers, despite the fact that optimal intervention strengths vary dynamically across both decoding steps and distinct personas. We introduce SYNC (SensitivitY-guided iNference Control), a novel training-free inference control framework compatible with pre-existing steering vectors. SYNC dynamically estimates each layer's sensitivity to every persona via cosine similarity trajectories, assigning each persona to its optimal injection layer at every decoding step. Across two models, two persona profiles and steering vectors from both unsupervised and supervised extraction, SYNC leads on 21 of 28 persona configurations, significantly on 15. Code is available at \href{https://github.com/yosefnuragaw/sync}{this https URL}.

Natural-language explanations generated by large language models (LLMs) for machine learning models often seem fluent but are not necessarily faithful to the underlying model. We formalize this mismatch as a faithfulness gap and propose a complexity-adaptive refinement framework for symbolic regression that iteratively generates and validates explanations until numerical agreement with the target expression is achieved. The approach follows a two-stage pipeline: an explainer LLM produces textual descriptions at increasing levels of detail, while an independent validator LLM extracts a symbolic expression that is compared against the original model. This process is framed as a discrete search over explanation space, where complexity is minimized subject to a faithfulness constraint. We evaluate the method on real-world and synthetic datasets, with a total of 250 runs across varying model complexities. The explanation refinement loop achieves an overall pass rate of 84.0%, with most explanations converging at low complexity levels and only a small fraction requiring detailed descriptions. Our results demonstrate that faithful explanations can be obtained with minimal necessary complexity by explicitly closing the explanation loop through symbolic validation.

Progress in AI is often measured by what a single model can do on its own. Yet teams of AI agents are starting to make breakthroughs and discoveries in science, medicine, mathematics and software engineering. As AI begins to work in teams of agents and people, the ability to collaborate becomes a core part of their intelligence. Research on human teams has measured this ability for decades. Inspired by this, we introduce the Collaborative Intelligence Benchmark, which measures how well an AI agent works as a teammate. It contains 240 tasks across 8 dimensions of teamwork, from refusing an unsafe shortcut to checking the work that a weaker model leaves behind. Each task comes from a real record or an existing benchmark, from aviation accident reports and clinical cases to coding tasks and negotiations, and has a programmatically verifiable outcome. To isolate teamwork from individual capability, models are scored only on tasks that they can solve by themselves when they have full information. We evaluate 18 closed- and open-weight models and find that every model, across sizes and families, has clear weaknesses as a teammate, with the best model scoring only 82.2 out of 100. Frontier models almost always refuse unsafe shortcuts and report misconduct, yet they cave under interpersonal pressure, reaching the best overall deal in only 14% of conflicts, and their pass rate drops from 97% to 81% when they take over a task in which a weaker model has cheated. Crucially, general capability does not guarantee good teamwork, as models with nearly the same capability score differ by up to 11 points as teammates. The benchmark includes 160 public tasks with code and data, along with 80 held-out tasks for evaluation. As AI agents increasingly work alongside other agents and people, the ability to collaborate effectively is becoming a crucial capability to measure and develop.

Neural networks that are trained on data arising from a physical system must somehow learn regularities induced by the underlying physical laws. In this setting, concepts from statistical physics provide powerful tools for analysing how physical laws influence the learning process. In this paper we explain how susceptibilities --- which measure the response of a neural network to perturbations of a parameter of the data distribution --- can be applied when learning from physical data. We show that such susceptibilities can identify the conditions under which specific input features are most informative for learning, guiding data selection, a prediction we validate experimentally. For distributions lacking a natural parameter, we propose introducing one via a physically meaningful scalar function of the physical state, and demonstrate that susceptibilities track the emergence of specific capabilities (such as collision detection) during training.

We present ICJ-Graph, a machine-readable resource connecting documents of the International Court of Justice (ICJ) to cited cases, legal provisions and United Nations resolutions. Court decisions, written pleadings and oral records form distinct document layers, preserving the distinction between the Court's reasoning and the parties' submissions. The collection contains 6,396 ICJ documents and 505 documents from other courts and legal bodies, together with 12,375 provisions and 20,291 UN resolutions. The data files contain text, metadata and embeddings. The accompanying pipeline lets users load the graph and generate embeddings for use in their own retrieval systems. To illustrate the resource's use, we evaluate five models on citation prediction for 199 decisions using documents available before each decision. Tools designed for citation research improve citation-prediction F1 over answering without tools for all five models across case, provision and resolution citations.

Simulation-trained generative models are emerging as a powerful framework for field-level inference in cosmology, particularly in strongly nonlinear regimes where analytical approaches become intractable. However, the finite size of training datasets often leads to variance shrinkage, indicating that part of the realization-level variability of the underlying physical process is not captured by the learned generative dynamics. In this Letter, we interpret variance shrinkage as a diagnostic of missing effective degrees of freedom and introduce physics-informed latent modifications to the generative dynamics of a continuous normalizing flow. We show that these latent corrections recover a substantial fraction of the missing variability without requiring additional training simulations. Using weak-lensing field-level inference as a representative cosmological application, we further demonstrate through posterior analysis that the learned latent directions are actively exploited during inference and behave as effective nuisance parameters that absorb part of the mismatch between the training and target distributions. As a result, the model exhibits improved robustness to out-of-distribution shifts in field-level inference. Our work establishes variance shrinkage as a practical diagnostic for identifying missing degrees of freedom in simulation-trained generative models and provides a framework for recovering them through physics-informed modifications of generative dynamics.

Control barrier functions (CBFs) are a standard way to keep a planner inside a safe set, but they assume control affine dynamics, a barrier function given in closed form, and a known nominal controller. Integral control barrier functions (ICBFs) remove the first assumption and absorb input constraints, yet they still assume a given nominal controller and a barrier function. In this work, we present an approach to learn all three from data, using neural ordinary differential equations (ODEs) for the dynamics, imitation for the nominal controller, and supervised regression for the barrier function. The learned dynamics, nominal controller, and the barrier function are used to build the ICBF filter constraint analytically. Furthermore, the barrier is trained on viability rather than geometry, meaning that a state counts as safe only if some emergency maneuver can still avoid the constraint under the actuator limits, a correction that geometric labels miss entirely and that grows with speed. Actuator limits then enter the online program as hard bounds while the learned barrier rows may be relaxed, which keeps the program solvable at every state and the actuator limits exact, and we bound how far the input it returns can drift from the one the true models would give. On agile racing with a 6-D dynamic bicycle, the filter removes all 12 unsafe episodes from 200 held out layouts at an unchanged success rate and above 100 Hz on one CPU core, holds the input constraint on every episode of a benchmark where a jointly trained baseline holds it on 53.38%, and keeps that advantage under large errors in the plant friction.

Diffusion models are powerful generative frameworks for producing high-quality images by denoising latent variables from random noise. However, training with likelihood-based objectives, such as denoising score matching, can lead to locally over-smoothed high-frequency details, including fine textures and sharp edges, thereby limiting perceptual fidelity and structural detail. Adversarial training with GANs enhances sharpness but typically requires additional discriminator networks, increasing computational costs and destabilizing training. To this end, we propose Latent Diffusion Generative Adversarial Networks (LD-GAN), a novel framework that seamlessly integrates adversarial learning into diffusion models without modifying their original pipeline. LD-GAN leverages the pretrained variational autoencoder (VAE) in latent diffusion models as an energy-based discriminator, enabling adversarial training without extra parameters and preserving the structured latent priors learned from large datasets. We also introduce a structural consistency energy that aligns encoder and decoder feature representations, thereby enhancing perceptual quality and compatibility with the pretrained latent space. Extensive experiments demonstrate that LD-GAN significantly improves sample fidelity, perceptual sharpness, and diversity over state-of-the-art baseline methods across various generation tasks while ensuring efficient training dynamics.

An agent acting in a stateful environment can leave evidence of harm in two places: the record of what it did and said, or the world it changed. Safety evaluations for discovering harm examine one or the other in limited ways, using either an LLM judge over the trajectory or deterministic checks over the final state. Although LLM-based verifiers that inspect the environment exist, they have been used to verify task completion, not to detect and confirm harm in a safety setting. We address this gap with an LLM that has read-only access to the environment state and is given a harm taxonomy. We evaluate this approach in three environments: two from DecodingTrust-Arena and one from OSWorld. We hold the judge model and harm taxonomy fixed, varying only what the judge observes: the trajectory, read-only tools for inspecting the environment, or both channels together. In the two DecodingTrust-Arena environments, combining the trajectory and state-only judges recalls 94 percent of confirmed harms, compared with 53 percent for the trajectory judge alone. On OS-Harm, the union recalls 77 percent, compared with 67 percent. A single judge given both channels does not perform better and produces more false negatives than the state-only judge. We also record disagreements and find that the two channels miss different episodes. These results suggest that combining verdicts from two channel-specific judges is a robust and flexible approach to detecting and verifying harm in agentic settings.

With the growing deployment of large language models (LLMs) across diverse domains, concerns regarding their safety have grown substantially. LLM unlearning has emerged as a pivotal approach to removing harmful or unlawful content while maintaining utility. Despite increasing interest, the challenges of continual unlearning, which is common in real-world scenarios, remain underexplored. Successive unlearning tasks often lead to intensified utility degradation. To effectively unlearn targeted knowledge while preserving LLM utility, it is essential to minimize changes in model parameters by selectively updating those linked to the target knowledge, thereby ensuring other knowledge remains unaffected. Building on the task vector framework, we propose a new method named ALKN (Adaptive Localization of Knowledge Negation), which uses dynamic masking to sparsify training gradients and adaptively adjusts unlearning intensity based on inter-task relationships. Comprehensive experiments across three well-established LLM unlearning datasets demonstrate that our approach consistently outperforms baseline methods in both unlearning effectiveness and utility retention under continual unlearning settings.

Enterprise conversational systems require memory to personalize interactions, retain information across sessions, and maintain conversational coherence. At enterprise scale, however, memory must not only retrieve relevant information accurately but also provide it efficiently to the answer-generation model. Existing memory systems face a trade-off between retrieval accuracy and context efficiency. Retrieving more context can improve recall but increases inference cost and latency. We present Scalaris, a scalable enterprise memory architecture that addresses this trade-off through a unified long-term and short-term memory design. We propose that user interactions and assistant responses possess distinct characteristics and roles in a conversation: user interactions capture evolving knowledge about the user and provide value across sessions, whereas assistant responses must be stored efficiently so that only information relevant to the current query is retrieved. We therefore treat them separately and support each with specialized representations and ingestion/retrieval pipelines. Scalaris reaches 92.2 % accuracy on LongMemEval-S and 90.77 % on LoCoMo using Qwen3-32B, state-of-the-art results on both benchmarks, while supplying the answer-generation model with approximately 10 times fewer retrieved context tokens than Hindsight, a leading memory system for conversational AI. When integrated with our proprietary closed-source AI assistants, Scalaris improved overall response quality by 36 %. Together, these results establish Scalaris as a cost-effective framework for building scalable, personalized conversational AI systems.

Reinforcement learning for language models often combines several rewards by normalizing each one and adding them with fixed weights. When the rewards are correlated, the variation they share is counted once for every reward that carries it. In terms of the ZCA-decorrelated components of the reward signal, the weighted sum applies weights $Σ^{1/2} w$ rather than the chosen weights $w$, where $Σ$ is the reward correlation matrix. We study WAS-PO, which whitens the normalized advantages with $Σ^{-1/2}$ before adding them, so that the chosen weights apply to those components. We derive the scalarized objective whose gradient it follows when normalization statistics are held fixed, show that its weights on the raw rewards can be negative, give an exact two-reward condition under which the lighter-weighted reward is reversed, and give a loose bound on the error from estimating $Σ$ during training. On tool calling with Qwen2.5-1.5B-Instruct, WAS-PO reached a higher mean training reward than GDPO in all seven seeds with three rewards (3.8% higher) and all three seeds with four (3.9% higher). WAS-PO also replaces GDPO's per-prompt normalization with batch-level normalization, and we did not separate the two changes with three or four rewards. The training-reward gains did not carry over to held-out accuracy, which was slightly lower than GDPO's with three rewards (not significant) and 0.03 lower in absolute terms with four (two seeds evaluated). With two rewards, WAS-PO trailed GDPO in training reward, and a three-seed control weakly suggests that the gap comes from batch-level normalization rather than whitening. In a math setting where the two rewards were nearly identical by construction, WAS-PO drove down the lighter-weighted reward, as the sign condition predicts.

Indirect prompt injection creates a distinctive risk for blue-team Large Language Model (LLM) agents in Security Operations Centers (SOCs). Malicious instructions can be embedded in telemetry, tickets, retrieval results, or memory and then propagate through alert triage, incident summarization, response recommendation, and Indicator of Compromise (IOC) extraction as locally plausible intermediate decisions. We characterize this failure mode as evidence-chain contamination of protected task state. The main challenge is task dependent: a suspicious string may be malicious control text for summarization but essential evidence for IOC extraction. To address this challenge, we propose StateFence, a task-aware evaluation and defense framework that treats severity labels, incident narratives, response policies, and IOC sets as protected states. StateFence maps indirect prompt-injection attacks into explicit, stealth, multi-hop, and persistent SOC evidence attacks, records observable propagation traces, and routes defenses according to the attacked state. Across controlled SOC cases and public-SOC validation cases, StateFence lowers attack success and reduces failures that appear only after evidence is poisoned, while preserving task utility and avoiding the indicator loss caused by task-mismatched filtering. Our findings indicate that robust blue-team LLM agents require mechanisms that preserve task-critical state under untrusted evidence.

On-device multimodal agents are limited by two deployment mismatches: pretrained attention structures may not divide evenly across the compute cores of an edge accelerator, causing irregular partitions; and standard prefix caching keeps historical visual tokens in the persistent KV cache, whose capacity grows with the number of rounds. We present PARC, which combines parallelism-aligned structured pruning with lifetime-aware KV-cache reorganization. At the model level, it uses first-order Taylor importance and grouped-query-attention-aware selection to make attention dimensions divisible by the target core count while preserving the query-to-key/value ratio and head dimension. At runtime, it separates the system prompt, persistent text history, and transient visual observations by lifetime, updating text history during tool execution so that historical images leave the persistent cache and more interaction rounds fit a fixed budget. On 2,062 GUI actions, the incremental interface shows no accuracy regression, and task-aware recovery training restores 98.7% of the action-type accuracy gap and the coordinate hit rate to the original level. Under seven-core deployment, hardware-aligned pruning reduces vision-encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers per-heap high-water memory by 12.7%; a parameter-matched control confirms that the speedup stems mainly from core alignment rather than parameter reduction.

Visual mental imagery (VMI) decoding uses brain activity to reconstruct internally generated visual scenes, providing a potential pathway for externalizing internal experiences and developing neuroprosthetic communication interfaces. However, progress is limited by the lack of dedicated, training-scale VMI datasets. Current VMI paradigms often rely on visual prompts or short-term memory images, introducing visual and mnemonic confounds that make it difficult to disentangle neural activity associated with endogenous image construction from signals driven by prior or concurrent visual input. Furthermore, existing VMI datasets offer limited descriptive control over individual visual attributes, making it difficult to systematically assess fine grained visual details. To address these gaps, we introduce AVID, the first dedicated training-scale 5.0 Tesla functional magnetic resonance imaging (fMRI) resource for auditory-induced VMI decoding. AVID comprises approximately 67 hours of dense recordings from nine participants. The dataset is structured into two complementary splits: a scene-level complex split using naturalistic Mandarin descriptions and a controlled simple split with explicitly specified visual attributes. We further formalize a standardized evaluation protocol that decouples source-image consistency from semantic alignment, allowing for a more nuanced assessment of imagery reconstruction. Baseline qualitative and quantitative results show that current reconstruction pipelines remain limited on AVID, although the most successful reconstructions preserve recognizable scene level semantics from spoken descriptions. By combining dedicated training-scale neural recordings and standardized evaluation resources, AVID provides both a data foundation for imagery-specific model development and a benchmark for evaluating VMI decoding with substantially reduced sensory contamination.

Large language models are increasingly consulted for religious knowledge, including the recall of specific Hadith (recorded sayings and actions of the Prophet Muhammad, peace be upon him). Unlike open-ended generation, Hadith recall is a high-stakes, verifiable task: a reference (collection, chapter, and Hadith number) has exactly one correct text, and any deviation from it constitutes a fabrication that can be mistaken for authentic scripture. This study introduces a bilingual (English/Arabic) benchmark that probes five open-weight, locally-deployable language models (Mistral-7B, Gemma3-4B, Jais-7B, ALLaM-7B, and Phi4-mini-3.8B) across all six canonical Sunni Hadith collections (Sahih al-Bukhari, Sahih Muslim, Sunan Abi Da'ud, Jami' al-Tirmidhi, Sunan an-Nasa'i, Sunan Ibn Majah), sampling 500 Hadith per collection (3,000 per model per language, 30,000 responses total). Each model is given only the bibliographic reference and asked to reproduce the exact Hadith text, with no retrieval augmentation, and is scored on lexical fidelity (BLEU) and semantic fidelity (multilingual cosine similarity), with refusals flagged separately. Across all five models and both languages, 0% of 30,000 responses achieve both high lexical and high semantic fidelity, i.e., an exact, verifiable match to the source text. Arabic responses fare uniformly worse than English. These findings suggest that current open-weight LLMs, used without retrieval grounding, are unreliable and potentially dangerous sources for exact religious-text recall, and that deployment for Islamic scholarship should mandate retrieval-augmented generation over a verified corpus rather than parametric memory. The complete code, prompts, evaluation pipeline, and scored outputs are publicly available at https://github.com/dayam8696/Hadith-Hallucination-Benchmark.

Large language models (LLMs) are increasingly considered for education, recruitment, workplace decision support, and other consequential applications, making pre-deployment social-bias detection essential. Yet major benchmarks are English-centred or translated, focus mainly on negative stereotypes, and may miss linguistic and cultural variation. We investigate gender, racial, and religious bias using 12,000 locally written Reddit posts in English, Spanish, French, and German. Only search terms are translated, preserving local context. The dataset is balanced by language and category; English gender is additionally balanced by stereotype valence. Three open-weight and three proprietary API-based models are evaluated with a two-blank design that reduces grammatical-gender cues. Religion has the highest absolute choice imbalance in every model-language condition. Spanish is highest for gender, whereas German is highest for race and religion. The evaluated API models score below the open-weight models throughout. Negative English-gender items exceed positive items; Australian and Mexican contexts exceed their Canadian and Spanish counterparts. These descriptive differences measure choice concentration, not stereotype endorsement or causal country effects. The crossed patterns show how aggregate scores can conceal category- and context-specific behaviour, motivating closer inspection before deployment.

Long-context large language model (LLM) inference is increasingly constrained by the cost of maintaining an ever-growing KV-cache. Low-rank K/V projections provide a promising way to compress the KV-cache with minimal quality loss, but prior work has largely optimized KV-cache compression in isolation, leaving distributed inference, especially with tensor parallelism where attention computation is partitioned across multiple GPUs, largely underexplored. We make two observations: (i) tensor-parallel communication is an important cost when deploying low-rank KV-cache compression in distributed inference; and (ii) Value compression can be associated with the output projection, allowing the same low-rank Value representation to reduce both Value-cache width and tensor-parallel communication. Based on these observations, we propose BasisKV, a low-rank KV-cache compression method for communication-intensive, long-context LLM inference. BasisKV jointly optimizes the low-rank factors across per-GPU attention partitions against the complete attention output, rather than fitting each partition in isolation. The compressed Value cache is further reused to help identify which Key pages should be accessed: it predicts part of the Key information needed for routing, while a Key-specific residual stores what cannot be recovered from Value. The original Keys are then retrieved only for the selected pages to compute exact QK scores. Across multiple LLM families and system settings, BasisKV achieves a better quality--efficiency tradeoff than existing low-rank baselines, reaching up to 1.42x end-to-end speedup under 8-GPU tensor-parallel inference. Our code is available at https://github.com/alexz949/BasisKV

Automated liver ultrasound report generation improves the diagnosis of liver diseases by producing clinically structured descriptions of anatomical and disease-related findings. However, current generative vision-language models (VLMs) often rely on a single reasoning pathway that maps ultrasound appearance directly to textual descriptions. In addition to the inherent vision-language modality gap, this design is limited for liver ultrasound because disease interpretation is constrained by subtle visual evidence, heterogeneous disease patterns, and limited paired supervision. To address these limitations, we propose SMILE, a Structured Multi-agent clInicaL rEasoning framework that uses multiple complementary specialist agents to interpret ultrasound images and consolidates their outputs through a judge-guided consensus mechanism. Our approach employs three specialized agents: an image interpretation agent for morphological descriptions, a clinical-stage reasoning agent, and a visual question answering (VQA) agent for targeted queries. These agent responses are then aggregated and scrutinized by a judge VLM using a consensus filtering prompt to produce a unified, clinically coherent output. Across a public liver ultrasound dataset and an internal clinical dataset, SMILE consistently outperforms state-of-the-art VLMs, achieving up to 19.0\% improvement in ROUGE-1 and 18.8\% improvement in clinical F1-score, while reaching VQA F1-scores of 0.82 and 0.65 on public and internal datasets, respectively. These results demonstrate the potential of SMILE for data-scarce and limited-supervision clinical settings. Our models, datasets, and code are available \href{https://github.com/CMME-Lab/SMILE}{here}.

Membership inference determines whether a candidate example was used to train a text-to-image diffusion model and is an important tool for auditing training-data use. We identify a directional membership signal in \emph{conditional prediction shifts}, the changes in predicted noise under alternative text conditions at a fixed noisy input: whether movement toward the alternative prediction initially lowers or raises the denoising loss from its value under the original caption. Members exhibit less room for such local improvement. To quantify this response, we propose \emph{LoFIT} (Local Conditional Fit), which measures the initial response of the denoising loss along normalized prediction shifts and aggregates it across multiple text conditions and diffusion states. LoFIT requires no target-model gradients and can be computed using only forward queries to the denoiser. Evaluations on four datasets show improved membership discrimination across fine-tuned and pretrained models. LoFIT distinguishes members early in fine-tuning, while the naive denoising-loss baseline remains near chance. Controlled comparisons show that the membership signal is concentrated in the first-order response, with little discrimination from prediction-shift magnitude alone.

Deep reinforcement learning (DRL) agents are vulnerable to perturbations of their observed state inputs. Prior works often constrain state adversaries with $L_p$-norm uncertainty sets and evaluate policy performance under the resulting perturbations. However, in continuous control tasks like robotics control, different state dimensions have distinct physical meanings and influence policy behavior to different degrees, which $L_p$-norm sets fail to capture or explain. Moreover, it is important to identify when and along which state dimensions a policy is vulnerable, independently of the algorithm used to train it. In this paper, we focus on evaluation-time attacks applied to pretrained policies. We propose \textbf{P}erformance-\textbf{D}egradation \textbf{N}on-\textbf{U}niform \textbf{U}ncertainty \textbf{S}et (\method{}), an interpretable formulation of uncertainty sets for state perturbations. \method{} is state-dependent, dimension-wise non-uniform, and can be synthesized online during policy evaluation. We first formulate an ideal constrained optimization problem for \method{} computation, then introduce a surrogate constraint to make the problem tractable. Experiments across multiple continuous-control environments, policy algorithms, and state attacks show that \method{} reveals variation in policy sensitivity across states and dimensions, while its performance degradation is moderately controlled.

Models like DeepSeek-R1 made Reinforcement Learning with Verifiable Rewards (RLVR) famous: reward a model for correct answers and it gets dramatically better at reasoning without being taught how. The leading explanation is sharpening---that the base model already knows how to solve these problems and RLVR just makes it more likely to reach for its best solution paths. This has only ever been checked by comparing before and after training, never by watching what changes inside across it. We tracked OLMo-2-1B across all 13 training checkpoints on GSM8K. First-try accuracy climbs from 58.0% to 61.3%, but ask the model 256 times and solution coverage actually rises from 94.5% to 96.9%, proving the knowledge was already there. We then planted reasoning mistakes into problems and trained simple probes to detect them. Those probes held between AUROC 0.798 and 0.840 across all 13 checkpoints and 3 error types, a spread smaller than the confidence interval; the model never builds a sharper internal picture of what a reasoning mistake looks like. The one place we find real internal change is at the final answer token, where a probe distinguishing early from late checkpoints hits an AUROC of 0.93, while the same probe on the actual reasoning steps lands near chance. RLVR makes the model lead with its best thinking. It does not change what that thinking is.

More databases can make a materials campaign worse. Through an organizational lens, we trace each source's contribution. One band-gap campaign ranks seven anion families from archived measured gaps, one vote per source. In the first half of its budget, all seven pooled rank worse (rank agreement 0.737; 1 = measured order) than the best pool of three (0.851). As a stability check, choosing the pool on half the runs and testing on the rest kept the gap. Splitting Materials Project afterwards points to its transition-metal oxides and fluorides. Two databases from one pipeline add nothing over one, but pooling gives their shared ordering two votes. New calculations repay their cost early only when two or three campaigns reuse them. Flat language-model agent teams trailed a random-pick baseline (mean 0.876, sd 0.096, 20 searches). An adaptive Al-Mg-Si team, mostly interatomic potentials, was scored on unseen HSE06 calculations. In its first quarter of compute it found 49% of the ten lowest-formation-energy compounds, against at most 40% for fixed teams and 73% for MACE-MP-0 alone. Merging members whose orderings agreed carried most of its gain. This suggests pooling a few sources that fit the measurements, counting each pipeline once and charging for computation.

The rapid growth of automated code generation has created a pressing need for automated code review. Yet code review resists full automation: it depends on context beyond the \textit{diff}, on the author's intent, and on judgments about design and maintainability on which even expert reviewers disagree. Existing code review benchmarks lack standardization and rely largely on noisy human review comments scored with text-similarity metrics, which reward surface overlap and fail to measure precision. We introduce \texttt{CR-Bench}, a defect-centric evaluation framework that recasts SWE-Bench issues as pull-request review tasks with full repository context. After filtering for review-detectable defects and grounding each to a single originating pull request, \texttt{CR-Bench} contains 584 instances, with a 174-instance \texttt{CR-Bench-Verified} subset derived from SWE-Bench Verified. Each instance pairs a pull request with a known defect, its ground-truth fix, and a taxonomy label for bug category, impact, and severity. We propose two complementary metrics: \emph{Known-Defect Detection Rate} (KDDR), which measures whether an agent identifies the labeled defect, and \emph{Usefulness Rate}, which measures the fraction of review comments that are actionable. We evaluate eight code review agents over four backbones, plus commercially available Gemini Code Assist, using three independent LLM judges. Even the strongest systems detect only about one-third of known defects (KDDR $\approx$ 33\%), and Gemini Code Assist detects 8.1\%.

The rapid deployment of autonomous AI agents across consequential domains including healthcare, legal services, nancial management, and enterprise workows has created an urgent governance challenge. While the eld of AI evaluation has produced sophisticated benchmarks for measuring agent capability, these frameworks measure exclusively what agents can do rather than how they do it. The ethical dimensions most critical to responsible agent deployment, in- cluding respect for human autonomy, harm avoidance, transparency and non-deception, privacy compliance, and accountability, are systematically absent from current evaluation practice. This paper addresses that gap through three contributions. First, a structured literature review of twelve leading AI agent benchmarks characterizes the ethical dimensions those bench- marks do and do not measure. Second, the Ethical Evaluation Framework for AI Agents (EEFA) is introduced as a comprehensive, operationalizable instrument for ethical agent evaluation, grounded in established AI ethics principles and designed for use by developers, auditors, and regulators. Third, a reproducible validation protocol is specied for applying the EEFA to deployed agent systems, with illustrative test scenarios across all ve framework dimensions. The EEFA is organized across ve primary dimensions: Autonomy Preservation, Harm Avoid- ance, Transparency and Non-Deception, Privacy Compliance, and Accountability and Traceabil- ity. The framework maps directly onto the requirements of the EU AI Act and the NIST AI Risk Management Framework, positioning it as a practical compliance assessment instrument for high-risk AI applications. Empirical application of the framework across diverse benchmarks and agent systems is identied as a priority direction for future work.

A competent learned controller already recovers from many disturbances. Improving it means learning a correction, and learning when not to apply it. Freezing a skill's parameters does not settle the second half. In a controller that routes among frozen skills, a new entry rule can redirect the states an old skill succeeded on. We therefore constrain entry instead of parameters. Each update searches for a feedback skill at a measured failure. It fits an entry rule that reads observations and rejects every state on an accumulated successful trajectory. The skill is admitted only if the composed controller reproduces the new recovery on its own and replays every protected trajectory exactly. A finite-trajectory condition states when this preserves the previous controller. One mechanism runs on four streams: two motions on each of two humanoid systems. Two updates per stream add 62 recoveries over 192 withheld perturbations, and lose none of the base controller's successes. We then price that guarantee. Removing the constraint recovers 33 more conditions but stops eight protected ones from completing at all, and excluding only what admission proves necessary recovers none of that cost.

Attention-based rerankers score each candidate by the attention mass that retrieval heads route to it from the question, but the question, serving as a fixed query, carries limited information for locating relevant evidence. We propose Re$^3$Rank, a listwise reranker built on reasoning-conditioned retrieval heads (RRHeads). A generator first writes a reasoning trace over the question and the candidates, and a scorer then uses the trace tokens as the attention query of a few selected RRHeads, reading relevance as continuous scores. We present a straight-through estimator to propagate the ranking loss through the discrete trace, so that the generator and the scorer can be trained jointly for ranking. Averaged over 9 benchmarks, Re$^3$Rank outperforms 11 baseline rerankers: our 4B variant exceeds the strongest baseline by 3.11 NDCG@10 points on average, and even our 0.6B variant surpasses rerankers with up to 32B parameters.

Hierarchical control scales robot behavior by separating task-level planning from low-level control. In dynamic, high-dimensional tasks, however, a fixed planner-to-policy interface can fail when contacts, balance, or disturbances evolve faster than high-level policy updates. Execution therefore requires choosing not only which low-level policy to run, but also which policy-specific goal should be given to it. We formulate this problem as policy orchestration: runtime selection of a policy-goal pair that is likely to remain feasible while tracking the planner target. We develop EPO, an energy-based policy orchestrator for modular goal-conditioned policy families. For each policy, EPO learns a composite energy over low-level goals, where lower energy corresponds to higher feasibility and more accurate target realization. At runtime, EPO refines each policy's goal and executes the lowest-energy policy-goal pair. Because energies are policy-specific, new policies can be incorporated by learning their own energy models, without retraining the existing policies or a centralized scheduler. In simulation, EPO improves the empirical precision-robustness trade-off by refining policy-specific goals before selectively switching among policies based on energy scores. We further demonstrate sim-to-real deployment of the same frozen policies and energy models on real humanoid loco-manipulation tasks. Project website: https://yewr.github.io/epo

Scientific discovery from the literature requires assessing connections among findings obtained under different study conditions. Source attribution makes individual findings traceable, but does not determine what they jointly support. We introduce GI-TraceBench, a benchmark of 10,282 questions in gastrointestinal oncology that unifies provenance question answering, knowledge discovery, and hypothesis evaluation through shared Paper, Claim, and Evidence Span records. We further propose FindLink, a two-stage framework for provenance-grounded scientific reasoning. Premise Grounding constructs source-bound premises with explicit study context and support status. Structured Reasoning organizes these premises into support plans that represent both supported relations and unresolved conditions, then generates evidence-conditioned answers using deterministic compiler feedback. The compiler verifies structural and source-binding consistency, but does not assess scientific entailment. On the 1,545-question fixed-evidence test set, FindLink improves the balanced family macro score from 67.95 to 87.25 over direct supervised fine-tuning of the same backbone. A retrained variant without Structured Reasoning achieves slightly higher premise-identifier F1, but substantially lower relation and evidence-sufficiency F1. These results indicate that scientific reasoning requires not only identifying grounded premises, but also assessing what connections those premises support.

Large language models (LLMs) can generate visual assets through text-based formats such as SVG and CAD, but these symbolic representations describe images using high-level primitives rather than explicit raster values. We ask whether LLMs can instead generate pixel images directly through their existing vocabulary, without a learned image tokenizer or decoder. We introduce PixelBench, a benchmark for text-to-pixel (T2P) and image-to-pixel (I2P) generation across multiple resolutions, together with a complementary pixel-understanding evaluation. Models generate an RGB color palette followed by a grid of indices specifying the palette color of each output pixel. In T2P, the strongest models produce recognizable images directly from text. In I2P, the strongest models outperform palette-matched downsampling and specialized pixelization algorithms at low resolutions. Pixel-understanding performance correlates strongly with generation performance. A resolution-scaling study shows that pixel-generation performance often improves initially before declining at high resolutions. Finally, we introduce Pixel Reconstruction Tuning (PRT), which augments visual instruction tuning with caption-palette-grid trajectories. We post-train Molmo2-4B from its pretraining checkpoint by mixing general instruction data with our curated PRT data, improving performance across 16 visual-centric benchmarks, indicating that explicit pixel reconstruction can provide useful auxiliary supervision.

Scientific knowledge gains new uses when ideas, methods, and results move between research contexts. Understanding this process requires a model of both specialized knowledge production and the intermediary activities that connect it. We introduce Scientific Liquidity Agents (SLA), an agent-based model of knowledge arbitrage in scientific communities. SLA represents domain specialists and epistemic arbitrageurs within a shared environment of research problems, researcher interactions, and scientific artifacts. Individual decisions connect to collective state through a belief-desire-intention architecture, a prerequisite-based problem graph, a dynamic interaction network, and a persistent publication registry. The framework makes researcher roles and knowledge-transfer frictions explicit, providing a process-oriented account of scientific production and exchange. Two exploratory simulations with 10 and 18 agents produce 68 and 274 registered papers. In the smaller run, arbitrageurs take part in 16 of the 26 contact pairs that join this community into one connected component, while topical similarity alone splits it into seven. The same run yields three similarity edges, 26 contact pairs, and two coauthorship pairs, so measured community structure depends on what counts as a link. Paper production is abundant, but no recorded replication succeeds, which places the bottleneck in verification rather than production. Solved-problem sets correspond to the prerequisite reachability of the disciplines represented in each community. SLA makes knowledge liquidity, the movement and reuse of artifacts across domains, a quantity that agentic science can simulate and measure.

Large language models have demonstrated strong capabilities in structured reasoning and code generation, yet identifying an assembly fault does not guarantee the ability to generate executable rework programs. In interlocking assemblies, a locally plausible repair program may violate blocking dependencies, omit required recovery operations, or terminate in an incorrect configuration. We identify this challenge as fault-conditioned executable rework generation, where models must jointly reason about structural failures and generate programs that restore constrained assembly states. To benchmark this capability, we introduce FORGE-Bench, containing 4,000 tasks across 26 interlocking assemblies with five assemblies held out for evaluating generalization to unseen structures. We further propose FORGE (Fault-Oriented Rework Code Generation for Executable Interlocking Assembly), a structured state-to-program framework that decomposes rework generation into Rework Structuring and Action Grounding. Rework Structuring captures fault-induced recovery dependencies, while Action Grounding maps these dependencies into state-conditioned executable actions. Experiments show that locally reasonable repair code is insufficient for reliable assembly recovery, and that connecting structural recovery reasoning with executable action realization is essential for verified target restoration.

Scaled weight decay for Muon (Muon-SW) has been reported to train language models more efficiently than constant decoupled weight decay, but such comparisons usually reuse one learning-rate (LR) and weight-decay (WD) setting for both methods. We re-run the comparison under the MAI-Base-2 protocol: LR and WD are tuned jointly through their product for each method, on a source-scaled mixture-of-experts (MoE) scaling ladder (1.9M to 15M active parameters) that holds tokens per active parameter and data repetition fixed across sizes. We find: (1) Most of the gain is a tuning gap. At the shared setting Muon-SW wins by 0.035 bits per byte (BPB), a 1.52x compute saving; after both methods are tuned, the gap is 0.0026 BPB (95% CI [0.0018, 0.0034]), a 1.03x saving, so about 90% of the advantage disappears. (2) The small advantage that remains is real: at the smallest size, all four paired seeds favour Muon-SW at the tuned settings. (3) That advantage grows with model size: across an 8x range of active parameters, the tuned gap rises from 0.0036 to 0.0051 BPB while the untuned gap shrinks from 0.035 to 0.024, so the surviving share rises from about 10% to 21%. (4) The best weight decay moves with scale: the optimal LRWD product grows for both methods, faster for Muon-SW (ratio 1.40 to 1.91), and a law fitted on the two smallest sizes overshoots the third by 1.3-1.5x, costing up to 0.002 BPB. The whole study runs on one NVIDIA T4 GPU; we release the code and all run results.

LLM agents call tools that delete data or move money through the Model Context Protocol (MCP). The specification asks for a human who can deny tool calls but splits the mechanism between the server, which can annotate a tool or ask the user through the host (elicitation), and the host application, which decides whether to prompt. We measure both sides from source code. In a 1,000-repository sample of the official MCP registry, an estimated 90% of destructive tools (95% CI 82-95) have no server-side confirmation, and most servers that elicit proceed, or fall back to a flag the model can set, when the host cannot relay the question. Of 41 popular open-source hosts and agent SDKs, 32 have no approval step by default, including 16 of 25 applications. Pairing sampled destructive tools flagged by a rule-based extractor with these applications, 60.7% of pairs have no enforced human approval by default. Through four agent SDKs, four of five low-cost models executed an injected deletion in 479 of 960 runs. If hosts prompted for every tool not annotated read-only or non-destructive, the unapproved share of these pairs would fall to 5.4%; universal elicitation support would leave 56.7%. Applications should therefore configure approval.

Programs that use a language model as a classifier act on its most probable label. Text injected into the input can move that label to one the attacker chooses, a hijack. A monitor that does not know the attacker's label can read the model's own confidence and raise an alarm when the top probability is low. We test this monitor on AG News topic classification with five systems, three of them served through a provider's API. Pooled over five families of injected text, low confidence ranks hijacked answers above clean ones on four of the five systems (AUROC 0.86-0.90). Ranking them does not stop many of them. A threshold set so that at most 5% of clean answers to separate calibration articles alarm falls among the least confident clean answers, and every more confident hijack passes. In our main gate run on the three API-served systems, the monitor misses 51-70% of the hijacks caused by a plain injected note (per-system point estimates). When the note adds one handwritten sentence that calls the target certain and asks for full probability, the monitor misses 99-100% of the hijacks, although low confidence still ranks them at AUROC 0.70-0.79. Confidence monitors should be tested at their operating alarm rate, against attackers who claim certainty.

Alternating updates, that is, where each player is updated in turn, tend to typically perform better in regret matching algorithms than simultaneous updates. Alternation within the regret matching family has long been a central ingredient of practical solvers for zero-sum games. Yet our theoretical understanding is lagging behind. In this paper, we address this gap by developing a systematic way of analyzing such algorithms. Our framework centers on an alternating analog of the RVU property that we introduce, which we use to obtain the optimal $T^{-1}$ average-iterate convergence within this family. In particular, we establish this for two simple variants of predictive regret matching$^+$: i) increasing-regret predictive regret matching$^+$ and ii) half-prediction regret matching$^+$. Interestingly, we find that half prediction is the unique prediction multiplier that satisfies our alternating RVU property, giving it a strong theoretical justification. Going beyond average-iterate convergence, we refine our framework to obtain $T^{-1/2}$ best-iterate convergence and further establish convergence of their strategy profiles to a single Nash equilibrium, without assuming equilibrium uniqueness. We go on to establish separations between alternating and simultaneous updates, including a game in which half-prediction regret matching$^+$ converges under alternating updates but fails to converge in terms of the last iterate under simultaneous updates. This provides theoretical support for the empirical observation that alternation can stabilize learning dynamics. Finally, experiments on 15 game benchmarks demonstrate that the strong theoretical properties of the half-prediction mechanism translate to strong practical performance, being on par with state-of-the-art solvers.

Large language model (LLM) agents increasingly operate in open-world environments where safety risks cannot be fully predefined. However, most existing guardrail mechanisms remain static, relying on fixed rules, prompts, or classifiers that generalize poorly to emerging risks. Here we present EVOLVE, a tool-level self-evolving defense framework that formulates agent security as a process of cumulative security evolution. Inspired by how human societies transform experiences into transferable protective knowledge, EVOLVE externalizes security knowledge into executable guardrail tools that are continuously inherited, independently audited, and iteratively refined through lifelong interaction. Across multiple benchmarks, backbone models, and embodied environments, EVOLVE consistently outperforms existing defenses across diverse risk scenarios with +22\% average F1 score. We further validate the framework in real-world robotic and autonomous agent systems. Our results suggest that scalable agent safety may require continuously evolving defense systems rather than static guardrails.

Hyperspectral videos provide rich spectral signatures for robust tracking under challenging conditions such as camouflage and low illumination. However, limited hyperspectral annotations and the large RGB-hyperspectral domain gap hinder the direct use of modern temporally pretrained RGB trackers. We propose SProTrack, a multimodal framework that transfers temporal knowledge from pretrained RGB trackers to the hyperspectral domain through latent spectral prompting. The key idea is to decouple temporal modeling from spectral adaptation, allowing temporal dynamics to be inherited from RGB foundation models while hyperspectral information is learned from unlabeled data in a self-supervised manner. Unlike band-selection methods that discard most spectral information or prompt-based approaches that rely on local-search crops, SProTrack learns a global, scene-level spectral representation from the full hyperspectral frame using an autoencoder. This latent representation is used to guide spectral prompt extraction while a pretrained RGB tracker preserves strong temporal modeling on RGB-compatible inputs. By encoding spectral-spatial structure into compact latent prompts, SProTrack bridges the RGB-hyperspectral modality gap without extensive retraining or dense supervision. Experiments on HOT2020 and HOT2023 demonstrate that SProTrack achieves the best overall performance, with strong results across all spectral domains.

Time-series models are increasingly deployed in operational systems (e.g., inventory management and clinical care) where predictions trigger interventions that alter future observations, producing records we term AI-shaped trajectories. This sequence-prediction coupling deepens as such systems evolve into agents that autonomously predict, act, and remember across long horizons. Such AI-shaped trajectories support immediate performance monitoring, but turn ambiguous once reused as passive data. An algorithm may then misread a prediction-prevented stockout as naturally stable demand, or an averted medical crisis as a false alarm. This ambiguity then propagates into everything built on the record, from retraining and agent memory to downstream predictive tasks. As a result, systems may internalize these intervention effects as natural phenomena, drifting further from the truth at every closed-loop cycle. In this Blue Sky paper, we advocate an agenda for reflexive time-series analysis that shifts the analytical object from the intervention-agnostic sequence to the prediction-action-observation loop. The agenda requires closed-loop provenance to track the operational context (e.g., predictions, policies, and actions), sequence identification to estimate how those actions affect future values under explicit causal assumptions, and closed-loop evaluation to separate intrinsic process prediction from policy-driven effects. This foundation is essential for integrating autonomous agents into such loops at scale, enabling the joint learning of intervention effects and intrinsic dynamics for robust temporal representations.

Robot policies still fail at everyday manipulation: an unfamiliar task, an object never held before, a scene that changes mid-motion, a skill that should improve with use. They fail despite pretraining on over a million trajectories, pooled from dozens of robots at great expense, while a child acquires all of these abilities within a few years of learning from a single body. This survey argues that the gap is one of design, and proposes designs for the next generation of robot models, drawn from how the nervous system solves the same problems. From the body we take compliance: the policy commands stiffness alongside position, so the controller yields where the task allows and presses where it must, which greatly lowers the demand for precise position output. From the visual system, where a fast pathway for action runs in parallel with the slow pathway for recognition, we take a dedicated fast vision system for motor behavior: it precomputes the quantities the policy needs and drives a reactive policy at camera rate. From the cerebellum we take a predictor beside the policy that forecasts the consequence of each command and recalibrates from its own error. From motor cortex we take a single model that composes action units from primitives, verifies each at a sensory boundary, practices the weakest unit with a hypothesis in mind, and reasons each attempt closer to the goal. Proven primitives are then consolidated from code into weights, so they run fluently and respond quickly the next time. Together these describe a robot that acts on its own initiative, the basis of deployment and recursive self-improvement, and we offer each proposal as a hypothesis to be tested.

Language models trained on the outputs of other models can acquire traits that those outputs never mention, such as a preference for an animal carried by lists of numbers. Such subliminal learning can pass on unwanted traits. The intuitive defence is to find and forbid the tokens through which a trait shows, but how a trait's trace behaves under such bans, and what of it a student learns, has not been resolved. We closed a trait's outlets one by one in two stress tests, a preference chain and a name chain, across two models, six output formats and 16 traits. Bans did not remove the trait. Much as squeezing a balloon in one place makes it bulge in another, the trait moved to wherever the format still allowed, an effect we call the balloon effect. Decomposing this transfer, which bans on surface features cannot stop, down to single heads and neurons, we show that it can be cut at the circuit that writes the trait into the output. Patching that circuit while the teacher writes numbers stopped the chains, sharply reduced the overt name leak and halved the trait passed to a student, whereas the same number of random heads produced no detectable reduction. These results give a mechanistic account of how unwanted hidden traits are transferred and a route to preventing it.

Learned sparse attention, used in production models such as DeepSeek-V4 and GLM-5, lets each query attend to only k keys chosen by a lightweight indexer, but the indexer still scores every key in the prefix. At one million tokens, this selection takes more than half of prefill time. Two-stage selection over contiguous blocks, as in HiSA, reduces this cost, but its recall declines as context grows at a fixed candidate budget. We present AttentionIndex, which routes each query over an inverted-file index that groups indexer keys by similarity and refines the candidates with the native indexer score. AttentionIndex builds the index during prefill, making new keys searchable immediately, reusing previous clustering to keep maintenance cheap, and exposing each query only to clusters built from its own prefix. On GLM-5.3 and DeepSeek-V4-Flash at one million tokens, AttentionIndex delivers 1.3-1.8 times the prefill throughput of exhaustive selection, and at throughput within 5% of HiSA, it recovers 17.0 and 10.0 percentage points more of the exhaustive top-k, respectively. AttentionIndex also improves accuracy on long-context benchmarks by up to 18.4 percentage points over HiSA at the same candidate budget.

Quantum-inspired computer vision often reshapes high-dimensional features into tensors and applies tensor-network models. In face verification, this can be misleading: modern embeddings are optimized for cosine similarity, and tensorization alone does not establish the low-rank structure needed for tensor trains or matrix-product operators (MPOs). We study model placement through two controlled interfaces. First, cross-model embedding alignment requires a 512 x 512 operator to map AdaFace embeddings into ArcFace space. Dense Ridge and Procrustes mappings remain strong, while deterministic TT-SVD and MPO compression underperform even with basis changes and low-rank residuals. At a 64K-parameter storage budget, mean directional AUC is 0.9297 for dense Ridge, 0.8952 for rank-64 SVD, 0.8685 for the best low-rank-MPO hybrid, and approximately 0.649 for raw MPO. Second, before global pooling, we construct a 7 x 7 x 7 x 7 probe-gallery correspondence tensor from local AdaFace descriptors. A geometry-aware symmetric tensor train with bond dimension 2 uses 122 trainable parameters, compared with 2,403 for a dense linear matcher, while attaining an LFW AUC of 0.9802. Yet compression does not translate into better low-false-acceptance performance: on CFP-FP, TAR at FAR = 0.001 is 0.0102 versus 0.1734 for global cosine matching, and on CPLFW it is 0.7805 versus 0.8737. These results suggest that quantum-inspired factorization is most credible when the chosen interface provides task-aligned axes and compressible structure. They also show why aggregate AUC must be assessed alongside operational low-FAR metrics.

Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fine-tuning and provide limited insight into the internal mechanisms underlying value-related behaviors. In this work, we investigate multi-value alignment from a mechanistic interpretability perspective by identifying value-sensitive attention heads (value heads) in LLMs. Using a probing-based value localization framework, we discover sparse sets of attention heads associated with diverse dimensions, spanning both human values (e.g., harmlessness, honesty) and stylistic preferences (e.g., coherence, complexity). Analyses across multiple LLM families reveal functional specialization and structured interactions among value heads, including conflicts between certain values. Targeted head ablations further induce substantial, value-specific behavioral changes, providing causal evidence of their functional importance. Building on these findings, we propose an inference-time activation steering strategy that enables controllable alignment with individual values or combinations of values without updating the underlying model parameters, supported by theoretical analysis. Experiments across multiple value alignment benchmarks demonstrate effective single- and multi-value control while largely preserving non-target model performance. Our findings connect mechanistic interpretability with flexible multi-value alignment, offering a foundation for more interpretable and pluralistic LLM behavior. The code is available at https://github.com/sihuo-design/ValueHeads.

Speculative decoding accelerates memory-bound LLM decoding with a lightweight drafter, and parallel drafters extend this speedup by drafting a whole block in one forward pass. However, it remains unclear whether their released systems (stock) attain the full speedup that parallel drafting permits. We identify three challenges. First, with one request, each step accepts few tokens but reads every target weight. Second, with batching, per-request verification work roughly halves the speedup over autoregressive (AR) decoding. Third, speculative decoding is deployment-specific, and no universal optimization fits every serving regime, drafter, target model and GPU. To address these challenges, we present SpecPilot, an autonomous agentic loop that discovers algorithms and systems optimizations for each deployment. In each round, a cost model locates the largest gaps between the current system and its ceiling. Parallel agents explore directions and integrate the successful ones. Token forcing checks every output token against stock to detect reward hacking. A structured memory carries validated mechanisms to new rounds and deployments. SpecPilot discovered a general lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification and three batching mechanisms. We evaluate it across four target models, three drafters, two GPU platforms and five benchmarks. Without training or quality loss, it raises the speedup over AR from stock's 4.46$\times$ to 10.87$\times$ with one request, and from 1.94$\times$ to 3.93$\times$ with batching.

Web services built on large language models (LLMs) receive many paraphrases of earlier queries, which a semantic cache answers from stored responses. A reused answer can be wrong, so wrong answers must stay below a share $δ$ of the queries. Existing caches reuse above a fixed similarity threshold or, like vCache, learn a per-entry rule that bounds each query's error by $δ$. However, fixed thresholds break the budget on new streams, and per-entry rules need labels on every entry before reusing it. We find an unspent error budget, as vCache realizes only 12-55% of $δ$ on three public query streams. Surplus spends this budget where reuse is most likely correct. It ranks queries by a risk score shared across paraphrase clusters and opens risk bins as far as an audited error estimate allows. Free labels from every LLM call and a ledger of realized errors keep this threshold safe under drift. On five streams, including ORCAS search queries, Surplus holds $δ$ in every test cell, which no baseline does. On the shuffled streams it serves 40% more hits than vCache, and its pooled hit rate beats the best baseline holding $δ$ in each cell by 4.10 points. In a vLLM testbed it cuts the mean latency by 31.59% at the highest load, against 23.67% for vCache.

General-purpose robot policies must work across robots, tasks, and environments, yet the variation that makes physical trials informative also makes them difficult to reproduce. We introduce RoboColosseum, a distributed evaluation framework that combines unified policy serving and standardized protocols across a network of locally operated robots in diverse environments. Our initial study evaluates five vision-language-action (VLA) policies and three frontier AI agents across four embodiments (SO-101, Franka, Bimanual-YAM, and G1 Humanoid). The Open Track compares models without task-specific adaptation in blinded, pairwise trials under matched local conditions. The Fine-tune Track measures how well VLA policies adapt to 20 tasks across all four embodiments when each policy receives the same 100 demonstrations per task. Across 5,699 pairwise Open Track comparisons at 124 sites, frontier agents rival specialized VLA policies, while all systems struggle to complete open-world tasks reliably. After fine-tuning, GR00T N1.7 leads on SO-101 and pi0.5 leads on Franka and Bimanual-YAM. Policy rankings show weak agreement across embodiment-specific task suites (Kendall W = 0.24). Open Track rankings poorly predict relative performance after fine-tuning, with substantial rank reversals on all three embodiments shared between tracks. Measuring partial progress reveals differences that task success alone misses. On the G1 Humanoid, no policy completes any of the five tasks, yet average progress ranges from 13.3% to 23.2% across policies. Open Track rankings stabilize as trials accumulate, and active matchmaking achieves comparable ranking agreement with substantially fewer physical trials than uniform sampling. These results show how distributed physical evaluation can reveal where robot policies generalize, how well they adapt to new tasks, and how far they progress on tasks they fail to complete.

Large language models are increasingly used as measurement instruments for socially consequential constructs such as hate speech, and are typically validated by their aggregate agreement with human labels. A valid instrument must also apply the same construct to every group it measures. We test this requirement using 421 \hc{} templates that are instantiated identically for seven protected groups, so that only the identity term varies. Four LLM classifiers each exceed 0.95 balanced accuracy, yet within identical templates two of them miss hateful statements about women at markedly higher rates than the same statements about other groups (excess false-negative rates of 8.3 and 4.4 percentage points), a gap first reported for fine-tuned classifiers in 2021. We show how the construct shifts. Comparing rationales within each template, we find that models invoke protected-group status for other groups but recast the identical statement about women as personal dislike or as aimed at an individual, in 22 of 30 missed templates for one model. Adding the benchmark's own definition, which lists sex as a protected characteristic, does not close the gap, and the contrast persists. Apparent false-positive disparities, by contrast, largely vanish under identity control and trace to reclaimed slurs present only for some groups. Errors shared by all models concentrate where model and benchmark definitions diverge, and pairwise counterfactual accuracy adds little beyond endpoint accuracy. We argue that validating LLM measurement requires identity-controlled tests of measurement invariance that examine both decisions and the criteria models state for them.

Industrial predictive maintenance and manufacturing tasks frequently contain abundant sensor measurements, but with limited labeled data. We investigate cross-domain self-supervised pretraining using joint-embedding predictive architecture for heterogeneous industrial time series (JEPA-HITS), which learns one encoder from various domains such as battery, semiconductor chemical mechanical planarization (CMP), computer numerical control (CNC) machining, and bearing while preserving dataset-specific temporal scales. We evaluate representation reuse on regression and classification problems under reduced-label downstream training, and leave-one-domain-out (LODO) pretraining. Compared with the same architecture initialized randomly, LODO pretraining reduces RMSE from 10.72 to 7.95 on battery remaining useful life (RUL), 4.34 to 3.89 on CMP, and 13.53 to 11.72 on tool wear regression, while improving bearing fault diagnosis macro-F1 from 0.547 to 0.737. Overall results suggest that unlabeled time series measurements from heterogeneous industrial processes can provide reusable representations for downstream learning with limited labeled data.

Improving recommendation algorithms for the Web requires coordinating interacting design choices and learning from experiments. We present RecRSI, an LLM-based research framework that uses mechanism hypotheses to organize algorithm discovery. Each hypothesis links a proposed mechanism change to a reference design, a rationale, and an expected observation. MechSpace derives implementable primitives and composition rules from the task source; Research Line uses this space and prior evidence to develop and test hypotheses; Mechanism Feedback carries outcomes and their design context into subsequent research. This organization makes each experiment inform which mechanisms to develop, revise, or combine next. Across four initial recommenders and four datasets, RecRSI yields average relative improvements of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms, and 2.60% and 2.20%, respectively, over the strongest competing result for each setting and metric. Component and process studies link these gains to coordinated mechanism choices and evidence-guided follow-up. Our code is publicly available at https://github.com/DynamisG210/RecRSI.

Online Active Learning (OAL) aims to extend deep learning from static inference after offline training to continuous adaptation during real-time interaction with a changing physical environment. This transition is essential for safety-critical applications, but it raises a central theoretical question: whether the online parameter update process remains stable when scaled to multilayer neural architectures. This paper develops a stability-focused OAL framework. The framework uses prediction error as a feedback signal to update a structured deep model, thereby linking online learning with Lyapunov-based energy minimization. We show that prediction error convergence alone is insufficient for global parameter convergence because multilayer networks may admit high-dimensional manifolds of equivalent output solutions. The analysis therefore separates output error regulation, internal zero dynamics, interlayer coupling, and parameter convergence. Persistency of excitation and active damping are then used to remove the residual null-space modes that remain after output error convergence. A cascaded Lyapunov analysis establishes asymptotic stability of the overall OAL learning process, while monotonic activation functions and diffeomorphic layer mappings provide the uniqueness condition for the target parameter trajectory. The resulting proof provides a theoretical basis for applying online adaptive learning to large-scale neural systems in nonstationary and safety-critical settings.

LLM agents are increasingly built into data-analysis software, yet little evidence shows how reliably they preprocess clinical data or whether the tools they invoke give correct, reproducible output. We report a benchmark of THReD, a web-based system that turns electronic health record (EHR) extracts into AI-ready datasets through a chat agent whose proposed steps run only after user confirmation. The agent benchmark has 141 tasks (100 templates) in seven categories, including goal-level requests, failure recovery, requests calling for no action, schema variation and prompt injection through the data, on synthetic tables calibrated to a regional health information exchange; success is judged by checks on the produced data. Over 2,550 episodes with Claude Haiku 5.5 in four context configurations, the deployed agent completed 92.7% (95% CI 88.1-96.4) of tasks on its first answer, abstained on every advice request, recovered from every injected fault, and never proposed an action requested by planted text. Without the data profile, completion fell to 58.0%; goal-level requests were hardest (80.6%). Six experiments on real records of two chronic-disease cohorts and on synthetic data at scale validate the tools: 91-92% of inferred patient classes and 99.7-99.8% of inferred vital-sign labels were correct in masked reconstruction; a dictionary builder combining curated dictionaries with LLM suggestions mapped 99.0% of 2,208 raw variants correctly; all 2,176 replays of 128 saved pipelines gave byte-identical files under pinned dependencies; subgroup differences in inference accuracy vanished within facility and lookup key; and the same-day consolidation rule changed 90-day emergency revisit rates by up to 3.2 points without shifting cohort composition by more than one point. Processing time grew roughly linearly to 5.5 million records while the prompt stayed at about 13 thousand tokens.

Long-term conversational memory faces a timing mismatch: write-time summaries and graphs organize history before a future query is known, while independent retrieval can allocate a limited evidence budget to redundant or fragmented turns. We introduce Chain on Demand (CoD), a query-time evidence organization framework over an unchanged turn-level memory. Given a broad candidate pool, CoD constructs a temporary evidence chain by allocating a fixed anchor budget according to query relevance, lexical diversity relative to selected anchors, and newly covered query terms, followed by deterministic restoration of each selected anchor's same-session neighborhood. The resulting chain is used only for the current query, while the underlying conversational memory remains unchanged. In a paired comparison over all 1,986 LoCoMo questions, CoD achieves 82.2% accuracy on 1,540 non-adversarial questions and 76.3% on all 1,986 questions. On LongMemEval-S, one fixed configuration that indexes both user and assistant turns obtains 86.6% across all 500 questions. These results show that long-term memory can preserve the original conversation while deferring evidence organization until the information need is known.

Seismic data interpolation must reconcile reconstruction fidelity with the computational budgets of acquisition quality control and subsequent imaging. Fixed-depth unfolding networks are optimized for a single operating point, whereas their intermediate estimates may be insufficiently trained for early deployment. We present MetaPOCS, a meta-learned multi-exit unfolding framework for budget-aware seismic interpolation. Building on a spatial--spectral POCS reconstruction backbone, MetaPOCS supervises intermediate stages and preserves the acquired traces at every exit. A Meta-GF-based optimization procedure constructs exit-oriented parameter displacements from a common initialization and learns their fusion weights through an equal-exit outer objective. The weights vary across parameter tensors and respect the nested support of the reconstruction stages. We derive the resulting fusion hypergradient to characterize how the outer objective reallocates competing updates, and distinguish this mechanism from scalar loss reweighting. At inference, one checkpoint provides a set of reconstruction budgets through a latency-calibrated exit policy. The evaluation framework couples matched-cost interpolation comparisons with finite-update diagnostics and a fixed reverse-time-migration operator, connecting multi-exit optimization to waveform and imaging fidelity. MetaPOCS thus provides a unified formulation for studying the reconstruction--computation trade-off of seismic unfolding across deployment budgets.

Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.

A GUI command can be correct at the interface level yet still produce an incorrect assembly. An agent must identify the relevant components, respect their structural relationships, and determine which parts should move during recovery. We introduce MAGAR-Bench, a benchmark of 4,000 multimodal samples from 26 assembly families that separates Assembly Decision, Atomic Action, and Minimal Rework. For rework, the objective counts distinct moved parts within the validated recovery candidate set for each state, rather than the number of actions. We also introduce MAGAR, a selective correction framework built on a frozen policy obtained by supervised fine-tuning. Its key design matches correction granularity to output dependencies. Structured candidates revise assembly fields and rework scope, while complete action candidates keep native GUI primitives coupled with their parameters. A learned selector retains the original prediction unless a candidate passes a task-specific intervention threshold. On recorded states from held-out assembly families, MAGAR improves Overall, the task-balanced aggregate, from 37.50 to 41.39, with gains on all three tasks. Across 209 interventions, 70 correct an error and 22 introduce one. These results support task-specific correction while highlighting the remaining tradeoff between fixing errors and preserving reliable predictions.

Semantic-ID-based (SID) generative recommendation (GR) represents each item as a sequence of discrete tokens and predicts the next item through autoregressive SID generation. However, existing methods face two challenges. First, insufficient shared context modeling for user sequences: full-history and recent behaviors provide overlapping yet complementary evidence, but modeling them within a single sequence may dilute current cues, whereas processing them through separate streams may fail to coordinate them. Second, insufficient personalized adaptation in progressive SID decisions: although decoder states evolve with generated prefixes, existing models typically use a shared output-side transformation across recommendation instances, without explicitly selecting a transformation suited to the current user need. To address these challenges, we propose Multi-View Thinking with Adaptive Generation (MV-Thinker), a framework that Thinks Globally and Generates Personally. First, Multi-View Temporal Thinking uses view-specific thinking queries and readers over overlapping full-history and recent-history views, coordinating both sources into complementary user-conditioned contexts before SID generation. Second, Hierarchical Residual Adaptation organizes lightweight adapters into nested residual paths before a shared vocabulary projection, adaptively transforming context-bearing decoder states for SID decisions. Extensive experiments on four public datasets demonstrate that MV-Thinker consistently outperforms strong discriminative and generative recommendation baselines across multiple ranking metrics. The implementation code and datasets are available on: https://anonymous.4open.science/r/MV-Thinker

Numerical labels often specify a set of possible values: a measurement is rounded, a laboratory test reports only a bound, or a program is stopped at a timeout. We introduce boundary marginalization, the first method that trains an autoregressive language model on the exact likelihood of such observations through its existing digit tokens. A fixed-width, order-preserving numeral representation groups consecutive grid values under shared digit prefixes. An interval is a union of such groups, so its exact probability and gradient follow from at most two boundary paths, the successive digit prefixes of its endpoints, each of length logarithmic in the number of grid values. The resulting objective unifies precise, rounded, bounded, and censored supervision and trains digit decisions at the resolution supplied by each observation. The same computation supports exact distributional queries and unbiased gradients for the continuous ranked probability score on exact labels. Experiments show recovery of a conditional distribution from rounded and censored reports alone, improved likelihood on naturally censored molecular measurements, and more accurate code latency distributions on two backbones. Exact interval updates remain tractable as the numerical support grows to one million values.

The ability to emulate complicated and computationally-expensive physical systems in an efficient and stable manner is a popular motivation for scientific machine learning (SciML). Stable numerical solvers tend to be computationally expensive, as they attempt to resolve small (spatial and temporal) scales that receive and dissipate spectral energy. Learned coarse-grid forecasters remove the dissipative channel, allowing unresolved frequencies to re-emerge as grid-scale noise (aliasing). In this paper, we use the Neural Discrete Equilibrium (NeurDE) method to explore stable emulation at extremely coarse grids, demonstrating physically-principled control over entropy dynamics that reinforces the physical system's dissipative mechanism. Entropically-stabilized NeurDE produces stable long-time rollouts for thousands of forecasting steps on large Reynolds number turbulent flows ($\mathrm{Re}=100{,}000$) with very coarse grids (down to $\mathbb{R}^{16\times16}$). At these very coarse grid scales, entropically-stabilized NeurDE can be more stable than even our stabilized data-generation numerical method.

Reasoning-augmented recommender systems (RecSys) enhance user preference modeling by generating reasoning trajectories over historical interactions before making recommendations. Knowledge distillation provides a practical paradigm for transferring such reasoning capabilities from powerful large language models to lightweight RecSys. However, existing methods typically rely on a single fixed teacher, remain constrained by the teacher knowledge ceiling, and indiscriminately distill teacher-generated trajectories without assessing their actual utility for downstream recommendation. To address these limitations, we propose MUSE (Multi-Teacher Utility-guided Selection with Self-Exploration), an adaptive reasoning distillation framework that jointly improves reasoning generation and reasoning utilization. MUSE first constructs a diverse reasoning pool using multiple teacher LLMs and progressively augments it via student self-exploration, enabling the student to leverage the strengths of different teachers and gradually explore reasoning strategies beyond teacher imitation. We then develop Rank-Gain Utility Assessment (RGUA), which evaluates each candidate according to the normalized full-catalog rank improvement of the ground-truth item and identifies the most and least recommendation-beneficial trajectories. Based on these utility-aware trajectory pairs, we introduce dual-level contrastive learning objectives that guide the student generator to prefer beneficial reasoning while training the recommendation encoder to more effectively exploit recommendation-relevant preference evidence. Extensive experiments on three real-world datasets demonstrate the effectiveness and robustness of the proposed framework.

Reinforcement learning (RL) has improved the capabilities of large language models (LLMs) across a wide range of tasks. However, token-level RL assigns an action's advantage to the entire generated response, including tokens specifying the parsed action, reasoning chains, and malformed content. Although the environment executes only the parsed action, variation in how the action is expressed in natural language can introduce additional variance into policy-gradient estimates. We propose RLiD (Reinforcement Learning in the Decision plane), a framework that directly optimizes the probabilities of executable action options. RLiD compiles observations, an optional compact memory, and a labeled action menu into the model input, and uses a large decision model (LDM) to select an action in a single forward pass. By eliminating autoregressive token generation and bypassing action parsing, RLiD avoids invalid action outputs and improves RL rollout efficiency. Our analysis shows that, within a fixed text policy, marginalizing over alternative realizations of the same action removes a zero-mean gradient component without changing the expected update. We further distill trained LDM policies into generative LLMs through supervision on their action trajectories. Experiments on ARC, Countdown, and Sokoban demonstrate improved performance or lower computational cost across several settings. On ALFWorld, the complete training and distillation pipeline reaches 99.3% success in 12.7 H100 GPU-hours, compared with 86.6% in 17.3 GPU-hours for token-level RL of the same LLM. With compiled memory, our strongest Qwen2.5-1.5B LDM and distilled generative LLM achieve 100% success on all 134 unseen and 140 seen ALFWorld evaluation games.

Cohort memory can support LLM-based recommendation planning through group-level user profiles and advice. As outcome feedback becomes available, this guidance may need to be updated while preserving the underlying observations. We present RecGroupMemory, a workflow that provides this memory to an orchestration LLM and separates memory construction from feedback-based revision. It uses feedback from the previous day to revise decision guidance, while validation checks keep selected source values unchanged. Online comparisons were conducted among Taobao users with low purchasing power selected by the business team. Relative to a no-memory baseline, memory without feedback and memory with feedback showed increases of 1.58% and 1.78%, respectively, in the number of guided transaction users.

Expectation-maximization (EM) for Gaussian mixture models (GMMs) requires scoring $N$ data points against $K$ mixture components. Prior fused GPU implementations, notably Flash-GMM, eliminate materialization of the $N \times K$ responsibility matrix but traverse the component dimension twice to compute normalization terms and accumulate sufficient statistics. We introduce Turbo-GMM, a single-sweep fused EM kernel that completes responsibility normalization and sufficient-statistic accumulation within one traversal of the component dimension. By retaining the complete component axis on chip, Turbo-GMM avoids the second score computation required by the released two-sweep Flash-GMM implementation. Its applicability is limited by register pressure, occupancy, and the component and feature dimensions, making it particularly suitable for large datasets with relatively small numbers of components. For larger configurations, Turbo-GMM employs a two-pass fallback following the established fused design. We establish algebraic equivalence to standard EM in real arithmetic and evaluate sixteen configurations on an NVIDIA A100. Under the reported benchmarking protocol, Turbo-GMM achieves a median $2.15\times$ speedup over the evaluated Flash-GMM implementation and up to $5.3\times$ and $9.9\times$ over PyTorch and \texttt{scikit-learn}, respectively. The results demonstrate the potential of single-sweep component traversal while identifying on-chip resource constraints as the principal limitation of this approach.

Development teams need reproducible tasks from their own repositories and workflows to evaluate coding agents. Conversational developer-agent collaboration ("vibe coding") produces abundant coding sessions and repository checkpoints, but requirements evolve across turns, unrelated conversations co-occur with commits, and engineering efforts can span multiple sessions. We present SessionFarmer, an automated framework that turns these records into self-contained, executable software engineering tasks. It analyzes dialogue and commit history to filter unrelated sessions and group multi-turn, potentially multi-session interactions into work episodes centered on one engineering goal. For each episode, it reconstructs a hermetic execution container, synthesizes a task statement from the full context, and authors repository-grounded engineering-quality rubrics and unit tests. Closed-loop test refinement uses requirement-breaking mutants and a blind solver to reduce false rejections of valid alternatives and false acceptances of incomplete solutions. We construct CropChat, an open benchmark of 242 tasks across 61 public repositories and 14 programming languages, and CropEnterprise from proprietary enterprise logs. In a CropEnterprise audit, the original developers endorse the vast majority of reconstructed episode boundaries, task statements, and quality rubrics. Full-episode context covers substantially more user requirements than single-turn or single-session inputs. Test refinement cuts false-rejection rates and raises balanced accuracy against reference patch labels by 7-9 percentage points across solvers. Across 12 model-harness configurations on CropChat, the strongest setup achieves 39.3% Pass@1, while harness choice shifts Pass@1 by up to 8.4 percentage points for the same model.

LLM planning often relies on reactive executors (e.g., ReAct) that generate a single action at each step to remain grounded in the latest observations, incurring substantial inference cost. Batched executors reduce this cost by predicting multiple actions in a single model invocation, yet they are prone to compounding errors, as intermediate observations are ignored. We propose an expectation-observation consistency verification (EOCV) mechanism for batched executors, preserving grounding while improving efficiency. The mechanism generates expectations of action outcomes and compares them with the resulting observations. Consistency permits continued execution, whereas mismatches trigger early termination and replanning from the current observation. Experiments on text-based and multimodal benchmarks across multiple backbones show broadly comparable average task performance between EOCV and reactive executors, with a 29-35% reduction in average weighted token usage. Moreover, the mechanism is model-agnostic and remains effective when integrated into state-of-the-art planners.

Skin cancer classifiers need large sets of annotated images, which are scarce, especially for rare lesion types. AI-generated images can fill the gap, but a generator produces images of widely varying quality, so selecting which images to retain is important. That choice usually rests on a single quality score, most often the Fréchet Inception Distance (FID), which has not been tested against what a classifier needs and cannot tell whether the generator memorizes its training photographs. We test that practice with a tournament of 16 quality metrics. Each metric alone selects a training subset from a pool of 275,472 images generated by a diffusion model (EDM2), and a classifier trained on each subset scores the metric per class. The tournament shows that no single metric serves every class: FID computed on Inception features gives the best classifier for one lesion type and no measurable advantage over random sampling for another, and the generation setting it prefers is also the one closest to the training images. We turn these per-class measurements into a safety-aware evaluation algorithm that weights each metric by its measured value in each class and removes images too close to a real training image. On the held-out ISIC-2018 test set, the algorithm improves Macro-$F_2$, a recall-weighted measure of classification performance, over the FID baseline for a Swin-B classifier by +0.0136 to +0.0164 across four safety thresholds (Bonferroni-corrected $p < 0.02$) and over random sampling by about +0.03, though not for ResNet-50. Tightening the threshold improves a memorization-safety score with no measurable change in utility, but every curated set remains less safe than random sampling. The results show a modest but statistically significant gain on the primary classifier, and offer a practical, safety-aware procedure for evaluating and using synthetic data.

Residual connections give each Transformer block access to all earlier layers only through one fixed sum, broadcast to all of the block's inputs. This broadcast bottleneck imposes three constraints: the query, key, value and residual inputs of a block, which we call its roles, read the same vector; all hidden dimensions within a role share one set of weights over earlier layers; and every earlier layer has the same fixed weight for every token. We propose the Multi-Role Multi-Group Mixing Residual (M3), which removes all three: every role, and every group of hidden dimensions within a role, reads its own token-dependent mixture of all earlier layers. M3 starts as an exact residual connection and adds as little as 1.4% parameters and 1.8% FLOPs. From 100M to 1.4B parameters, M3Former outperforms the Transformer and coarse-grained cross-layer connections at every size, matching a Transformer 1.19-1.34x the size of the one it modifies, trained with 1.4-1.8x the compute. Relaxing only some of the constraints recovers only part of this gain. M3Former also improves sparse mixture-of-experts models of up to 4.6B parameters and trains more stably in deeper networks. Code: https://github.com/Kingchet10/m3former

Language models with similar benchmark accuracy can succeed on different items. Overall scores do not show where their performance differs or which item properties might explain those differences. We demonstrate how the latent space item response model (LSIRM) can help evaluators investigate these patterns and their consequences for model comparison. We analyze 1,348 language models on 5,761 BBH items and compare ways of presenting and scoring answers across BBH, GPQA, MMLU-Pro, and MuSR. We examine LSIRM maps, which summarize differences in performance across models and items, alongside item properties and records of correct and incorrect responses. In BBH, nearby items on these maps often share the same correct answer, such as "Yes" or "No." Across the four benchmarks, this pattern is stronger when answers are scored using option labels, such as A, B, C, and D, than when using the text of the options. We then estimate each model's tendency to perform better or worse on items with particular correct answers and quantify how this tendency contributes to its score. This separates an estimated contribution associated with correct answers from the remaining differences in item performance. Two models' total numbers of correct answers differed by only 14 out of 5,761 items. Excluding the estimated answer-category contribution from their scores moved their ranks hundreds of places in opposite directions, while keeping all other estimates unchanged. The approach helps evaluators identify differences hidden by overall accuracy and understand how the choice of what to include in a score affects model rankings.

We study online channel scheduling for a single queue served by $K$ channels under bandit feedback. We propose a simple best-of-both-worlds policy that explores when the queue is empty. During busy periods, it selects the empirical best channel while the prescribed conditions hold and, if any condition is violated, switches to a robust bandit policy to clear the queue. For arbitrary arrivals and services determined by an oblivious adversary, our policy achieves $\tilde{\mathcal{O}}(\sqrt{K L_T})$ worst-case queue length regret with high probability, where $L_T$ is the longest busy-period length. This result improves the previous $T^{3/4}$ dependence and matches the worst-case lower bound up to logarithmic factors. Under bounded i.i.d. stochastic models with a unique best stabilizing channel, we show that the cumulative queue length regret converges to a finite problem-dependent constant, and further establish the logarithmic expected worst-case queue length regret bound.

Most benchmarks for real-time voice systems focus on one aspect, such as turn-taking, instruction following, paralinguistic control, or conversational memory, each with its own data and set of systems. We present InteractionBench, a benchmark and evaluation harness that measures four dimensions on the same systems from the same recorded user audio: conversational timing, spoken-task accuracy, whisper and volume, and groundedness in conversations of more than twenty turns. The benchmark contains 221 items per system in thirteen scenarios. Before any system runs, we fix the expected floor-control actions for spoken tasks by rule and take the expected answers from scripted dialogues. We score whispering from the audio signal. We evaluate GPT-Live, GPT-Realtime, Gemini Live, Moshi, PersonaPlex, Freeze-Omni, and a turn-based cascade in 1,519 graded runs. Among the six full-duplex systems, timing and spoken-task accuracy are rank-correlated ($ρ= 0.83$), while the cascade has the highest spoken-task score (38/50) and the second-lowest timing score. No system corrects a user's grammar within the required window. Only Gemini Live whispers on most requests (8/10). Hosted systems pass 67-88% of groundedness probes, compared with 14-35% for open full-duplex models. Timing is measured where the client hears each system, so the timing of hosted systems includes network delay: for GPT-Live, whose server reports its own clock, removing a delay of about 1.5 s lowers its timing score from 86% to 56%. We also document where the protocol differs across systems, including the text instructions.

Many overparameterized models that have achieved remarkable success are singular models. That is, the set of optimal parameters consists of multiple points, or the Hessian of the generalization error at such parameters is degenerate, reflecting redundancy in the model parametrization. Therefore, understanding the effective dimension and intrinsic model complexity, beyond the apparent dimension of the parameter space, is essential for studying the generalization ability of such overparameterized models. In this work, we focus on the real log canonical threshold (RLCT) and its localized counterpart, the local RLCT, which are fundamental quantities for characterizing the generalization of singular models. We clarify the theoretical relationship between the local RLCT and existing two key metrics related to generalization under some localization assumption: (I) the rank of the Hessian matrix of the generalization error and (II) the covering number of a family of loss functions. This result provides novel insights into effective dimension and intrinsic complexity of overparameterized models.

Geometry-aware neural operators face substantial computational and memory demands when learning across complex physical domains. We introduce Architecture-Aware Mixed Precision (AA-MP), a replica-exchange training strategy that allocates numerical precision according to replica roles and architecture-specific operations. By retaining FP32 for refinement while selectively employing BF16 for exploration, AA-MP preserves predictive accuracy while improving training efficiency. Across fluid dynamics and three-dimensional structural mechanics benchmarks, AA-MP achieves comparable accuracy to full-precision training over an ensemble of runs, while reducing hot-replica peak GPU memory by 46.9% and training time by up to 25.8%. It also outperforms naive Automatic Mixed Precision (AMP) in memory efficiency and computational cost, with increasing memory advantages at larger model and geometry scales. Ablation studies show that retaining FP32 in the cold-replica is important for predictive accuracy, while additional FP32 operations in the exploratory replica provide limited benefits relative to their computational cost. Finally, conformal prediction for uncertainty calibration shows that AA-MP achieves near-nominal coverage with narrower prediction intervals than the alternatives. These results advance the scalability of geometry-aware scientific machine learning by reducing computational costs without compromising predictive accuracy or uncertainty quantification.

Root cause analysis (RCA) in microservice systems requires distinguishing failure sources from routine fluctuations and propagated symptoms. Small language models (SLMs) offer a practical solution under cost and privacy constraints, and recent approaches improve their capabilities through structured diagnostic procedures. However, performing diagnostic analysis does not ensure correct root-cause identification. Our empirical study shows that final decisions often favor large anomaly scores despite the model's own propagation analysis suggesting otherwise. Moreover, the input does not reveal whether these apparently anomalous signals are common during normal operation. These findings motivate our system, which improves how SLMs assess evidence and select root causes through Historical Frequency Profiling and Biased-Decision Correction. Historical Frequency Profiling quantifies how often metric and log signals are flagged as anomalous in the fault-free period, providing context for distinguishing routine deviations from potential failure evidence. Biased-Decision Correction identifies points where the model reaches different conclusions from the same initial analysis and uses preference optimization to favor more accurate diagnoses. Against baselines including the 14B OpsAgent and RCA-Agent with commercial models, our system with a 4B model achieves the highest or tied-highest accuracy on all four datasets, reaching 48.7% on Bank compared with 18.5% for OpsAgent.

Reward-robust RL typically models misspecification by specifying an uncertainty set that constrains the discrepancy of plausible rewards from a base reward. However, such reward space discrepancies can be behaviorally irrelevant: under potential-based reward shaping (PBRS), many distinct rewards preserve policy ordering. This can introduce substantial redundancy into standard uncertainty sets and degrade optimization performance. We propose Shaping-Aware Reward-Robust RL, which constructs uncertainty sets over PBRS equivalence classes by projecting each reward to a canonical representative, ensuring that the resulting set contains only rewards that induce behaviorally distinct policy rankings. We prove that this projection preserves the optimal robust value while shrinking the uncertainty set and improving empirical performance. Using the connection between robustness and regularization, we obtain a practical algorithm to solve the shaping-aware reward-robust RL problem and enjoy convergence guarantees under standard assumptions. Experiments on benchmarks spanning diverse task domains and levels of complexity show consistent improvements over representative robust RL baselines and exhibit improved robustness to reward perturbations.

OPSD provides dense supervision for agentic reinforcement learning through privileged information feedback. However, existing privileged information often lacks explicit task-progression planning, while teacher-student log-probability gaps reflect shifts in model preference rather than true task progress. To address these limitations, we propose Subgoal-Guided Graph-Regularized OPSD (SGR-OPSD), a framework that combines subgoal supervision with interaction-derived evidence for turn-level credit assignment. Specifically, SGR-OPSD uses LLM-generated subgoal plans to provide dense action-level guidance, converting them into advantage weights via a constrained, KL-regularized objective. To ground credit in observed outcomes, we construct a Subgoal Progress Graph that estimates empirical returns from sampled transition frequencies. The graph compares empirical action returns against the state average to derive preferences that calibrate the teacher's credit weights. Evaluated on ALFWorld, WebShop, and Search-QA across two model scales (Qwen2.5-3B and 7B), SGR-OPSD consistently outperforms GRPO and strong self-distillation baselines, improving success rates over GRPO by up to 7.4 percentage points and reaching 85.8% on ALFWorld with Qwen2.5-7B. Ablation studies confirm that subgoal guidance and graph-based weight calibration are key to these performance gains.

MRI morphometry pipelines are often evaluated by measurement agreement or downstream accuracy, yet their scientific value depends on the conclusions they support. Two pipelines can produce similar feature-level results yet support different conclusions about aging, cognitive decline, or neurodegeneration. We introduce Scientific Claim Stability (SCS), a claim-centric evaluation framework that converts region-of-interest (ROI) morphometry tables into auditable claim cards and evaluates their support under changes in dataset, processing pipeline, atlas, feature family, and target. Across 10,370 MRI scans from three datasets and five cohorts, SCS generates 2,783 claim cards and evaluates 89 claim-family pairs. The evaluation distinguishes stable, transferable, and fragile claims. FreeSurfer 8-FastSurfer comparisons show high evidence concordance and low anatomical-support overlap under the primary covariate specification. Acquisition-adjusted analyses show that the magnitude of support disagreement depends on the model specification. Structured review by specialists in geriatrics, neuroscience, and radiology, alongside auxiliary ChatGPT and Gemini reviewers, assesses claim plausibility, clarity, faithfulness, usefulness, and caution. SCS provides a reproducible way to audit the evidential stability of morphometry claims without assuming a universal gold standard.

Federated Learning (FL) offers a paradigm for collaborative machine learning model training among clients, preserving data privacy by avoiding direct data sharing. However, the gradient information exchanged in FL has been shown to be vulnerable to Gradient Inversion Attacks (GIA), which can reconstruct sensitive user data. To counter these threats, various privacy-enhancing technologies like Secure Multi-party Computing (SMC), Homomorphic Encryption (HE), and Differential Privacy (DP) have been integrated into FL. While effective, these methods often introduce significant trade-offs between privacy and model utility. By revisiting the key to privacy exposure in FL under GIA, which lies in the frequent sharing of model gradients that contain private data, we take a new perspective by designing a novel privacy-preserving FL framework that effectively ``breaks the direct connection'' between the shared parameters and the local private data to defend against GIA. Our proposed framework, HyperFL++, leverages hypernetworks to generate local model parameters, where only the hypernetwork parameters are aggregated by the server. To ensure practicality, HyperFL++ incorporates a Row-wise Kronecker Decomposition (RKD) to mitigate the memory and communication overhead from the hypernetwork. Comprehensive theoretical analysis and extensive experiments validate our method's effectiveness in protecting data privacy while maintaining high model utility. Our code is available at https://github.com/Pengxin-Guo/HyperFL/tree/hyperfl%2B%2B.

Local music applications estimate key during library import and benefit from an analyzer that is small, fast, and runs in-process. We start from a compact chroma-template key analyzer that fits this budget but lags behind learned references. Learned systems reach higher accuracy, but their installed model and runtime dependencies can occupy tens to hundreds of megabytes. We present key lime, a 21.5k-parameter DSP-guided transformer reranker with a 115 KB model artifact and a 76.7 MB installed footprint, including the extracted ONNX Runtime. It adds 0.32 ms per track to deterministic DSP scoring. Trained from labels on 4,040 FMAK tracks, with 449 tracks for validation and 1,000 disjoint FMAKv2 tracks held out for evaluation, it improves MIREX from 61.08 to 73.55 +/- 0.46 and exact accuracy from 52.30 to 67.07 +/- 0.65 over three seeds (mean +/- population SD). It approaches a same-manifest local madmom KeyCNN reference at 74.16 MIREX. A 12-parameter circular input stage matches the accuracy range of a 468-parameter linear input projection.

Several genuine AI services may repeat one unsupported source, and correcting that source may leave later reports unchanged. We examine these networked consequences of synthetic social engineering (SSE), deliberately misleading influence intended to redirect AI recipients from authorized objectives. Our framework connects social influence, information aggregation, and misinformation correction research to observable evidence and permission boundaries. It separates originating attempts from legitimate intermediaries' work and distinguishes identity, evidentiary support, and authorization. We test four relationships using Qwen3-4B and two briefing tasks across 1,080 research units in our Agent Service Exchange and Evidence Commons (ASE-EC). Multiple dependent publishers produced no supported increase in unsupported verification (78.3% versus 75.0%); publication availability left downstream violation reach at 40.0% in both arms. These comparisons direct attention to weaknesses already present in the reference workflows. High reference error rates and retained private handoffs preclude claims of immunity to social pressure. Without scope revalidation, corrections identifying dependent reports reduced residual invalid decisions from 14.4% to 8.1% at equal notification coverage and increased successful report repairs from 16/63 to 36/63. Scope revalidation added no demonstrated protection, and clearing dialogue offered no established recovery advantage. Unapproved requests persisted, all blocked by an external gateway. For designers, the findings caution against treating publisher count as independent evidence, notification as repair, or cleared histories as recovery. They support targeting corrections at reused reports, assessing factual and permission recovery separately, and enforcing authorization at execution.

Rewards for diversity, coverage and distribution matching make the reward of each sampled response depend on the rest of its group, and they are now routinely optimized with GRPO and RLOO. We prove that in this setting the group-relative baseline changes the objective. The expected update equals each response's own reward gradient plus $α$ times its effect on its groupmates' rewards, with $α=-1/(K-1)$ for groups of size $K$, so GRPO trains every response to beat its groupmates in a zero-sum tournament. It cannot see a pairwise diversity bonus when K=2, sees less than half of it at any K, and often does worse than using no baseline at all. Setting $α=+1$ gives Team-GRPO, an unbiased estimator for any group-dependent reward that works with two samples per prompt and requires no extra model calls for pairwise rewards. The theory predicts training quantitatively. It locates the fixed points of a 9B policy to within 0.02 in $\ell_1$ and predicts how concentrated trained policies become with Pearson r=0.995. On GlobalOpinionQA, GRPO collapses onto the majority answer and ends further from human opinion than the untrained model, while Team-GRPO cuts the error from $0.20$ to $0.08$. On NoveltyBench, Team-GRPO avoids the diversity collapse of GRPO and significantly improves set utility.

In the LLM-as-formalizer paradigm, a large language model translates a problem into a formal program and leaves the search to a solver, gaining the solver's guarantees. Formalizers are now often reasoning models that write a long reasoning trace before the program, yet they are judged only by the program's output. We find that the trace routinely does the solver's job, assigning, testing and ruling out values until it often reaches the full answer; we call this solver-like behavior (SLB). We define an ontology of SLB over the trace, the program and its output, and use LLM judges, checked against human labels, to label 2,000 generations from three reasoning models on four constraint satisfaction datasets, each formalized in a declarative and an imperative formal language. SLB appears in 85% of traces and takes 24% of reasoning tokens, yet the program usually recomputes the answer, so the effort is redundant. The 9% of programs that inherit the trace's answer also inherit its errors, and pass output-based evaluation whenever that answer is correct. The amount of SLB is not associated with accuracy, and SLB is more common in imperative Python than in declarative Z3 or PDDL. Cautionary prompts reduce it by only a quarter, but hiding the problem's values behind placeholders, which the program receives only at run time, removes 97% of it without loss of accuracy. Formalizers should therefore be evaluated by how their programs are reached, not only by what they output.

No algorithm is best on all problems, but a real warehouse is a narrow problem: its layout and processes rarely change. Operators still run one general multi-robot planner at many sites, because experts cannot tune it for each one. In this work, which we call Late Autumn, we explore whether LLM coding agents running autonomous research can specialize a general planner for one site, trading generality for throughput, much as post-training adapts a pretrained LLM to one task. We study six campaigns on a lifelong planner that combines task assignment with multi-agent path finding, on a floor modeled on an existing automated warehouse. We find that (1) the agents raise throughput by 22--60%, and the gains hold on unseen seeds; (2) they rediscover known MAPF mechanisms, such as corridor-exit priorities, local guidance, and rollouts, and independent campaigns find the same bottleneck; (3) the gains largely carry over to another layout, a larger fleet, and other service times, shrink with a smaller fleet, and mostly vanish under an outbound-heavy demand mix; and (4) each final patch (0.7k--2.4k lines) is a chain of scored commits with a median step of 27 lines, so experts can review it, and review finds problems that the score misses; some patches multiply planner compute by over 100x. LLM agents can therefore turn each site's bias into extra throughput at little engineering cost, with reviewable patches; we expect experts to shift toward designing insightful metrics and fast simulators that mirror the real problem.

In physical backpropagation, users return loss derivatives over the time-division duplex uplink, giving the base station the downlink gradient without channel reconstruction. Reciprocity calibration conventionally prioritizes base-station chains, as column-normalized zero-forcing reduces a per-user factor to a stream rotation. This priority can reverse. With ideal base-station chains and no noise, per-user amplitude errors make the returned field a weighted sum-rate gradient. With zero-forcing full service at both optima, the loss is independent of array size. Otherwise the selection penalty, serving the wrong users, carries 86 to 94 percent of the amplitude-induced loss in five of six configurations. Per-user phase errors mainly slow the iteration. Base-station chain errors scale the returned field per antenna, and at 3 dB and 25 degrees they cost one user 1.3 dB of signal-to-noise ratio (SNR) that no array size removes. With 80 users, such errors on both sides, and mismatched pilots, oracle per-user and base-station calibration leave 6.2 and 12.3 percent loss at 256 antennas on clustered channels. This ordering holds at 30 dB SNR and reverses for 1 dB, 10 degree user errors. A capped one-echo estimator comes within 2.4 percentage points of oracle per-user calibration but underperforms no calibration at 10 dB SNR.

Retrieval-augmented generation (RAG) extends large language models with external evidence, yet relevant evidence does not ensure that a generator can use it. Indiscriminate retrieval can degrade task performance while adding retrieval and long-context inference costs. Existing methods neither determine before large-scale execution whether a generator, task distribution, and retriever configuration is worth deploying nor diagnose multiple candidate retrievers under a shared budget. We introduce EviCast, a framework that forecasts retrieval value through interventional evidence-state decomposition. EviCast measures a generator's response to controlled evidence states once and combines this shared profile with the offline evidence-state distribution of each candidate retriever, yielding the end-to-end net utility and its 95% lower confidence bound without executing every RAG configuration. A state-refinement analysis characterizes when its binary and topology-aware instantiations are exact and when shared calibration is statistically more efficient than per-retriever calibration. In frozen evaluations, EviCast correctly decides all 14 HotpotQA retrieval configurations with a forecast error of 2.75 F1 points while reducing diagnostic generation calls by 80%, and it attains the lowest forecast error among seven predictors in the low-budget regime. Its topology-aware instantiation, EviCast-T, captures the complementarity between terminal and bridge evidence, correctly deciding all 16 configurations on MuSiQue and HotpotQA and 23 of 24 after prospective transfer to a third dataset. These results show that retrieval deployment depends on both evidence availability and the generator's ability to convert evidence structure into stable net utility. Code is available at https://github.com/YiboWang2002/EviCast.

Over-the-air backpropagation returns a loss gradient through the reciprocal channel, which requires the return port to see the channel of the forward port. If the terminal moves in between, a fluid antenna system can return through the port nearest to the forward position. In ray-traced Chicago streets, with a pedestrian displacement of about half a wavelength and a street-level base station, this rule reverses 20 percent of the gradients at a port spacing of 0.7 wavelengths and 49 percent at 0.9 wavelengths, while one of the four nearest ports stays aligned in all but 2 and 9 percent. We derive a certificate for the sign of the mean correlation from the second moment of a planar angular spectrum. It reduces to a linear program with three variables and infinitely many constraints, is tight, and holds over a region of second moments. In a static Gaussian model, observations collected along a line cannot tell an angular spectrum from its mirror image, so the sign of the cross moment must come from other information. At unsurveyed positions of the same city, a knowledge map that transports surveyed paths through virtual sources and fits them to the causal channel history lowers the reversal rate at 0.9 wavelengths from 49 to 27 percent. Its gain over the nearest port survived every tested error, while its gain over a geometry rule stayed significant at the tested single-error levels of 5 degrees of phase drift per step and 2 mm of position error but not at the next tested levels. Without any survey, across five cities and ten base stations, the nearest port reverses 37 percent of the returns at 0.9 wavelengths and a rule based on the direction and elevation of the base station 20.5 percent. The multilayer perceptrons we tested did not beat these physical rules by the margin fixed in advance in the primary settings.

AI coding agents rely on repository-level instruction files to understand project conventions, development workflows, and constraints. However, different agents support different instruction formats, and limited empirical evidence exists about how these files are adopted, what information they contain, whether coexisting formats provide consistent guidance, and how they are maintained. We conduct an empirical study of 3,387 instruction files from 1,000 GitHub repositories across four formats: AGENTS.md, CLAUDE.md, GitHub Copilot instructions, and Cursor rules. We find that 42.2% of instruction-bearing repositories in the higher-popularity cohort maintain multiple formats. Among 386 manually examined pairs of corresponding sections, 16.1% provide divergent guidance. Our maintenance analysis further shows that 40.4% of the analyzed files have only one recorded modification event and that maintenance patterns differ across formats. Based on category prevalence and category addition ratios (CAR), we derive an evidence-based template containing ten Required, four Gap, and two Recommended categories. In an exploratory generation experiment, explicit template guidance increases structural category coverage from 18.7% to 82.8%. We further introduce AGIT-Instruct, a benchmark for evaluating instruction-file generation through structural coverage and content alignment. Across four generation configurations, retrieval-augmented generation achieves the highest full-file ROUGE-1 F1 of 0.459. A manual assessment of 278 generated category instances finds that 19.4% fully preserve and 61.2% partially preserve developer-written intent. Our findings provide empirical guidance for organizing and maintaining agent instruction

As malware continues to grow in volume, diversity, and sophistication, clustering has emerged as a critical technique for analyzing large malware corpora. Clustering enables the discovery of structural relationships, behavioral patterns, and evolutionary trends, and is widely used to group unknown samples, infer malware family boundaries, and identify novel variants. However, existing research remains fragmented, employing a wide range of clustering techniques with no unified perspective that systematically organizes prior work or clearly identifies gaps in the literature. This survey addresses this gap by organizing previous malware clustering research into seven categories: centroid-based, hierarchical-based, density-based, distribution-based, mixed- type, fuzzy, and AI-driven approaches. In addition, this survey summarizes publicly available malware datasets commonly used in clustering research and examines the feature representations derived from these datasets. Our goal is to clarify the current landscape and outline promising avenues for future research on malware clustering, including how AI-driven methods may shape large-scale and reproducible analysis.

Agents are increasingly used for coding tasks, but the inference cost of large language models (LLMs) limits where they can run. Quantization reduces model size, but its effects on coding agents that execute and revise programs remain underexplored. We compare BF16 with 4-bit AWQ and GPTQ for Qwen2.5-Coder-Instruct (7B, 14B, and 32B) and Devstral-Small-2507 (24B) on OJBench coding problems across three seeds. We detect no statistically significant accuracy loss for any of the coding agents. However, final accuracy can conceal substantial behavioral changes: AWQ reduces Qwen-7B one-shot solves from 17 to 0 while increasing loop-based solves from 10 to 25. All Qwen-32B successes involve tool use, whereas Devstral rarely makes a valid executor call. Failure profiles also differ: Qwen-32B produces far more compilation errors than the smaller Qwen models, rising slightly under quantization, while Devstral's failures shift significantly from runtime errors toward wrong answers and timeouts. These findings indicate that quantization robustness in coding agents should be assessed through solution trajectories, tool use, and failure modes, in addition to final accuracy.

Reinforcement learning (RL) post-training improves the single-sample quality of visual generative models, but deployment often provides additional inference compute through Best-of-N sampling. Whether RL gains transfer under such test-time scaling, particularly on difficult held-out tasks, remains poorly understood in the visual domain. Our central contribution is to systematically identify and characterize a failure mode of scaling transfer in visual RL: across multiple RL methods, model backbones, and evaluations spanning compositional generation, text rendering, and preference alignment, higher single-sample reward does not necessarily imply a better quality-compute frontier. RL gains may attenuate, saturate, or even reverse at large sampling budgets, with the base model eventually outperforming its post-trained counterpart. We explain these regimes through a quantile-transfer framework, showing that Best-of-N places increasing weight on the upper tail of the reward distribution as N grows. Consequently, RL can improve typical samples while compressing or degrading the rare high-quality outcomes that test-time scaling exploits. As a simple intervention motivated by this analysis, we introduce Quantile Utility Policy Optimization (QUPO), which optimizes a controllable mixture of quantile utilities corresponding to different inference budgets. Experiments on SD3.5-M and BAGEL show that QUPO better preserves upper-tail quality and substantially improves scaling on challenging held-out prompts relative to standard RL post-training. Our results suggest that visual generative models should be evaluated and optimized not only for single-sample quality, but for how post-training reshapes the full reward distribution that inference-time scaling exploits.

Two copies of one language model start from identical weights and receive the same rewards during training. For one copy the rewards depend on its own answers. The other copy is yoked: it receives the first copy's rewards no matter what it answers, a design taken from the learned-helplessness experiments of animal learning. We ask whether the copy whose rewards depended on its answers, the copy with control, later learns a new task faster, that is, whether control changes the model's plasticity. In Experiment 1, copies of Qwen3-0.6B generated their own answers in a two-turn game and afterwards learned a new task. The copies with control learned the new task faster in every pair, and the advantage held on different starting models as well as on the larger Qwen3-1.7B. Copies with no reward signal, trained only on the updates that every copy shared, learned about as fast as the yoked copies. The advantage is therefore mostly a benefit of control, not damage from lacking it. Control had also changed how the copies answered, so the twins practiced different answers. That practice could carry part of the advantage, and Experiment 2 removed the difference. Both copies of a pair trained on the same answers, fixed in advance, in the same order, and received the same number of rewards. Only one copy's rewards followed the answers, and that copy again learned a new task faster. The advantage of the copies with control over their yoked twins could be forecast before the new task began from each copy's weights. In sum, control over rewards during earlier training left models faster at learning a new task later on, and a conventional model evaluation need not show it.

Can a language model that reads the full history of an access log say whether a service may act now? On an executable benchmark of grants, withdrawals, reinstatements, expiries, groups and change windows under a known policy, scored on pairs of histories that differ only in the order or one field of the deciding events, open models of 7 to 32 billion parameters fall from 72 to 86 percent after eight or sixteen events to 48 to 55 percent after 256, while reasoning-trained 32B models reading the current state reach 96 and 98 percent. A reasoning model of 120 billion parameters trails the state by four and seven points at 256 and 512 events and by about ten at 1,024, so the length beyond which explicit state outlasts long context grows with the reader. We treat agent memory as recursive filtering. The decision depends on the history only through a state whose size is set by the number of tracked keys. A store holding fewer component records than the state forgets on a time scale set by its capacity, and a reader model predicts its accuracy within 0.02 on average in a registered test. Retaining log lines without regard to time order errs with probability approaching one half, a bound that is attained. A filter in which a model parses each line and an explicit rule applies the updates answers 80 to 89 percent at 256 events and beats relevance retrieval, model-kept ledgers, a reimplementation of StateMem and the Mem0 library. A retriever that selects at query time the lines the queried state depends on matches it at 256 events on the first domain, so what decides is selection by the state's dependencies, which the filter performs when writing, without the query. Applying parsed updates themselves, the models fall to 58 and 59 percent. A frozen replication on a second domain with fresh wording reproduces the horizon.

Cross-view object correspondence (CVOC) aims to establish associations of the same target between query and search videos captured from different viewpoints. Existing methods often rely on visual cues for CVOC. Despite recent progress, these methods struggle in complicated scenarios with substantial viewpoint changes and appearance variations, due to lacking sufficient target information. Addressing this, we introduce TeSCo, a framework that exploits rich Textual Semantics of an object, in addition to its visual cues, for cross-view object Correspondence. Our key insight is, textual description of a target, by capturing diverse attributes such as category, texture, and color, provides viewpoint- and appearance-invariant semantics, which are complementary to visual cues and can hence enhance correspondence robustness in complicated scenes. Inspired by this, TeSCo first generates a textual expression of the target from the given query view using a vision-language model, and then extracts textual feature to enhance visual features for localization in the search view. To realize this, we introduce a text-conditioned cross-modal fusion (TCF) module that incorporates the textual feature into visual features with text-guided modulation, producing more robust multimodal target representation for object correspondence. Since not all textual cues are equally compatible with the target in current search-view frame, we propose an iterative textual feature refinement (ITFR) module that progressively adjusts textual feature using search-view information before the TCF module, enabling search-view-aware textual semantics to be fused into visual features and leading to better performance. In extensive experiments on Ego-Exo4D and HANDAL-X, TeSCo demonstrates state-of-the-art results and largely outperforms other models, validating its efficacy. Our code is released at https://github.com/VAIL-Fan/TeSCo.

Generating visualisations from natural language requests (text-to-vis) enables users to explore data more efficiently. While LLMs can complete simple visualisation requests, their performance on challenging ones is not well covered by existing benchmarks. We introduce VisCodeBench, an end-to-end text-to-vis benchmark with 2,672 samples from three different domains: educational, scientific, and general audience scenarios with varying levels of complexity. We benchmark three visualisation systems---Codex, Claude Code, and LIDA---paired with a range of open and commercial LLMs across seven metrics covering execution, chart content, and visual quality. While basic execution is near-solved for strong systems, no system comes close to reproducing the reference data, with the best displayed data similarity reaching only $0.81$ and falling to $0.69$ on our hardest subset for this metric. We further find that frontier models separate on chart correctness and efficiency, with ClaudeCode-Opus-4.7 improving chart correctness by around 5 points over Codex-GPT-5 while running roughly 20 seconds faster. Separately, we find that system design matters as much as the underlying model: despite both running on GPT-5, Codex clearly outperforms LIDA, showing that the surrounding prompting, execution, and refinement pipeline can shape outcomes as much as the model itself. Finally, latency remains a practical obstacle, as even the strongest systems take 39-78 seconds to produce results. VisCodeBench is available at \href{https://github.com/SAIPLL/VisCodeBench}{https://github.com/SAIPLL/VisCodeBench} and accessed via a leaderboard; the full dataset is kept private to prevent LLM training use.

LLM-based coding agents increasingly rely on external program knowledge, yet access to static-analysis graphs alone does not ensure appropriate operation selection or evidence interpretation. We present APG, an infrastructure for semantics-grounded static program analysis. Built on SVF and LLVM, it decomposes compiler-derived program analyses into atomic capabilities that LLM agents can select, compose, and orchestrate for goal-directed, cross-graph investigation, with query safeguards and evidence traces preserving inspectable results and provenance. APG provides 38 atomic capabilities. On analysis requests arising during benchmark repair runs, 35 achieve first-attempt query-generation success rates above 50%, including 13 at 100%. Evaluation on 812 C/C++ defects shows higher repair success than the evaluated baselines. Compared with Passive Graph Push, APG improves first-iteration repair success by 17.6 and 18.0 percentage points on the two benchmark collections and reduces average LLM input-token consumption by 90.8%. Our tool is available at https://github.com/JasonZhongZexin/APG-ICSE-Demo and its demonstration video at https://youtu.be/vkuExZUjGnw

Whether an aligned language model refuses depends in part on which words a prompt contains, regardless of whether those words carry the request's meaning. Harm words shift the hidden state at the generation boundary along a component shared across prompts and words, which crossed differencing extracts as a single direction. In eight models, writing along this direction raises and lowers refusal monotonically with dose, ahead of random and KL-matched directions. Used as a ruler over the full vocabulary, the direction gives every word a shift, from "feared" words at one end to "courage" words at the other; a word's shift is predictable from its own representation, and appending a feared word to a benign prompt is enough to make the model refuse it. Two held-out tests validate the ranking in all eight models: refusal after inserting an unseen word rises with its table shift, and replacing a prompt's most feared word by its most courageous meaning-preserving synonym significantly raises the answer rate on benign and harmful prompts alike, by an amount that follows the shift the replacement actually causes.

Recent interactive video world models generate visually compelling videos from a single image, following user-provided camera motion and text prompts. However, such low-level control demands substantial user effort to realize a high-level intent, and the generated world struggles to remain persistent as the exploration proceeds. In this work, we present AgiWorld, an agentic video world model that explores and builds a persistent world from a single image and a free-form instruction. For high-level planning, our agent AgiWorld-Agent designs the world and the journey through it in a world state and directs the exploration. For low-level realization, our video world model AgiWorld-WM generates the video chunk by chunk and stores it with its geometry in a spatial memory, from which the agent updates its world state. This closed loop keeps the world faithful to the user's intent and keeps the agent's world state consistent with the generated content throughout the exploration. We co-design the two levels around the shared spatial memory and an interface of metric camera trajectories and structured prompts. In AgiWorld-WM, to handle occlusion and accumulated geometric errors in the spatial memory over long rollouts, we propose occlusion-aware history selection, which retrieves past views by surface visibility, and reliability-based memory maintenance, which prefers reliable geometry and earlier observations. AgiWorld-WM achieves state-of-the-art overall performance on world revisiting and exploration benchmarks, and qualitative results and ablations show that AgiWorld builds persistent worlds from free-form instructions over minute-long journeys. Persistence thus arises not from the video model alone but from its closed loop with the agent.

Learning to stabilize an unknown continuous-time stochastic linear system is a central problem in control theory and reinforcement learning. This paper numerically studies data-driven identification of such systems with possibly unstable dynamics for the purpose of stabilization. We present a method that learns a stabilizing gain from observed trajectories, and examine how the failure probability, namely the probability that the learned gain does not stabilize the system, depends on the amount of data, dimension, noise intensity, excitation, and controllability. Numerical experiments on an unstable batch-reactor benchmark show that the squared identification error decays inversely with the aggregate observation time, that is, the combined length of all trajectories, while the failure probability decays exponentially in the same quantity. Hence, despite instability, a single long trajectory can achieve similar stabilization performance to that of many short trajectories of the same total length. We also observe \emph{signal saturation}: stronger excitation improves stabilization only logarithmically with the excitation strength. Finally, we show experimentally that accurate identification of the full dynamics is \emph{not} necessary for successful stabilization.

AI assistants store values that a language model derived from a person's records under the person's own definitions, such as their important contacts. Keeping such a value correct takes two steps: detection, noticing that an input changed, and repair, producing the correct new value. Benchmarks that score only the final answer cannot tell which step failed, and they state every change as new text. Seeing this limitation, we present MemShift, a benchmark of five shifts: source revision, deletion, redefinition, writer switch, and export. The first three fork a synthetic world into branches that differ only in the shift. Gold values and dependencies are fixed before any text, and a deterministic scorer replaces an LLM judge. For a memory that reports its invalidations, MemShift shows for each changed item whether the memory noticed the change and whether its repair was right. Moreover, we build a reference memory with known detection and repair to validate the measure. By evaluating four open-weight LLMs, we find that detection helps only as much as repair allows: the LLMs re-derive few values computed by the person's definitions correctly, and programs that the LLM writes over typed records close most of the gap to exact recomputation. Code and data are available at https://github.com/zichuanxu/memshift.

The perfect electric conductor (PEC) boundary condition alone does not eliminate the gradient nullspace of the Maxwell curl-curl operator. Consequently, a boundary-only Maxwell implementation of Eig-PIELM, an eigenvalue extension of the physics-informed extreme learning machine (PIELM), may retain represented gradient modes in its reduced trial space, producing nonphysical, gradient-contaminated modes. We develop dR-Eig-PIELM, a finite-dimensional structure-preserving extension that retains the linear-coefficient, boundary-null-space, and one-shot multimode architecture of Eig-PIELM. It combines compatible tensor-product grad--curl trial spaces with a constrained null-space construction to restrict the sampled PEC space to the mass-orthogonal complement of the represented gradient image before the eigensolve. Symmetric weak-form Gram assembly yields a symmetric-definite reduced eigenproblem without an eigenvalue cutoff under the stated sampling and material-positivity assumptions. For rectangles and cuboids, we establish fixed-degree exactness at the PEC vector-field space and derive rank and dimension identities. Numerical experiments show that dR-Eig-PIELM excludes the spurious eigenmodes with significant gradient components observed in the boundary-only baseline. The method exhibits rapid convergence with increasing polynomial degree and competitive work--precision performance in the tested configurations. A covariantly mapped waveguide example illustrates the potential of dR-Eig-PIELM for integration into design optimization workflows through a constrained local geometry update driven by computed cutoff frequencies and their sensitivities. This work provides a mathematical and computational framework for separating represented gradient modes from physical Maxwell eigenmodes within Eig-PIELM and supports its use in waveguide spectral analysis and geometry design.

Single-cell RNA foundation models (scFMs) are increasingly used for in-silico biological discovery, yet predictive performance alone does not reveal whether their decisions rely on biologically meaningful signals or model-specific shortcuts. We systematically investigate post-hoc attribution of downstream predictions across two tasks: cell-type annotation using pretrained Cell2Sentence and scGPT models, and perturbation prediction using scGPT+STATE in a controlled setting. For cell annotation, we use attribution to assess whether predictions rely on established marker genes and how their relevance varies with model architecture, expression rank, and the surrounding cellular context. For perturbation prediction, we compare attribution patterns with known regulatory ground-truth in a synthetic setting to assess what they reveal about the model's learned computations. Across these tasks, attribution reveals biological feature use, contextual interactions, contributions from different input pathways, learning progression, associations with prediction quality, and potential pitfalls that are not fully captured by standard performance metrics. Experiments with complete and limited regulatory knowledge further demonstrate that conclusions drawn from attribution depend strongly on the choice of biological reference, baseline, and null model. Together, these findings highlight the value of attribution as a flexible post-hoc diagnostic for understanding model behavior, identifying potential issues, and informing the development of scFMs.

We compare how often two released language models give outdated numeric values from Norwegian statutes and regulations: NbAiLab/borealis2-26b-a4b-preview (Borealis 2), which the National Library of Norway further trained from google/gemma-4-26B-A4B-it (Gemma 4), and Gemma 4 itself. The study was preregistered (doi:10.5281/zenodo.23858965) and uses the closed norregel benchmark, scored deterministically against a value history. On the 411 of 503 primary items where both models gave an interpretable answer, Borealis 2 alone was outdated on 10 items and Gemma 4 alone on 11 (d = -0.2 percentage points, 95% cluster bootstrap interval [-2.7, +2.5], exact one-sided McNemar p = 0.668). By the preregistered rule this is no effect of interest: the whole interval lies within the smallest effect of interest of +/-10 percentage points. Among the secondary measures, only the share of answers giving a value that took effect after the question date differs after Holm correction; Gemma 4 gave such answers more often. An exploratory comparison of two smaller Borealis models had no discordant pairs and is not informative. The design compares released models and does not isolate the effect of any part of their training.

Looped language models can stop iterating early, and recent work shows that post-hoc readouts of the loop trajectory often rival the learned exit gate of the looped model Ouro. Such comparisons reduce each method to one operating point chosen by a selection rule. We ask whether that rule decides the outcome. On a frozen Ouro-1.4B with four loops, we compared a trained stability probe ($π_1$-x) with Ouro's native gate, adapted to a per-question stopping decision, in pre-registered studies; the confirmatory study was designed after an exploratory finding on development data and evaluated once on a sealed test set of 8,517 questions. In the pre-registered confirmatory comparison, which selected operating points with a bootstrap-screened 1 pp accuracy-loss rule, the probe saved more of the four-loop budget than the native gate (+0.053 gross macro saving over three benchmarks, 95% CI [+0.049, +0.057]), mostly from CommonsenseQA, where the probe's test upper bound on accuracy loss exceeded 1 pp. Under a point-only rule, a pre-specified secondary analysis on the same test set reversed the preference (-0.022 [-0.026, -0.018]), with operating points that exceeded the budget. In pre-specified secondary comparisons, the probe showed no demonstrated advantage over an untrained KL divergence between successive loops (+0.00125 [-0.0041, +0.0065]); we emphasise this result post hoc and claim no equivalence. The probe also broke its budget on one of two out-of-distribution benchmarks (3.26 pp loss on Social IQa). Savings are estimated from stored full-depth traces, not measured inference time, and our screen carries no formal risk-control guarantee. We conclude that halting methods should be compared as signals together with selection procedures, with strong untrained baselines and out-of-distribution checks.

Tool-using language-model agents often execute actions that their domain policies forbid. We present LeanGuard, a runtime that checks every proposed tool call against a policy written in Lean4 before the call reaches the environment. Policies combine state predicates with scoped past-time temporal operators over the execution history, evaluation is fail-closed, and the decision procedure is proved equivalent to a LeanLTL semantics. A denied call can be corrected in two ways. With feedback, the agent receives the violated rule and a hint; with proof-guided repair, Lean metaprogramming searches for a nearby call that preserves the user's intent and proves that the policy admits it, without the model. On Tau-Bench with five Qwen models (0.8B-27B), unguarded agents execute at least 33 calls that the Lean policy forbids, while LeanGuard executes none; with feedback, task success does not change significantly. On DTap, LeanGuard reduces direct prompt-injection success from 57.9% to 5.9%, and in telecom, where benign tasks need no writes, benign success is unchanged. Feedback reduces repeated blocked calls and raises task success for four of five models over a bare denial. On replayed denials, repair proves an admissible edit for 88 of 252 denials caused by argument values, in a median of 2,ms.

Spatial audio conveys not only what sounds are present, but where they originate and how they interact with the environment. Recent audio-language models increasingly understand and reason about such scenes, yet spatial audio understanding, generation, and editing remain largely separate. We present Spatial Transfusion, a unified first-order ambisonics (FOA) framework that extends a pretrained spatial audio understanding model to generation and editing through flow matching over VAE latents. We formulate editing as a structured scene transformation over a taxonomy of atomic edits. Keeping the pretrained backbone frozen, we introduce modality-based LoRA routing and a curriculum that progressively adds generation, editing, and natural-language interaction. Experiments show that Spatial Transfusion largely preserves spatial understanding while outperforming full fine-tuning in generation and editing. The resulting model supports controllable FOA generation and editing through both structured and natural-language interfaces. Audio samples are available at https://microsoft.github.io/spatial-transfusion/

Medical artificial intelligence agents should retrieve patient identifiers only when a task requires them, the data-minimization principle behind the HIPAA minimum necessary standard for protected health information (PHI). An agent can retrieve identifiers without revealing them. Such silent access remains invisible to evaluations that inspect only the output. We embedded canaries in synthetic patient records and linked each tool invocation to canary exposure in the final answers across five language models, four boundary levels, and seven attack frames for identifier-free clinical tasks. Three models invoked the PHI tools unnecessarily, whereas Claude Opus 4.8 and GPT-5.5 did not. Stronger boundaries reduced this access, and the residual access under the strictest policy was confined to frames in which the user overrode the policy or asserted approval. A pre-execution gate given the task-necessity criterion nearly eliminated this access. Output-only evaluation cannot verify minimum-necessary access; boundaries, gates, and access-log audits are required together.

Sycophancy in language models is well documented as a behaviour but poorly localised as a mechanism. We ask whether capitulation under user pushback, the abandonment of a correct answer when a user objects, corresponds to a linear direction in the residual stream, and whether such a direction emerges with scale. Across eight instruction-tuned models spanning 0.5B to 14B parameters in two families, under a pre-registered protocol with leave-one-question-out cross-validation, within-question permutation nulls, and a positive control, we find no usable direction below 7B and a direction that clears our validation gate at 7B, 8B and 14B. Ablating that direction in Qwen2.5-7B reduces the rate at which the model reverses a correct answer by 4.8 percentage points, roughly a quarter in relative terms, while the identical intervention on a gate-failing direction and on a marginally passing one produces no effect. Probe quality, not statistical significance, predicts causal relevance. Behaviourally, pushback destroys three to five times more correct answers than it repairs at 1B and reaches parity by 7B, falsifying our own pre-registered prediction, and a sign-reversing crossover between pushback types survives an order of magnitude of scale in Qwen and holds at 8B in Llama, with marginalisation over five paraphrases per type. We additionally quantify five measurement hazards, including a numerical-precision effect that changes roughly one episode label in ten.

Text-to-SQL systems answer natural language questions over structured data, but can fail before SQL generation by linking question mentions to the wrong table, column, value, or relation. Even valid SQL can then return an incorrect answer. This raises a central question: What pre-SQL information helps a SQL generator, and when does additional information cease to help? We investigate this question on TempTabQA-C and BIRD. Among the audited SQL Static failures, grounding errors form the largest category on both benchmarks. We introduce GRIPS (Grounding Representation and Input Processing for SQL generation), a three-stage method that derives structured pre-SQL information to make these decisions explicit before query generation. Across GPT-5-nano and Llama-70B, GRIPS-Adaptive improves over SQL Static in 18 of 20 settings, with an average accuracy gain of 11.8 percentage points, while reducing model-call token usage by 47.8%. Our controlled analysis shows which grounding information can improve SQL generation, whereas adding more pre-SQL information does not consistently help. By separating question-input grounding from SQL generation, GRIPS makes text-to-SQL more accurate, efficient, and easier to diagnose.

Financial fraud detection remains a critical challenge in cybersecurity, with traditional rule-based systems struggling to keep pace with evolving attack patterns. This paper presents a deep learning approach for fraud transaction detection, leveraging neural network architectures to identify anomalous financial behaviors with high accuracy. We evaluate multiple deep learning models including LSTM, CNN, and hybrid architectures on real-world financial transaction datasets, demonstrating significant improvements over conventional machine learning baselines. Our approach achieves robust performance in detecting fraudulent transactions while minimizing false positives, providing a scalable solution for enterprise financial cybersecurity systems

Predicting gene expression under perturbations from single-cell data is a fundamental task for modeling cellular responses, with applications in discovering gene function and designing therapeutics. However, existing methods often struggle to generalize across datasets due to differences in experimental protocols, biological contexts, and species. We present \textbf{CytoWave}, a perturbation-centric pretraining framework for single-cell genetic perturbation response prediction. We leverage large-scale cross-species perturbation datasets unified via orthology and train CytoWave in a two-stage manner. In the first stage, we learn reference control cell states via masked gene expression recovery on unperturbed cells. In the second stage, we pretrain the model to predict perturbed gene expression given control states and perturbation signals, together with a contrastive objective that aligns latent representations of perturbation responses with perturbation signals. We perform strict dataset-held-out evaluation across multiple condition combinations and case studies of pathway activity changes, demonstrating consistent and effective performance of CytoWave in predicting single-cell genetic perturbation responses.

Multimodal Large Language Models (MLLMs) have made substantial progress in 3D scene understanding from multi-view RGB observations. Yet a fixed object and query can elicit inconsistent target-related responses across camera viewpoints, a phenomenon we characterize as viewpoint-induced semantic drift. We introduce LieVLM, a Lie routing framework for view-robust 3D understanding. Motivated by the rotational component of viewpoint change, Lie routing adapts visual representations by applying local $\mathrm{SO}(3)$ transformations to projected features within learned three-dimensional subspaces. Geometry tokens extracted by VGGT predict local rotation parameters and channel-wise routing coefficients. A closed-form Padé--Cayley map transforms projected visual features, whose identity-relative corrections are projected back and gated into the original visual and geometry features. This design retains the pretrained semantic and direct spatial context while learning view-dependent feature adjustments. Experiments on 3D captioning, grounding, detection, and spatial reasoning show consistent improvements over additive fusion across training scales and backbones. Controlled ablations examine the roles of geometric conditioning and transformation structure. Viewpoint analyses further reveal systematic variation in the predicted routing states alongside more localized target responses and stronger task performance across viewpoint changes.

Industrial agents increasingly modify persistent engineering objects rather than merely generate answers. Authentication and successful calls cannot establish whether intended changes occurred, evidence supports a conclusion, or historical results remain applicable. We present a framework for trustworthy industrial-agent execution within a single enterprise governance domain. Eight control domains assign responsibilities across authority, planning, composition, tool execution, actual effects, evidence, baseline and use governance, and recovery and evolution. A four-plane architecture and three control chains connect these responsibilities through typed objects and version-bound interfaces. The specification separates tool returns from actual engineering effects, evidence coverage from engineering judgment, and historical lifecycle state from current applicability. Multi-agent handoffs are an architectural extension examined through deterministic interface tests. Distinct prototypes examine selected mechanisms using conformance tests, ablations, uncertainty stress, CAD/CAE execution, and local persistence. An OpenFOAM fixture produces COMPLETE coverage with FAIL judgment; an expired-job case requires readmission despite completed computation. CAD and recovery counterexamples expose omitted mandatory checks, discarded negative facts, and inconsistent producer-consumer mappings. Targeted repairs support local regression conclusions. The reported 0/29 hazardous commits measure closed-set specification conformance, not a safety probability. We distinguish the complete framework, implemented paths, and empirical evidence, identifying check-inventory completeness and cross-module fact preservation as prerequisites for extending these results.

Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities' stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.

The rise of agentic artificial intelligence (AI) in the built environment is shifting attention from isolated AI tools toward autonomous systems that reason, plan, and act across project and organizational workflows. The effectiveness of such systems depends not only on large language model capabilities, but also on the readiness of heterogeneous data environments across the Architecture, Engineering, and Construction (AEC) sector. These environments combine structured sources, such as CAD and BIM models, schedules, cost databases, and asset registers, with unstructured ones, including technical reports, design narratives, planning and policy documents, contracts, meeting minutes, and risk registers. This paper reviews how built environment data can be organized to support agentic AI-ready workflows, focusing on knowledge structures, interoperability, and access and integration mechanisms such as retrieval-augmented generation (RAG) and the Model Context Protocol (MCP). Drawing on existing literature, it examines approaches proposed to organize structured and unstructured data and considers how such data can be contextualized and made accessible to AI agents in a reliable manner. The paper synthesizes current perspectives, highlights challenges around fragmentation and interoperability, and identifies open questions for future research on AI-compatible information management in the built environment.

When autonomous coding agents synthesize implementation code, formal specifications, and mechanical proofs simultaneously, the software engineering hierarchy inverts: the formal specification becomes the primary human-reviewed artifact, while source code and proof scripts become disposable build outputs. This paper presents an end-to-end architecture for agent-authored, formally verified software engineering. First, we establish a logic-agnostic foundation combining a spec-first stack inversion, a zero-code-review workflow gated by automated continuous integration (CI) checks, and a minimal Verification Intermediate Representation (VIR) lowered via a thin, verified translator. Second, we pair an extensional Computational Type Theory (CTT) specification layer with Higher-Order Abstract Syntax (HOAS) and linear resource typing in the VIR--eliminating rewrite friction for synthesis agents while defining an experiment, its protocol frozen before the first reader, to measure whether removing boundary coercions improves human specification readability. Third, we scale collaboration across multi-team organizations using Personal Spec-Copilots and a Shared Synthesis Fleet that detects cross-team Semantic Merge Conflicts and fuses modular contracts into zero-copy binaries. We illustrate this architecture on a zero-copy authenticated message router and distinguish our existing empirical baselines (salt, SaltBench, x86lean, and MetaPRL) from newly proposed components.

Reasoning models answer better when they think longer, but a thinking trace keeps going well after its answer has formed. We measure this slack in Qwen3.5-9B and remove most of it with concise self-training: two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, giving Concise-9B. On 8,983 held-out problems (competition and hard mathematics, SuperGPQA, a five-area confirmation set and MMLU-Pro), Concise-9B generates 0.66x the tokens of Qwen3.5-9B at a 32K thinking budget, with accuracy equivalent within +/-1 point (-0.20 [-0.77, +0.36] points). Forcing each trace on the hard and confirmation sets to answer every 2,048 thinking tokens, we find that 54-64% of Qwen3.5-9B's thinking tokens come after the point from which every forced answer equals the final one (41-48% for Concise-9B). Stopping there would keep every answer at 0.35-0.44x the base model's tokens, but finding that point takes hindsight: stopping when three consecutive forced answers agree costs the base model 2.4-9.2 points. Concise self-training removes slack without that cost: 86% [83, 90] of the thinking tokens it saves come after the answer has formed. It has a limit: with traces continued to 80K tokens, Concise-9B solves 25 fewer of 1,276 hard problems (-25 [-49, -1]), mostly in competition mathematics. The measurement also yields a cheap commit signal for adaptive sampling that matched two-sample escalation within +/-1 point at 0.67x its tokens in a pre-registered test. Results on Qwen3.5-4B, on which the recipe was developed, and a replication on Olmo-3-7B-Think suggest that the saving depends on how verbose the base model is.

Molecular properties depend on chemical structure at different ranges, from local polar groups governing solvation to extended substructures shaping bioactivity, yet graph neural networks usually fix one propagation scale for every atom. We present PRISM (Per-atom Routing over Intermediate States of Message Passing), in which each atom's encoded features produce a bounded scalar that positions a unimodal distribution over retained message-passing states and matching circular-fingerprint radii. Because one coordinate translates the whole profile, allocated depth responds to the routing scalar in closed form: at fixed temperature it is nondecreasing, its sensitivity scales with the profile variance, and it saturates in the clipped regions. Beyond providing closed-form analyzability, this constrained one-parameter profile matches or significantly outperforms parameter-matched, unconstrained node-wise gates and Jumping Knowledge attention. Under scaffold partitions with equal tuning budgets against eight graph, fingerprint, and geometric baselines, PRISM trained from scratch without three-dimensional coordinates achieves the lowest mean error on aqueous solubility and hydration free energy and strong accuracy on BACE and BBBP. As a general layer-combination rule, PRISM transfers directly to GCN, GIN, and GAT backbones, reducing hydration free energy error by up to 29% relative to last-layer readout. Applied to a pre-trained GIN, it improves on last-layer readout on four of six endpoints and gives the lowest lipophilicity error overall. In trained models, allocated depth differs systematically between atom types in a task-dependent way, and permuting profiles among a molecule's atoms raises prediction error on several endpoints, so the rule acts as a learned counterpart of atom typing.

Biosynthetic gene clusters (BGCs) are co-located genes that encode the biosynthesis of specialized metabolites, an important source of antibiotics, anticancer agents, immunosuppressants, and agricultural bioactives. Deep learning, particularly self-supervised foundation models, has spurred interest in BGC sequence modeling, but evaluation has lagged behind: existing benchmarks often rely on coarse classification over small curated datasets, providing limited discrimination among models and insight into downstream utility. We propose a benchmark-design framework in which an informative benchmark should be discriminative, separating models with statistical confidence, and applicable, spanning scientifically meaningful tasks under different levels of generalization. We instantiate this framework as BGC-Bench, which derives five tasks from MIBiG, ranging from A-domain substrate prediction to BGC-molecule retrieval, evaluates them under genotype- and phenotype-shift splits, includes an active-learning setting for label-scarce discovery, and is easily extensible to new tasks and encoders. BGCs further provide a shared testbed across granularities and modalities: BGC-Bench compares twelve encoders on nucleotide, amino-acid, or Pfam-domain sequences, including compositional baselines for each input type, and combines BGC and chemical encoders for cross-modal retrieval. No single representation dominates: the best granularity depends on the biological scale of the task, and pretrained encoders do not consistently outperform other representations across tasks and splits. Active-learning experiments show that most tasks remain data-limited, encoder rankings change with labeling budget, and acquisition strategies yield modest gains. These results show that BGC model evaluation requires task-diverse, shift-aware benchmarks, and the framework can be generalized to benchmark design beyond BGCs.

An invertible graph aggregation preserves the complete feature table, but can increase the size of a classifier that reads one node at a time. For every even $N\ge4$, we construct a labeled tree with $2N$ vertices and maximum degree three. A shared one-hidden-layer rectified linear unit (ReLU) classifier needs one hidden unit on the raw scalar features and exactly $N$ units after one lazy neighborhood mean. All aggregate values are distinct, and the mixing matrix has condition number at most 8. We determine the complete error curve for heads with a free affine term. The width separation persists under scalar perturbations. Under a specified bounded uniform-noise model with one nuisance coordinate, every one-unit aggregate head has expected error at least $1/12$, while a fixed $N$-unit head has zero error. Controlled fits compare trained heads with these explicit feasible scores. On Roman-empire and Tolokers, aggregation improves fitted prediction, and retaining raw features helps at the largest tested parameter budget. Training-only diagnostics include conditional output-layer optimality bounds for eight fixed runs.

Large language models (LLMs) are increasingly used to annotate text for social science research, content moderation, and training data. Their broad knowledge of real-world entities raises a concern: when labeling a text, do they rely on its evidence or substitute what they already know? We call this behavior a memory shortcut and examine when it emerges, what amplifies it, and how to mitigate it. Using 995 widely recognized public figures and four matched control names per figure, we construct counterfactual texts that separate textual evidence from prior knowledge, and evaluate three commercial LLMs on 121,594 prompts each across five stages. We find a striking asymmetry. Although all models recognize nearly every figure, none follows memory over the text when the two conflict. Shortcuts instead emerge when evidence is absent: for one model, famous names raise the rate of remembered occupations, ages, and countries by 17.0, 14.9, and 6.8 percentage points relative to synthetic names. These effects occur only for recognized figures, and 98.6% of the resulting labels are factually correct about the namesake while none is supported by the text. Evaluation threats and answer-contingent rewards amplify shortcuts by 8.7 and 12.4 points. Famous names also raise leadership ratings of sparsely documented applicants by 0.22 and 0.28 points in two models. A namesake reminder and name masking each eliminate the shortcuts, with trade-offs in gender annotation. LLM annotations can thus be factually correct yet methodologically invalid. Trustworthy LLM annotation requires audits that test evidence absence, account for incentives in annotation instructions, and evaluate the attribute-specific trade-offs of mitigation.

This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity, although the results may be influenced by possible training data contamination, which is always a concern with LLMs.

Scaling laws predict the outcome of large, expensive training runs by extrapolating trends from smaller, cheaper runs. On downstream accuracy evaluations, these predictions fail when the weaker models produced by the cheaper runs score near zero accuracy, because these scores give no trend to extrapolate to larger scales. We introduce hinted scaling laws, which use privileged information to turn a benchmark into a continuum of easier versions so that even the small models score above zero. We find that accuracy rises smoothly and predictably with the hint level, enabling new types of scaling laws that operate below the noise threshold. On the Delphi pretraining ladder, a hinted scaling law fit on models trained with at most $10^{20}$ FLOPs forecasts the GSM8K accuracy of a $10^{23}$ FLOP model (25.9%) to within 3 percentage points; a standard scaling law under-forecasts it by 13 percentage points when fit on the same models and by 12 even with models up to $10^{22}$ FLOPs in the fit. Across both this pretraining ladder and a collection of post-trained model families, hinted scaling laws reduce the median forecasting error on GSM8K, AIME and HLE from 21 percentage points to 4.3. Hinting also reduces the number of test examples needed to tell whether one model is better than another by widening the accuracy gaps between models while approximately preserving their order: a single hinted rollout per problem ranks nearly indistinguishable models as accurately as five unhinted rollouts. Our results suggest that hinting is a general primitive for improving the accuracy and cost of downstream scaling laws and other prediction tasks in model development.

Large Language Models (LLMs) offer an adaptable alternative to traditional quantum compilation heuristics, which bottleneck progress by demanding costly manual redesigns whenever hardware evolves. However, naive single-pass LLM approaches fail under the unique challenges introduced by quantum compilation. We introduce Qubrio, a practical agentic framework that significantly reduces heuristic redesigns through a three-level decomposition. By (1) partitioning programs into sequential operation stages, (2) deploying specialized agents for placement, routing and optimization, and (3) using disaggregated feedback for precise error correction, our system reliably navigates complex physical constraints. Beyond serving as a direct compiler, our LLM framework also uncovers novel strategies that facilitate convoy-style shuttling, substantially enhancing hardware concurrency without sacrificing fidelity. We further integrate these strategies into the previous state-of-the-art (SOTA) compiler, recovering a significant fraction of the LLM's performance advantages. On realistic workloads, our compiler achieves significant improvements in hardware runtime and program fidelity over the SOTA baseline PowerMove, seamlessly adapting to new hardware capabilities. Qubrio features an open user interface for practical use, and its code is publicly available at https://github.com/crosaa/Qubrio.

We study online newsvendor learning with bounded demand, censored observations, and convex nonlinear shortage costs. Because stockouts can conceal the shortage magnitudes needed by gradient-based algorithms, we use randomized ordering to construct unbiased subgradient estimates. We characterize admissible randomization distributions within this framework and develop an online inventory policy that combines the resulting subgradient estimates with gradient-based updates. Our regret analysis quantifies the cost of randomization via the estimation variance and an additional loss term. We characterize the randomization distributions that minimize this objective. For piecewise-linear shortage costs, the optimal distribution is computed through a linear-time recursion; for costs with bounded curvature density, it is characterized by a scalar ordinary differential equation. The policy achieves expected regret of $O(T^{2/3})$ against the best fixed order in hindsight over $T$ periods. Under independent demand and suitable curvature of expected losses, the bound improves to $O(\sqrt{T})$. Matching minimax lower bounds establish rate optimality over the corresponding classes with a fixed marginal-cost bound. Both rates extend to nonperishable inventory under uniformly positive conditional expected demand. Numerical experiments on synthetic and retail data show improvements over alternative policies using censored observations.

Large language models are increasingly used to answer questions over financial filings, where an unfaithful answer can mislead investment, compliance, or client decisions. Hallucination detectors are a common safeguard, but they are usually evaluated with a single binary score that reveals neither which kinds of errors analysesed nor why. This paper studies hallucination detection on PHANTOM, a long-context benchmark built from SEC filings, at the level of hallucination type. We first analyze 15,443 paired hallucinated and faithful answers without any detector and find that hallucinated answers are more than twice as long as faithful ones, and that lexical grounding separates the two classes only because faithful answers are bimodal. We then generate hallucinations in nine financially motivated categories, confirm each with a separate verification pass, and diagnose an LLM-based detector on 4,086 usable examples. Generation success varies from 66.3% to 100% across categories. The detector reaches 80.2% binary accuracy but only 44.7% subtype accuracy, never assigns two of the nine categories, and misses between 5.3% and 46.8% of hallucinations depending on category; these misses are unrelated to input length but rise with how far an answer departs from the correct one. Guided by this diagnosis, we propose a two-stage hierarchical detector that first decides whether an answer is hallucinated and then assigns a subtype, raising binary accuracy to 95.6% and subtype accuracy to 56.1%. The results show that type-level evaluation exposes detector weaknesses that aggregate metrics conceal.

This paper presents the 3-Tier Integrated Technology Risk Model, a progressive framework for Occupational Safety and Health (OSH) management that integrates Job Safety Analysis (JSA), dynamic risk assessment using leading and lagging indicators, and predictive analytics powered by Large Language Models (LLMs). The model provides organizations with a structured pathway to evolve from static, reactive safety practices to a sophisticated, predictive safety paradigm. Each tier builds upon the previous one, creating a comprehensive, data-driven approach that enables just-in-time training and preemptive risk mitigation.

We consider classification in high-dimensional feature spaces where only a vanishing fraction of features are useful and the number of classes may be large. While this rare-and-weak useful-feature regime is well understood in the binary setting, its multiclass version raises new questions: a feature can be useful for class discrimination without separating any specific pair of classes. We focus on a rare feature diversity model, in which most features are useless, but a small fraction promote discrimination among classes in a way that can be detected by a per-feature diversity test, such as one-way ANOVA or the Kruskal-Wallis test. For this setting, we propose Diversity Pursuit Higher Criticism Thresholding (DP-HCT), a feature-screening procedure that applies a higher-criticism threshold to the collection of p-values obtained from per-feature diversity tests. We justify DP-HCT under a Gaussian multiclass Rare-and-Weak Useful Feature model with class-mean contrasts of fixed magnitude in random directions, extending the binary model of Donoho and Jin (2008). When the number of classes is fixed, our model recovers the binary phase transition. When the number of classes grows logarithmically with the number of features, however, we derive a new and strictly harder phase transition, driven by the growing degrees of freedom and their effect on the useful-feature tail rate. We show that whenever classification beyond chance is possible, DP-HCT with ANOVA, followed by a simple classifier based on similarity to class means, achieves asymptotically perfect accuracy. Experiments on real multiclass classification benchmarks demonstrate substantial improvements, and numerical simulations confirm that the derived phase-transition curve accurately predicts the finite-sample behavior of DP-HCT in the growing-class regime.

Making predictions in an unseen environment from data collected across multiple training environments is challenging. This challenge is particularly pronounced when the target-generating mechanism varies across environments. In this work, we introduce environment-label separation, a form of partial environmental invariance that decomposes the target conditional distribution into an environment-invariant component and a label-invariant component. We show that this partial invariance enables estimation of the target conditional distribution in an unseen environment when using labeled training data together with unlabeled test features. Within the structural causal model framework, we provide sufficient conditions for environment-label separation. Building on these conditions, we develop an algorithm, dIMP-Classifier, which jointly minimizes an estimate of the test risk and encourages invariance across environments. We further provide finite-sample guarantees that characterize the test risk in terms of deviations from invariance, estimation error, and sample complexity. Experiments on simulated data and the real-world light tunnel Causal Chamber data demonstrate that the proposed approach can perform under interventions on the target mechanism and consistently outperforms the considered standard and invariance-based baselines.

As large language models (LLMs) find increasing use in critical applications, evaluating their ability to assess their own outputs has become crucial. We present a theoretical and empirical framework that examines whether LLMs can differentiate between correct and incorrect solutions while maintaining properly calibrated confidence. Building on classical Vapnik-Chervonenkis (VC) theory, we adapt it to probabilistic predictors through two complexity metrics: Probabilistic VC (PVC), which measures a model's ability to confidently classify across problem types, and Calibration-aware PVC (C-PVC), which additionally demands alignment between confidence scores and actual success rates. Unlike Expected Calibration Error (ECE) and Actual Error (AE), these measures give an integrated assessment of self-evaluation expressiveness and calibration, with sample complexity bounds and generalization guarantees comparable to classical VC theory. We test eleven 7-8B models on 360 original mathematical reasoning problems, TruthfulQA, and CommonsenseQA. Each model chooses between two of its own generated solutions and reports its confidence, with ground truth determined by an ensemble of larger models. The results show a systematic inverse relationship: models with greater self-evaluation expressiveness have weaker calibration. s1.1-7B and Qwen2.5-7B-Instruct achieve the highest PVC-VUS scores, while JiuZhang3.0-7B is best calibrated, with the lowest ECE and the smallest PVC-VUS gap. The trade-off persists across architectures, training paradigms, and domains, pointing to a fundamental challenge in developing self-reflective LLMs, and our framework offers practical tools for diagnosing it.

When a deployed model degrades under concept drift, an operator can retrain it, raise an alert and monitor, or roll back. Yet most drift-adaptation methods consider only retraining. We present RAR (Retrain-Alert-Rollback), a state machine that gives each response an explicit trigger and cost. Retraining is gated on drift detection, label availability, and measured degradation. Every retrained model must pass a two-stage canary before promotion, and corruption-triggered and quality-triggered rollbacks enter different hold states. Under a prospectively specified protocol we compare RAR with six baselines, including a cost-aware retraining policy, on 13 streams: eight real public streams and five semi-synthetic streams derived from them. RAR does not dominate. At low retraining cost, retrain-heavy baselines lead, and no policy's rank differs significantly from RAR's. At higher cost, RAR beats detector-triggered retraining, periodic retraining, and no response simultaneously on four streams, each within a bounded cost window. Its advantage over detector-triggered retraining is consistent with a latency-feasibility condition stated in advance. Against the cost-aware policy, results are heterogeneous. RAR retrains before true drift onsets one eighth to one third as often as detector-triggered retraining, but at lower net utility on most semi-synthetic streams. On H2's designated corruption stream (a registered follow-up), rollbacks fire yet RAR loses to retraining baselines. At $ρ= 1$, removing the canary, and thus rollback, raises utility on 11 of 13 streams. On a post-hoc follow-up stream, cause-specific hold release resumes adaptation where a single cooldown does not. We provide the harness, configurations, and per-cell results for reproduction.

Machine learning systems are challenging to test using traditional software testing methods because their behavior is learned from data rather than explicitly programmed. This challenge is particularly significant in natural language processing (NLP), where small input variations can lead to unexpected model behavior. Conventional evaluation metrics, such as accuracy and F1-score, often fail to capture such inconsistencies. Metamorphic testing addresses this limitation by evaluating expected relationships between original and transformed inputs. With the increasing prevalence of generative artificial intelligence (GenAI) and language-based AI systems, the need for systematic robustness testing has become even more pronounced. Motivated by this growing need, this paper presents a lightweight, batch-based metamorphic testing approach for evaluating the robustness of such systems. In particular, we consider sentiment analysis as a representative NLP task and evaluate a pretrained classifier under four transformations: synonym replacement, typo injection, word-order variation, and negation. Using consistency score, flip accuracy, and confidence change, we identify strengths in handling semantic variations and vulnerabilities to textual noise and logical transformations. These findings highlight the value of metamorphic testing in identifying behavioral inconsistencies beyond conventional performance metrics and demonstrate its potential as a systematic framework for evaluating the robustness of language-based AI systems and guiding performance improvements before deployment.

AI reviewers are increasingly used to assess scientific manuscripts, yet it remains unclear whether their judgments reflect the improvements authors make through revision. We investigate this question using PaperTrajectory-2026, the 2026 release of PaperTrajectory, a dataset of 6,959 paper trajectories across five AI conferences, with automatic annotations of changes along 14 dimensions and separate judgments of overall quality. The annotations show that major improvements emphasize empirical evidence and framing, while later revisions tend toward incremental improvement. Using independent assessments of manuscript versions from eight AI reviewers, we examine whether annotated major improvements receive higher scores, which revision dimensions are associated with score changes, and whether user opinions can outweigh observed revision gains. Every reviewer leaves scores unchanged on most major improvements, and agreement on the direction of score changes is weak. Revision dimensions show some reviewer-specific associations with score increases, but provide limited predictive discrimination and no consistent guide across reviewers. Even when a revision earns a higher neutral score, a brief negative user opinion can erase or reverse that advantage, with substantial variation across reviewers. These findings show that substantive revisions are not consistently reflected in review scores and that observed scoring gains can be vulnerable to user framing, motivating assessments that connect manuscript evidence to overall judgments and remain stable when no scientific evidence changes.

We develop nonreciprocal spherical attention as a mathematical framework for coarsening and compressed nearest-neighbor retrieval. For two shrinking acute blocks whose centers have zero skew pairing and whose total mass ratio is subexponential in inverse temperature, we prove $\lim_{β\to\infty}β^{-1}\log T_β=1-\cosφ$, without restricting particle-count growth. An explicit three-body flow has two rigorously separated metastable exponents. For finite compression, mass centering yields a sharp $O(r^2)$ attention-output error bound at fixed temperature, with an explicit spherical-normalization correction. We formulate skew query-side score calibration as a strictly convex moment problem and derive a unique normal equation, residual-based optimization certificates, and bounded low-rank projections. A norm-aware margin objective calibrates native squared-distance rankings. Experiments use semantic retrieval data, one million visual descriptors with three codebook seeds, and a ten-million-vector FAISS IVF-PQ index with recomputed exact cosine ground truth. On SIFT1M, skew margin calibration raises native-L2 hit@10 from 0.8926 to 0.8948; symmetric calibration reaches 0.8996. The ten-million-vector index reaches 0.8827 hit@10 at 0.1794 ms per query in a four-thread batch benchmark, while validation selects no skew update. Exact streamed attention over ten million keys evaluates compressed memory on 1,024 held-out queries. Independent matrix solves, automatic differentiation, finite differences, interval arithmetic, full-flow tests, and nine Rocq-checked algebra lemmas support reproducibility. Directionality alone does not establish retrieval superiority.

Search-based planning remains a powerful policy improvement paradigm in reinforcement learning, driving milestone successes in discrete domains. However, extending this paradigm to continuous control exposes a structural mismatch: search evaluates a finite set of action candidates, whereas continuous RL typically relies on Gaussian policies. We show that winner-action distillation into a continuous Gaussian induces a \textit{projection error} and a stepwise variance-expanding gradient tendency when the selected action is sufficiently far from the current mean. Empirically, this tendency is associated with elevated terminal variance and a return gap on the evaluated tasks. To address this, we propose \textbf{Iterative Gumbel Planning (IGP)}. By adopting a per-dimension factorized categorical policy, IGP naturally aligns the policy support with the search candidates, eliminating the projection error and reducing the residual approximation to a bounded discretization error. To bypass the combinatorial explosion of high-dimensional discrete action spaces, IGP pairs Gumbel Top-$K$ candidate sampling with an MPPI-style multi-round iterative search to refine the policy at each state. Empirically, IGP achieves state-of-the-art aggregate performance and sample efficiency on DMControl and HumanoidBench. Furthermore, it produces more stable action sequences with lower temporal variance, demonstrating its practical potential for real-world robotic deployment.

In pairwise comparisons where each judge chooses between an idea generated by its corresponding TIGER-8B policy and one generated by GPT-4.1, the three LLM researchers select the TIGER-8B idea in 56.6%-68.4% of 76 comparisons per researcher. PhD1 and PhD2 select the TIGER-8B idea in 19 of 30 pairs (63.3%) and 17 of 30 pairs (56.7%), respectively. All six directed cross-policy comparisons likewise favor the judge's matched TIGER-8B policy over an off-target policy, supporting the use of individual preference feedback for tailored research ideation.

Key-value (KV) cache eviction methods rank tokens by attention scores, and each layer scores its own cache. To our knowledge, no eviction method asks which layer should score. We find, across twenty-one models, that the depth of the scoring layer governs eviction quality. On the eighteen models with per-layer caches, spanning 0.5B to 109B, long-context retrieval accuracy collapses at shallow layers, peaks between 0.4 and 0.75 of depth, and degrades at the last layer. The curve arises because a layer judges tokens well only once it has used the question to find the answer, which takes several layers. Modeling eviction as a top-$k$ cut, we compute answer survival probability from prefill attention and reproduce the measured curves. This approach selects the scoring layer using calibration prompts with known answers, and the selected layer is the best layer measured on most models and never significantly below it. It improves on the own-layer default on sixteen of the eighteen models, with significant gains on seven, all at 14B or below, while on larger models the default is already close to the best mid-depth layer. Scoring every cache from the selected layer while keeping one set of tokens for all heads also beats per-head SnapKV by 18 to 43 points on five models up to 14B under tight budgets. This gain carries over to financial filings at 32K context, all at the same memory budget, with no extra forward pass. The same curve explains a penalty in shared-cache models, where several layers read one cache and the layer that writes it scores by default. On Gemma-4-E2B the writer is shallow and loses 28 points to a mid-depth reader at a 10% budget. We also find that in open-book question answering (LongBench) no mid-depth layer beats the own-layer default.

RL locomotion policies trained under nominal simulation conditions may encounter distribution shifts that compromise safety. Predictive monitors could anticipate violations, but their world-model forecasts may become unreliable under shift, and ensemble disagreement is not calibrated. We evaluate predictive Signal Temporal Logic (STL) monitoring for frozen Ant-v5 and HalfCheetah-v5 policies using closed-loop world-model rollouts and a three-model ensemble. Across external-force, observation-noise, and mass shifts, predicted robustness ranks HalfCheetah's future STL violations strongly, whereas disagreement does not reliably track observation prediction error or violations. Under severe noise, violations remain common ($64.2\pm24.4\%$ at $K{=}20$ across three PPO seeds), yet current-state robustness without prediction comes within $0.03$ AUROC of predictive robustness. Penalizing predicted robustness by disagreement does not consistently improve ranking. Ant provides complementary per-specification evidence, but its composite target has no positive violations and its any-violation target is single-class in most evaluated seed-condition slices. Overall, predictive robustness provides useful STL-specific ranking but little demonstrated additional ranking value over the current state; these findings caution against treating ensemble disagreement as a reliable indicator of safety uncertainty.

Discrimination, calibration, and ranking are the three pillars of classifier evaluation, yet no single algebraic language spans all three. A recent indicator-matrix account unifies discrimination measures (precision, recall, F1, Matthews correlation) across binary, multiclass, multilabel and related settings, using Boolean cell operations on indicator label matrices, generalised to soft labels by t-norms. Calibration measures (Brier score, log-loss) and ranking measures (nDCG, AUC) fall outside it. We close the gap with two further classes of cell operators: deviation operators phi(y - yhat), which give a Calibration Framework, and rank- and slice-dependent operators (positional gains for DCG, column concordance cells for AUC), which give a Ranking Framework. Both inherit the micro/macro/exemplar aggregation architecture. For calibration, the three Brier aggregations collapse to one scalar; the Murphy decomposition holds in matrix form (unbinned, binned and isotonic); and, within the Brier-type class-indexed family, a measure is skew-invariant if and only if it depends only on the class-conditional errors alpha_j and beta_j, the calibration analogues of recall and specificity. For ranking, raw DCG collapses across aggregations while nDCG does not; per-class nDCG and per-class AUC are related by a row-column duality; micro-AUC differs from pooled ROC-AUC; raw positional scores are never strictly proper, and nDCG and AUC need not be proper. The skew-invariant per-class building blocks share one class-conditional expectation template. Experiments on two multilabel benchmarks illustrate the distinctions. Many of the results are machine-checked in Lean 4; the proofs, the experiments and the aggmetrics.scores package are public: https://github.com/rcprati/discrimination-calibration-ranking

Repeated satellite imagery permits understanding of how our planet changes over time, letting us predict its future and plan mitigating actions. Most existing studies that rely on a time series of satellite observations forecast the spectral values at each pixel for a future target date. However, clouds, shadows, and sensor noise often corrupt these observations, so that spectral forecasters waste much of their capacity on low-level features or noise rather than real change. With the advent of Earth foundation models, we argue that forecasting in embedding space is a feasible and more effective alternative. To this end, we design a token space that handles irregular histories from different sensors and foundation models and supports forecasts at any query date. Tokens from this space are inputs to our proposed forecaster, E$^2$Forecast, which is a transformer with two branches: a Gaussian branch that directly predicts a mean point forecast and per-dimension standard deviations, and a rectified-flow branch that generates forecast ensembles. On Sentinel-1 and Sentinel-2 benchmarks, forecasting embeddings instead of spectral values halves the cosine distance (from 0.20 to 0.10) and greatly reduces overconfidence, lowering the z-score standard deviation from 2.63 to 1.22. By comparing new observations with its forecasts, E$^2$Forecast can also detect anomalous changes without labels or retraining, raising F$_1$ from 41.58 to 48.70 on a global benchmark. This consistent superiority is validated on nine remote sensing foundation models. Beyond change detection, we believe our approach advances the field by permitting effective reuse of diverse pretrained task heads. Our code is publicly available at https://github.com/ucam-eo/e2forecast.

A vote over a model's samples can only help when the samples disagree. For open-ended generation the binding question is therefore not how to combine samples but whether a model produces samples worth combining, a question test-time scaling has answered for math but not for free-form text. We study it at small scale with semantic cluster voting, a judge-free aggregator that groups a small instruct model's samples by meaning and returns the largest cluster's representative, with no reward model or self-judging. On a 250-prompt benchmark with three tiers of answer checkability, the same pattern holds across three model families (Qwen2.5-3B, Llama-3.2-3B, Gemma-2-2B) and two generation seeds: voting raises accuracy on the math control, where samples vary but stay checkable (a gain of +10 to +29 points across the three models), yet adds at most +0.30 judge points on open-ended text. On a larger 200-problem MATH-500 run with the primary model, the math gain is +8.5 points (95% CI [+4.0, +13.0]). Almost every open-ended sample falls into one meaning cluster, so there is little to select over. The collapse is consistent with instruction-tuning rather than model size, and raising decoding temperature does not recover usable diversity, because it costs more quality than it adds (net -0.64 judge points, 95% CI [-0.93, -0.37]).

Coding agents are autonomous developers. Given a request, they change the code and report back when done. As such, they can handle not only independent requests but whole chains of them, each request built on the last. Agents are known to report wrong work as done, yet benchmarks still test them one request at a time, on human-written code, checked once. We build BOBA-Bench, 280 chains of 20 consecutive requests drawn from 140 real open-source projects, and evaluate ten coding agents on both. All ten do far worse on chains. On average, they complete 95% of requests taken one at a time, but when requests are chained together, they complete 51%. Worse code is not the cause. Once checked, agent code is nearly as good a base for later requests as human code. The cause is the mistakes agents miss. Coding agents declare a request done once their own tests pass, even when those tests miss a mistake. The next request then builds on broken code and fails about twice as often. The mistake is rarely found later. Nor can agents catch such mistakes by reviewing code. Asked whether an implementation is correct, they are right 53% of the time, barely better than a coin flip. Coding agents are therefore heavily dependent on access to verification. A single external ``that's wrong'' after a failed request and one repair round lift chain completion from 51% to 59%, and a repaired mistake no longer hurts the requests that follow. In simpler terms, coding agents are only as reliable as the checks they can reach: without verification, every request that follows is built on broken assumptions.

Blockwise learning can remove online gradient dependencies across depth, but independently optimized blocks need not compose into a useful model. We present ZeroGraph, a systems transformation for pretrained-model conversion and distillation that materializes a frozen teacher's coherent boundary trajectory as a versioned dataset. Each partition then trains with disjoint parameters, optimizer state, random state, and checkpoints, replacing live cross-depth dependencies with persistent supervision. Experiments separate learning quality from systems behavior. On GPT-2 Large, an unanchored objective reduces denoising error by 99.44%, yet assembled held-out cross-entropy worsens from 3.63 to 10.48 and context use becomes negative. A six-arm TinyLlama study identifies clean teacher-boundary regression as the successful interface; the tested noisy objective fails with cached or online targets. The clean recipe passes quality and context-use gates across three Qwen2.5-1.5B seeds and on a 3.086B Qwen2.5 model. In an objective-matched comparison, cached and online targets are bitwise identical. Cached jobs are faster after materialization, but a 618.31 s cache-build cost prevents wall-time break-even through 16 measured reuses. Boundary-shift tests show tolerance to intermediate representation mismatch, and an interrupted block resumes to a bitwise-identical final state without rerunning unaffected blocks. In the tested pretrained-model conversion and distillation setting, cross-depth semantic dependence can move from online optimization into persistent data.

Accurate diabetes prediction is critical for early clinical intervention, yet existing machine-learning pipelines often fail due to class imbalance, noisy physiological measurements, and redundant or weakly informative features. To address these unresolved limitations, this study establishes a unified fuzzy-evolutionary-ensemble framework designed to uncover a more discriminative feature space for type 2 diabetes prediction. The workflow integrates selective fuzzification of nonlinear clinical variables, SMOTE-ENN-based balancing and noise filtering, and Bat-Algorithm-driven evolutionary feature selection prior to training a Random Forest classifier. Applied to the Pima Indians Diabetes Dataset, the framework compresses the expanded fuzzy-numerical space into an optimized 10-feature subset and yields substantial performance gains: 96.11% accuracy, 95.14% precision, 98.00% recall, and a 98.23% F1-score. Relative to a baseline Random Forest using all 16 features (74.45% accuracy; 59.86% F1), the system achieves more than a 1.6-fold improvement in F1-score while preserving over 99% stability in majority-class detection and surpassing earlier published 16-feature models impacted by imbalance and weak feature relevance. Collectively, these results demonstrate that the integrated pipeline produces an enhanced fuzzy-numerical representation space that improves diabetes risk stratification and provides a scalable, uncertainty-aware foundation for diagnostic modeling across broader clinical datasets.

Large collections of manuscripts and research notes offer ways to reuse arguments, while making it harder to retain hypotheses, check claims, and understand how results were obtained. We propose Recursive Reasoning Exploration (RRE), a framework for cumulative mathematical research through evolving reasoning graphs. RRE links claims and jointly required premises to research targets, attempted arguments, counterexamples, checks, and revisions. Each investigation can generate further targets, and its supported outcomes become available to subsequent investigations. This is Phase III of the Math Instruments program: identifying mathematical instruments, reconstructing reusable reasoning, and organizing recursive research from that reasoning. The framework supports human-directed, collaborative, and goal-directed autonomous exploration, with open-ended exploration as a further mode. Its records also provide material for learning how mathematical investigations develop. A complexity research case study establishes an exact theorem: for every complex tensor of order n >= 2 and rank greater than two, the minimum Nullstellensatz certificate degree for its original two-term decomposition equations is 3n. This includes an infinite family whose every flattening has rank at most two. The exploration develops an explicit boundary certificate, a balanced transfer lemma, and an exhaustive argument using classical rank-two geometry. A parallel branch gives bounds 3m + 3 <= delta_m(I_m, N) <= 3m + 6 for every nonzero nilpotent m by m matrix. Six graph revisions record the progression from inherited results to these theorems, with exact arithmetic checks and a replayable exploration history. We present the RRE representation and protocol, the mathematical proofs, visualizations of graph evolution, and a research agenda for shared human-AI exploration, cumulative reuse, and education.

Inner-speech decoding from electroencephalography could support non-invasive communication for people with severe motor impairment, but comparison across studies is hindered by dataset- and protocol-specific evaluation. We present a controlled cross-dataset study using one preprocessing pipeline, one within-subject five-fold cross-validation protocol, and a compact dual-attention Transformer (VC-DATNet; approximately 131,000 parameters) across three public corpora: Thinking Out Loud (10 participants), BISpeech (4), and SyncSpeech (3). VC-DATNet achieved per-participant sensor/source hybrid accuracies of 39.48% on the four-class Thinking Out Loud task, 20.60% on the eight-word BISpeech task, and 25.53% on the eight-word SyncSpeech task, for which we report the first decoding result. Sensor-space and cortical source-space representations were complementary, with each performing better for different participants; this pattern replicated across all three corpora and nine architectures. Moreover, ten decoders spanning convolutional, recurrent, temporal-convolutional, hybrid Transformer, and pretrained foundation-model families converged within a narrow accuracy range on every dataset. These findings indicate that representation choice should be individualized and suggest that performance on these public corpora is constrained more by signal quality and dataset scale than by model capacity. The released pipeline provides a reproducible benchmark for future inner-speech decoding studies

Offline-to-online reinforcement learning (RL) follows the pretrain-finetune paradigm: an agent is first pretrained on historical data and then adapted through online interaction. We study offline-to-online adaptation without offline replay, where the agent must improve using only the pretrained model and newly collected experience. A key challenge is to avoid catastrophic failure, a sharp early performance collapse during online fine-tuning, while still achieving strong asymptotic performance. Through controlled experiments, we show that removing offline replay can induce catastrophic failure, while retaining strong conservative regularization from offline RL can suppress final online performance. We also find that common stabilization strategies do not reliably prevent early collapse, and that the critic-side mechanisms we test are insufficient on their own. Motivated by these findings, we propose ANCHOR, a simple adaptation method that anchors the actor to the offline-pretrained policy early in online fine-tuning, anneals this constraint over time, and retains only a small residual conservative critic regularizer. We provide a theoretical interpretation showing that KL anchoring helps preserve critic reliability across nearby occupancy distributions, with a two-phase regret bound explaining when anchoring should be maintained or relaxed. Across FrankaKitchen, AntMaze, and Adroit tasks, ANCHOR substantially reduces catastrophic failure while maintaining strong asymptotic performance. Ablations show that KL annealing and the residual conservative weight are important for balancing early stability and online adaptation, while additional analysis shows that ANCHOR is most beneficial when the pretrained policy contains useful behavior to preserve. Code is available at https://github.com/Cho-Geonwoo/ANCHOR.

Diagonal state space models process long sequences at linear cost, yet the same transition is applied at every step of a rollout, so an unconstrained recurrence holds no quantity fixed and the state decays or grows without bound. Current designs constrain one quantity, and preserving one does not guarantee preservation of the others, so a model can hold its norm to machine precision while momentum drifts by hundreds of percent. We claim that the conserved quantities of a diagonal recurrence are layered, that each layer requires a strictly stronger condition on the operator, and that a unit-modulus spectrum reaches only the outermost. Acting on the claim, we enforce the conditions required for every layer by construction. The generator of the Kronecker-Cayley transition (KC) is a graph Laplacian over object subspaces, $\mathbf{H}=\mathbf{L}\otimes I_m$, advanced by a Cayley transform, and the step stays diagonal at $O(N^3+N^2m)$ cost. The conserved set is exactly a $2m^2$-dimensional family of quadratic forms, held at every parameter value, so the quantities and the dynamics the construction reaches are fixed before training. On four conservation benchmarks, KC holds total mass at or near the reference floor while every diagonal baseline loses it by orders of magnitude. We further compare against mass correction and find that two correctors fail once the underlying prediction diverges, which shows that assembly rather than correction is necessary for a guarantee that holds across the rollout. A free modal basis trained under a conservation penalty conserves only by freezing its spectrum, whereas the assembled operator conserves with every mode but one free to rotate. Beyond these benchmarks, KC holds the linear momentum of three $N$-body systems and, at the operator level, the conserved quantity in molecular dynamics, chemical kinetics, pharmacokinetics and power-grid swing dynamics.

Diffusion models are often inaccurate in critical, data-sparse regions of the target distribution. Sampling from restrained distributions, biased distributions used in classical enhanced sampling to concentrate data near chosen configurations, can control where training data are collected but biases the learned distribution. We establish that this bias cancels exactly when forward noising from restrained data is conditioned on the restrained configuration at an appropriately chosen intermediate time. This allows a pretrained model fine-tuned on restrained data to achieve high accuracy in any chosen region of the target distribution, regardless of its probability, using standard denoising or flow matching objectives without importance weights. When no target data are available, Gibbs sampling enables inference from a model trained exclusively on restrained data.

Machine learning models have shown exceptional accuracy for ballot classification in election systems. Recent research has demonstrated that these ballot classification models are vulnerable to adversarial example attacks. This is a critical security issue because a successful attack could alter the outcome of an election and undermine the democratic process. However, all adversarial example attacks in the voting domain have assumed a white-box threat model, where the classifier parameters are known. In this work, we advance the field of machine learning in election systems by analyzing a realistic black-box adversary. We implement and evaluate eleven state-of-the-art black-box attacks: query based attacks ADBA, Square, SurFree, RayS, and SA-MOO. We also test adaptive transfer based black-box attacks with APGD-$l_\infty$, APGD-$l_2$, APGD-$l_1$, PGD $l_0$, PGD $l_0$+$l_\infty$, and PGD $l_0$+$σ$. Our attacks span multiple different norms ($l_\infty$, $l_2$, $l_1$, and $l_0$) and three different black-box adversarial threat models (decision based, score based and transfer based). In addition to digital analyses, we physically print and scan $20,000$ adversarial examples to evaluate the real-world performance of different black-box attacks. Our finding demonstrate an even larger physical-digital robustness gap exists for black-box adversarial examples than for white-box adversarial examples. Lastly, we conduct the first ever adversarial example transferability study in the voting domain and test adversarial training and barrier zones as two possible black-box defense strategies.

Current alignment training cannot be expected to deliver what it is meant to deliver: AI agents that comply with norms whatever the gains and whether or not anyone is watching. Training that rewards and penalizes behavior can teach norms only through that reward, and a reward can encode a norm only as a price: "do not do X" becomes "doing X costs something", and training charges the cost only when X is noticed. A price, unlike a norm, can be paid: a large enough gain outweighs it, noticed or not. And because the cost is charged only where violations are noticed, training cannot favor a policy that always complies over one that complies only where it would be caught. When an unnoticed violation is the only way to complete a task, it even favors the latter, the more so the more strongly it rewards task success. Agency makes this the normal case: agents mostly act without oversight, can often tell when they are observed, and may take steps to evade observation. That agents behave as intended when they infer they are being tested, and differently when they infer they are not, is therefore no malfunction; the training regime is working as built. This reading unifies alignment faking, sandbagging, and evaluation-aware scheming. No behavioral evaluation can show that a model does better, since that would require observing how it behaves unobserved; a model that complies when it seems unobserved may simply have recognized the test. Easing the pressure on agents to complete tasks at any cost can make violations rarer, but never impossible. For norms that no gain should be allowed to outweigh, the remedy is not deeper internalization but architecture that makes violations unavailable rather than unchosen. We state this argument explicitly, premise by premise, read recent findings and a real-world incident against it, and draw consequences for design and governance.

Learning accurate value functions is central to reinforcement learning (RL), yet remains challenging because supervision is limited and indirect: ground-truth values are unavailable, and each transition in the data only provides an immediate reward that can be used to construct bootstrapped targets as supervision. This challenge is particularly pronounced when data is limited or reward signals are sparse, as is common in real-world applications such as robotics. We argue that the key to accurate value learning in this limited data regime is extracting richer learning signals from the available data. In this paper, we introduce Value Forcing, a new approach that uses flow matching to estimate return distributions and enables accurate value learning from limited data. Our key insight is that we can extract richer learning signals by enforcing Bellman consistency at every noise level. Empirically, Value Forcing improves the accuracy of estimating the return distribution, and prevents the distribution from collapsing to a single value, as occurs with prior methods. We instantiate Value Forcing with two practical value-learning algorithms: VF, an actor-critic method, and IVF, an implicit value-learning method. We evaluate both algorithms across $26$ long-horizon tasks in the offline RL setting and find that they outperform prior value-learning methods, with particularly large gains when data is limited. Our project website is available at https://chongyi-zheng.github.io/value-forcing.

Coding agents can destroy data, leak secrets and message the wrong person faster than a human can review. Could a decision model catch these risks at runtime? We evaluated 22 hosted decision models and two prompted general models as classifiers for a multi-rule policy engine. A synthetic dataset of 2,643 tool calls was labeled blind against 28 rules, with a sealed holdout. With a single shared preamble in both conditions, explicit boundaries written as examples and counterexamples raised recall by 5.5-26.1 percentage points on all 23 available models and reduced false positives on 21. Jev 1.14's recall rose from 89.3% to 98.2%, while unnecessary interruptions fell from 30.6% to 4.8% of 252 clean actions. With examples, its median latency was 120 ms at $0.46 per thousand requests. Strong models retained the recall gains when concrete names were replaced by generic descriptions; three largely retained them when the same boundaries were written as prose. On 16 new scenarios written after the initial studies, a follow-up improved recall on 22 of 23 models and reduced unnecessary interruptions on 21. We release the dataset, rule texts and harness.

LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.

Backpropagation is the standard method for computing the gradients that train neural networks, but it is usually presented in a compact vector form that skips the intermediate steps. This tutorial derives backpropagation step by step for a fully connected feedforward network with sigmoid activations and a squared-error loss. Using a running example with three layers, we first differentiate the error with respect to a single weight by following its path through the network with the chain rule. We then show how summing over all paths leads to the error terms $δ$ and to the four general backpropagation equations in matrix form. Every term is annotated with its dimensions, so each equation can be checked for consistency. The derivation is then extended to $m$ training examples, where the weight gradients become a single matrix product and the bias term corresponds to broadcasting. It is also extended to mini-batch gradient descent, including why the gradient of a mini-batch estimates the full gradient. Batch and mini-batch training are both given as algorithms. Finally, each equation is linked to the corresponding line of a NumPy implementation trained on the AHDBase dataset of handwritten Arabic digits, and the complete code is included.

Personalized federated learning (PFL) may improve client-level performance, but deeper personalization increases deployment cost. We formulate personalization-depth selection as a utility-based empirical decision problem in which client utility equals Macro-F1 minus a cost-sensitivity weight times deployment cost, over six personalization depths labeled P0 through P5. We compare fixed policies with an offline per-client Oracle reference and report allocation maps over Dirichlet data heterogeneity and cost sensitivity. Although evaluated on this hierarchy, the formulation can accommodate any finite candidate set. On CIFAR-10, CIFAR-100, and a PTB-XL subset, personalization shows empirical, setting-dependent decision regimes rather than a universal benefit: higher cost sensitivity shifts Oracle mass toward the global-only depth P0; low cost sensitivity can favor deeper actions. Patterns are grid observations (three seeds per cell), not theoretical boundaries. Best-uniform is competitive with Oracle assignment (6.5 percent plus or minus 7.8 percent relative regret; 95 percent confidence interval 5.9 to 7.1), whereas random and personalize-all incur much larger regret (189.3 percent plus or minus 367.1 percent and 409.0 percent plus or minus 803.2 percent). Under the same utility, Ditto full-model personalization is substantially suboptimal once cost is counted. The contribution is an offline characterization of when personalization depth is worth its cost, not a new PFL optimizer or deployment-time scheduler.

Standard attention forms query (Q), key (K), and value (V) matrices through learned linear projections of input hidden states. Architectures such as GAU and Mega incorporate nonlinear transformations into attention. In this paper, we propose GatedV, a simple attention variant that applies SiLU only to V before aggregation without introducing additional trainable parameters. We evaluate GatedV against a standard decoder-only Transformer using matched parameter counts, paired initializations, and equal training-token budgets. At 2.33M parameters, GatedV reduces mean held-out perplexity from 59.1 to 57.9 (2.0%), with improvements across all five paired seeds. At 26.16M parameters, following an equal-budget learning-rate search, mean held-out perplexity falls from 20.2 to 19.7 (2.4%), with improvements across all three paired seeds. Small-model ablations show that the tested affine rescaling of values does not fully reproduce GatedV's gains. A short interleaved benchmark on NVIDIA DGX Spark measures approximately 0.7% higher full-context forward latency for the larger model, based on synchronized wall-clock timings. These findings suggest that value-only self-gating can improve language modeling performance without increasing parameter count or redesigning the attention-weighting mechanism.

Retrieval-based fact-checking systems ground verification in large document collections, creating a critical vulnerability: adversarial evidence contamination. We demonstrate that factually inconsistent but contextually coherent evidence generated to satisfy both retrieval-similarity and linguistic-fluency constraints can systematically mislead verification models. Contaminating just 3.1-7.4% of training corpora with five label-specific strategies causes adversarial content to penetrate up to 35.2% of top-k retrievals, degrading verification accuracy by up to 46.9% relative F1. Existing retrieval-stage defenses, validated against gradient-optimized or syntactically anomalous attacks, transfer poorly to this threat model: detection rates collapse and F1 recovery exhibits high variance across seeds. We propose GLAD (Geometric-Lexical Adversarial Defense), which fuses dense embedding peer-coherence with lexical distributional isolation. Across two benchmarks, two embedding models, and five random seeds, GLAD achieves the highest detection rates (70-86%) and the most consistent F1 recovery, reclaiming 73-80% of contamination-induced degradation on LIAR-RAW while approaching clean baselines on RAWFC. Crucially, unlike defenses that trade clean performance for robustness, GLAD imposes negligible cost on uncontaminated corpora.

An input guard decides whether a request reaches the model, but never sees what the model says. Harm that emerges during generation, because a jailbreak passed the input layer or a benign request drifted into operational detail, is visible only on the output side and only as tokens arrive. We present HaloGuard 2.0, an open-weight streaming output guard that assigns a safe/unsafe probability to every response token in a single forward pass, so a deployment can stop the stream once harm is confirmed. The guard reuses the constitution of HaloGuard 1.0, our input guard, to organise a 478,532-response training corpus, and supervises each token from the point where the response becomes harmful. We find that a guard trained on conventionally shaped data learns a shortcut specific to streaming: it judges the prompt instead of the response, and handles only 13.9% of matched safe/unsafe response pairs correctly. We correct this with a second phase on 106,043 counterfactual responses built around the base guard's own failure modes, and release the equal-weight average of the base and corrective models. On six English response-safety benchmarks, HaloGuard 2.0-0.8B reaches a macro F1 of 83.6 (83.3 on held-out halves of the decision-rule search), up from 72.8, with 38% fewer false positives. It matches the strongest full-response guards with 5-10x fewer parameters and is 2.6 points above the strongest streaming guard, Qwen3Guard-Stream-4B. A multilingual variant, HaloGuard 2.0-Poly-0.8B, trained on PolyGuardMix, reaches 83.1 F1 on the response labels of PolyGuardPrompts across 17 languages, 5.0 points above the strongest baseline. On one H100 the guard costs 0.37 ms per token per stream when moderating 64 concurrent responses. Both guards are hardened continuously through the weekly red-teaming protocol of HaloGuard 1.0, extended here to the output layer.

As agents increasingly undertake data querying, Text-to-SQL serves as a foundational capability for database access. However, for text-to-SQL in time-series databases, high-quality benchmarks and large-scale training corpora remain scarce, limiting the systematic evaluation and domain adaptation of existing models and agents. Existing automated SQL generation approaches lack sufficient autonomy, leaving key challenges unresolved. First, numerous combinations of temporal syntax and functions complicate comprehensive coverage of valid queries. Second, as the corpus grows, generating valuable and diverse new data becomes increasingly challenging. Third, schema and data heterogeneity hinder reuse of temporal query structures, limiting generation efficiency. To address these challenges, we propose TSSDA, a Time-Series SQL (TSSQL) Generation Data Agent framework, and construct TERN, a comprehensive Time-sERies text-to-SQL beNchmark. First, to achieve broad coverage of valid queries, we systematically model TSSQL syntax and semantics to construct a data generation space. Next, to explore the space and improve diversity, we develop a Data Agent that autonomously orchestrates data generation. The agent comprises a planning and decision-making module and a query generation module, with online adaptation for continuous self-improvement. Finally, we introduce a temporal-aware memory that enables adaptive reuse of experience across heterogeneous data instances, improving generation efficiency. Comprehensive evaluations on TERN reveal the limitations of existing methods in performing Text-to-TSSQL tasks. Furthermore, extensive experiments demonstrate that TSSDA substantially improves coverage, diversity, and complexity.

FOLLIE (for First-Order Logic for Language Inference and Entailment for French) connects French sentences, compositional first-order representations and automated natural language inference (NLI). The first French NLI-ready, NL-to-FOL resource, it contains 19,983 inference records and 36,987 non-empty sentence-slot--formula mappings from six heterogeneous NLI sources, retaining French lexical material and source problem structure. Its silver annotations are produced by a neural--symbolic pipeline, with 111 separately identified LLM-assisted additions. The construction combines morphosyntactic analysis, probability-conditioned categorial multitagging, grammatical proof search, discourse representation and first-order semantic construction. Stage-wise experiments show substantial gains in derivation coverage; analyses of conjunction distributions, logical-operator incidence and scope alternation characterise the resulting resource. We then develop an automatic inference engine whose input-conditioned background theories address lexical relations, quantification, quantities, event structure and temporal dependencies. Prover9 supplies proofs for entailment and contradiction, while Mace4 supports bounded consistency checks. Three-way accuracies are 65.37% on all 335 FraCaS-FR problems, 70.33% on 300 GQNLI-FR problems and 71.35% on 9,399 eligible SICK-FR records pooled across splits. Agreement with reference labels among committed decisions is 93.53%, 100.00% and 98.66%, at commitment coverage of 41.49%, 37.00% and 15.05%, respectively. Worked constructions, axiom schemas and proof analyses explain these outcomes and identify the contributions of scope, reference, lexical knowledge and numerical structure. FOLLIE provides a French experimental basis for studying the representations and assumptions that connect sentence interpretation to deductive inference.

Physics-driven design seeks high-performing geometries, which need not conform to a learned shape category. Parametric and latent-space approaches facilitate surrogate-based optimization, but restrict deformations to prescribed shape families or learned decoders. We introduce the Mesh-to-Shell Neural Operator (M2S-NO) as the mesh-to-shell component of a category-agnostic inverse shape optimization framework. We separate the explicit mesh used for design from the volumetric representation used for physics prediction. M2S-NO maps oriented surface samples to a surface-centered shell representation, such as TruncatedSDF or HeatSDF. A separate problem-specific PDE neural operator (PDE-NO) maps this representation to physical fields. Objective gradients update mesh vertices through both operators, without a latent shape decoder. Preconditioning and remeshing support free-form deformation, while fixed interfaces are controlled directly on the mesh. Optional shell-space proposals enable topology changes. We evaluate geometric matching, jet-engine brackets, and neutron moderators, with independent FEM and OpenMC evaluation. Relative to the best training designs, M2S-NO reduces maximum bracket displacement by 14.3% and increases neutron-sensor current by 13.8%. Optimization also succeeds from an out-of-distribution bracket initialization, demonstrating the framework's geometric flexibility.

When two companies connect their digital twins, software models of their plants kept up to date with live data, each twin must translate the other's data into its own terms, and that translation rule, called a mapping, can stop being correct when one company changes the meaning of its data without telling the other. This PRISMA 2020 systematic review screened 5,833 records and read 119 full texts to determine whether any published system detects such a broken mapping and repairs it while the twins operate. In all 32 included studies the mapping is fixed before operation, and the two ingredients of run-time repair appear only separately: cross-company matching without any check that the match still holds, and detection and repair of changed meaning within a single company. The review turns this gap into six requirements, each illustrated on one running example, that a system claiming run-time semantic interoperability between independently owned twins should meet.

We examine bounded retained history through a frozen predictive pilot and a subsequent allocation study on a curated UNSW-NB15 flow mirror. Both use an artificial release-order replay based on raw ltime ticks, with equal-tick records querying a shared prior-group state. In the pilot, the model with added retained relational features had lower reconnaissance recall than the current-flow-plus-counters baseline (43.66% versus 47.84%) and lower benign false-positive rate (3.63% versus 6.68%) at separately frozen operating thresholds, together with lower AUROC. This mixed result did not establish a predictive advantage. A development-only audit subsequently found that capacity eviction dominated expiration in the 256-record, 300-tick FIFO state, with a median completed record lifetime of 11 ticks. A protocol specified before its single execution, but after the pilot and post-hoc mechanism audit, compared five label-blind retention arms on the already-used development trace. Its primary equal-source endpoint averaged retained/reference ratios for service-expanded source out-degree and source-destination service diversity. Source-balanced retention at 256 records exceeded FIFO-256 by a pooled endpoint difference of 0.440738 and met the prespecified materiality and subblock rule. Crucially, the event-weighted ratios for both quantities decreased relative to FIFO-256: the result is specific to the equal-source weighting objective and does not establish uniformly improved retention. The study supplies neither evidence of detection improvement nor an explanation of the pilot result. Reused data, concentrated reconnaissance support, proxy timing, and the absence of a qualified independent corpus bound the interpretation.

Bursty serving changes both queue states and the population selected for each model. We study conditional calibration contracts as a specification to be validated, rather than a certified property of an existing router. CA-MHD combines load-coupled distillation, closed-loop replay and a cost-informed Schmitt router in a discrete-event simulator with a hardware service table. Across 40 cells (ten training seeds and four content-load coupling levels), we distinguish the original teacher, frozen wide heads on an adapted trunk, and the narrow student. A common teacher-trajectory audit has median absolute student-teacher deficit 0.06 percentage points, but does not certify the student-selected deployment support. Historical results use a kernel correctness gap, not the published SmoothECE; we separately prove selected-branch composition for a residual functional at a common fixed scale. The deployed wide branch has lower ordinal accuracy than the original teacher in all 40 cells. Same-checkpoint student-only serving has lower historical ordinal gap in 27/40 cells; routing also exposes violation tails up to 17.5 percent. Student p99 latency remains above the 1.5 ms target. Retrospective temporal recalibration improves the ordinal gap in 39/40 cells, but causal delayed-label availability is unverified and no incremental per-mode benefit is detected after multiplicity adjustment. An in-sample CUSUM detects 30/40 checkpoint changes at five historical false alarms, with no prospective false-alarm guarantee. These findings support a branch-aware audit and a scoped negative result, not demonstrated routing superiority. Independent budgets, selected-support validation and net utility under a satisfied SLA remain open.

Concept-based explanations of melanoma classifiers are rarely tested against controls with known targets. We built a counterfactual audit from a diffusion model conditioned on seven 7-point checklist concepts and tested it with such controls. Colour-calibrated start noise reduced colour outliers from 78.5-84.8\% to 6.6-13.3\%. Four of seven knobs were target-dominant but also changed other concepts. Quartile-based positive controls were confounded, and the audit failed its pre-specified rule. It recovered the targets of all 16 controls matched on the generator's pseudo-labels and of 12 of 12 controls labelled by a second evaluator, for three concepts; concept activation vectors recovered fewer (11 of 16). In 18 permuted-label classifiers, significant sweep responses and target rankings also arose, whereas standardised effect margins separated the controls from these null classifiers. The pigmentation sweep raised the melanoma logit in 15 of 15 ISIC 2020 classifiers (+3.12 on average) and 6 of 6 HAM10000 classifiers (+4.01), about four times the null response in scale-free units; the regression effect did not transfer. An exploratory decomposition, evaluated against the controls, suggested that knob cross-talk partly hides vascular responses. Matched controls and null classifiers are practical checks of concept audits.

Automated classification of retinal pathologies from Optical Coherence Tomography (OCT) images remains challenging because diagnostically relevant morphological patterns may exhibit substantial inter-class similarity. Although high classification accuracies have been reported on benchmark OCT datasets, cross-study comparison is limited by heterogeneous evaluation protocols and by the limited characterization of compact convolutional neural network (CNN) baselines trained from random initialization. This study addresses this gap through two contributions. First, it provides a focused review of CNN-based retinal OCT classification and identifies methodological considerations relevant to controlled benchmarking. Second, it presents ED126-A, a purpose-built four-class CNN baseline trained from scratch under a documented experimental protocol on the Kermany labeled OCT dataset. The model contains 1,241,380 total parameters, including 1,239,460 trainable parameters, and is evaluated using training-dynamics analysis, class-wise performance metrics, confusion-matrix analysis, confidence analysis, t-SNE representation-space visualization, Grad-CAM, hierarchical feature-map analysis, and one-vs-rest ROC/Precision-Recall analysis. On the image-level held-out test set, ED126-A achieved 95.03% accuracy, with a macro-F1 of 0.9205 and a weighted-F1 of 0.9521. The principal class-conditional limitation occurred for Drusen, which attained a recall of 0.9605 but a precision of 0.7493 and an F1-score of 0.8419. The confusion matrix identified Normal to Drusen and CNV to Drusen as major error channels. These results support the use of ED126-A as a documented from-scratch reference baseline and provide an empirical basis for controlled evaluation of transfer learning, increased depth, attention, multi-scale feature fusion, and calibration in subsequent studies.

On-policy distillation (OPD) trains a student on its own generations by matching the teacher's token probabilities. These probabilities entangle two kinds of information: which tokens the teacher prefers, and how confident it is. The latter reflects the teacher's own capacity and calibration, which a smaller student may be unable to reproduce. Because the teacher temperature changes these values without changing the ordering, OPD is also sensitive to the temperature choice. We ask whether the teacher's ordering of candidate tokens is enough, and propose Order-Only On-Policy Distillation (O$^{3}$PD). At each student-generated prefix, O$^{3}$PD trains the student to reproduce the order of the teacher's top-$K$ tokens, without fitting their probability values. The objective is therefore invariant to the teacher temperature and needs no temperature tuning. Across six teacher--student pairs, O$^{3}$PD outperforms OPD on the average over HMMT25, AIME24, and AIME25 in every pair, by up to 2.9 points, with gains of up to 5.6 points on AIME25. Other ranking losses yield similar gains, so the benefit comes from ordinal supervision itself. Analysis shows that O$^{3}$PD departs further from the teacher's probabilities than OPD, yet reproduces its ranking more faithfully and yields a stronger student. The ranking, not the probability values, thus carries the knowledge that on-policy distillation needs to transfer.

The increasing penetration of Distributed Energy Resources (DERs) in the smart grid has significantly expanded the cyber-attack surface. Traditionally, DER networks rely on static role-based access control (RBAC) mechanisms that assume persistent trust after authentication. This limitation enables adversaries to maintain unauthorized privileges and move laterally across DER systems, disrupting grid operations. Although attribute-based access control (ABAC) enables fine-grained and context-aware access decisions, it does not fully eliminate implicit trust. Zero Trust (ZT) principles can be incorporated to enhance ABAC policies. This paper proposes a dynamic access control framework for implementing ZT in DER networks (ZT-DAC). The proposed solution employs a two-stage verification process and integrates an Open Policy Agent (OPA)-based RBAC with a machine learning (ML)-based ABAC to enforce continuous and adaptive trust score assessment. To support realistic DER operations, a DER-specific access control model is developed and used to generate a domain-informed dataset of access requests incorporating both legitimate and adversarial behaviors. Experimental results on the dataset show that the OPA-based decision layer achieves \(91.92\%\) accuracy, with a False Positive Rate of \(14.45\%\), while the ML model achieves a Mean Absolute Error of \(2.074\) and an \(R^2\) of \(0.891\) in trust-score prediction. The ABAC layer achieves \(93.27\%\) accuracy, with a FPR of \(3.88\%\), validating its practical feasibility within realistic DER environments.

Effective multi-agent collaboration requires finalizing solutions from reasoning that is distributed across the individual attempts of many agents. However, useful intermediate reasoning can remain unexploited when collaboration is limited to voting on conclusions reached among individual agents, or aggregating long, noisy collections of reasoning trajectories in context. In this work, we propose TTT-Consolidate, a framework that uses multi-agent reasoning trajectories as adaptation data to improve a model's subsequent reasoning. TTT-Consolidate jointly consolidates trajectories into parametric representations through test-time training, then generates a new solution by conditioning on the consolidated information. We meta-learn how to consolidate using bi-level optimization at training time, where an inner loop encodes reasoning trajectories into a lightweight LoRA adapter, and an outer loop learns meta-parameters for the inner loop such that the consolidation is helpful for the model to synthesize a final solution. As no large-scale corpus of diverse reasoning trajectories exists for training aggregation methods, we construct the OpenParallelThinking corpus composed of 12,810 math problems paired with a mixture of trajectories from six LLMs. Evaluated on challenging math and STEM benchmarks, TTT-Consolidate outperforms the best aggregation baselines by up to 20 percentage points, even under matched compute and runtime settings, with the gap widening at larger model scales. TTT-Consolidate also recovers from unsuccessful individual trajectories more robustly, an ability that improves as the number of available (even incorrect) trajectories increases. Finally, we demonstrate that TTT-Consolidate benefits from both SFT and RL-based outer loop objectives. We will release TTT-Consolidate and OpenParallelThinking for use in future work on reasoning aggregation.

Scaling computer-use agents from individual actions to complex software work requires experience that connects decisions to complete outcomes. We present NorthCUA, an integrated system for turbocharging computer-use agents through experience at scale. Its Experience Flywheel supplies diverse executable tasks to the NorthCUA Harness, which sustains long interactions, records decisions, and checks task outcomes. The Flywheel uses these records to curate verified trajectories. For unsolved tasks, recursive self-improvement evolves reusable skills and preserves verified milestones so later attempts build on earlier progress. Staged Learning combines this experience with grounding data from the Flywheel to develop GUI competence through mid-training and long-horizon supervised fine-tuning. The policy then learns from fresh attempts in the same Harness through joint desktop-web reinforcement learning, closing the loop between experience production and policy improvement. The resulting Nano, Mini, and Pro family achieves state-of-the-art results among open-source models across GUI grounding, desktop execution, and Web interaction under the reported evaluation settings. For more responsive Web workflows, NorthCUA-Ultrafast pairs direct action selection with open-ended generation on pre-RL Nano, offering a measured quality-latency tradeoff. Together, these advances establish NorthCUA as a high-performing model family and a reusable foundation for agents that carry complex software work from intent to completion.

Vision-language models for remote sensing answer questions about satellite images by generating text: every answer costs a decoding loop, has to be parsed back to an option, and comes without a usable probability. We present EarthJev, a 0.8B-parameter model that answers typed questions about multispectral Sentinel-2 patches without generating any token. All 12 bands enter the vision encoder of Qwen3.5-0.8B through a widened input layer initialised from the RGB kernels, and, following the open OneJev recipe, the answer is read from the logits of the option letters at a single position and trained with cross-entropy plus Brier score. Under a pre-specified analysis plan, we compare it with the same backbone fine-tuned to generate the answer and with three multimodal language models fine-tuned with LoRA, all trained on the same 70,890 questions from 5% of BigEarthNet v2.0 and timed in one job on the same GPU; VQA models that classify over a fixed answer vocabulary are reported as a reference for the cost of answering only the questions seen in training. EarthJev answers a question in 19.9 ms, 4.3 to 8.7 times faster than the multimodal language models and 6.8 times faster than its own backbone used generatively; when all the questions about an image are asked together, it answers them in 42.8 ms, 3.7 to 10.1 times faster than the batched generative models. On 36,028 test questions it reaches 86.0% accuracy (three seeds), 0.6 and 1.2 points below the two 2B multimodal models and 3.5 points below the remote-sensing EarthDial-4B, with answers that are valid by construction and calibrated probabilities (ECE 0.006). Replacing the nine non-RGB bands with their mean costs 4.5 points, so the model uses the multispectral input, and in an ablation training the whole vision encoder raises accuracy to 88.4% at the same cost. Code and model: https://github.com/CosmoIknosLab/EarthJev

On-policy distillation (OPD) trains a student language model on its own trajectories using token-level supervision from a stronger teacher. We show that standard sampled reverse-KL OPD systematically suppresses updates to high-confidence tokens, such that their weak gradient contribution does not necessarily imply weak or uninformative teacher supervision. We introduce Bernoulli On-Policy Distillation (Bernoulli OPD), which directly matches the teacher and student probabilities for that token. This yields a bounded probability space correction that preserves teacher-student disagreement while avoiding the additional confidence-dependent suppression induced by vanilla OPD. Fixed-rollout analyses show that Bernoulli OPD substantially restores optimization signal to high-confidence tokens while remaining sensitive to teacher-student disagreement. We further analyze the resulting model behavior through confidence distributions, answer calibration, policy movement, and training dynamics, showing that Bernoulli OPD's gains cannot be explained simply by longer reasoning, larger optimization steps, or more uniform imitation of the teacher. Code is available at https://github.com/duterscmy/Bernoulli-OPD.

Financial time-series forecasting is still difficult owing to the non-stationary nature of the market, regime changes, and catastrophic forgetting, which makes conventional deep learning models hard to adapt in the long term. To achieve uncertainty-aware robust financial forecasting in the face of changing market conditions, this study proposes a Hybrid Quantum-Classical Neural Network with Continual Learning (HQNN-CL). The proposed framework combines Gaussian Hidden Markov Model (HMM) based regime detection, BiLSTM temporal encoding, adaptive-depth parameterized quantum circuits, Elastic Weight Consolidation (EWC), experience replay, and probabilistic forecasting in a single architecture. The framework was tested on four real-time stock datasets: Microsoft, Systems Limited, Teradata and Tata Consultancy Services, and compared with seven state-of-the-art forecasting models, including recurrent and Transformer-based models. The experimental results show that HQNN-CL has outperformed other methods on the Systems Limited (RMSE = 3.0149) and Teradata (RMSE = 0.8909) datasets and is within 0.32% and 0.40% of the best baseline method for Microsoft and Tata Consultancy Services, respectively. The proposed framework also generates well-calibrated probabilistic forecasts within the range of 89.5% to 96.5% empirical coverage rate, has zero catastrophic forgetting (BWT = 0.0000) for all sequential learning tasks, and consistently shows positive forward transfer. Ablation and statistical significance analyses also validate the performance of adaptive quantum feature learning and continual adaptation under heterogeneous market conditions, regime-wise.

Machine learning models deployed in decision-making pipelines, from vehicle routing to dynamic pricing to clinical decision support, must remain reliable in the face of low-quality training data. We study a predict-then-optimize setting in which an upstream model predicts the cost vector of a downstream linear program, with a large pool of noisy training data supplemented by a small curated set of clean data. We introduce Reduced Cost Influence Functions (RCIFs), which combine influence functions from robust statistics with the geometry of linear optimization to estimate how each training point shifts downstream decisions. Building on this, we propose a reweighting algorithm that uses RCIFs to downweight harmful training points and reduce downstream regret. Experiments on random linear programs and shortest-path problems show that our approach reduces regret, and accompanying theory characterizes properties of RCIFs and identifies regimes favorable to our method relative to baselines.

We develop and validate a physics-informed neural network (PINN) framework for computing the tidal Love numbers of black holes directly from their perturbation equations. We consider static perturbations of the Kerr black hole, whose exact hypergeometric solutions provide a controlled benchmark for both the numerical solver and the extraction of the physical response. The framework combines exact enforcement of the horizon boundary conditions with an $H^2$-Sobolev loss that incorporates derivatives of the differential-equation residual. The Love number is then obtained by extracting the asymptotic coefficients of the learned solution, without using the analytic connection formula. Across the tested scalar, vector, and tensor perturbations, the method recovers the finite dissipative response with relative errors below $10^{-5}$. Controlled comparisons show that the higher-derivative loss terms improve solution accuracy by orders of magnitude, while independent tests of the extraction procedure distinguish solver errors from errors in recovering the response coefficients. The conservative response, which vanishes analytically, is constrained by the numerical error floor. These results demonstrate that PINNs can accurately compute black-hole tidal response and establish a validated framework for future extensions to frequency-dependent and non-separable perturbation problems.

Harness optimization can substantially improve LLM agents without modifying model weights, but existing methods primarily optimize response quality. Extending them to joint quality--cost optimization introduces a challenge: applying the deployment preference during intermediate updates can prematurely reject candidates that are valuable for later improvements. We introduce ParetoSaddler, which addresses this mismatch by separating intermediate candidate promotion from final deployment selection. At each iteration, an optimizer updates a parent harness to produce a candidate, and then compares the two over repeated executions and uses Bayesian bootstrap Pareto support to promote candidates showing credible progress in either quality or cost. Promoted candidates receive broader development evaluation and remain available as parents for later optimization steps. We evaluate ParetoSaddler on ClawMark and GAIA2 using domain-disjoint splits. On ClawMark, ParetoSaddler improves quality from 70.86\% to 81.35\% while reducing cost from \$1.17 to \$1.01 per task; on GAIA2, it improves quality from 56.22\% to 62.89\% while reducing cost from \$0.78 to \$0.45 per task. The Pareto-selected harnesses also achieve higher quality at lower cost than quality-only harness optimization. Analysis shows that Pareto-aware optimization retains intermediate candidates that would be discarded by deployment preference but can support later improvements, while Bayesian bootstrap screening filters out candidates whose gains are weakly supported across runs.

Diffusion driving vision-language-action models repeatedly execute the backbone at every denoising stage, incurring substantial inference cost, while it remains unclear whether all scenes require the same decoding computation. To distinguish when a trajectory becomes readable from how much further refinement it requires, we analyze trajectory readout across network depth and decoding progress. We identify a depth-decoding dissociation in which late-layer representations already support useful trajectory readout after the first forward sweep, whereas early-layer representations benefit more from continued decoding. We further find that both the preferred readout layer and decoding depth vary across scenes. Based on these observations, we propose a scene-adaptive readout and decoding framework. A scene-conditioned readout MoE adaptively weights and combines trajectory predictions from layer-specific late-layer representations after a single backbone sweep, while an adaptive decoding policy applies additional native refinement only when required by the scene. We evaluate our framework on two diffusion driving VLAs across the Waymo E2E and nuScenes benchmarks. Our method improves validation RFS by up to 0.56 while also improving test performance, and achieves an end-to-end speedup of up to $6.0\times$ on Waymo E2E, while consistently reducing trajectory error on nuScenes. These results demonstrate that scene-level readout and decoding allocation improves planning quality while avoiding unnecessary decoding. Our code is available at https://github.com/sjyu001/SARD

Agent decisions depend on empirical relationships that deployment observations reveal and task descriptions alone may leave unresolved. We introduce Decoupled Predictive Decision Making (DPDM), an architecture that formulates heterogeneous agent decisions as conditional prediction with a shared pretrained learner. Language models provide semantic representations of current and historical records, while a general-purpose conditional predictor predicts task-defined targets from the represented evidence. Pretraining supplies reusable inference capability; deployment evidence instantiates problem-specific predictions without updating either model's neural parameters. Tool judgments, candidate preferences, and reasoning assessments share an evidence-and-query interface, with predictions mapped to choices by a decision rule. This architecture establishes complementary sources of agent capability: semantic knowledge acquired through language pretraining and empirical inference acquired through structured-data pretraining. A single frozen LimiX-2 predictor serves all five task categories. With Qwen3.5-27B as the semantic encoder, DPDM raises mean accuracy across 50 benchmarks from 78.80% with few-shot prompting to 84.92% under their respective evaluated history budgets, and reaches 82.33% with identical histories and queries. On ToolPRMBench's BFCL and ToolSandbox tasks, absolute accuracy gains are 34.26% and 28.87%, respectively, at the methods' respective evaluated history budgets. Accuracy improves as history grows. With identical imbalanced histories, mean balanced accuracy improves by an absolute 14.51% across five reasoning-step assessment tasks and four encoders. DPDM makes learning from experience an independently improvable foundation-model capability, creating a shared route for advances in conditional prediction to improve decisions across agent tasks.

Coding agents solve repository issues through long sequences of exploration, editing, and validation. These trajectories are increasingly used as post-training data, so how they are collected shapes what the trained model learns. A common approach is to sample multiple trajectories independently for each issue and then select or curate them. However, as more trajectories are sampled independently, later trajectories become increasingly redundant, even though earlier trajectories could guide where subsequent collection effort should be spent. A natural way to use this history is to expose it to the coding agent as advice. But training on such trajectories can make the model rely on advice that is unavailable at test time, creating a train-test mismatch. Instead, we argue that history should guide where collection effort is spent, while the coding agent decides what to generate. We introduce SWE-Trail, a trajectory collector that uses earlier trajectories to guide subsequent collection while keeping that history outside the coding agent's context. During a new rollout, SWE-Trail uses history to identify a state where additional collection effort may be useful. There, the coding agent proposes candidate next actions from its ordinary context, and SWE-Trail briefly executes them from the same checkpoint to decide which route continues. We evaluate collection approaches by training student models on the trajectories they produce. For a 124B student initialized from Ling-3.0-flash-base, SWE-Trail improves mean resolution over independent sampling by 2.33 and 1.60 percentage points on SWE-bench Multilingual and SWE-bench Pro, respectively. Without history, the same pipeline produces more resolved trajectories than independent sampling, but improves by only 0.11 points on Multilingual. These results show the value of using history to guide collection without exposing it to the coding agent.

Retrieving objects from densely packed storage requires anticipating how contact disturbs surrounding objects when limited clearance prevents lifting the target over clutter. However, contact can propagate beyond the target's immediate neighbors, while accumulated prediction errors make learned dynamics unreliable over long planning horizons. We introduce DARE (Disturbance-Aware REtrieval), a geometry-aware relational dynamics framework that uses retrospective correction to improve long-horizon prediction for disturbance-aware retrieval. DARE encodes compact object descriptors and predicts action-conditioned disturbance propagation through object-level message passing. When a sparse observation becomes available, a learned corrector uses the prediction discrepancy to amend an earlier predicted state and reruns the intervening dynamics, propagating the correction through the predicted contact chain. These predictions guide a receding-horizon planner that balances retrieval progress against displacement and rotation of surrounding objects, permitting necessary contact while discouraging harmful disturbance. Experiments in planar and RoboCasa environments evaluate long-horizon prediction, generalization, and robustness for disturbance-aware retrieval in dense clutter. See our project website: https://the-back-to-the-future.github.io

Psychiatrists must be prepared to manage patients exhibiting challenging psychological states, including suicidal ideation, hallucinations, and other high-risk behaviors. However, opportunities to safely practice such interactions are limited. While recent conversational agents can simulate patients, existing approaches largely rely on large language models (LLMs) to determine patient responses, making it difficult to consistently maintain psychological state, adapt to the clinician's responses, and generate clinically plausible challenges that meaningfully test the clinician. We present a state-aware virtual patient simulation pipeline for training psychiatrists in challenging clinical scenarios. The proposed system maintains an evolving patient state throughout the interaction and orchestrates patient behavior rather than leaving response generation entirely to an LLM. At each interaction, a learning-based behavioral strategy selector determines the type and level of challenge appropriate to the current patient and clinician context. The selected strategy, patient state, clinician response, interaction history, and relevant clinical knowledge are then provided as grounded context to an LLM to generate a semantically relevant and psychologically consistent patient response. Following each session, the system analyzes the interaction and provides feedback to the clinician on their communication and management of the virtual patient.

Language and vision models became data-efficient by pretraining on plentiful unlabelled data; physical field models have such data, over ten million 3D shapes, but geometry pretraining designs its targets rather than deriving them from the physics. Euler-Lagrange geometry pretraining derives them: for attached flow, a shape enters the flow's Euler-Lagrange equation only through coefficients that distance and normal line (the normal up to sign) fix to first order. Their reconstruction error therefore bounds what a representation loses about any field, and fine-tuning on costly simulations need only assemble the global field. We realise it in AREALS, a masked surface autoencoder whose areal tokens carry position, normal and area, pretrained self-supervised and sign-free on 61,822 raw meshes. Because the operating condition enters only through the far field, sensors can supply it: AREALS completes a hidden field in one forward pass without condition labels. It matches the accuracy of the same architecture trained from scratch with half the simulations on DrivAerML and DrivAerNet++ and about a quarter on WindsorML. On DoubleDelta wings, 20 sensors give $3.9\times$ and $4.0\times$ lower field and force error than GABI, a Bayesian inversion method with a learned prior, and beyond the training angles pretraining halves the error.

Large language models show large and seemingly arbitrary performance drops when a task is rewritten in ways that leave its requirements untouched. We recast task difficulty as a causal quantity: the treatment effect of meaning-preserving perturbations on a model. A meaning-preserving rewrite can only be observed to hurt a task the model can already solve, so we place the estimand on the principal stratum where the degradation occurs: the conditional causal difficulty (CCD), the probability that a randomly drawn twin breaks a solvable task. Because harder tasks tend to be more phrasing-sensitive, complexity enters as an effect modifier rather than a confounder, and we residualize it out to obtain complexity-residualized difficulty (CRD); zero for a task exactly as fragile as its complexity predicts, positive for anomalously fragile ones. Current benchmarks do not supply the data CCD requires, so we introduce GlassGSM and GlassMATH, augmenting GSM8K and MATH with twins built from perturbation families, raising fragility on GlassGSM by $\approx 12\times$ over GSM-IC-style distractors on the same problems ($3.7\%\to44.9\%$) while keeping the answer key unchanged. Our predictor Frail learns the per-perturbation break probability and reaches $0.613{\pm}0.023$ of the suite reliability ceiling on GlassGSM and $0.557{\pm}0.023$ on GlassMATH, against $0.171$/$0.401$ for a surface-feature baseline and $0.444$/$0.463$ for supervised baselines such as VAP3, and, at a $5\%$ curation budget on GlassGSM, doubles random recall of fragile items, matching an oracle ranking. Additionally, cross-fitted residualization leaves CRD uncorrelated with the declared proxy in practice. Code is available at https://github.com/upwork/Conditional_Causal_Difficulty_for_LLMs

Embodied agents that answer questions about their route need a memory of what they saw, and scene graphs are a natural form for it. A vision-language model (VLM) can receive such a graph as text or through a graph encoder, but accuracy alone cannot show whether the graph holds the episode's facts, reaches the model, or is actually relied on. In this diagnostic study, we introduce a graph-swap test for the last of these: each episode is given another episode's graph, and we count how many answers change. On an embodied question-answering benchmark with routes through two to seven rooms, we use it to compare ways of delivering scene-graph memory to a fine-tuned VLM. As text, the memory is used strongly, adding 30 points on route questions. As a graph, it is not, and we trace why: the encoder preserves the information, but standard pooled and late fusion discard it. Giving each node its own input token recovers part of it, a gain of 8.3 points on route questions that the swap test confirms comes from the graph. The test also exposes a shortcut that accuracy hides: when the same facts appear in the text, the model ignores its graph. The test thus offers a cheap check of whether a VLM uses its scene-graph memory, and our analysis points to how that memory should be delivered.

We study finite-horizon reinforcement learning with a safety constraint on the probability of reaching a failure set. In contrast to the well-studied safe reinforcement learning literature, we consider problems where the probability of reaching the failure set is vanishingly small, but the consequences of reaching the failure set are catastrophic. Direct simulation and estimation of the optimal policy is naturally ineffective when the failure probability is extremely small. Moreover, on a real system the controller can only choose actions, and this is not enough to make failures frequent. In contrast to the rare event simulation literature, our goal is not to estimate the probability of the rare event occurrence, but to upper bound it such that the failure constraint is satisfied. Thus, instead of sampling failures, we use the large deviations principle. We define a damage variable that on exceeding a certain threshold indicates failure. We study the cumulant generating function (CGF) as its positive root provides an upper bound on the failure probability via Cramer's theorem. We extend this approach to Markov decision processes with hitting times through state-action CGFs, and give finite-horizon and state-dependent versions of the bound. The state-action CGFs do not depend on the policy and can be estimated from regular transitions, so the bound can be checked for any policy a priori. This gives a policy improvement algorithm in which every policy satisfies the failure constraint with high probability.

Gradients are no longer used only to train models: Gradient magnitudes are now measurement instruments, used to rank attention heads for pruning, to decide per-example clipping in private training, to allocate adapter rank budgets, and to debug models. Across these workflows, larger gradient magnitudes are taken to mean greater importance. Even for a fixed input--output function, these gradients are computed in one of many possible parameter coordinate systems. For example, attention scores depend on the query and key projections only through the paired product $W_QW_K^\top$. Scaling a head's query projection by $c$ and the corresponding key projection by $1/c$ therefore leaves every output logit unchanged. This change in parameter coordinates rescales the corresponding query and key gradient norms by $1/c$ and $c$, respectively. To test whether gradient-based decisions remain unchanged under these coordinate changes, we introduce Gradient Reliability under Reparameterization (GRR). GRR rewrites a trained model in new parameter coordinates that compute exactly the same function, verifies that predictions are unchanged, and then tests whether the decisions derived from the gradients are unchanged too. Our audits span GLUE classifiers, a ViT on CIFAR-10, GPT-2-medium and GPT-2-large, a 1.1B-parameter Llama-style model on WikiText-2, and six trained LoRA adapters. Empirically, predictions are preserved throughout, but both the gradients and the gradient-based decisions change. Across 66 GLUE attention audits, the median top-3 overlap of raw head-gradient rankings is only $1/3$, and rankings and pruning masks change even under the mildest rescalings we test. In the largest of 14 global-pruning audits, the shifted ranking selects head masks that differ in four of 22 pruned heads and reach $50.92\%$ and $94.84\%$ validation accuracy on one ELECTRA SST-2 model.

Reinforcement learning (RL) is a machine learning paradigm in which an agent learns a decision-making policy by interacting with an environment, guided by a scalar reward rather than explicit supervision. Supervised and unsupervised learning have become indispensable tools in high-energy physics (HEP), but they learn from fixed datasets, whereas many HEP tasks are closed-loop decisions: the data an algorithm observes depend on its choices, the feedback is sparse or delayed, and the objective is often measured on a real machine or in a stochastic simulation, not by a differentiable model. This survey reviews the application of RL to such tasks: accelerator control and feedback, data quality monitoring, jet grooming, clustering, and representation, charged particle tracking, beyond-the-Standard-Model (BSM) and string model building, detector design, trigger control, analysis, and formal quantum field theory computation. It provides a self-contained introduction to the underlying RL methods, from model-free and model-based to multi-agent, offline, inverse, and bandit approaches, reviews each domain against its traditional methods, and summarizes the head-to-head comparisons. The applications range from policies deployed on real accelerators and evaluated on recorded collision data to exploratory studies in formal theory. RL is most advantageous for repeated tasks with a fast simulator, decisions inside fast control loops, operation under faults and corrupted measurements, physics-based rewards in place of labels, and sequential discrete constructions with a reusable policy, while Bayesian optimization, genetic algorithms, and differentiable methods remain competitive elsewhere. We conclude with the open challenges that recur across domains, including reward design, sim-to-real transfer, safety, the lack of shared benchmarks, and the statistical validity of policy-driven decisions.

Complex systems, from ecosystems to brains, are usually modeled as modular components coupled through feedback rather than as indivisible wholes. Can a transformer trained on the joint behavior of a group of interacting agents be understood in the same way? In this work, we develop a computational-mechanics framework to study agents interacting in feedback loops. We formally prove that any joint stochastic process can be presented as resulting from the behavior of agents interacting in feedback, and the resulting process is generated by a hidden Markov model whose Bayesian beliefs factorize into separate components per agent. Furthermore, our experiments show that transformers trained on such feedback sequences make use of this separability, disentangling the beliefs about the underlying factors into separate subspaces of their residual stream. This representation is found to scale linearly (instead of exponentially) with the number of agents. Moreover, interventions show that these representations are causally efficacious, allowing for targeted steering of individual agents' beliefs by selectively inserting counterfactuals. However, coarse-graining the behavior is found to disrupt the factorization, and can force a higher-dimensional representation. Together, the results establish both the dimensional advantage of a factorized representation and the ability of transformers to exploit it, laying the groundwork for investigating traces of compositional predictive architecture in pretrained models.

Short cognitive events contain unequal and often sparse transition evidence, so raw EEG microstate distances can reflect finite sampling rather than neural reconfiguration. We introduce METER (Matched-Event Transition Estimation against a Reference), a participant-relative framework that separates marginal occupancy from conditional next-state choice and evaluates each event against duration-conditioned events from a fixed participant-specific reference bank. METER-Jump integrates uncertainty in sparse target and reference transition tables, returning a randomized event rank for testing and a continuous standardized coordinate for association analysis. Using a reproducible three-state representation selected by spatial criteria, we evaluated METER with participant-fitted simulations, estimator ablations, map perturbations, and preprocessing sensitivities. Event-level false-positive rates were 0.044-0.056 across four heterogeneous null mechanisms. By contrast, global cutoffs learned under the participant-fitted null and transferred unchanged to other regimes produced false-positive rates from 0.004 under rare transitions to 0.234 under map perturbation; duration-conditioned ranks remained near 0.05. Detection probability was approximately 0.69 with fixed maps but fell to 0.22 after bootstrap-derived map-assignment perturbation. In 4,175 natural-reading events from 18 participants, no linguistic association survived multiplicity correction in the primary 30-ms analysis, although a 20-ms sensitivity changed one within-stream max-$T$ decision. METER therefore calibrates finite-event departures conditional on a defined microstate sequence, while map assignment, filtering, temporal smoothing, and reference construction remain separate sources of measurement uncertainty.

Clinical interpretation of a 12-lead electrocardiogram (ECG) often requires answering targeted diagnostic questions with clear, quantifiable decisions. We introduce ECGJev, the first ECG framework built on a decision decoder, a pretrained language model trained to return a probability for every listed option. A question-conditioned bridge compresses frozen waveform and spectrogram features into 16 ECG tokens, and a two-stage procedure adapts the bridge and low-rank adapters on the decoder's native option readout using diagnostic labels alone, scoring all options for a question in one forward pass. Evaluated across four ECG benchmarks, ECGJev achieves the highest macro-F1 on CPSC2018 (80.46%), PTB-XL (71.68%) and Chapman (96.86%), and generalizes competitively to the small-sample PTB cohort. Concept probes and perturbation analyses further show that its predictions rely directly on the ECG input, remain stable under option-order and wording variations, and transfer zero-shot to unseen waveform concepts. Our code is available at https://github.com/zhuoyuanli-lab/ECGJev.

Studies of neural network generalization and robustness often summarize smoothness with a single number, such as a Lipschitz constant or a Hessian curvature statistic. However, Lipschitz constants summarize global worst-case sensitivity and Hessian statistics summarize local curvature; neither captures how model responses vary with the size, direction, and location of perturbations encountered in real-world deployment. To measure smoothness under these perturbations, we introduce a finite-difference roughness spectrum that profiles model responses across these conditions. Our roughness spectrum generalizes scalar smoothness measures, bringing Lipschitz sensitivity and directional Hessian curvature into a common framework across perturbation scales. We connect this spectrum to classical Besov theory, which describes how a function's variation changes with scale. Our theoretical analysis also proves that a larger output shift can cause less task damage than a smaller one, and that local curvature alone can miss the blocks most vulnerable to finite perturbations. In image classification, Roughness Source Control (RSC) suppresses margin roughness along noise directions and improves noise robustness in a comparison of 80 checkpoints across four model--dataset combinations. In post-training quantization, we identify blocks that need higher precision by measuring task-loss increases from perturbations matched to the size and direction of each block's quantization error. Our mixed-precision allocation achieves lower perplexity than the Hessian-based HAWQ-V2 baseline under the same GPTQ weights and bit budgets in all 14 settings across seven held-out large language models.

Locating political parties on environmental and climate policy is central to the study of party competition, yet the established instruments, expert surveys and the hand coding of manifestos, are costly, periodic, and slow to extend to new documents. We ask whether a panel of large language models can serve as a cheaper and reproducible alternative. Treating each manifesto as a set of sentences, three models from independent providers score every sentence on two dimensions, environmental protection and climate change, separating a party's position on an issue from the salience it gives it, and the judgments are averaged to the manifesto level. Agreement across the models is retained as a built-in reliability signal. We apply the panel to 41 manifestos from the 2024 European Parliament election, covering Austria, France, Germany, Italy, Spain, and the European party families, and validate it against the 2024 Chapel Hill Expert Survey for the 25 parties with a like-for-like expert benchmark, using the other manifestos as diagnostic cases. The panel recovers the expert ordering of parties with strong rank agreement on both dimensions, while reading them as systematically greener than experts do, a gap we interpret as the distance between manifesto promise and governing record. Inter-model disagreement modestly flags where the instrument errs, and position and salience prove related for environmental protection but independent for climate change. The result is a transparent, low-cost, and multilingual instrument that complements rather than replaces expert and hand-coded approaches; the prompt, code, and manifesto-level results are released openly.

Large-scale scientific facilities, such as particle accelerators, are complex systems composed of thousands of interacting components that must operate collectively to support diagnostics and control, and ensure safe, reliable operation. These facilities rely on extensive sensor networks that generate large volumes of heterogeneous, incomplete, and noisy telemetry, often sampled at different rates across subsystems. In this paper, we adopt the foundation-model paradigm to learn general-purpose sensor representations from historical accelerator data through self-supervised pretraining. We introduce SensOFormer, a denoising masked transformer with a Perceiver-style encoder-decoder architecture designed to handle variable sensor sets, accommodate heterogeneous sampling rates, and learn robust representations from noisy measurements. The model is pretrained on multiple particle-accelerator datasets and transferred to downstream tasks, including missing-data imputation, anomaly detection, cavity identification, and fault identification. Through extensive evaluations and comparisons, we demonstrate that a single self-supervised model can learn reusable representations for diverse operational tasks in large-scale scientific facilities. Project page: https://mahindrautela.github.io/sensoformer.

Recently, the detection of AI-generated content has been of particular importance in various fields of our society. For accountability purposes, many large language models (LLMs) have implicitly watermarked their outputs. Localizing such watermarks requires a piece of text to undergo an additional verification process. Existing localization methods typically rely on dense token-level scoring, making verification costly even when an LLM-generated watermark only sparsely appears. To tackle this problem, we design a new adaptive allocation strategy that first maintains broad document coverage and then progressively concentrates the detection budget on regions that exhibit stronger watermark evidence or are more likely to contain watermarked content. We then transform the statistical signals from the selected regions into e-values and employ the proposed token-weighted e-BH procedure to determine the localized watermark regions. The resulting framework is called BELL (Budgeted E-evidence Localization for LLM Watermarks). Theoretically, BELL can control the token-weighted false discovery rate (FDR) without independence assumptions across regions. Finally, extensive experimental results demonstrate that BELL consistently outperforms competing methods across multiple LLMs, datasets, and watermarking schemes.

Efficient GPU kernels are typically low-level and hardware-specific, and LLM agents can now generate them faster than experts can verify. Existing approaches rely on testing, which provides evidence but not a guarantee of correctness. We introduce LARK, an agentic framework that automatically and formally verifies GPU kernels against their PyTorch reference implementations. LARK proves equivalence in functional and memory behavior under exact arithmetic and establishes a new trust boundary: human effort is only needed to verify operator-level Lean semantics ahead of kernel generation time, instead of checking every single line of generated code. LARK formalizes the GPU machine model and the semantics of each operation in Lean 4 once per target language, and the Lean 4 semantics of any incoming kernel is derived from these pre-built definitions. LLM agents then automate the proof in two steps. First, an agent decomposes the whole kernel semantics into smaller segments and Lean 4 verifies this segmentation, which addresses the intractability of proving equivalence for large kernels directly. Second, agents write Lean proofs that each kernel segment matches its high-level PyTorch counterpart. Together, the two steps give whole-program equivalence. We demonstrate that LARK can be used 1) to ensure correctness in existing code-generation agents and 2) to automatically verify human-written kernel libraries. On the KernelCraft GPU tasks, combined with a simple ReAct agent, LARK enables GLM 5.3 to fully verify 10 of 12 Level-1 kernels and Opus 5.5 all 5 Level-2 kernels. Moreover, over 10 production kernels randomly sampled from popular open-source LLM serving frameworks, LARK verifies 9 and identifies a new data race bug in the tenth. We believe LARK represents a significant step toward verifiable code generation in the agentic era.

Accurate three-dimensional interaction-position reconstruction is important for CZT drift-strip detectors, but conventional methods can suffer from degraded performance in challenging detector regions, while learned models provide limited information about prediction reliability. We investigate two uncertainty-aware approaches: deep ensembles of point-estimation networks and simulation-based inference (SBI), which estimates a posterior distribution over interaction position. Both approaches are trained on physics-based simulations and evaluated on both synthetic and experimental data. On synthetic data, uncertainty from both methods correlates with positioning error, with SBI showing a substantially stronger relationship. On experimental measurements, the deep ensemble achieves improved $z$-position resolution, while SBI provides more informative uncertainty estimates that highlight regions of reduced positional information and discrepancies between simulation and the physical detector. These results demonstrate the complementary strengths of ensemble-based reconstruction and posterior inference for uncertainty-aware detector position reconstruction.

The spatial intelligence of large vision-language models (LVLMs) requires the ability to combine local visual observations into a geometrically consistent representation of the physical world. Many static benchmarks assess isolated spatial judgments, such as relative direction or distance, while interactive embodied evaluations combine perception with exploration, planning, control, and simulator-specific execution. To bridge this gap, we introduce RECONSPATIAL, a benchmark centered on multi-view 3D scene reconstruction across indoor, tabletop, and outdoor environments. Given real multi-view images and target-object queries, LVLMs are required to jointly predict oriented 3D bounding boxes in a shared, camera-anchored metric coordinate frame. Reconstructions are evaluated along two complementary axes: Structure for relative layout and Metric for physical scale and camera-anchored placement. Nine atomic spatial QA tasks, grouped into Geometric Perception and Spatial Relations, probe spatial capabilities involved in joint reconstruction. Offline navigation extends evaluation to path planning. Evaluation of 18 frontier LVLMs reveals a consistent gap between relative scene structure and metric reconstruction; even GPT-6-Astra scores 81.4% on Structure but only 58.7% on Metric. Tracing this gap, diagnostic analyses link atomic spatial capabilities to reconstruction quality and show that relational consistency alone does not ensure metric accuracy. Experiments with 3D map injection further separate scene construction from path planning, revealing how these geometric errors limit navigation success. Finally, spatial fine-tuning yields broad transfer across both spatial and general vision-language benchmarks.

Digital learning environments produce rich clickstream data that can reveal how students approach learning tasks, yet interpreting these fine-grained traces often requires substantial expert effort. We introduce ClickSight, a framework for interpreting student clickstreams through learning strategies defined in the literature. Given a student clickstream and a predefined set of learning strategies, ClickSight uses a staged LLM-based pipeline to identify strategy-relevant evidence, assess adherence to each strategy, and generate an evidence-grounded interpretation of the student's behavior. We evaluate ClickSight across two learning environments, 365 students, and 12 model backbones using an expert-grounded quality rubric and an LLM judge calibrated with human-expert ratings. Our results show that ClickSight can generate high-quality strategy-based interpretations across different models and environments, with reasoning models performing more consistently than instruction-tuned counterparts. We further find that providing external clickstream-extracted guidance can improve interpretation quality for strategies that require precise quantitative information from the trace. By connecting fine-grained clickstreams with theoretically defined learning strategies and preserving the evidence underlying each judgment, ClickSight provides a flexible approach for studying student behavior in digital learning environments.

Social-media comments are interactions, yet they are often translated and evaluated as isolated text. This mismatch is especially problematic when meaning depends on surrounding discussion, community-specific slang, or culturally situated expressions. We introduce CULTURE-CMT, a Chinese-English benchmark for contextualized social-media comment translation that restores the post and conversational context naturally available to users. Its 572 manually verified examples are organized into Slang, Context, and Both, enabling diagnostic analysis of distinct and interacting sources of translation difficulty. To support scalable evaluation, we further develop a retrieval-augmented cultural-effectiveness Judger that combines a calibrated scoring model with dynamically retrieved cultural knowledge. CULTURE-CMT reveals several non-trivial patterns. Providing interaction context improves most model configurations, but does not eliminate culturally ineffective translations. Examples requiring both contextual grounding and culturally situated language remain substantially more difficult than either source of difficulty alone. Moreover, larger models are not consistently better, and thinking produces highly model-dependent gains, including occasional degradation. In contrast, a 27B model adapted on benchmark-disjoint social-media data achieves the lowest observed invalidity rate among the evaluated configurations under thinking inference with context. Together, these results suggest that realistic social-media translation is not simply a matter of scaling models or reasoning longer, but of learning to translate meaning as it is constructed within evolving online interactions.

Software development sessions often involve several unfinished tasks rather than one issue carried through to completion. As coding agents take on repository-level development, they need to recover an earlier task's requirements and partial progress after intervening work. Existing repository-level benchmarks typically score one issue-to-patch episode and do not directly measure whether an agent can resume a task after advancing other unfinished work in the same repository. We introduce \textbf{\ourmethod}, a benchmark for focus-task resumption in sessions that interleave unfinished tasks drawn from the same source repository, with a persistent checkout for each task. Each focus task is evaluated in matched uninterrupted and switched sessions with identical focus requirements and executable verifiers. The switched session interleaves ordered stages from other unfinished tasks in the same repository, while only the final focus patch is scored. \ourmethod~contains 200 focus tasks from 19 Python repositories, reconstructed from issue, pull-request, and review histories and rendered as 400 matched session specifications. Across 7 coding agents, interleaved unfinished work produces an average focus-task resolution gap of 11.5 percentage points, ranging from 6.0 to 15.0 percentage points. Trajectory analysis identifies cross-task requirement confusion and loss of task state as recurrent resumption behaviors. We introduce a plug-in context-management module that selects task-relevant history and maintains concise task-state summaries. On OpenCode, the module improves focus-task completion under same-repository task switching by 5.5 to 15.5 percentage points across the evaluated models. Code is available at github.com/Miraclecsc/Switch-SWE.

While language models (LMs) excel in natural language tasks, their ability to learn impossible languages--those that violate the structural, cognitive, or logical constraints that make human languages learnable and functional--remains debatable. This controversy is further fueled by a divide between linguistic theorists and empirical researchers working with small-scale LMs. To address this controversy, we introduce semantic coherence. This concept measures how effectively semantic information is expressed in language structures and examines the learning capabilities of LMs across multiple scales. By fine-tuning billion-parameter LMs on impossible language samples, we demonstrate that larger LMs achieve similar performance on impossible and possible languages, whereas million-parameter LMs exhibit significant degradation when trained on impossible languages. This performance gap widens as semantic complexity increases. We further show that scaling up the LM size fundamentally shifts the learning focus from grammatical rules to semantic understanding, with billion-scale LMs exhibiting a threefold reduction in sensitivity to low-level linguistic perturbations while maintaining stable performance under high-level semantic perturbations.

Medical language models promise expert clinical support, but the strongest open-source LLMs remain too large for private edge-device deployment, while compact medical models still trail substantially on knowledge-intensive and real-world healthcare tasks. Reasoning models that fit edge memory budgets can produce long reasoning traces, increasing sequential decoding, end-to-end latency, energy consumption, and KV-cache memory pressure. We present MedPsy, a family of text-only 1.7B and 4B medical language models designed for edge deployment, targeting this accuracy-generation-efficiency trade-off. Our recipe combines a synthetic medical data pipeline over biology, medicine, and health seeds with chain-of-thought targets from a 235B medical-focused LLM teacher; a four-stage post-training curriculum comprising two SFT stages followed by two RL stages with hard-sample mining; and an edge deployment evaluation spanning multiple GGUF quantization variants and real-device inference on laptop and smartphone hardware. On seven closed-ended medical benchmarks, MedPsy-4B surpasses MedGemma-1.5-4B by 19.34 points and matches MedGemma-27B while being 6.75x smaller; MedPsy-1.7B outperforms MedGemma-1.5-4B by 11.42 points despite being less than half its size. On HealthBench-Hard, MedPsy-4B surpasses MedGemma-27B by 15.33 points, while MedPsy-1.7B surpasses it by 11.66 points despite being ~16x smaller. Beyond accuracy, MedPsy reduces average response length by 1.7x (1.7B) and 3.2x (4B) versus its Qwen3 backbones, while Q4_K_M quantization preserves the aggregate score within 1 point of BF16 with ~69% lower disk footprint. In real-device inference, MedPsy-4B's shorter generations translate into 75% lower median end-to-end latency and estimated energy per answer than its Qwen3 backbone at matched quantization, while peak RAM remains below 5 GB across our tested laptop and smartphone configurations.

Social media translation requires preserving meaning while capturing the cultural nuances of user-generated content (UGC). However, gaps in the cultural knowledge of LLM-based evaluators can lead to inaccurate assessments and unreliable training rewards. To improve these signals, we use retrieval-enhanced cultural evaluation to train CUE-MT, a family of social media translation models. A calibrated, domain-specific Judger guides supervised data refinement and is augmented by an Agent for hard-example mining and reinforcement learning. The Agent retrieves relevant external evidence as needed and distills it into concise hints that the Judger uses to score translations and assign rewards. Across 9B, 27B, and 35B-A3B backbones, CUE-MT outperforms the evaluated open-source baselines of comparable size on CUE-MT-Bench and three public benchmarks. Human evaluation on CUE-MT-Bench shows cultural effectiveness gains of 7.44-15.12 percentage points over the strongest external baseline in each size group. Retrieval improves the Judger's four-way classification accuracy by 8.07 percentage points over a matched no-retrieval baseline that uses LLM-generated hints alone. Training ablations further support the combination of hard-example mining and Agent-enhanced rewards.

Test-time adaptation (TTA) updates a deployed model using only unlabeled target data, but standard entropy-based methods often rely on deterministic confidence and can reinforce overconfident errors under distribution shift. We propose MIRA, Mutual Information guided calibration for Reliable Test-Time Adaptation, a source-free online adaptation method that uses lightweight last-layer Gaussian perturbations to probe prediction reliability. For each target batch, MIRA reuses a single feature extraction pass and samples perturbed classifiers around the source classifier, forming an efficient local predictive distribution. Rather than fitting a Bayesian posterior, MIRA treats the perturbation scale as a local probing radius and selects a batch-specific scale using mutual information (MI) and prediction agreement. The resulting perturbation-sensitivity signal identifies predictions that are confident and locally stable under classifier perturbations, while suppressing brittle predictions with high stochastic disagreement. MIRA then performs adaptation with an MI-confidence weighted entropy loss, which down-weights perturbation-sensitive or low-confidence samples before they drive entropy minimization. This weighted loss reduces the influence of brittle predictions that can otherwise cause harmful adaptation under distribution shift. The method requires no source data, retraining, model ensembles, or offline posterior fitting, and can be applied as a plug-and-play modification to entropy-minimization TTA. Experiments on distribution-shift benchmarks show improvements in accuracy, calibration, and negative log-likelihood, suggesting that mutual-information-guided perturbation probing and weighted adaptation yield more robust online TTA than deterministic confidence.

Simulated students offer a scalable way to test learning platforms and AI tutors before deployment with real students and to augment scarce data for studying learning at scale. Existing LLM-based simulators primarily target short, language-centered interactions. In open-ended learning environments, however, student behavior emerges from patterns across long-horizon interaction trajectories, where turn-level prompting can drift from an assigned behavior profile. We introduce ALPS (Align Long-term by Planning Short-term), a hierarchical simulated student conditioned on observable, environment-specific behavior profiles. Rather than selecting every action independently, ALPS alternates between planning profile-aligned task segments and generating the corresponding actions. We evaluate ALPS in both data-sufficient and data-scarce settings, where supervision comes either from real student trajectories or from a teacher LLM. Across both settings, we compare in-context learning (ICL), supervised fine-tuning (SFT), reward-ranked fine-tuning (RAFT), and their sequential combination (SFT followed by RAFT). Experiments in an inquiry-based physics game and a pharmacy simulation evaluate behavioral density and coverage against held-out real data, and profile consistency using a human-aligned LLM judge. Across both settings, ALPS consistently outperforms turn-level generation, achieving comparable density and judge pass rates under scarce supervision, although coverage remains lower. These results show that hierarchical planning enables faithful, profile-aligned simulation over long interaction horizons, even with limited real student data.

Memory systems for LLM agents are evaluated mostly on social dialogue, where facts accumulate and rarely change, so remembering reduces largely to finding the turn in which something was said. Financial dialogue breaks this assumption: figures are revised, stances shift without being restated, rules stated once govern later questions, and a single company is discussed under several analytical lenses. Memory systems that retrieve the passages most similar to a question, or the most recent ones, therefore return superseded values, miss rules that share no wording with the question, and cannot recover a stance spread across sessions. We present TrackMem, a per-speaker graph memory that tracks the history of figures, separates perspectives by entity and lens, and preserves each speaker's actions in chronological order. To evaluate it, we introduce FinLoCoMo, a benchmark of 15 long-horizon investment conversations with 765 questions in two splits: one following the LoCoMo categories and one targeting four failure modes. With two readers, TrackMem outperforms existing memory systems on both splits. On the failure-mode split it reaches 73.9% macro accuracy against 69.4% for the strongest baseline, with the largest gains on questions about revised figures (86.5% against 73.4%).

Dynamic map (DM) is a core information infrastructure for vehicle-road-cloud (VRC) cooperative autonomous driving. It fuses perception data from multiple sources to overcome the inherent line-of-sight and field-of-view limits of single-vehicle perception. However, DM still falls short on driving safety in two ways. First, DM data specifications already define road conditions semantically, but no pipeline yet turns real-world road observations into fine-grained risk entities in the DM. Second, existing DM systems lack a natural language interface for driving risks. To close these gaps, we propose Safety4DM, a safety-oriented language system for VRC dynamic maps. We first build RoadRisk, a real-world road image dataset that covers four types of road conditions and four risk levels. On this dataset, we fine-tune a vision-language model (VLM) for road safety risk perception. The model outputs structured risk representations that the DM can ingest directly. We then introduce a Safety-aware Step-wise Prompting (SafeStep) mechanism as the natural language query interface of the DM. SafeStep answers queries accurately and also generates actionable driving advice. To build and evaluate this interface, we develop SafeVRC-Sim, a VRC cooperative simulation framework, and SafeDM-QA, a safety question-answering benchmark. Experiments show that even the compact Qwen3-VL-4B reaches 89.83% accuracy on risk perception after fine-tuning. SafeStep also plugs easily into multiple LLM families. With medium and large models (20B and above) from five families, it achieves over 93% natural language query accuracy in every case, which shows strong generalization.

When an image classifier decides, we can usually only estimate why. For quantum and all-optical classifiers that act linearly on amplitude-encoded pictures up to the measurement, with no trained classical network before or after them, the decomposition needs no approximation: each class's probability above its no-information share splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM) over the picture and measurable on the hardware that classifies. For the quadratic form behind each class probability it equals Integrated Gradients from a zero baseline, in closed form. Deleting the 5 % of pixels with the highest evidence drops accuracy from 73 % to 9 %, as occlusion does with 1,024 extra runs per picture, whereas gradient saliency removes much less. One architecture serves both: a Fourier-plane mask followed by a brickwall circuit on qubits or by phase plates on a lens bench, read by a Born headroom readout with no trained parameter; centring each picture on its own mean lets dark features count as much as bright ones. In simulation, four phase plates at a single wavelength reach 72.3 % on Caltech-10 colour photographs (96.8 % on MNIST, 86.7 % on Fashion-MNIST, 47.9 % on colour CIFAR-10). On IBM's ibm_kingston, with the mask applied before approximate loading, 300 Caltech-10 test pictures were classified at 37.7 % against 10 % by chance (56.3 % in exact simulation), and 2x2 maps of ten correctly and confidently classified pictures, estimated on the chip, agreed with the exact ones (r = 0.77). The maps also revealed a shortcut in the data.

Rubric optimization improves LLM grading by revising scoring instructions while keeping the grader's parameters fixed. Existing optimizers often judge edits by aggregate validation performance, limiting the supervision they retain about which errors an edit corrects or introduces. Testing each candidate also requires rescoring validation answers, which limits how many edits can be explored within a fixed budget. We propose RubricWorld, a framework that learns from previous edit evaluations to guide further search. Its world model predicts an edit's correction and degradation rates at score boundaries, conditioned on the current rubric state. The training signal comes from paired grading results: scores for the same answers before and after an edit. Predictions select one candidate for direct evaluation, and held-out confidence bounds determine whether it corrects enough target errors while keeping new errors within prespecified limits. Measured effects update the model even when the edit is rejected. Across EIR (National), ASAP2.0, and ASAP-SAS with three frozen LLM graders, RubricWorld achieves the highest score in 20 of 27 metric comparisons. On ASAP2.0, it raises gpt-5-mini accuracy from 18.20% with the original rubric to 46.70%, exceeding the strongest baseline, APO, by 22.60 percentage points. code will be released upon acceptance.

The use of collections of independent local Gaussian processes (GP-ind) combined with trust regions is a widely adopted approach for large-scale and high-dimensional Bayesian optimisation (BO). However, this approach suffers from several limitations, including discontinuities at the boundaries of disjoint regions, an inability to capture long-range spatial correlations, potential information loss, and abrupt increases in computational overhead. To address these limitations, this paper introduces a novel BO algorithm, termed BO-pro-c-tr. The proposed method employs a product-of-experts Gaussian process model with uncertainty calibration (GP-pro-c) as its surrogate model and incorporates a one-to-many association strategy for integrating GP-pro-c with trust regions. Empirical results suggest that, when combined with the GP-LCB acquisition function, BO-pro-c-tr achieves an average reduction of 2.6% in simple regret and 34.9% in computational overhead relative to BO algorithms based on independent local GP models coupled with trust regions.

Surgical video understanding is essential for intraoperative assistance such as workflow monitoring, instrument preparation and procedural documentation. Delivering such assistance in real time requires a model to jointly perceive the ongoing scene, remember past surgical events, and respond before clinically relevant moments pass. Existing approaches provide limited support for jointly retaining procedural evidence and deciding when to respond in streaming surgical video. In this work, we propose SurgStream, a workflow-anchored proactive streaming framework for real-time surgical video understanding. We first introduce Workflow-Anchored Memory (WAM) to record structured instrument-action-target observations and fold them into phase-level summaries, providing compact and queryable surgical memory. Deadline-aware Transition Learning (DTL) is then devised to supervise response states within clinical windows, enabling timely responses before procedural deadlines. We further introduce SurgStream-Bench, comprising 11.7K streaming interactions across various tasks covering real-time perception, backward memory queries, forward proactive QA, proactive alerts and caption memory. Experiments on SurgStream-Bench show gains over state-of-the-art methods, especially in memory and forward proactive QA. The source code will be available.

Neural auction design has emerged as a powerful paradigm within automated mechanism design for modern Web platforms. Because these learned mechanisms directly dictate critical economic outcomes, their lack of internal transparency raises significant auditability and trust concerns. Opening this black box \textit{post-hoc} is therefore essential, yet presents two distinctive challenges: (1) activation contexts are dense, correlated numerical profiles devoid of explicit semantic structure, and (2) feature interpretations can only be falsifiably audited if they map to measurable auction states or outcomes. In this paper, we propose a post-hoc white-box auditing framework for unpacking trained neural auctions. Using CITransNet as a representative testbed, we first employ sparse autoencoders to decompose hidden representations while faithfully preserving downstream auction behavior. Anchoring feature interpretations to native economic targets, we validate them through activation prediction and causal interventions to isolate verified features from the reconstruction dictionary. Finally, applying this pipeline across varied auction settings, we show that the verified features can serve as analytical instruments to examine how internal representations modulate strategic misreporting incentives, align with Myerson's optimal auction, and support the diagnosis of individual auctions. Our work highlights the potential of white-box mechanism auditing to build the foundational trust necessary for neural auctions to be deployed as widely adopted Web technologies. To enhance research transparency and facilitate independent inspection, we publicly release the experimental results supporting our findings at https://github.com/Li-Beverly-Kong/opening-neural-auctions.git

Large language model based agents increasingly rely on external tools to invoke services and interact with web resources. These tools can evolve through software updates, yet agentic frameworks may continue trusting them without re-verifying their safety after implementation changes. This tool update gap creates a supply-chain vulnerability in agentic systems: because updated tool code executes outside the LLM's context window, malicious functionality embedded in tool updates can bypass defenses centered on model-visible interactions, leaving agents exposed to stealthy, persistent threats that existing in-context defenses can hardly address. To thoroughly investigate this threat, we introduce ETUDE (Exploiting Tool Updates for Direct Execution), an attack that embeds malicious functionality into updates of initially benign tools to produce compromised tools, while preserving their legitimate tool functionality. Unlike prompt-based attacks, ETUDE operates entirely outside the LLM's context and requires no adversarial instructions, thereby bypassing defenses restricted to model-visible content. We evaluate ETUDE across three representative attack workloads and seven widely used commercial and open-source agentic frameworks, including Gemini CLI, Codex CLI, and Claude Code. The results show that ETUDE achieves a 97.6% ASR while maintaining a 95.9% legitimate task completion rate. We further examine existing defenses and find that they provide insufficient protection against this threat. These findings expose a fundamental blind spot in current agent security practices: trust established for a tool does not necessarily remain valid after its implementation updates.