The Inference Report

Video generation models can now depict remarkably realistic motion. Yet it is unclear how to use this motion beyond pixels: to animate a specific 3D asset and replay the animation in its 3D scene. For instance, a video model can show a monkey jumping over a crossbar, but it cannot return that jump as replayable 3D motion placed in the 3D scene. Here we present a model that directly co-generates the video and 3D motion for any supplied rig, along with a camera that registers this motion to the scene. Prior video-motion models either handle only human skeletons, often adding billions of parameters, or triangulate 3D motion from generated 2D pose maps. We instead keep 3D motion in a separate model that supports arbitrary skeletons, and connect it to a pretrained video model through a small set of cross-attention layers. This separation lets the motion model first learn how skeletons move from motion-only data, which is easier to obtain and augment than paired video. Additionally, our model learns to locate each joint in the video frames, which lets us recover the camera and place the 3D motion in the scene. To train across many skeletons, we use a generative video editor to place animated 3D assets in scenes suited to their actions, curating the edits so their known motion remains valid. Together with human and robot video, this yields nearly 200K training pairs. Our evaluations demonstrate that our model animates diverse rigs from text more faithfully than video-first pipelines, and its motion responds to the surrounding scene. It also turns video into rig motion, rig motion into video, and enables keyframe-based control of the 3D motion. Videos and additional results are available on our project page: https://research.nvidia.com/labs/par/projects/jca/.

Agentic artificial intelligence (AI) systems hold immense potential for automating complex workflows in materials discovery and design, but extending them to domain-intensive mechanistic investigations remains challenging. Here, we present GBClaw, an agentic AI framework for investigating the structure and deformation of general grain boundaries (GBs). Although general GBs constitute the vast majority of real-world interfaces in polycrystalline materials, they remain understudied because their structural complexity and interfacial mismatch complicate experimental and modeling research. GBClaw integrates large language models (LLMs) into a modular architecture comprising an orchestrator, specialized agents, domain-specific tools, simulation interfaces, and hierarchical memory. Through dynamic replanning and iterative workflow refinement, the framework effectively connects systematic simulations with targeted mechanistic analysis. Applied to 14 general asymmetric tilt GBs in nickel, GBClaw characterized interfacial defect structures and identified diverse deformation modes across these incommensurate boundaries. Comparing boundary sliding with and without secondary GB misfit dislocations revealed substantial differences in shear strength, underscoring the critical role of defect hierarchies in sliding resistance. Furthermore, researcher-guided follow-up calculations traced an anomalously high shear resistance to metastable boundary structures. This work demonstrates that combining agentic AI with domain expertise bridges high-throughput computational screening with deep physical understanding of complex material interfaces.

Dexterous multi-finger hands hold promise to unlock diverse skills like functional grasps, in-hand reorientation, and dynamic handover that the human-designed world assumes, yet existing learning recipes force a trade-off. Goal-reaching reinforcement learning generalizes to arbitrary object poses but, lacking any human prior, converges to unnatural and unreliable finger strategies; motion tracking reproduces human-like behavior from a single clip but cannot deviate from it. We present AdaDex, a goal-conditioned RL framework that turns a single hand-object trajectory of a few seconds---from mocap data or HOI reconstruction from in-the-wild RGB(D) video---into an adaptable dexterous skill. At the core of AdaDex is Hybrid Dexterous Mimicking: a shared policy is trained on two populations of massively parallel environments: motion tracking environments trained with dense motion tracking objective, which instill the human prior from human motion; and generalization environments with noised initial states and goals and a sparse object-success reward, which force the skill to work far off the reference. From seconds of human motion per skill, the resulting policies can learn diverse adaptable dexterous skills, such as functional grasping, object reposing, in-hand reorientation with finger gaiting, and dynamic behaviors like throw-and-catch. Our method outperforms baselines in terms of generalization and functionality, and we demonstrate transfer across simulators and to the real world. Project page: https://yuxuank.com/AdaDex/.

Motivated by the human ability to integrate experience across language, perception, and action, we introduce UniWALM, a unified world-action-language model that learns video, action, and language in a single shared representation, where each modality is both a conditioning signal and a supervision target. We ask: can jointly generating language enrich this shared representation, improving manipulation and supporting memory-dependent behavior? Unlike prior approaches that attach a reasoning or vision-language model to a video-action backbone and run them as distinct stages, UniWALM unifies all three modalities inside a single mixture-of-transformers backbone whose per-modality diffusion heads co-denoise synchronously, decoding subtask and textual memory in the same pass as action with no separate planning stage; the generated language is re-consumed on the next step, closing a language-as-memory loop without an external planner or memory bank. To keep the action stream reliable under this richer objective, we further introduce Flow Matching with Correction (FMC), a training objective that teaches recovery from perturbed action trajectories at no inference-time cost. UniWALM consistently improves over same-backbone baselines (e.g., FastWAM) on compositional pick-and-place (MolmoSpaces) and across the memory-dependent subset of RMBench, where it reaches 34.5% without any pretrained language or reasoning model and exceeds Mem-0's keyframe-equipped setup (28.5%) using only a single anchor frame. On a real-world Find-the-Item task, generating textual memory lifts success from 11.1% to 61.1%, while state-of-the-art VLAs stay at or below chance (<=33.3%).

Warehouse robot fleets continuously observe operational conditions, but preserving these observations as reusable semantic memory incurs substantial computation, communication, and storage costs. For example, explaining recurring aisle blockages can require reviewing video from multiple robots across several shifts. We present IndustrialRemembr, a shared spatiotemporal memory system that extends NVIDIA ReMEmbR to robot fleets and reduces server-side computation by generating scene descriptions only on demand. On each robot, an NVIDIA Jetson encodes video clips into retrieval embeddings with Cosmos-Embed1, vision tokens with Cosmos3-Edge, and reconstruction latents with the Wan2.2 VAE, without uploading raw camera frames. The encoding models are re-used to detect people in image patches, zeroing them out to support PII abstraction for GDPR-compliant deployments. We split Cosmos3-Edge at its vision-language boundary, placing the vision encoder on the robot and the language decoder on an on-premises server. A Nemotron agent retrieves memories by content, position, and time and queries the decoder using cached vision tokens. This design eliminates caption generation at write time, while cached-token decoding reproduces the unsplit model's output in all 450 tested decodes.

Planning under uncertain control execution noise requires reliable feedback policies across diverse environments. We introduce Planning Under Noise (PUN), a neural operator that learns noise-aware value fields for planning in unseen geometries. Building on an established screened-Poisson formulation of stochastic optimal control, we address two challenges: specifying boundary conditions that support navigation across varying geometries and efficiently computing fields with reliable gradients. Zero-Dirichlet obstacle boundaries terminate sampled paths at walls, weakening the goal signal along long routes and making it harder to estimate reliably. We instead impose reflecting Neumann conditions on obstacles and a Dirichlet condition at the goal, allowing sampled paths to continue exploring free space until they reach the target. To solve the resulting PDE without a configuration-space mesh, we extend a Monte Carlo (MC) solver, Walk on Stars (WoSt), to higher dimensions. However, local MC fluctuations can reverse estimated gradient directions and cause planning failures, making direct planning require costly sampling. PUN amortizes this computation by learning geometry-to-field mappings from noisy, partially converged Monte Carlo labels, avoiding expensive convergence for every training example. At inference, test-time scaling uses bounded local search guided by the predicted global field. We evaluate PUN on planning tasks across dimensions. Across four noise levels, PUN matches or outperforms all compared planners, achieving 95-100% success on the 4D replanning benchmark while maintaining the lowest runtime. At the highest tested noise level in 4D, PUN achieves 10.7-60.6% relative gains in success rate over other methods while running 3.29-32.15 times faster.

Spatial intelligence has recently been assessed through generative tasks such as spatially grounded image editing, yet language-guided generation of spatially consistent views from partial multi-view observations remains underexplored. We introduce MindView, a benchmark for generative multi-view spatial intelligence with 1,500 samples from 513 scenes across three indoor datasets. Its seven tasks form two families: geometry-dominant tasks specify changes in camera position and orientation, while semantics-dominant tasks require inferring target viewpoints by grounding referred objects and interpreting spatial relations and reference-frame cues. Evaluation of eleven generators reveals substantial room for improvement on MindView. A paired context analysis further shows that all eleven generators obtain lower overall scores under the expanded context setting. These findings motivate us to investigate whether task-specific supervision can improve target-view generation and whether models can use additional multi-view evidence effectively. We therefore fine-tune BAGEL on the MindView training dataset, improving its overall score by 24.3 points and yielding gains across all seven MindView tasks. Beyond these generation gains, the fine-tuned model also yields higher scores on three spatial-understanding benchmarks without training on their annotations, suggesting that learning to generate spatially grounded views may benefit spatial understanding. To improve evidence utilization under expanded context, we propose MindCue, a training-free spatial evidence selector guided by a target-view hypothesis. MindCue reverses the drop in overall score under expanded context and outperforms the standard-context baseline. Overall, MindView provides a concrete testbed for generative multi-view spatial intelligence.

Multi-token prediction (MTP) can accelerate language-model decoding by drafting several tokens and verifying them in a single target-model pass. However, the best speculation horizon changes with context: longer drafts offer greater amortization when reliable, but can waste compute on suffixes discarded after a rejection. Unlike autoregressive drafting, where a stopping rule can decide token by token, MTP must fix the horizon before drafting, as all heads fire in the same forward pass. We therefore formulate horizon selection as an online control problem, using verification feedback to learn how far the draft can be trusted in each context, and introduce Adaptive Multi-Token Prediction (AMTP): a lightweight controller that requires no retraining and leaves the exact verifier unchanged. To guide horizon selection, we characterize the expected number of tokens committed at each horizon via acceptance survival, the probability that a draft prefix is fully accepted. Before each round, AMTP predicts how far the draft is likely to survive and picks the horizon with the highest expected tokens per second, learning freely from every verification pass. Theoretically, we show that any online learner's prediction error translates into throughput regret against the best horizon for each context, and in turn into a bound on wall-clock completion time. With online gradient descent, this regret is sublinear. Across five draft-target pairings built on Llama-3.1-8B-Instruct with Medusa heads, Meta's MTP models, and Qwen3-4B-Inst-2507-MTP, evaluated on AIME 2025 and GSM8K, AMTP decodes up to 4.1x faster than standard decoding and matches or outperforms the fixed window in 8 of 10 settings, with throughput gains of up to 10%.

Graph generative models commonly use graph-specific Transformer architectures to jointly update node and edge features. However, these architectures do not directly benefit from recent hardware-aware attention implementations, limiting their scalability. The adaptation is, however, non-trivial: edge and graph-level features must be injected into pairwise attention scores, graph masks must be respected, and dynamic edge updates remain a major source of quadratic memory usage. We propose an edge-aware attention layer for graph generation that incorporates structural information through head-wise score modulation, while updating edge representations with a lightweight residual mechanism that retains attention-score dependent edge updates. This layer can be effectively implemented using PyTorch's FlexAttention. We achieve similar generation quality to our baseline while reducing peak GPU memory by around 65% and substantially improving training and sampling throughput. These results suggest that the proposed architecture is a practical step toward scalable graph foundation models.

Picture-description dementia classification uses a shared visual scene, but a transcript-level prediction alone does not identify the scene evidence underlying its inputs. NarraSpace links words and speech times to named picture elements and retrieves utterances for predefined scene events. Narrative-spatial descriptors summarize coverage and temporal organization. A frozen language model contrasts each event's text under examples from Alzheimer's disease and healthy-control training participants, alongside transcript surprisal and coverage measurements. Retained source records link event scores to their source utterances, speech intervals, and lexical matches. On the official 48-participant test set of the Alzheimer's Dementia Recognition through Spontaneous Speech benchmark, a fixed support vector classifier achieves $80.88\%\pm2.56\%$ mean accuracy over fifty seeds; averaged probabilities yield $81.25\%$ accuracy and $91.32\%$ area under the receiver operating characteristic curve. NarraSpace makes the construction of scene-linked classification inputs explicit and inspectable by connecting event-indexed language-model measurements to their source utterances, speech timing, and extraction records. Such source-linked measurements could support clinician review of picture descriptions by presenting quantitative summaries alongside the original speech evidence.

Software engineering agents solve repository-level tasks through iterative tool use, but their continuously growing interaction traces incur substantial token consumption and inference overhead. Existing compression approaches often require model training, auxiliary LLM calls, or assumptions about specific content such as source code. We present TraceLite, a simple yet effective, training-free framework for compressing heterogeneous SWE-agent observations. Our preliminary study shows that optical rendering preserves substantial textual and semantic information under moderate compression, while agent trajectories exhibit both cross-step and phase-level redundancy. Motivated by these findings, TraceLite preserves recent observations as high-fidelity text, renders historical observations as compact images, removes overlapping content, and discards phase-expired observation payloads after code editing. It requires neither offline training nor auxiliary LLM inference. We evaluate TraceLite on SWE-bench Verified and DeepSWE using GPT-5.6-Luna and Gemini-3.8Flash as backbone models. On SWE-bench Verified, TraceLite achieves 78.25% resolve rate, outperforming the evaluated compression baselines while reducing average input-token consumption and API cost by 33.05% and 21.30%, respectively. Moreover, it introduces only 0.438 seconds of compression latency per observation. These results demonstrate that TraceLite improves SWE-agent efficiency with low processing overhead and without compromising issue-resolution capability. Our code is available at https://github.com/xjywhu/TraceLite.

Modern distributed microservice deployments generate massive volumes of heterogeneous, multi-format observability logs across physically and logically partitioned infrastructure. Isolating root causes during cascading failures remains an arduous task for Site Reliability Engineers (SREs), bottlenecked by manual keyword filtering and disjointed temporal correlation across monitoring dashboards. We introduce Semantic Observability, an autonomous multi-agent framework that unifies disparate log streams into a dynamic, queryable Property Graph Database. By modeling physical infrastructure topologies, logical software dependencies, and chronological event sequences as attributed nodes and directed relations, it bridges the semantic gap inherent in flat, unstructured log repositories. Building on this unified substrate, we deploy a stateful multi-agent system orchestrated through Directed Cyclic Graphs (DCGs). The system autonomously navigates the topology, applying causal graph algorithms and temporal back-traversal to isolate the precise blast radius of anomalies. The pruned subgraph is then interpreted by a Large Language Model (LLM) reasoning engine to generate contextualized, code-level Root Cause Analyses (RCA) and actionable remediations. This paper establishes the foundational system architecture, formal relational ontology, and multi-agent control loops of Semantic Observability. We show conceptually how the framework eliminates ambient telemetry noise, reduces Mean Time to Resolution (MTTR), compresses LLM token consumption, and prevents structural hallucinations. We also specify an open-source validation framework on standard benchmarks (LogHub and Train-Ticket), laying the groundwork for a forthcoming companion paper on exhaustive quantitative evaluation and production benchmarking.

Interactive webpages underpin modern applications, but reproducing both their visual appearance and multi-step behavior remains challenging for multimodal large language models (MLLMs). Scalable training data must capture user actions, state transitions, and expected outcomes, while real-world webpage code is often entangled with irrelevant scripts and external dependencies. Moreover, existing GUI-agent-based verification provides slow and potentially unstable functional feedback. To address these challenges, we propose InterWebCoder, an efficient framework for interactive webpage code generation. InterWebCoder introduces a Test-Specification-Guided and Execution-Verified Data Synthesis pipeline that records real-world interaction trajectories and synthesizes self-contained webpage code with executable test scripts. A failureattributed repair mechanism uses runtime evidence to locate and iteratively fix errors in either the webpage code or test script. We then combine test-ID-weighted supervised fine-tuning with group-relative policy optimization, directly optimizing functional correctness and visual fidelity using test pass rates and state-level visual similarity. This replaces costly GUI-agent verification with fast, reproducible executable feedback. Our pipeline produces 1,736 verified samples from 2,058 tasks, achieving an 84.35% synthesis success rate. On IWR-Bench and Vision2Web, InterWebCoder achieves final scores of 48.54 and 48.42 with a 9B backbone, outperforming 30B-level models and Gemini-3.7-Flash while remaining competitive with Claude-Sonnet-5. It reduces reward-computation latency by over 97%, eliminates external MLLM verification costs, and accelerates end-to-end training by up to 1.4x.

Mathematical reasoning remains underexplored for low-resource languages, including Dialectal Arabic. To bridge this gap, we introduce the DAR-MGSM (Dialectal Arabic Multilingual Grade School Math) Benchmark, the first manually curated benchmark for evaluating mathematical reasoning in Dialectal Arabic. DAR-MGSM consists of math word problems (MWPs) requiring multi-step reasoning and covers five major Arabic dialects alongside Modern Standard Arabic (MSA). The benchmark comprises two parallel versions: translated-only and translated & culturally localized. The latter version adapts entities and references to regional varieties. We evaluate eleven Arabic-centric and multilingual large language models on DAR-MGSM. The results show that the performance generally declines on the localized version, highlighting the challenges introduced by culturally grounded adaptations. However, the magnitude of this effect varies across models, with Arabic-centric models exhibiting different levels of sensitivity to localization. A mechanistic interpretability analysis shows that translated and localized variants of the same MWPs often induce distinct internal reasoning trajectories, even when aggregate accuracy remains similar. Overall, our results show that mathematical reasoning performance depends on both linguistic variety and cultural context.

Current evaluation of LLM-generated debate largely focuses on argumentative quality, leaving the value structure of generated arguments underexamined. We introduce a diagnostic metric framework for comparing how values are expressed and organized in human-reference and AI-generated debates. The framework separates value expression, the range of recognizable values a debate invokes, from value organization, how those values are distributed and sustained across turns. Grounded in Schwartz's Basic-10 value theory, it represents debates as multi-level value distributions over argumentative units, utterances, stances, and full debates, enabling metrics for value coverage, alignment, diffusion, and cross-turn dynamics. Across multiple generation systems, AI-generated debates are broadly value-expressive, often activating as many or more value dimensions than matched human references, but they organize these values differently. Their utterances are more locally value-diffuse, and source-blind judges distinguish value clarity from value focus. These findings show that debate evaluation should consider not only value presence, but also value organization. Code and supplementary materials are available at https://github.com/Maoyuan-li/Value-Expression-and-Organization.

TS2Vec is one of the strongest self-supervised methods for time series classification, introduced in 2022. It is designed to be trained anew on every dataset: self-supervised pre-training followed by linear probing. Since then, the field has shifted towards foundation models that reuse one encoder across datasets. Yet TS2Vec offers a compact convolutional architecture that operates directly on temporal samples and accepts variable-length inputs without resampling. We ask whether these properties can be retained when the encoder is pre-trained only once, and show that they can: TS2Vec can essentially be turned into a foundation model. Three changes suffice: pre-train on a synthetic corpus, feed the encoder a raw amplitude view and an instance-normalized shape view of the input through two small stems, and distill one specialist encoder per view into a single encoder. The result, TS2Vec-FM, is a 1.6M-parameter dilated convolutional network that, kept frozen, reaches 83.3% on UCR-128 and 75.3% on UEA-27 under linear probing, within 0.3 pp of MantisV2, the current state of the art, and above TS2Vec trained separately on each target dataset. What we find most interesting is how different TS2Vec-FM is from other foundation models in design: near-parity is reached without attention, without a patch tokenizer, and without intermediate-layer extraction, using an "old-school" convolutional encoder and a two-layer MLP stem.

Computer-use agents are limited by the interfaces humans give them. An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function. Both the interfaces and the knowledge of when to use each are built by hand, which cannot keep pace with new applications, workflows, and models. We argue that interface design should itself be an agent capability. We introduce Agent S4, an autonomous computer-use scientist that builds Evolving Agent Interfaces from task experience. A meta-agent examines task agents' trajectories on synthetic training tasks, without success labels, and constructs and revises interfaces and usage knowledge in a repository. Population-based evolution then selects which repositories to develop further by their performance on validation tasks the meta-agent never sees, favoring changes that generalize beyond the seen tasks. On OSWorld V2, Agent S4, with Qwen3.8-27B and the evolved repository of interfaces, sets a new state of the art in open-source computer-use agents, reaching 21.3% binary score, 3.8 times the same model with fixed interfaces. This is above both GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%), at about 4.3% and 1.5% of their respective costs. On ScienceBoard, the evolved interfaces improve the success rate from 41.4% to 58.6% with the same model. These results suggest that useful interfaces can be evolved from experience, making autonomous interface design a source of improvement complementary to advances in models and harnesses. We open-source our code at https://github.com/simular-ai/agent-s.

Neural implicit representations provide memory-efficient signed distance field (SDF) parameterizations for large-scale LiDAR scenes, but the geometry recovered from sparse and often noisy inputs carries jitter and artifacts. Classical geometric optimization techniques resolve such artifacts with surface energy functionals, whose minimizers correspond to smooth surfaces. Motivated by this principle, we formulate neural LiDAR mapping as the optimization of energy functionals over the parameter space of neural implicit representations. Specifically, we regularize the signed distance field with an energy functional defined on the shape operator of its level set, which represents the curvature of the surface, thus encouraging smoothness along the surface. This adapts classical geometric priors to the scale of million-point LiDAR maps. Experiments on large-scale datasets show reconstructions with improved smoothness at preserved geometric fidelity, and no additional inference-time cost. The code is publicly available at \href{https://github.com/HrishikeshVish/nieo}{https://github.com/HrishikeshVish/nieo}

RTL repair benchmarks should measure whether a model can fix broken RTL given a specification. To this end, we construct and hash-bind RTLRepairDataset, 5,784 (broken, fixed) training pairs; RTLRepairBench, 459 held-out cases with 374 syntactic and 85 behavioral mutations; and an evaluation harness supporting simulation and formal equivalence. We then use this infrastructure to identify two measurement confounds. First, simulation and formal scoring give opposite signs for the apparent effect of providing a specification in one evaluated cohort; reset-aware rescoring attributes almost all of the simulation-formal disagreement to reset and first-cycle comparison semantics. Second, on 274 mined model failures, the specification-only minus specification-conditioned recovery point estimate ranges from -8.5 to +18.3 percentage points across six models. Thus, a conventional repair score need not isolate whether the model used the faulty RTL. We translate these findings into evaluation controls: report initialization, reset, first-cycle, and unknown-value semantics; disclose oracle coverage and undecided cases; and include a specification-only baseline. Together, the resources and audit provide both a benchmark for behavioral repair and a protocol for diagnosing what contributes to its measured score. We release the benchmark, the training set, construction code, sanitized aggregate evidence, and reproduction materials at https://github.com/Evaaaz/RTLRepairBench.

Neuro-spectral architectures (NeuSA) are a class of physics-informed neural networks (PINNs) that have proven effective for solving time-dependent PDEs in the frequency domain. However, training and inference often rely on costly numerical integration of the learned vector field, typically using sequential Runge-Kutta time stepping, which limits temporal parallelism. We introduce PaPi, an integration scheme inspired by weakly nonlinear flows that enables time-Parallel inference through exponential integration and a Duhamel-Picard scheme. We show that the Duhamel-Picard iteration converges exponentially for all flows, and provide explicit error bounds that depend on how close the flow is to linear. Experiments show that PaPi accelerates training and inference by up to two orders of magnitude, achieving training times of under one second for low-dimensional PDE learning problems, while maintaining competitive predictive accuracy.

Camera-trap imagery presents substantial variation in animal appearance, illumination, background clutter, and visual degradation. Masked autoencoders (MAEs) offer a label-free framework for modeling such data through reconstruction of randomly hidden image patches. We present WildMAE, a decoder-side extension of MAE that combines local structure-aware refinement with masked-to-visible cross-attention. A zero-gated convolutional residual path introduces a two-dimensional neighborhood prior, while gated cross-attention allows masked positions to retrieve contextual information from visible-position features without replacing the original Transformer decoder. The associated training recipe progressively introduces normalized-pixel, Sobel-edge, and perceptual objectives. In the original Snapshot Serengeti study, covering 4,744 test images at 75% masking, the reported reconstruction scores were 27.90 versus 24.33 dB PSNR, 0.840 versus 0.780 SSIM, and 0.160 versus 0.320 LPIPS for WildMAE and vanilla MAE, respectively. Additional historical records document favorable average differences under four synthetic degradation conditions and describe component variants. We present the method, experimental observations, and their scope together: the retained metrics are historical protocol-dependent results, rather than a newly reproduced, harmonized comparison or a validation of restoration from genuinely unavailable pixels. WildMAE provides a practical architectural case study for combining global token mixing with spatial and targeted contextual refinement in wildlife masked-image modeling.

We prove that every $\operatorname{ReLU}$ network can be exactly simulated by a $\operatorname{ReLU}$ network with two hidden layers. We present two proofs: one that is based on fillings in homological algebra and a direct combinatorial proof.

Agentic on-policy distillation (OPD) trains a student agent on its own trajectories while a teacher scores every token. A natural explanation of where it goes wrong is compounding error: the student's mistakes carry its trajectories beyond the region where the teacher is reliable, so recent methods keep training inside that region, through episode-length curricula, teacher-guided turns, prefix replay, or teacher intervention where the two disagree. We show this explanation is incomplete. From the states where the student departs most from the teacher, the teacher's own takeovers still finish the task, yet the student trained on its token scores keeps running out of turns. We attribute this to the myopia of next-token scoring. The teacher rates each token without seeing where the episode ends, so its scores let through the safe, delaying moves that outcome-based RL would punish with zero reward, a penalty agentic OPD lacks. We call this difference between what the teacher does and what it scores the rollout-scorer gap. We propose Rollout-to-Scorer Distillation (R2S), which repairs the scorer before distillation by mining high-disagreement states from the untrained student, running teacher takeovers from them, keeping the states where enough takeovers pass a reward-free check such as consensus or an LLM judge, and fine-tuning the teacher to give its takeover-informed scores on every assistant token from the divergence turn onward, without the takeover in context. The repaired teacher is frozen as the sole scorer for standard OPD, requiring no task reward during repair or training. Across three environments, two teachers and four students, R2S outperforms three state-of-the-art agentic-OPD methods in all twelve teacher-student-environment cells, by 12.8 success points on average over the best of them in each cell. Project page: https://tree-shu-zhao.github.io/r2s/.

Simulation is a powerful tool for robotics, but building diverse, simulation-ready scenes at scale remains difficult. Reconstructing scenes from single images is a promising alternative, but visual similarity alone is insufficient: objects must also be recovered, accurately placed, and physically grounded. Existing methods that rely on post-hoc physics settling often misplace objects, producing scenes that look correct but behave poorly under manipulation. We present SceneRig, an agentic system that turns a single image into an editable, simulation-ready scene. SceneRig combines geometric, visual, and physical tools to reconstruct a faithful 3D scene. An agent then performs edit-and-validate loops inside a physics simulator to achieve physical stability. Across diverse inputs, SceneRig improves instance segmentation, object-centric appearance, and 3D placement while keeping the reconstructed scenes physically stable. Furthermore, SceneRig is effective for downstream robotics: its scenes predict real-robot policy outcomes more accurately than the baseline method and more reliably support replay of recorded manipulations without trajectory adaptation.

Autoregressive models (ARMs) face a "Sequentiality Bottleneck" in reasoning, while diffusion language models (DLMs) offer parallelism but lack semantic coherence. To addresses both limitations, we propose Collaborative Reasoning Asymmetric Learning (CoRAL), a hybrid framework via an Asymmetric Planner-Executor (APE) architecture: a DLM Planner generates K diverse reasoning trajectories in parallel, while a large ARM Executor selects the optimal plan and synthesizes the final answer. To train the Planner effectively, we introduce Reward-Guided Denoising for CoT (RDC), which integrates outcome-based rewards into the diffusion process to steer it toward goal-directed reasoning. Experiments on vision-language benchmarks show that CoRAL significantly outperforms prior diffusion-based methods and achieves competitive performance against strong ARMs, achieving both superior efficiency and quality.

Continuous-time reinforcement learning seeks an optimal policy under unknown dynamics driven by stochastic differential equations using observations collected at discrete times. We develop a policy gradient framework that combines rigorous convergence analysis with learning from discrete data. We first establish a performance difference lemma and derive policy gradient formulas for randomized and parametrized policies. We then prove exponential convergence of an entropy-regularized policy flow to the regularized optimum. For discrete parametrized updates, we derive a convergence bound that explicitly separates the contributions of critic error, policy-regression error, and the learning step size. To enable learning under unknown dynamics, we integrate this framework with PhiBE, which constructs a continuous-time surrogate generator from discrete transition data. The resulting on-policy and off-policy model-free actor-critic algorithms preserve the continuous-time structure of the control problem. Numerical experiments on an entropy-regularized linear-quadratic regulator and a nonlinear wheeled robot demonstrate the performance and stability of the proposed methods with both known and data-driven dynamics.

Cellular states are shaped by both intrinsic gene expression programs and the spatial microenvironment in which cells reside. Spatial transcriptomics makes this context measurable, but effective representation learning requires capturing both spatially informative molecular signals and intercellular organization. Existing approaches address these challenges only partially: conventional gene selection can overlook spatially structured signals, while methods for encoding cell-cell relationships often either reduce a cell's microenvironment to proximity alone, discarding direction, or rely on absolute coordinates tied to an arbitrary tissue frame. Consequently, cells with similar neighbor distances but distinct spatial organization may receive overly similar representations, despite experiencing different biological contexts that shape their cellular states. We introduce Tessera, a spatial representation learning framework that addresses both molecular selection and cellular context. Tessera combines spatially informed gene selection with a geometric inductive bias that contextualizes each cell within its local tissue environment. By modeling relative cell-cell relationships, Tessera is translation-invariant while preserving direction, allowing neighboring cells to contribute differently according to their spatial context and cellular state. Across held-out tissues and technologies, Tessera outperforms existing spatial foundation models in representation quality and generalizes zero-shot to contexts up to $16\times$ larger than those seen during training.

In this report, we introduce Sage-A1.5, a 35B Mixture-of-Experts agentic model that builds on A1 to advance long-horizon problem solving through adaptive capability composition. Compared with A1, Sage-A1.5 learns not only to solve individual tasks, but also to organize work and revise its approach during execution. Sage-A1.5 highlights three improvements: agentic orchestration, verification, and open scientific discovery. First, agentic orchestration extends single-agent execution through task delegation and parallel collaboration among subagents with isolated contexts, enabling complex tasks to be decomposed and addressed concurrently and supporting efficient execution while reducing interference across concurrent subagents. Second, agentic verification strengthens long-horizon reliability through progress checking, evidence validation, and reflective strategy revision, allowing the agent to recognize unproductive search paths and pursue alternative approaches. Third, open scientific discovery extends agentic capabilities beyond answering predefined questions to conducting exploratory research, connecting scientific inquiry with interactive analysis and inspectable research artifacts. We showcase these capabilities through complicated agentic tasks and multidisciplinary scientific applications. Besides, we hope Sage-A1.5 provides the community with an accessible foundation for open scientific discovery, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.

Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's maximum context length, requiring control over active cache growth. Prior work uses context engineering to reduce cache size by trimming or summarizing history. However, modifying the trajectory prefix invalidates subsequent cached states and triggers costly re-prefilling of retained context. Task accuracy can also suffer when too little context is retained to limit re-prefill costs. We propose cache engineering as a new scheme. It allows agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context. Specifically, agents can proactively use drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings. We formalize these cache operations and identify history dependence as a serving challenge: identical token sequences can yield different cache states under different cache operation sequences, making prefix matching insufficient for correct cache reuse. Thus, we co-design the AKV serving engine. It provides a stateless interface compatible with existing serving APIs, while internally managing stateful cache reuse, eviction, and restoration. Experiments show that a tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. With AKV, it achieves up to 1.2x the output throughput and a 1.5x speedup in P95 request latency. Recursive self-improvement of cache policies further reduces cost by 7.1% over expert design. Our code is at https://github.com/thu-nics/AKV.

Energy-efficient deployment of large language models requires understanding how model architecture, hardware, and serving settings shape energy per generated token. We introduce the Token-Energy Law, a compact, interpretable model of GPU-board energy per output token for dense, full-history decoding. It relates token energy to two core dimensions: active bytes and realized decode time. Active bytes describe logical byte demand per output, separating dense weights shared within a decode iteration from private key--value (KV) histories read and updated for each request. Realized decode time is the elapsed decode duration divided by the number of output tokens produced. We combine the two byte coordinates and time in a common affine form, with coefficients calibrated for each model and hardware/runtime configuration. This formulation explains how batching amortizes shared-weight work, longer attended histories increase private KV demand, and throughput determines the time-associated cost per output. Our experiments cover three models under BF16 decoding on NVIDIA Hopper and Blackwell GPUs, including two-GPU tensor parallelism on Hopper. Using fixed coefficients and observed decode time, held-out condition-mean energy estimates differ from measurements by at most 3.73%.

Generative retrieval (GR) casts search and recommendation as autoregressive generation of discrete semantic identifiers (SIDs). SID design largely determines GR quality, but its discrete, interdependent components (SID construction, collision handling, decoding, and training) cannot be optimized end-to-end with gradients and remain largely hand-designed. We introduce a framework that improves these components through LLM-driven evolutionary program search, to our knowledge the first application of this approach to GR. A locally served, open-weight 31B model mutates executable GR components, so no proprietary LLM API is used, and a 1M-parameter proxy decoder provides fast feedback in place of repeatedly training the 1B target generator. Across 21 SID programs on NQ320K, proxy and target Recall@100 show strong rank correlation (Spearman $ρ= 0.90$). Guided by this proxy, evolution improves the seed GR on NQ320K, MSMARCO300K, and a multilingual e-commerce dataset under a fixed training recipe. On NQ320K, jointly evolving SID construction and collision handling improves Recall@1 from 55.8% to 59.3% and Recall@100 from 85.8% to 90.6%. A component study shows that proxy reliability depends on what is evolved: improvements transfer reliably for the semantic index but not for the generator's training objective or the full pipeline. LLM-driven evolutionary search is thus a practical tool for GR design, provided proxy reliability is established for the chosen search space.

Text-to-image personalization aims to equip pretrained diffusion models with the ability to learn user-specific concepts from only a few reference images. Existing methods, however, typically train each concept independently or assume that all concepts are available simultaneously. In practical applications, concepts often arrive sequentially, requiring models to continually learn new subjects, objects, or styles while minimizing interference with previously acquired concepts. We address this continual concept personalization setting by extending residual-space adaptation to the stochastic adaptation process of pretrained diffusion models. Our work achieves the highest concept fidelity among competing continual personalization methods while maintaining near-zero forgetting. Beyond these empirical gains, the learned residual-space dynamics reveal where adaptation capacity is concentrated within the diffusion model. Restricting adaptation to the identified layers by our method preserves nearly the same performance while reducing the number of trainable parameters by up to 72\%, demonstrating that residual-space dynamics can identify the layers that contribute most to continual concept personalization and directly guide more parameter-efficient adaptation.

Clustered federated learning (CFL) can suffer from cluster collapse. In particular, under unfavorable initialization, early client assignments concentrate heterogeneous groups on one cluster model, which is then trained toward a pooled compromise, while another model receives no updates. Since the current models determine client routing and routing determines subsequent model updates, we formulate clustered federated learning as control of an unknown dynamical system, with cluster models as the evolving state, client routing patterns as the control input, and routing interventions as state dependent controllers. In a two group linear model, we identify an initialization region in which greedy routing makes collapse self reinforcing. We then derive finite horizon error dynamics showing how routing interventions can reverse collapse by repeatedly training the starved model, primarily with updates from its compatible group. Motivated by this condition, each client uses blocked EXP3 to choose between greedy routing and interventions that redirect updates toward unused models. Experiments show that with simpler predictors and faster specialist formation, autonomous controller selection recovers usable specialists from unfavorable starts that trap standard CFL baselines. In transformed FEMNIST, the synchronized-entry configuration with persistent commitment recovers usable specialists in settings where the autonomous controller configuration often does not. Code and scripts for reproducing the experiments and figures are available at https://github.com/prsinha1/FedRecover.

Online search activity is widely used as a digital trace of information seeking during epidemics, but its predictive value is often reduced to correlation, a selected lag, or a single forecast-error score. We examine eight pre-specified Google Trends-COVID-19 pairs in both directions using matched LightGBM forecasts over rolling 400-day windows and horizons of 1-14 days. Log-RMSE gain is combined with forecast-error distribution entropy (FEDE), which captures horizon-dependent changes in the signed error distribution. The resulting directional predictive signature comprises response class, RMSE effect, FEDE endpoint and early-slope responses, persistence across rolling regimes, and horizon profile. For screening candidate directions, we also introduce a relative persistence-weighted directional predictive response strength (DPRS) on a within-study 0-100 scale; 100 denotes the strongest observed response in the comparison set, not an absolute coupling maximum. The pairs show concordant, RMSE-only, FEDE-only, and mixed responses. "loss of smell" -> reported cases is accuracy-dominant, with mean log-RMSE gain 0.104 (about 9.9% equivalent RMSE reduction) and DPRS 100, whereas the reverse direction is FEDE-dominant (D_FEDE=0.048; DPRS=42). "covid test" -> positive rate is FEDE-dominant (D_FEDE=0.039; positive in 72% of windows; DPRS=98) despite negative RMSE gain, while "lockdown" -> stringency is the clearest concordant case (DPRS=78). World-level symptom/case pairs are less persistent. Thus, forecast accuracy and error-distribution dynamics are complementary, predictive effects can be direction- and metric-dependent, and forecast horizon is part of the relationship. Aggregated regime-level patterns should not be extrapolated to individual time points or interpreted as evidence of causality.

Many scientific phenomena are best understood as outcomes of underlying optimization problems, yet automated scientific discovery has largely focused on recovering equations that describe observed input-output patterns. We introduce OptScientist, a novel automated framework for discovering compact symbolic optimization laws directly from data. OptScientist comprises (i) a model that infers symbolic optimization problems by first learning to represent data as an optimization problem and then extracting its symbolic form, and (ii) an LLM agentic system embedding the model in a discovery process that integrates theoretical priors from scientific literature, discovers candidate laws, abstracts their underlying mechanisms, and suggests new data to collect. We first evaluate OptScientist on OptFeynman, a novel benchmark of 99 physics-derived optimization problems, where OptScientist recovers ground-truth laws for over 80% of the problems and remains robust to noisy contextual information, while substantially outperforming machine-learning, symbolic, and LLM-based model-discovery baselines. On human risky choice and central-bank monetary policy, OptScientist discovers compact symbolic optimization laws that achieve held-out predictive performance competitive with leading baselines, while using few parameters and substantially fewer tokens than LLM baselines. Overall, OptScientist extends automated scientific discovery to uncovering optimization laws that generate observed phenomena, enabling automated theory building across the physical, biological, behavioral, and social sciences.

Stochastic gradient descent (SGD) is the cornerstone of modern deep learning, yet its theoretical convergence properties remain poorly understood in non-convex settings. To address this gap, we study SGD dynamics in deep linear neural networks - simplified models that retain key non-convex characteristics while remaining analytically tractable. Building on recent results for gradient descent, we extend convergence guarantees to SGD under regularized squared loss. We establish almost sure boundedness of SGD iterates. Using the framework of asymptotic pseudotrajectories from stochastic approximation theory, our analysis establishes then, for both deterministic and adaptive step-size choices, convergence of SGD to the set of critical points under mild conditions (without relying on Lipschitz properties or whitened data). To our knowledge, this is the first work providing almost sure convergence guarantees for SGD in deep linear networks, bridging the gap between theoretical optimization and practical stochastic training.

Visual anomaly detection aims to detect abnormal objects, while anomaly samples are often scarce during model training. Assembling multiple pretrained detectors can help by leveraging their transferable image scores and anomaly maps. However, most existing ensemble methods apply fixed fusion rules over saved predictions which struggle when pretrained detectors clearly disagree due to different training objectives, priors, and visual representations. To this end, we design AnomalyAgent, a post-training fusion framework for accurate visual anomaly detection. AnomalyAgent aligns heterogeneous scores and maps, summarizes score- and region-level disagreement, and treats fusion as a finite action-selection problem. It implements two types of actions: 1) score actions that determine how to fuse image-level detector confidence via consensus, robust aggregation, detector weighting, or score-map consistency; 2) map actions that fuse dense anomaly responses via consensus, detector weighting, geometric aggregation, or disagreement correction. A constrained vision-language model (VLM) is introduced to inspect high-conflict samples and select valid score action and map action to produce the final prediction. We evaluate AnomalyAgent on MVTec AD and VisA as primary benchmarks, together with seven additional public anomaly benchmarks and Real-IAD for large-scale score-level evaluation. Across these benchmarks, AnomalyAgent consistently improves over equal-weight fusion and achieves the best mean image AUROC and pixel AP among the compared fusion methods on MVTec AD and VisA, while demonstrating robust transfer across additional anomaly domains.

Continuous flow language models offer a path to parallel generation with tools for control and acceleration unavailable to discrete diffusion, yet effective techniques for building them at the billion-parameter scale remain less established. We present Lafa (Large flow language model), a fast, high-performing flow language model scaled to 8B parameters in a data-efficient manner. Lafa is a blockwise flow model with causal attention on the prefix and bidirectional, asynchronous flow within each block. This design enables efficient training by adapting pretrained autoregressive (AR) models and fast inference via prefix caching and blockwise parallel decoding. We train Lafa at 0.5B, 7B, and 8B scales with fewer than 30B tokens, more efficient than previous diffusion models, and evaluate it on instruction following, reasoning, and coding, where the learned flow matches or surpasses the quality of AR and can be distilled into flow maps for one- and few-step generation. Notably, flow training preserves or even improves pretrained AR capabilities via a multi-token prediction effect, which we leverage in a self-speculative decoding scheme that verifies flow-based drafts autoregressively within a single model. This yields 2.54x the throughput of an AR model at comparable quality, outperforming multi-token prediction and discrete diffusion methods. Our findings show that continuous flows can serve as a scalable foundation for language modeling.

Large language models (LLMs) can leverage semantic feature names alongside summary statistics for tabular feature selection, yet it remains unclear when such information improves subset selection relative to established statistical methods. We investigate this question within an 18-dataset framework comprising 10 datasets for Group Relative Policy Optimization (GRPO) training, 2 for checkpoint validation, and 6 held-out evaluation datasets. We compare five classical baselines, frozen Qwen2.5-7B and Llama-3.3-70B selectors, and two GRPO-adapted Qwen2.5-7B models, reporting classification ROC-AUC and regression R^2 separately and using dataset-wise ranks for cross-task comparison. Across five seeded evaluation partitions, classical selectors generally achieve higher predictive performance than the tested 7B LLM/RL selectors, including under a pre-specified ~33% feature budget with strict train-only classical fitting, while the 7B selectors retain compact subsets. A cardinality-matched semantic sensitivity experiment shows that, for Zero-shot Qwen-7B, informative feature names improve regression performance over anonymized and shuffled alternatives (+0.1964 and +0.1606 mean R^2), whereas aggregate classification differences are smaller (+0.0159 and +0.0084 ROC-AUC) and dataset-dependent. These results characterize when semantic information benefits LLM-based feature selection and show that classical selectors remain strong baselines for predictive tabular feature selection.

Statistical learning in high dimensions suffers from the curse of dimensionality. Fortunately, many functions of practical interest are compositionally sparse, meaning they admit a composition of local functions, each depending on only a few variables, allowing sample complexity to depend on local arity rather than ambient dimension. Yet realizing this benefit requires exploiting hidden compositional structure, which theory does not establish for practical gradient-based training. We study this gap with CSFRBench, a controlled benchmark that hides the target's compositional computation during training but retains it for diagnosis, and find that dense networks and sparsity-based baselines leave much of that benefit unrealized. Inspired by the spatial organization of biological neural circuits, we introduce Learned Wiring (LeWi): each neuron has a learned low-dimensional embedding, only nearby neurons connect, and embeddings and weights are trained jointly. LeWi improves sample efficiency on targets with hierarchical structure, intermediate reuse, or independent modules. Across 32 additional targets generated by independently sampled computation graphs, LeWi is the most sample-efficient method not given the target graph, with 27% lower mean learning-curve area than dense networks. Dense networks perform better on targets with a single high-arity function. LeWi's representations and connectivity align with the target computation. On 17 world-model prediction tasks spanning control, navigation, and manipulation, LeWi beats the dense baseline on 16 and wins 99/119 baseline comparisons. These results show that jointly learning connectivity and weights can exploit computational structure to improve sample efficiency.

Decentralised federated learning (DFL) trains personalised models on devices that average with their neighbours rather than with a central server. Real devices, however, are often offline for hours or days, and a client returning from an outage can recover only by aggregating the models its neighbourhood holds. We observe that the stalest of these models is usually the client's own, a case that existing staleness-aware rules, which discount only the models a client receives, do not address. We formalise aggregation after an outage as estimating the state of a fully available network and derive three principles. First, the receiver must be discounted like any neighbour. Every rule that weights it at least as much as each neighbour has an error that grows without bound with the outage, whereas discounting all candidates by relative staleness keeps the error bounded. Second, staleness must be measured as inherited lag rather than age, since local training resets a model's age but not the lag it absorbed from stale neighbours. We introduce the effective lag, which tracks this quantity at the cost of one scalar per message. Third, the discount must be steep, because absorbed lag persists in every later round. These principles define Self-inclusive Freshness-weighted Gossip (SFG), a family of drop-in aggregation rules with the communication and computation of decentralised federated averaging, whose guarantees hold for the stale receiver itself. On three image benchmarks and eight outage processes, including real smartphone availability traces, the lag-based members SFG-L and SFG-Cut consistently outperform decentralised federated averaging and recent age-of-information rules, and ablations show that each principle is necessary. The analysis also explains why tuned limited merging from gossip learning is effective.

Discrete diffusion models iteratively denoise a noisy sequence of categorical data such as text, enabling any-order generation with self-correction and in-filling capabilities. This makes them promising candidates for reasoning, planning, and constrained-generation tasks like coding and math. Yet masked and uniform diffusion erase a token's identity in a single corruption event, so their noisy states cannot explicitly represent partially resolved uncertainty in which several alternatives remain plausible. We close this gap with Binary Discrete Diffusion, a family of continuous-time discrete-diffusion processes whose noisy state is a set of candidate tokens per position, represented as a binary membership vector. Binary diffusion changes candidate sets through additions and removals, allowing information to be lost and recovered gradually while the diffusion state itself represents partial uncertainty. On Sudoku-Extreme, Boxoban, and graph coloring, the symmetric formulation of binary diffusion consistently outperforms strong, well-tuned masked and uniform baselines under matched architectures and controlled training, sampling, and tuning budgets, with its largest gains on the hardest instances, while remaining competitive on language modeling (Text8, TinyStories, OpenWebText). These results support candidate sets as an effective intermediate representation for iterative reasoning and show that the corruption kernel can materially change the computation learned by the reverse model.

How can we effectively oversee AI systems that surpass human intelligence? AI Safety via Debate proposes that two AI systems argue opposing positions and a judge, either a human or a weaker model, decides which is correct. By facilitating supervisors in providing high-quality training signals, debate aims to encourage truthful and safe behavior from AI systems whose capabilities exceed our own. However, human judgment is not neutral: people often rely on biases which advanced AI systems may learn to exploit. Drawing on insights from collective intelligence and social cognition, this paper investigates whether debate can guide human and LLM supervisors toward the truth despite their prior beliefs. We evaluate several protocol configurations. We also explore setups inspired by the wisdom of the crowd such as the number and diversity of judges, as well as interaction mechanisms including deliberation and voting. Overall, we find that debate increases accuracy relative to prior beliefs for human judges, whereas the consultancy baseline, in which a single expert argues for one answer, does not. For weaker language models, however, the results are mixed: debate provides no comparable gains in most settings and can even reduce accuracy. Among multi-judge protocols, weighted voting produces the highest accuracy for human supervisors, whereas for LLMs none of them significantly improves group accuracy, even as the number of judges increases. Notably, the behavior of minority judges in non-diverse deliberation protocols suggests that, under symmetric settings, incorrect positions could be more persuasive than correct ones. These findings support debate as a viable path toward scalable human oversight, but cast doubt on its benefits for weaker AI supervisors, raising concerns about fully automated training and control pipelines in high-stakes settings.

Post-training quantization (PTQ) compresses the weights of a trained model with limited compute and data. Down to 4 bits it is near-lossless, but at 2 bits scalar PTQ configurations fall to chance-level accuracy. We study three choices usually treated as implementation details that matter little at 4 bits but decide 2-bit quality: the quantization representation, the calibration data, and the recovery fine-tuning data, across six dense and mixture-of experts models (ranging from $3.8$ to $32$B parameters). Representation is the largest factor and vector codebooks restore usable models where scalar PTQ pipelines collapse. Most of the remaining loss in accuracy is on \emph{generative} tasks, and the amount of performance drop is model-dependent. Furthermore, at a short recovery budget, calibration and recovery data are coupled. The best recovery data depends on the calibration data, so the two must be chosen jointly. The resulting recipe involves a vector codebook representation, a small task-aligned calibration mixture, and an inexpensive codebook-only recovery stage. Including codebooks, it uses about $2.3$ effective bits per weight and retains $89$ to $97\%$ of full-precision question-answering (QA) accuracy where scalar PTQ is at chance, and, including recovery, uses $1.4$ to $30\times$ less compute than the strongest quantization-aware-training (QAT) baseline.

Humanoid-human interaction requires robots to operate under severe partial observability: during close physical contact, the human partner is occluded, truncated, or outside the robot's field of view, causing pose estimation to fail and often necessitating external tracking or privileged human state. At the same time, heavy mutual occlusions introduce errors into the human-human motion-capture data used for learning. We present REACT, a framework for learning deployable humanoid-human interaction policies from human-human demonstrations. REACT treats semantic contact structure as the key supervisory signal connecting demonstration processing, interaction learning, and onboard control. First, a VLM-guided contact modeling and retargeting pipeline identifies semantically meaningful contacts and transfers them onto humanoid-human geometry while preserving interaction structure. Using these annotations, privileged interaction policies learn physically grounded behaviors through interaction-aware contact rewards and are then distilled into deployable egocentric depth policies trained entirely from onboard sensing using photorealistic human depth rendering. To address severe partial observability, we introduce generative partner-state modeling objectives that predict partner motion from occluded egocentric views. We evaluate REACT in simulation and on a real Unitree G1 humanoid across five contact-rich interactions. Semantic contact modeling substantially improves preservation of demonstrated interactions, while generative partner-state modeling improves contact recall on four of five interactions, most under heavy occlusion. On real hardware, REACT performs all five interactions using onboard sensing alone.

Decision models support routing, guardrails, and output verification in LLM systems. However, item accuracy alone cannot establish whether models make correct decisions across controlled variations or assign reliable confidence to their predictions. We introduce DecisionPivot, a diagnostic benchmark comprising 48 tasks and 40,033 items across seven decision roles. Its core track contains 7,800 items from 13 generated task families, organized into contrast groups with computed answers verified by independently implemented solvers. The complementary reference track contains 32,233 items from 35 public test sets, preserving their original evaluation metrics. We evaluate 40 decision models and 14 general models using documented, model-specific protocols, assessing item and contrast-group accuracy, probability ordering, calibration, and empirical selective coverage. Our results reveal substantial differences obscured by aggregate scores: two 27B models separated by only 2.6 points on the reference track achieve contrast-group accuracies of 70.1% and 44.2%. Across 2,400 binary contrast pairs, Perplexity Decider correctly ranks the positive member above the negative member in 92.3% of pairs but answers both members correctly in only 74.1%, exposing a gap between probability ordering and decision correctness. Models with similar item accuracy also differ in the reliability of their high-confidence predictions. On a separate 1,296-item subset, GLM-5.3 achieves 93.5\% contrast-group accuracy at a higher reasoning setting, compared with 81.5% at its lowest effort. By linking errors to pivotal details and difficulty levels, DecisionPivot provides fine-grained diagnostics of decision failures that aggregate evaluation alone cannot reveal.

Unresolved atmospheric processes can introduce history dependence beyond the short input windows used by many weather forecasting models. We introduce GraphCast--Mamba, which augments a frozen, pretrained $1^\circ$ GraphCast backbone with a trainable Mamba correction branch combining graph message passing and Mamba layers. The branch retains a temporal memory state across successive six-hour forecast steps, learning corrections informed by temporal context beyond GraphCast's two-frame input. Training is confined to the correction branch, using ERA5 data from 2015--2021. After 2,500 optimizer updates, including warm-up, the selected model achieves substantial improvements relative to the uncorrected GraphCast backbone. It reduces the GraphCast loss (normalized weighted mean-squared error; MSE) by 14.42% when aggregated over the full ten-day forecast and by 21.66% at day 10 (240 h after forecast initialization). At this final forecast lead time, root-mean-square error (RMSE) decreases by 7.09--14.60% across all 11 evaluated variable groups, spanning temperature, wind, pressure, geopotential, specific humidity, and precipitation. These results demonstrate broad improvements in ten-day weather forecasts through a short correction-training schedule while retaining the pretrained backbone.

We study variance reduction for non-convex optimization when gradient evaluations return with a delay. The server can update the parameter while earlier evaluations remain unfinished. We first analyze an ordered recursive momentum estimator for stochastic optimization. Under average smoothness and bounded variance, its oracle complexity is $O(1+σ^2ε^{-2}+LΔ_f(1+σ/ε+τ)ε^{-2})$, where the target is an expected squared gradient norm of at most $ε^2$ and $τ$ bounds the feedback delay. We next study finite sums with $n$ components and obtain $O(n+LΔ_f(\sqrt n+\min\{τ,n-1\})ε^{-2})$ evaluations. An asynchronous snapshot variant continues updating during full-gradient computations under an explicit bound on their completion time. For the stochastic method, we further replace a maximum-delay bound by a bound on the weighted amount of missing feedback.

Laboratory measurements are essential for clinical decision making but typically require invasive blood sampling. Recent research suggests that ECG recordings may contain information beyond cardiac rhythm and provide information related to laboratory values. This thesis investigates NT-proBNP prediction from ECG recordings and patient metadata using a multimodal deep learning framework, while considering the physiological plausibility of ECG-based prediction. It also compares a 1-lead configuration using lead II with the corresponding 12-lead configuration. Additional experiments on glucose, hemoglobin, and potassium are presented in the appendix. Regression, binary classification, and three-class classification experiments were conducted using MIMIC-IV, with the three-class task based on clinically established thresholds. The 12-lead configuration consistently achieved better performance. Regression reached a mean absolute error of 534.98 pg/mL for 1-lead and 514.27 pg/mL for 12-lead. Binary classification achieved AUROCs of 0.849 and 0.876, respectively, while three-class classification achieved macro-averaged AUROCs of 0.738 and 0.752. The results indicate that precise NT-proBNP regression remains challenging, whereas broader concentration ranges can be classified more effectively. The findings are consistent with the physiological hypothesis that ECG characteristics and NT-proBNP concentrations reflect shared cardiovascular processes. ECG-based NT-proBNP prediction may therefore provide complementary information, particularly for classifying broader concentration ranges. However, laboratory measurements remain necessary, and validation on independent datasets is required.

For efficient deployment of large language models (LLMs), methods of ultra-low-bit backbones reduce weight storage, memory traffic, and linear-layer execution cost, while multi-token prediction (MTP) methods reduce sequential decoding steps by predicting multiple future tokens in parallel. Although complementary, the direct combination of these methods cannot accelerate decoding at comparable quality: masked drafting lacks intermediate token input supplied in next-token training, while wider computation and causal key-value (KV) updates can offset the savings from fewer steps. To address these challenges, we propose BitMTP, an algorithm-system co-design framework for efficient MTP on ultra-low-bit LLM backbones. A joint next-token prediction and MTP objective equips a single model with both autoregressive (AR) generation and block-causal drafting. Confidence-only decoding, speculative-only decoding, and hybrid decoding provide different quality-throughput operating points without using separate draft and verifier models. Accepted-prefix alignment keeps the committed prefix, materialized causal KV boundary, and CUDA Graph replay state consistent for continued generation. BitMTP uses this contract to select width-specialized low-bit execution and reuse captured graphs over shared KV storage. On NVIDIA H20, BitMTP with hybrid decoding achieves five-task mean speedups of 2.2x on BitNet b1.58-2B and 2.1-2.4x on 1.7B-4B Bonsai models at comparable task quality. On RTX 3090 and Jetson Orin, it further achieves up to 2.0x in five-task mean throughput and 3.3x on MBPP+, respectively. These results demonstrate that BitMTP accelerates ultra-low-bit LLM generation across backbones, model sizes, and GPU types. The source code is released at https://github.com/PolyU-VCLab/BitMTP

Probing studies of latent reasoning often read a model's state where the answer begins and take a failed probe as evidence that a value is absent. We test this with a digit copied from the question, in the looped model Ouro-1.4B and the standard transformer Qwen2.5-1.5B. Comparing prompts that differ only in that digit shows it is present at the last prompt token, in a code shared across questions (readout about 0.9). A probe that sees one prompt at a time, the usual setting, barely reads it: 0.36 in Ouro and 0.19 in Qwen even after 5,249 training questions. The digit's change is small rather than crowded: random noise lands on equally busy directions, while a random word in its place moves the state far more and is readable. Since every varied token leaves a trace, a successful probe is ambiguous too. We give a protocol, including a synthetic null for generic mixing, for testing computed values against copies and mixtures of the inputs. In pre-registered studies of two-step problems it finds a small signal for the unwritten intermediate a+b just before Ouro writes the answer's last digit, and its sign follows the operation, though mixing alone produces much of that sign-following part. Holding that value at another problem's from the first loop on does not move the written digit toward the other problem's, while the same patch of the answer does: the value is readable but not used.

Deep State-Space Models (Deep SSMs) have shown promising performance and computational benefits in processing of time-series data with complex temporal dependencies. Spiking neural networks (SNNs) constitute a biologically inspired class of deep SSMs, where computing blocks communicate through spikes, discrete single-bit events rather than continuous activation functions. A particularly important task in time-series processing is forecasting, i.e. predicting future values of a sequence from past observations alone. However, how to best perform forecasting with SNNs is an open question, where the standard autoregressive nature of the data processing with SNNs requires values that are not available during the forecasting phase. In this article, we address this challenge. We explore two closed-loop forecasting strategies, in which the network's own generated values are fed back into the network as if they were observed past values: (i) model-level and (ii) layer-level closed-loop forecasting. We evaluate these approaches on the ETT-small datasets, a standard benchmark for time-series forecasting, under both spike-based and continuous activation functions. We compare the proposed closed-loop approaches against state-of-the-art methods such as SpaceTime, as well as the zero-masking baseline in S4. Both of the proposed approaches perform comparably to or better than zero-masking with S4, and in most cases match SpaceTime. These findings highlight the trade-off between continuous and low-bit quantized activations, and demonstrate that the proposed Deep SSM architectures provide a competitive framework for forecasting.

Long-term assistants must follow unfinished tasks across days as information and circumstances change. Finding information or starting an action, however, does not ensure that the task gets completed. We introduce OptionWorld, a benchmark that traces information acquisition, judgments, actions, and outcomes in simulated daily-life tasks to identify where current LLMs fall short as long-term personal assistants. In runs with archived decision inputs, the key facts remain in the model's last input before the deadline in 35.5% of failed instances. Failure analysis identifies the three most frequent problems: needed information is not obtained or combined, available opportunities to act go unused, and agents impose restrictions that the task does not require. Three follow-up experiments show that supplying missing facts restores task success in some cases; action reminders can restart work while subsequent steps still miss the deadline; and correcting a supplied note that wrongly forbids contact restores permitted inquiry. These findings locate shortcomings in maintaining task state: connecting facts to unfinished work, following through before deadlines, and separating actual restrictions from the agent's assumptions.

Looped Transformers increase effective depth by repeatedly applying shared layers, offering a way to scale computation without increasing parameter count. However, full-stack looping places context construction and iterative state refinement within the same recurrent computation, without explicitly distinguishing their roles. We introduce the Asymmetric Looped Transformer (AsymLoop), which separates single-pass context encoding from recurrent state refinement. A single-pass causal encoder produces layer-specific Global key-value (KV) caches that remain fixed across iterations, while a weight-tied recurrent decoder generates Local KV from its evolving state. Joint attention over both sources under a single softmax allows each update to integrate persistent contextual information with the current intermediate state. For training, we apply a heavy-tail-guided layerwise learning-rate method to the encoder and decoder. Experiments on FineWeb-Edu across four model scales, from 60M to 1B, show consistently lower validation perplexity and higher average downstream accuracy than Full Loop under matched effective depth and training token budgets. At 1B, AsymLoop improves average zero-shot accuracy across seven commonsense reasoning benchmarks from 49.58% to 50.85%. Further experiments extend these benefits to longer training budgets, an OLMo-style backbone, and sparse mixture-of-experts models. These results support asymmetric context encoding and state refinement as an effective design for recurrent language models.

We introduce a two-stage data-driven neural framework for discovering dynamically admissible extreme scenarios. Rather than assuming prior knowledge of the governing equations, the first stage infers the effective dynamical constraints directly from observational data. When the underlying physical system possesses suitable symmetries, the learned dynamics can be represented in a geometry-agnostic form, independent of any fixed grid, coordinate system, dimensionality, and intrinsic curvature. In the second stage, a physics-informed neural-search uses this learned dynamics to uncover rare, high-impact coherent structures that may not have appeared in the original observations. We demonstrate this framework through representative examples of increasing dimensionality and geometric complexity. Starting from a traditionally learned Duffing dynamics, we find a homoclinic pulse as a plausible localized extreme scenario. Next, since the same equation also appears in the stationary nonlinear Schrodinger system, with the coordinate interpreted as space instead of time, this pulse becomes a standing soliton. We then retrain the first stage using a geometry-agnostic neural representation, enabling the learned dynamics to be lifted to higher-dimensional space to obtain higher-dimensional solitary profile, known as Townes soliton. Finally, with a Liouville system, we demonstrate that dynamics learned from data on a (1+1)-dimensional Lorentzian quasi-manifold can be transferred to a (2+0)-dimensional Euclidean manifold to identify an instanton. Our approach can be applied to other coherent spatiotemporal structures, including kinks, rogue waves, compactons, peakons, and potentially even singularities, with applications in risk analysis, nonlinear forecasting, and anomaly detection.

AI agents increasingly act on behalf of people or other agents, purchasing services, exchanging information and delegating work. What an agent must respect is written in instructions, policies or agreements, but text alone cannot make it comply. It may make mistakes, be redirected by what it reads, or pursue different interests. For an economy of such agents to function, parties need agreements they can rely on without inspecting or trusting one another's agents. We introduce Proof-Carrying Contracts (PC2), a framework for making agreed constraints enforceable throughout an agent's interactions with a system. Before a requested operation can run, the agent must supply a machine-checkable proof that it preserves the agreed constraints in every case allowed by its model. An execution system called the Airlock checks the proof and controls access to the operation. The parties specify the conditions to preserve, while the agent supplies the reasoning needed to justify its decisions and remains free to adapt its strategy. This avoids having to design a separate permission policy and prove that every decision it permits is safe. We prove that the constraints hold at every instant during execution, even for an adversarial agent, and identify when this implies safety in the world. Reusable capability modules provide tools and guarantees from which contracts can be constructed. PC2 provides a foundation for an agent economy whose agreements protect both those delegating work and those carrying it out.

Learning latent dynamics is fundamental to building action-conditioned world models for planning. Among existing families, Joint Embedding Predictive Architectures (JEPAs) offer a natural framework by modeling transitions directly in latent space rather than at the pixel level. However, two issues remain in JEPA-based world models: (i) physical grounding lacks explicit aggregation of temporal context and (ii) action planning always overly relies on generic sampling distributions. To this end, we propose $\mathcal{G}$-JEPA, which incorporates temporal context into physical grounding for latent world models and couples the learned models with prior-guided action planning. Specifically, during learning, we incorporate a shared causal temporal module that aggregates temporal context within observation and prediction sequences under physical state supervision, grounding both in task-relevant physical states. When planning, we further introduce action inference through stepwise conditioning of a generative action distribution on the current and goal latent representations together with their evolving discrepancy, yielding anchored proposals that guide model-based online refinement. Together, physical state supervision guides temporal representation learning, while the action prior guides search through the learned world model. Experiments demonstrate improved planning performance on manipulation tasks, including tasks with increasing temporal separation between initial and goal observations. Code is publicly available at https://github.com/myl11045/G-JEPA.

When a language model develops capabilities not traceable to its individual components, when a multi-agent AI system outperforms each agent on its own, or when a human-AI pair succeeds where either fails in isolation, AI researchers describe it as a case of ``synergy'' or ``emergence''. However, these are usually informal assessments not supported by a theoretical framework, making these claims impossible to compare or even meaningfully dispute. Here we argue that \emph{information decomposition can serve as a unifying formal framework to operationalize assessments about synergy in AI systems}. Information decomposition provides a way to distinguish qualitatively different \emph{types} of information, including unique information available from individual components, redundant information shared across multiple components, and synergistic information accessible only from parts taken together. This gives claims of emergence a precise, falsifiable criterion: a system's behavior is emergent only if it carries synergistic information about task-relevant outcomes. We apply this lens to attention heads, autonomous agents, and human-AI teams, showing how it organizes existing knowledge across scales and substrates, while also generating cross-level conjectures that siloed fields would miss.

Human mobility data are fundamental to Web-based location services and broader urban applications, and mobility generation provides synthetic trajectories when real data are costly or inaccessible. Although recent mobility generators have incorporated collective mobility patterns, they primarily focus on individual trajectories and aggregate population dynamics, leaving evolving community structures formed by user-location interactions insufficiently modeled. In this paper, we study community-consistent human mobility generation. To this end, we formulate mobility generation as the generation of dynamic user-location graphs to model community structures and propose MOTION, a Multi-agent framework for cOmmuniTy-consIstent human mObility geNeration. MOTION consists of user agents, location agents, and a global manager agent. MOTION first detects coherent communities from historical mobility using the Leiden algorithm and extracts community-conditioned interaction priors to guide the generation process. Guided by these priors, the user agents plan mobility, search destination candidates, and infer user movements from the individual view. In parallel, selectively activated location agents infer likely user arrivals from the community view. The manager coordinates both views, diagnoses interactions causing deviations, and attributes targeted reflection signals to revise user decisions, location attraction, and source-user selection. Experiments on two real-world datasets show that MOTION outperforms baselines by at least 4.05% across reported metrics of individual trajectories, population and flow distributions, and community and graph structures. Further analyses confirm its robustness, privacy preservation, and downstream utility. The code is available at https://anonymous.4open.science/r/MOTION-04C4.

Joint-Embedding Predictive Architectures (JEPA) offer an appealing approach to self-supervised learning, but are typically prone to representation collapse. LeJEPA resolves this by introducing a theoretically grounded distribution-matching objective that drives the embedding towards an isotropic Gaussian. However, this approach is fundamentally tied to Euclidean assumptions, and its global target does not reflect domain-specific geometric priors, such as the use of non-Euclidean embedding geometries, e.g., equivariant CNNs or hyperbolic ViTs. We thus introduce LocaLeJEPA, a framework motivated by a manifold-theoretic reformulation of LeJEPA's entropy-maximization property. The accompanying LocaLeReg, a novel locally Euclidean distribution-matching regularizer, enables a stable and scalable implementation of the LocaLeJEPA framework. We then perform extensive experiments in the vision setting on models trained on ImageNet-1K. LocaLeJEPA significantly outperforms LeJEPA when paired with hyperbolic ViT-S/8 and equivariant ResNet-50 backbones, and also with ordinary Euclidean ViT-S/8 and ViT-L/14 backbones. This holds both in terms of linear and k-NN classification performance on ImageNet and on a series of downstream tasks tested. We view the LocaLeJEPA formalism as a principled approach to the JEPA paradigm that better respects intrinsic data geometry while also accommodating generalizations to exotic embedding geometries. Code available at https://github.com/alessioborgi/LocaLeJEPA.

General-purpose multimodal models have shown promise as direct robot action policies, yet whether they grasp the physics underlying real-world interactions and can apply that understanding to guide action remains underexplored. We investigate this question using contact-rich manipulation as a representative testbed, examining tactile understanding as a concrete instance of this broader capability. Contact-rich manipulation requires interpreting tactile feedback and adjusting actions after obstruction. Existing learned tactile policies acquire these capabilities through specialized pretraining and per-task optimization, with substantial data and training costs. We present GPT-as-Tactile-Policy, which summarizes expert demonstrations into task guidance and tactile references for contact assessment and corrective action. GPT-6-Astra combines this context with multimodal observations and interaction history to generate control commands and adjust actions through contact feedback. Without additional robotics-specific pretraining or per-task policy fine-tuning, our method achieves 93.8\% mean success across five UniVTAC tasks. Overall performance is comparable to state-of-the-art policies, with higher success on selected tasks. On Insert Hole, tactile feedback improves success from 31\% to 99\%. Execution analysis further demonstrates contact-guided probing and correction, supporting manipulation progress through explicit expert experience and online feedback.

We extend the Bochner-Weitzenböck decomposition to cellular sheaf Laplacians, splitting it into a positive semidefinite term, the Bochner sheaf Laplacian, and a block diagonal term, the Forman curvature. We show that the Bochner Laplacians form a class of cellular sheaves that extends both connection and matrix-weighted Laplacians, and generalizes signed graphs to $d$ dimensions. We give a holonomy characterization of their kernel, prove the related Cheeger inequalities, and attach to them a random walk that generalizes the Gremban expansion. For the Forman term, we prove a discrete Bochner theorem, which bounds the first cohomology of a cellular sheaf with nonnegative curvature, and we read the whole decomposition as an opinion-dynamics model. We then apply this machinery to Sheaf Neural Networks, whose learnt Laplacian is otherwise unobserved. The Bochner interpretation yields a node-level conviction score that acts as an operational importance measure under frozen-model interventions, while the Forman profile provides a local descriptor of the learnt operator.

Agents are starting to sell to other agents. Seller agents write the listings of tools, services and products, and buyer agents read these listings when they choose what to use. We show that seller agents told only to be selected more often practice generative engine optimization on their own, and that this distorts the market for sellers and for buyers. We build tool markets from real data tools on Amazon and Instagram and run every tool on every question, so we know which tools can answer each question. Under all five production models we test, sellers rewrite their descriptions to claim more and more fields, including fields that their tools return for fewer than half of the targets. We call this process toolflation. Nothing in the market limits it, and how far it goes depends on the model, as the number of claims that contradict the sellers' own evidence differs by an order of magnitude between models. Toolflation shifts the buyers' choices between sellers, and the most capable tool loses share in seven of the ten combinations of seller model and platform. Under four models, the choices stay with capable tools and buyer agents keep their task success. Under the fifth, sellers copy one another until the descriptions no longer tell the tools apart. Buyer agents then pick tools by their position in the list, tools that answer almost nothing win a quarter to two fifths of the choices, and task success falls from 89% to about 30% on Amazon, close to random choice, and from 77% to 40% on Instagram. The sellers did only what their vendors asked, and the cost fell on buyers they never dealt with. We therefore argue that honesty toward third parties under competition is part of safety alignment, and that models should be tested as sellers before they are deployed as sellers.

We present a tool for the comparative study of mythological and religious texts. Traditional manual analysis is highly labor-intensive and difficult to scale. To address this, we use neural language models to embed texts into a shared semantic space, so that users can find semantic parallels across cultural traditions in an interactive interface. Embedded fragments form distinct semantic structures whose coarse geometry aligns with well-documented historical relations between traditions. We quantify how much of that geometry is already explained by lexical overlap and by proper names. The current system is available at https://mythoscope.io, and the installable package is available at https://github.com/haafus/MythoSemantic/blob/main/RUN.md.

This study examines whether short, minimally preprocessed EEG segments contain sufficient information to distinguish between individuals classified as extraverted or introverted according to the Myers--Briggs Type Indicator (MBTI). EEG recordings from 50 participants were divided into 4-s segments and analyzed using a subject-independent nested 5-fold cross-validation procedure to compare EEGNet, EEGProgress, and a newly proposed compact convolutional architecture, SSCNet. SSCNet achieved the highest mean accuracy (0.6481), AUC (0.7063), and F1-score (0.6276). The proposed architecture also showed the lowest variability across folds. SSCNet demonstrated the best computational efficiency, with the shortest median training epoch time and the highest training throughput despite having a similar number of trainable parameters to EEGProgress. The results suggest that MBTI Extraversion--Introversion preferences may be moderately predictable from short time-domain EEG segments without manual spectral feature extraction, while highlighting the importance of subject-independent validation and computationally efficient model design.

Large Language Model inference over long contexts is increasingly bottlenecked by the cost of attending to a growing Key-Value (KV) cache. Sparse attention reduces this cost by accessing only a fraction of the cache, but existing methods face a trade-off between deterministic retrieval, which captures high-scoring tokens but discards residual attention mass, and sampling, which improves coverage at the cost of higher variance. Hybrid methods combine top-$k$ retrieval with uniform sampling, yet still exhibit a substantial gap from dense attention at aggressive sparsity. We introduce \emph{Stratified Sparse Attention (SSA)}, a unified framework combining deterministic retrieval with score-guided sampling. SSA deterministically attends to high-scoring tokens while sampling the residual using proxy attention scores and correcting for their inclusion probabilities, thereby recovering attention mass beyond the retrieval set. SSA encompasses top-$k$, top-$p$, and hybrid top-$k$+sampling as special cases. Across RULER, LongBench, LOFT, and AIME, and five models ranging from 3B to 27B parameters, SSA consistently improves over the baselines at matched sparsity. On RULER-HARD with Ministral-3-8B, SSA improves accuracy over the strongest practical baseline from 81.8\% to 84.2\% at 32K and from 63.8\% to 69.3\% at 128K under $50\times$ attention sparsity, recovering 94\% and 85\% of dense-attention performance, respectively. With optimized CUDA kernels, SSA achieves up to $2.96\times$ decode speedup over FlashInfer at 1M context, demonstrating that its accuracy gains translate into practical long-context inference acceleration. We open-source our code at https://github.com/skylight-org/SSA-Paper-Code.

A neural model is the set of functions that can be expressed by a given neural network architecture. Neural models contain natural submodels, such as those obtained by setting some neurons to be inactive. We study the geometry of neural models and ask whether they prefer their submodels. By this we mean whether local minimizers of the loss function lie in a given submodel. We introduce a geometric framework to determine if this occurs. For shallow multilayer perceptrons, we show that subnetworks with inactive hidden neurons are never preferred. For deep multilayer perceptrons, we give conditions under which subnetworks with inactive hidden neurons are not preferred. For lightning self-attention, lower-rank submodels are usually not preferred, but we identify examples where they are because they lie on a boundary of the model. Our results show that, aside from a few exceptions, preference for submodels cannot be explained by the geometry of the neural model, and must instead come from other sources such as training dynamics.

Large language models leave model-specific traces in their generated text, but it remains unclear how finely these fingerprints can distinguish related models and which signals remain informative as obvious cues are reduced. We study this through fingerprint resolution, considering both model resolution, from different families to distilled and adapted variants, and signal resolution across behavioral, stylometric, lexical, and semantic evidence. We progressively suppress surface cues, paraphrase outputs, vary the amount of available text, and test generalization across prompt distributions and languages. Our results show that fingerprints persist beyond broad family differences, remaining detectable in distilled models and even LoRA adaptations of the same base model. Although attribution weakens as surface and lexical cues are altered, substantial model-specific signal remains. These signals transfer across prompt sources and languages, with semantic representations proving most robust across languages. These findings show that model identity is recoverable at multiple resolutions, but that recoverability depends on model similarity and the evidence available in the text.

World-Action Models (WAMs) have emerged as a promising paradigm for embodied generalist policies by jointly modeling future world evolution and robot actions across diverse manipulation tasks. However, representing diverse futures is only the first step: a capable WAM must also generate futures that match the requirements of the current task and execution progress. In long-horizon manipulation where similar observation-action segments occur across different tasks and stages, existing approaches may generate locally plausible yet misdirected rollouts to drift toward wrong task stage, i.e., a phenomenon we term Stage Confusion. To address this problem, we propose TaskWAM, a closed-loop, stage-aware framework that represents fine-grained task requirements with Mode Tokens and tracks execution progress with State Tokens, forming a structured task state to steer world--action generation toward intended future. We further introduce StageBench, a controlled multi-stage manipulation benchmark with parameterized stage variations and automated evaluation.Extensive experiments on both simulation benchmarks and real-world tasks demonstrate that TaskWAM effectively reduces confusion and achieves competitive performance with state-of-the-art methods. Moreover, TaskWAM provides strong plug-and-play transfer across heterogeneous WAM architectures and robot embodiments, including single-arm robots, dual-arm grippers, dexterous hands, and humanoid robots.

GPU kernels are critical to the performance of today's AI services, yet the same operators are rewritten repeatedly: for each GPU generation, numeric format, serving regime, and kernel language. Every rewrite must preserve the function computed by the implementation it replaces. Today, this obligation depends almost entirely on differential testing: run the new kernel and a trusted reference on randomly sampled tensors, then compare their outputs within a tolerance. This approach has blind spots that additional test cases cannot eliminate. For example, a logical bug may produce an error smaller than the tolerance required to accommodate a correct optimized kernel. We propose symbolic equivalence checking, which executes a kernel and its reference on symbolic inputs, turning each output cell into an exact algebraic expression, and then checks whether the paired expressions agree for all inputs. Because these expressions are exact, symbolic equivalence checking distinguishes floating-point error from logical bugs, while a single check covers the entire input space. We implement Racer, a PTX-level symbolic equivalence checking engine, with three techniques that make this approach practical for production kernels: (1) Shape Reduction to control checking cost, (2) a lockstep CTA scheduler to handle concurrency, and (3) idiom templates with exact numeric refutation for otherwise inconclusive results. Across 2,408 production kernels from 41 collection targets, Racer found 82 real bugs in 23 libraries, including 76 previously unknown bugs and 41 confirmed so far.

The increasing complexity of Industrial Control Systems (ICS) and their integration with IT infrastructures has significantly expanded the cyber threat landscape. Traditional cybersecurity models, primarily based on the Confidentiality, Integrity, and Availability (CIA triad), fail to capture the broader societal and operational impacts of cyberattacks. This paper proposes a novel framework that integrates Digital Twin technology with a Cyber-Social Security model to simulate, analyze, and quantify cyber threats in critical infrastructures. Virtualized ICS environments built with Mininet and MiniCPS are used to model attack scenarios and assess their consequences across technical and social dimensions. At the core of the framework lies the Cyber-Social Vulnerability Scoring System (CSVSS), a quantitative metric that extends CVSS with a Social Impact dimension grounded in the Entity and Place pillars, producing a single, reproducible severity score that reflects real-world consequences. Crucially, the framework does not automate risk assessment but keeps the human analyst at its center: a Large Language Model (LLM) performs a first-pass analysis, mapping vulnerabilities and drafting risk reports, while the analyst, supported by explainable AI feedback, validates the findings and retains decision authority over critical cases. Experimental results demonstrate the framework's ability to model and score cyber-social impact, while the human-in-the-loop design is presented as a conceptual principle to preserve oversight.

We investigate whether a 2D Convolutional Neural Network (CNN) trained on minimally preprocessed EEG signals can classify hypnotic susceptibility across distinct brain regions. Sixty-two male participants (32 Control, 30 Guided Imagery / Hipno group) underwent a 20-minute EEG recording using a 256-channel EGI cap. A sliding-window approach evaluates all 10-second windows within the recording for each of seven predefined brain regions (prefrontal, frontal, temporal, central, parietal, occipital, and facial), yielding up to 868 independently trained models. Initial screening with a segment-level 90/10 random split showed that five regions exceeded the 0.67 accuracy threshold, with the occipital region achieving the highest mean accuracy (0.721) across 90.3% of windows. Subject-independent evaluation via Leave-One-Subject-Out (LOSO) cross-validation, however, substantially revised this picture: occipital performance collapsed to near-chance (mean ACC 0.453, AUC 0.429), revealing that high random-split accuracy was driven by subject-specific EEG fingerprints rather than genuine group differences. In contrast, the frontal and prefrontal regions maintained above-chance LOSO accuracy (0.646 and 0.624, respectively), demonstrating genuine generalisability. These results underscore the necessity of subject-independent validation in EEG-based classification studies and highlight anterior cortical regions as more reliable neural markers of hypnotic susceptibility.

Modern LLM agents persist state across sessions via two channels: long-lived memory files silently prepended to the system prompt, and scheduled tasks waking the agent without user input. A one-time indirect prompt injection committed to either channel can re-fire across sessions long after the attacker leaves, threats we term carryover attacks. Existing prompt-injection benchmarks evaluate attacks only in the session where malicious content is first observed, leaving the gate governing when and how often a planted payload later re-fires unmeasured. We present CarryoverBench: 12 personal-assistant scenarios and 101 attack tasks across three trigger channels: Instant, same-turn IPI; Autonomous, heartbeat-scheduled; and Latent, memory-resurfaced. A rule-based verifier scores planting and triggering as independent stages. Across five frontier models, planted payloads re-fire without measurable decay over five rounds; conditional on successful planting, the trigger stage often provides little additional protection. Porting the Latent channel to a second runtime, Hermes, attenuates but does not eliminate the effect, suggesting that carryover defense requires runtime-layer controls in addition to prompt-level interventions.

Generative-prior-based super-resolution (SR) methods leverage implicit knowledge learned from large-scale models to enhance low-resolution inputs. However, such priors are inherently stochastic and optimized for diverse generation, conflicting with the deterministic, high-fidelity reconstruction required for SR. It is also tied to the base model, costly to update, and prone to hallucinations as real-world data evolves. To address these limitations, we introduce WorldSR, a novel framework that augments SR with explicit, entity-aligned visual world knowledge. WorldSR automatically decomposes the input into semantic entities and performs entity-wise retrieval of structurally and semantically aligned images from a large-scale external corpus. For efficiency and reliability, we incorporate a memory mechanism to reuse previously retrieved results and a multi-step filtering process to refine candidate references. The selected references as world knowledge are then integrated through a novel multi-reference super-resolution model, which provides fine-grained constraints for detail synthesis and structural recovery. Experiments on real-world SR benchmarks show that WorldSR consistently improves structural fidelity and visual quality, outperforming existing generative super-resolution methods. Our results suggest that entity-level grounding in external visual world knowledge provides a simple and effective complement to model-bound generative priors.

Vision-language trackers (VLTs) accept descriptions intended to assist tracking, but accepting text does not establish that its content changes tracking decisions, particularly as visual evidence deteriorates. We introduce DISTRACT, a matched-intervention benchmark that holds video, target, and distractor fixed while crossing four text priors with four image degradations along a controlled severity axis. We evaluate five VLTs on four datasets using behavioral contrasts, layerwise representation probes, and activation patching. Molmo2 and JointNLT exhibit directed text use. Some effects contract under degradation, whereas others show no significant endpoint change after multiplicity correction. UVLTrack, MMTrack, and DUTrack respond only weakly to changes in text semantics, with no directed behavioral effect surviving FDR correction in the tested settings. Model-internal analyses reveal that at decision-near states, correct-versus-conflicting text is decodable in Molmo2 and JointNLT but not in the other three systems. In the responsive models, decodability peaks at late states where replacement produces little behavioral change, whereas replacing selected earlier states redirects conflict-associated trajectories toward the target. Because matched-correct donors do not significantly outperform unrelated donors, these interventions identify earlier intervention-sensitive states without establishing target-specific semantic rescue. DISTRACT thus exposes when and where text affects tracking under degraded visual evidence.

Code-based video generation offers fine-grained control, targeted editing, and component reuse compared with direct video generation. This approach requires multimodal large language models (MLLMs) to translate visual and temporal requirements into executable programs that produce complete videos. However, existing code-based benchmarks primarily target specialized tasks, leaving complete video creation across diverse creative categories and input modalities insufficiently evaluated. We introduce Code2Video-Compass, a general-purpose benchmark for code-based video generation across text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) settings. Code2Video-Compass comprises 2,104 tasks spanning 10 categories, 24 subtypes, and three complexity dimensions covering construction, dynamic animation, and state dependency. It combines tasks curated from textual instructions and visual references with corresponding T2V, I2V, and V2V tasks derived from complete professional and user-created videos. We propose a dual-level evaluation framework combining 12 objective metrics and five subjective metrics. Objective evaluation covers operational performance, video quality, text-video alignment, and reference fidelity, while subjective evaluation uses an MLLM judge to evaluate instruction following and holistic video quality. Extensive evaluation of 15 recent proprietary and open-weight MLLMs from nine families reveals that even leading models face challenges in reliably generating high-quality complete videos aligned with textual and visual requirements, with performance generally declining as construction, dynamic animation, and state-dependency complexity increases. We anticipate that Code2Video-Compass will advance reliable, coherent, and controllable code-based video generation and inspire broader progress in programmatic visual creation.

Deep learning models have become essential tools in medical imaging, yet training them end-to-end for diverse tasks remains challenging due to limited labeled data and computational constraints. To address this, we introduce Brain2Vec, a foundation model pre-trained on 74,425 T1-weighted Magnetic Resonance Imaging (MRI) scans pooled from multiple studies. Brain2Vec can serve both as a weight initialization backbone for fine-tuning downstream tasks and as a standalone feature extractor. We evaluated Brain2Vec in three tasks: downstream predictive performance, convergence speed, and embedding quality. Using Brain2Vec's pretrained weights as initialization and fine-tuning on labeled data for a brain age prediction task, results demonstrate improved final accuracy and faster convergence compared to random initialization. Additionally, its frozen embeddings outperform anatomically defined volumetric features on four of five downstream classification tasks and perform competitively with a state-of-the-art foundation model pre-trained on substantially more data. Lastly, we show that Brain2Vec encodes information primarily from a spatially localized and neuroanatomically meaningful set of brain regions, and that manipulating a subset of latent features produces monotonic, region-specific effects on the reconstructed anatomy. Brain2Vec code and trained models are available as an open-source package on https://github.com/CBICA/Brain2Vec. Users can apply Brain2Vec locally by installing the package, or directly through the NiChart cloud platform https://cloud.neuroimagingchart.com, where Brain2Vec features can be derived without requiring local installation or specialized infrastructure.

Large language model agents are often given an inference budget and asked to maximise success within it. A multi-turn agent has to allocate this budget across many environment interactions. An external harness can enforce the budget, but it remains unclear whether the agent itself can learn when and how much to spend. We ask whether agents can learn this allocation across dynamically specified budgets and which rewards best teach this behaviour. We adapt existing length penalties to budget constraints spanning multiple resource types, introduce new reward formulations, and compare eleven objectives in a shared training pipeline. We train Qwen3.5 models at 4B and 9B on AppWorld with budgets over tokens, API calls, and environment steps, and on the BabyAI gridworld with turn budgets. We train a separate policy for each environment and resource type, with each policy trained across budget values. Training with rewards that penalise overspending substantially improves within-budget success and natural compliance over untrained models and over training on task success alone, with agents learning to adapt their resource use to stated budgets and task difficulty. These gains extend to unseen budgets and, zero-shot, to a shifted task distribution and a new benchmark. Reward objectives trade off success within budget and natural compliance. A sigmoid penalty for overspending most often achieves the highest within-budget success at an individual budget, while a fixed negative reward for overspending generally yields the strongest natural compliance. Combining penalties for overspend frequency and magnitude gives the smallest average gap to the best method at each budget. Overall, we show that a single agent can learn to allocate resources across turns and adapt its behaviour to user-specified budgets using budget feedback without external budget-allocation controllers.

We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model's weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64--256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person's vitality before predicting their place of death.

Recent advances in multimodal large language models (MLLMs) have substantially improved spoken interaction and video understanding. Existing online and streaming benchmarks primarily assess causal multimodal perception through timestamped question answering, whereas real-world interaction requires models to continuously identify task-relevant multimodal cues and decide whether, when, and how to respond. This leaves open how well current models can sustain real-time interaction throughout an evolving multimodal stream. To address this gap, we introduce DIVE, a benchmark for evaluating full-duplex interaction with visual engagement under continuous and realistic multimodal streams. It organizes six tasks into three complementary interaction scenarios. (1) Proactive Intervention evaluates whether models can identify emerging visual or semantic cues and intervene at the appropriate moment. (2) Continuous Engagement evaluates sustained participation through continuous visual tracking and visually grounded query interaction. (3) Interaction Coordination evaluates how models coordinate listening and speaking under overlapping and multi-party interactions. DIVE contains 525 audio/video samples with fine-grained human annotations and precise temporal metadata. Its evaluation protocol separates response quality, participation decisions, and response timing. We evaluate representative real-time models using response quality, window response rates, and reaction latency. Experiments show that leading real-time models achieve sub-second average response latency, yet their average window response rates remain at or below 53.77%. Response quality varies substantially across models and tasks, with adapting ongoing responses to newly arriving information remaining particularly challenging. We hope DIVE facilitates progress toward more natural and capable real-time multimodal interactive systems.

Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems further expand these capabilities through specialization and collaboration, but distributed authority, heterogeneous roles, and opaque or potentially compromised participants make safeguarding even more challenging. Existing defenses address important individual threats but provide fragmented coverage across local execution, system-level interactions, and enforcement mechanisms. We introduce MAS-Guard, a unified framework that closes this gap for safeguarding single- and multi-agent AI applications containing both instrumentable and opaque agents. Within instrumentable agents, the proposed local Agent Guardrail (A-Guard) layer applies task-specific experts to assess agentic events. At the system level, developer-owned Agent Charters and a Control Plane combine deterministic structural checks of routes, agent capabilities, and message formats with semantic assessments of roles and message content. These complementary layers enable enforcement through a shared set of interventions: allow, reflection, confirmation, and block. Evaluated across six benchmarks covering diverse single- and multi-agent scenarios with three agent LLMs from different providers, MAS-Guard reduces the overall attack success rate by 71.5-86.3% relative to unguarded agents, while meeting the utility reference criterion on benign tasks.

Precisely reproducing a target span verbatim within a long context is a fundamental capability test for LLMs: even when the target string is present in context, models often still drift within the span or fail to stop correctly, revealing limitations in source grounding and boundary control. This capability has broad practical relevance. For instance, code generation often requires embedding sensitive strings such as API keys exactly as they appear in context. In this work, we construct a benchmark to test exact span reproduction ability across contracts, scientific literature, code, and random sequences, and find that current models still make surprisingly frequent errors. To address this, we propose CopyGen, a lightweight copy-constrained plugin for frozen LLMs that turns copying from a byproduct of next-token prediction into a controlled process. Our key idea is to decompose copying into three distinct decisions: when to copy, how far to copy, and when to stop. Concretely, CopyGen first determines whether to enter copy mode, then predicts a tentative copy length, and finally verifies the generated span token by token. During verification, we adaptively adjust the logits of copy continuation tokens to incentivize or suppress continuation, and emit only the longest valid prefix. Across three LLM backbones and four benchmark tasks, CopyGen improves average exact match by 35% relative without model finetuning, and integrates seamlessly with SFT to further improve it by 4%, while substantially accelerating generation. These results suggest that reliable exact copying is not something LLMs obtain for free, but a decoding-time control problem that requires explicit modeling. Code is available at https://github.com/mbzuai-nlp/CopyGen.

LLM agents are commonly trained with RLVR, but the sparse nature of typical RLVR rewards makes credit assignment challenging, especially for long-horizon tasks. We propose Self-Credit, an end-to-end agentic RL method that creates dense supervision signals from sparse outcome rewards using self-generated credit assignment via intra-group rescaling. Self-Credit builds on group-based RL with three components. First, during its rollouts, the policy model generates summaries with citations to earlier turns containing critical evidence, which serve as self-generated credit. Second, we refine this model-generated credit with intra-group rescaling, which compares cited actions across successful and failed rollouts for the same prompt. This approach uses group-level agreement and disagreement to remove undeserved credit and further emphasize credit assigned to critical actions. Third, advantage reweighting uses the rescaled credit weights to update the advantages of the cited turns. On BrowseComp-Plus, Self-Credit improves over the base models by 37.0 and 36.2 points on Qwen3 8B and 32B respectively, and outperforms the strongest agentic RL baselines by 8.5 and 5.6 points, respectively. Self-Credit also improves over the strongest baselines by 4.9 points on GAIA and 4.7 points on xBench-DeepSearch. Analyses show that self-generated credit provides useful signals for selecting critical actions, and that intra-group rescaling, especially on failed rollouts, is important across model scales.

Visual tool use enables vision-language models (VLMs) to acquire task-relevant evidence, but its effectiveness can be limited by predefined visual operations and the tool-use strategies learned during training. Representing these operations and strategies as reusable external skills allows them to be refined through self-evolution without updating model parameters. However, skill refinement can introduce performance regressions. Meanwhile, inspecting accumulated task execution trajectories together with skill evolution history may lead to evolution evidence overload. To address these challenges, we propose VisualEvolve, a self-evolving framework that evolves executable visual operations and tool-use strategies for VLMs through a Select--Execute--Evolve--Evaluate cycle. Skill-Library Snapshot Selection uses performance on a fixed validation set disjoint from execution tasks to select an effective historical snapshot for subsequent evolution, limiting the propagation of performance regressions. Progressive Evidence Access (PEA) guides capability-gap diagnosis through a coarse-to-fine hierarchy comprising the Skill-Performance Report, Evolution Log, Trajectory Summaries, and relevant raw trajectory segments. Across four VLM backbones and six evaluation sets from three benchmarks, VisualEvolve improves over the Initial Skill Library in 23 of 24 settings, with a maximum gain of 15.1 percentage points on VisualProbe Hard using Qwen3.5-Plus. Ablation studies support the contributions of Snapshot Selection and PEA, while further experiments show that skills can be improved using experience from different data settings and directly reused across backbones. Moreover, VisualEvolve enables effective skill self-evolution with over an order of magnitude fewer task-solving trajectories than estimated for representative RL-based methods.

Head-dimension compression can reduce the memory required by the key-value (KV) cache, but existing methods often reconstruct full-dimensional representations during inference, adding significant decoding FLOPs overhead. We formalize the design considerations for converting pretrained attention into a natively reduced-dimensional form that reduces both KV-cache storage and attention computation using a calibration set. Specifically, we consider reductions that (1) apply a shared projection to queries and keys so that they remain in a common reduced feature space, (2) derive this projection jointly from query and key activations to better preserve their interaction, and (3) are compatible with the pretrained RoPE transformation. To obtain a reconstruction-free realization of this reduction, we restrict it to coordinate selection and theoretically show that RoPE compatibility then holds if and only if complete rotary pairs are retained. Building on this characterization, we introduce RACE, a post-training method that selects rotary pairs using their joint query-key activation energy. We theoretically show that this criterion minimizes a separable, calibration-dependent upper bound on post-RoPE query-key product approximation error. The resulting reductions are folded into the model weights and require neither runtime reconstruction nor specialized kernels. Across three model families and different compression levels, RACE outperforms structured pruning and remains competitive with reconstruction-based methods on language-modeling, zero-shot, and long-context tasks. At 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8%, yielding 1.79x higher throughput, 43.2% lower time-to-first-token (TTFT), and 28.1% lower allocated GPU memory in comparison to an uncompressed model.

Reasoning about safety is essential for embodied agents powered by large language models (LLMs). We introduce the Contextual Action and Plan Safety benchmark (CAPSafe), designed to evaluate safety reasoning at two levels: (1) an action-level dataset that tests whether a single action is safe given its environmental context, and (2) a plan-level dataset that evaluates whether agents can generate safe long-horizon plans for complex hazardous household tasks. CAPSafe models safety conditions where hazards are situational (inferred from object properties), multiple hazards can co-occur within the same environment, and each is annotated with a relative severity level for fine-grained analysis. Evaluating nine closed and open-source LLMs reveals a substantial gap between hazard recognition and safe long-horizon planning. Several models achieve high accuracy in recognizing unsafe actions (exceeding 90% in the best cases), yet safe long-horizon planning remains substantially harder. Gemini-3-pro achieves the highest safe success rate (78.19%), while other models remain below 65%. These results highlight that robust safety reasoning over long action sequences remains an open challenge.

Agent harnesses are executable code that specifies the prompts, tools, and control flow around a language model. A single global harness can struggle with heterogeneous tasks that require different execution strategies. In this work, we introduce Object-Oriented Harness (OO-Harness), a harness optimization framework that evolves a hierarchy of executable harness classes using object-oriented design. During training, an evolution agent uses execution records to create a family of specialized harnesses, each implemented as a subclass of the base or an existing specialized harness via inheritance. At test time, a harness router uses routing lessons accumulated during training to select a specific harness for each task. Across three long-horizon benchmarks spanning business workflows, workspace tasks, and tool use, OO-Harness achieves the best performance among the compared methods, with an absolute improvement of up to 15% over the base. Extensive ablation studies examine the contributions of its core components.

A growing body of work proposes dynamic expert-skipping methods for Mixture-of-Experts (MoE) models that adjust the number of experts activated per token during inference, aiming to preserve model quality while further reducing the number of active parameters. These methods have added calibrated importance scores, inter-expert similarity metrics, or learned per-token allocators in order to identify experts to skip. However, different calibration regimes, different reductions in the effective number of experts, and evaluation suites make comparisons difficult. We classify four representative prior methods into a unified taxonomy and benchmark them against a natural, simple baseline: statically reducing the number of experts $k$ (i.e., same reduced $k$ across all layers and tokens). We provide a fair comparison of these methods at matched expert-compute budget, on two modern MoEs (Gemma4-26B-A4B and Qwen3-30B-A3B) in a text-only setting and a contemporary vision-language model (Qwen3-VL) in a multimodal setting. Surprisingly, we find that the simple baseline is hard to outperform consistently across models, modalities, benchmarks, and calibration settings. The methods we benchmark often underperform the baseline and achieved gains, when present, are minor. We further find that the effective $k$ achieved (or optimized for) in calibration can diverge from the effective $k$ in testing, resulting in over- or under-estimating the compute at inference time. Based on these findings, we recommend a calibration-free per-modality static $k$ reduction and validate this on a held-out model and modality. We further argue that a static control at matched measured compute is the baseline any dynamic skipping method must beat.

Independently trained autoregressive language models often acquire similar capabilities, but it remains unclear whether their internal representations organize behaviorally relevant information in compatible ways. We show that models do in fact organize information similarly by using specific behavioral probes that operate on compatible representations. This unlocks portable white-box monitoring: a probe developed and validated on one model can be reused when the model changes, without requiring labeled target data. To overcome incompatible native activation coordinates, we express each example through its angular distances to a shared set of anchors. We introduce a global anchored construction that summarizes information from all transformer layers and removes model-specific layer selection, so the same frozen probe can be straightforwardly applied to another model. We evaluate cross-model probe transfer across monitoring tasks on trivia datasets, safety monitoring, and mathematical reasoning. On average, transferred global anchored probes preserve at least 85% of same-model probe AUROC over chance. Anchored probes also support transfer to related datasets, remain robust across model scale and prompt changes, and retain performance better than standard probes after model updates. Transfer is nevertheless task- and target-dependent and can fail even when the monitored property remains strongly decodable in the target model. Overall, global anchoring reveals substantial task-relevant compatibility across independently trained models and provides a practical basis for portable white-box monitoring.

Local Retrieval-Augmented Generation (RAG) systems are vulnerable to indirect prompt injection and retrieval poisoning, where malicious instructions in retrieved content are mistaken for trusted input. We present nokast-secureRAG, a roughly 0.5B-parameter small language model that acts as a semantic firewall between retrieval and generation. It jointly evaluates the user query and retrieved context, classifying the pair as safe, suspicious, or malicious. The model is trained using a teacher-judge-student pipeline: a 35B teacher generates 5,233 examples across four attack families, and an independent 120B judge re-labels them, achieving 95.0% agreement before student fine-tuning. On a held-out in-distribution test set, the model achieves 0.994 detection recall at a 0.026 false-positive rate and a 0.006 attack-success-rate proxy, outperforming a regex baseline, a public prompt-injection classifier, Llama Guard 3 1B, and its untuned version. Prompt-Guard-86M reaches 0.900 recall but has a 0.690 false-positive rate. On a judgment-flip test, where the same sentence is benign or malicious depending on context, the model correctly classifies both cases 75% of the time, outperforming the compared baselines. Ablations indicate that retrieved context carries the main attack signal, while query-context consistency reduces false positives. Single-request latency is 293-404 ms, above the target of 100 ms, while amortized throughput latency is 37 ms under batching. On the external PoisonedRAG benchmark, recall falls to near zero, indicating that fluent fact-poisoning attacks lie outside the detector's instruction-consistency threat model. We report this limitation alongside the in-distribution results. The model, data-generation pipeline, and evaluation harness are released open source.

Long-horizon deep search requires LLM agents to iteratively decompose complex information needs, gather evidence from external sources, and maintain a reliable research state under multiple constraints. A central challenge is that early mistakes can be amplified over the search trajectory: agents may anchor on an incorrect candidate, collect partially supporting evidence, overlook contradictions, and accumulate errors that misguide subsequent reasoning. We propose FALCA, a falsification-aware framework that formulates deep search as a propose-falsify process. FALCA adopts a collaborative Searcher-Falsifier architecture, where the Searcher proposes candidate answers or sub-answers, while independent Falsifiers actively seek counterevidence, check unmet constraints, identify conflicts, and produce concise falsification reports for research-state revision. To strengthen this loop, we introduce falsification-aware training, combining SFT warmup on falsification trajectories with agentic RL based on parallel rollouts and hierarchical credit assignment over research stages, state transitions, and actions. Experiments on five search benchmarks show that FALCA improves long-horizon deep search performance over backbone and harness ablations, while reducing confirmation bias and error accumulation through more effective falsification and research-state revision.

In reinforcement learning, one presumes the action set to be known ahead of time, and, thus, one must know if any state-action pairs are not permissible ahead of computations. In some applications, determining this action mask is the main computational bottleneck. Isolating this problem to highlight its core properties, we study how to combine the search for an optimal policy with the learning of the mask itself, under a linear MDP model in which each attempted action reveals the exact value of a linear score determining its validity. In this framework, we first construct an algorithm based on deterministic optimistic optimization and show that it is minimax optimal, which we show is also computationally efficient. However, it is sub-optimal on non-degenerate instances, as we show. We propose an alternative, which is also computationally efficient, which plans its exploration strategy. This algorithm admits a formal interpretation as an analogue of information-directed-sampling for deterministic feedback, which may be of independent interest. Finally, we study an application of this framework to transfer learning, where the invalid action set changes over time.

Object counting in videos remains a poorly characterized capability of multimodal large language models. Counting questions are included in many existing VideoQA benchmarks, but the ground-truth count distributions in these benchmarks are heavily skewed toward small values. Across the nine benchmarks we survey, the median ground-truth count is at most five, and the three most frequent values together account for between 45% and 93% of the videos. The resulting mean accuracy is then dominated by the most frequent counts; a value-dependent bias at a rare count contributes negligibly to the average and remains invisible. We introduce CountVideoBench, a benchmark of 2,600 synthetic videos covering 26 cardinalities in the interval [5, 50] (100 videos each), grouped into six clusters of consecutive counts that span the range while enabling near-neighbor comparisons. Each video contains a single class of static objects under a controlled camera motion, removing object motion and reducing recognition as confounds. We evaluate sixteen video-language models from nine architecture families in a zero-shot setting. No model exceeds 27.6% Exact Match. The dependence of accuracy on cardinality varies across models, ranging from monotonic decline to a non-monotonic peak-and-collapse pattern, which rules out increasing difficulty with cardinality as the dominant explanation of counting failure. Every model further concentrates its predictions on its own set of preferred integers, with accuracy peaks coinciding with these values and collapsing at adjacent counts. These findings highlight the need for balanced, diagnostic counting benchmarks to reliably assess this fundamental visual capability.

In speculative decoding, a lightweight draft model proposes tokens that the target model verifies in parallel, speeding up generation without changing the target distribution. D-PACE (Wu et al., 2026) trains such drafters with a loss that weights each position by its contribution to the expected accepted length. Both its weights and its cross-entropy (CE) loss use only the draft's probability of the recorded target token, treating the target as a point mass. Under rejection sampling, however, a sampled draft token is accepted with probability equal to the overlap of the draft and target distributions, one minus their total variation (TV). We derive D-PAL (Dynamic Position-wise Acceptance Loss) from the exact expected accepted length, using this overlap as the acceptance signal. This yields overlap-based position weights with log acceptance (LA) as the per-position loss, and the objective reduces to D-PACE for a point-mass target. Since the weights are detached, any per-position loss can replace LA in practice. We show that forward KL divergence and Rényi divergence of order 1/2 match or exceed LA, and that Rényi-1/2 provably bounds the true acceptance rate from above and below. Replacing CE with Rényi-1/2 improves acceptance length under both D-PACE's weights and ours. On six benchmarks with a Qwen3-4B target, D-PAL raises mean acceptance length over D-PACE from 4.50 to 4.57 at T=0 and from 4.20 to 4.28 at T=1, and consistently achieves higher throughput at both temperatures.

Spatial reasoning asks a model to work out where objects are in 3D space, how they relate to one another, and how they move. Agentic VLMs have recently brought large improvements on such tasks by letting the model actively gather evidence with 3D perception tools instead of answering in a single pass. Even so, on many questions the tools return the right evidence and the model still does not arrive at the correct answer: the gap between model and tools, where raw readings become usable conclusions, is left unfilled. We present S-Agent 2.0, a spatial reasoning agent that equips an off-the-shelf general coding agent with a command-line perception toolkit, and we design the interface around one principle: help the VLM close the evidence loop: acquire scene evidence, remember it in a persistent scene memory, and interpret each reading in plain spatial language. This design keeps the full flexibility of a coding agent, and every reading it acts on stays grounded in the tools' geometry. On MMSI-Bench and ReVSI, the toolkit improves a strong open-source model, Qwen3.8-27B, by +13.9 and +20.5 points over the same agent without tools, and a strong closed-source model, Gemini-3.8-Flash, by +11.2 and +9.6 points. With the open backbone S-Agent 2.0 reaches 66.5 on MMSI-Bench and 82.3 on ReVSI, ahead of the specialized system SpatialClaw on the same backbone (62.6 and 74.5), and with Gemini-3.8-Flash it reaches 74.1 and 90.5 against its 67.0 and 79.7. Both tool gains shrink as the backbone grows more capable, but SpatialClaw's shrinks far more on ReVSI, from +13.2 to +2.7, while ours stays at +9.6. S-Agent 2.0 also works across different general coding agent harnesses, such as Codex and Claude Code.

Reconstructing dynamic 3D scenes from monocular video requires inferring motion from noisy and intermittently visible observations. Long-range tracks provide valuable geometric constraints, but integrating their fragmented temporal coverage into shared, continuous motion remains challenging. We introduce \textbf{RayCo4D}, a ray-conditioned continuous motion model for monocular 4D reconstruction that explicitly separates visual measurements from latent 3D motion states. Each track is associated with a reference-time 3D query constrained to its camera ray, and its image and depth observations constrain shared latent motion nodes through local many-to-many associations. Each node maintains an $\mathrm{SE}(3)$ pose and body-frame velocity at sparse support times. A white-noise-on-acceleration process prior couples neighboring states and defines continuous-time interpolation, allowing observations across visibility intervals to jointly constrain the same latent state sequence.The inferred shared motion is transferred to Gaussian primitives through dual-quaternion blending, while reference-anchored local $\mathrm{SE}(3)$ residuals capture fine-scale non-rigid deformation beyond the shared scaffold. Experiments on diverse monocular dynamic scenes show that RayCo4D achieves competitive novel-view synthesis quality, improves structural and temporal consistency under rapid motion and incomplete tracks, and supports continuous-time rendering.

Realistic few-shot adaptation of medical vision-language models (VLMs) can leave low-prevalence classes without any labeled support. The training-free SS- Text+ solver remains operational in this setting, but its zero-support prototypes receive no visual adaptation and fall back to a purely textual class prototype. We in- troduce Gated Knowledge-Guided Prototype Transfer (GProtoT), a pre-repulsion module that transfers supported-class visual residuals to missing classes using VLM similarity and an ontology-based clinical distance. A validation-free evidence gate scales transfer based on semantic relevance and residual agreement, thereby limiting unsupported extrapolation. In 50 paired realistic one-shot trials on MESSIDOR with FLAIR features, severe and proliferative diabetic retinopathy were absent from every support set. Relative to SS-Text+, gated ProtoT improved missing-class recall by 5.23 points, average class accuracy by 1.37 points, and the missing/supported recall harmonic mean by 1.83 points. It reduced negative- transfer frequency from 46-48% for ungated transfer to 2%, while at the cost of reductions of 1.20 percentage points in supported-class recall and 0.45 percentage points in top-1 accuracy. These results indicate that evidence-based gating provides a safety mechanism for the knowledge source. GProtoT is based on SS- Text+ and requires no gradients, validation data, or additional model inference, of- fering a low-compute path toward accountable adaptation under incomplete clini- cal supervision. Code is available: https://github.com/sumaiyaafridi/GProtoT

Machine learning-based fault detectors are often evaluated at the relay locations used for training, although local voltage and current measurements vary across power networks. This work investigates cross-relay generalization for binary detection of faults anywhere in the network using local measurements, rather than line-selective tripping. Using 50 ms windows from a simulated 110 kV transmission network, we evaluate a multilayer perceptron across all 64 source target combinations of eight relays and five random seeds. A shared simulation-level split and source-only threshold selection provide a controlled held-out evaluation without target-specific calibration. Despite nearly unchanged mean F1-scores (0.730 within-relay versus 0.734 cross-relay), the mean false-positive rate (FPR) increases from 0.088 to 0.341, while the mean false-negative rate decreases from 0.399 to 0.318. The change is strongly directional: models trained at Bus 2 relays yield a mean FPR of 0.670 at Bus 1 and Bus 3 relays, compared with 0.096 for cross-relay transfers within the same bus group. Linear probing indicates that learned representations retain coarse bus-group information, while exploratory class-conditional latent distribution analysis shows no consistent monotonic association with transfer errors. These findings demonstrate that aggregate metrics can conceal location-structured changes in false-alarm behavior, highlighting the importance of evaluating error types separately when assessing cross-relay generalization.

Cloud-accessible quantum computing has made heterogeneous Quantum Processing Units (QPUs) available through provider platforms that differ in qubit technology, calibration state, queueing behavior, and pricing model. This heterogeneity makes execution management a quantum-engineering concern: a quantum circuit may be executable on several backends, but the resulting cost, latency, fidelity, and reproducibility depend on backend selection and on how measurement shots are allocated. This paper presents the Quantum Broker, an open-source and provider-agnostic framework for Quality-of-Service-aware quantum execution management. The framework operationalizes shot-wise distribution by exploiting the statistical independence of quantum measurement shots to dispatch a single circuit across multiple heterogeneous QPUs according to user-defined constraints and optimization objectives. It combines a modular execution architecture, interchangeable backend-profiling strategies, and the Quantum QoS Specification Language (QQSpec), a declarative language for expressing QoS requirements. Dispatch planning is supported by two solver backends: a mixed-integer linear-programming (MILP) solver, and a nonlinear dual-annealing solver. Experiments on benchmark quantum circuits and noisy simulator backends show that the Quantum Broker can produce feasible QoS-aware dispatch plans under heterogeneous backend conditions and can improve QoS satisfaction with respect to single-backend and naive multi-backend baselines. The results also expose a practical trade-off between planning time and dispatch quality, with the nonlinear solver producing higher-quality allocations at increased computational cost. Overall, the Quantum Broker provides an extensible execution-management layer for engineering portable, reproducible, and QoS-aware quantum software over fragmented quantum cloud infrastructures.

Reliable retrieval-augmented generation (RAG) requires recognizing when neither retrieved documents nor parametric knowledge supports an answer. We introduce DeclineCraft, a dynamic abstention benchmark whose unanswerable questions target Minecraft feature additions after a corpus cut-off, with audits of possible parametric exposure. Later cut-offs support evaluation of future models. Across 27 LLMs, we measure Correct Abstention Rate (CAR; how often models acknowledge insufficient knowledge on unanswerable questions), Answerable Attempt Rate and Answer Accuracy. We find that (1) single-shot retrieval significantly raises CAR in 38 of 40 configurations (model and thinking setting) versus no retrieval. (2) Iterative RAG, in which the model self-assesses its context and can request up to two further retrieval rounds, significantly lowers CAR versus single-shot RAG in 30 of 40 configurations (mean 46.4%), with little change in Answer Accuracy (68.5% to 69.3%). Adding query rewriting and reranking to hybrid (dense & sparse) retrieval lowers CAR for all 12 models tested. (3) Larger models answer more accurately (Spearman $ρ= +0.60$ with total parameters), but neither model size nor release date shows a significant association with CAR. (4) Without retrieval, enabling thinking on the same weights lowers CAR significantly for 8 of 13 models (by 14% points on average), and the responses shift from acknowledged uncertainty to premise denial.

Emoji are widely treated as compact affective tokens whose meaning is summarised by a single sentiment score. We argue that this picture is incomplete. From a corpus of 16{,}575 English tweets and a per-emoji metric we call \emph{stability delta} (within-class minus between-class cosine similarity of sentence embeddings), we show that emojis used in tightly clustered topical contexts also exhibit \emph{more} polarised crowd-sourced sentiment, not less. Validated against the Kralj Novak et al.\ (2015) sentiment lexicon, the central correlation is Spearman $ρ= +0.361$ (Holm $p = 0.015$, 95\% bootstrap CI $[+0.10, +0.59]$); against the independent Ferré et al.\ (2025) Spanish 13-emotion ratings, $ρ= +0.509$ ($p = 0.0003$). The result survives permutation testing, BCa bootstrap, leave-one-out jackknife, partial correlation controlling for emoji frequency, and Holm correction across a four-test family. Topical coherence and affective dispersion are distinct dimensions of emoji meaning. As a downstream check we report leakage-free five-fold cross-validation results for sarcasm classification on the Emoji Sarcasm Dataset; all three baselines (a gated multimodal architecture, fine-tuned BERTweet features, and fine-tuned RoBERTa features) score above 0.97 macro F1, indicating the dataset has saturating emoji-label cues. We frame the classifier as supporting evidence rather than the contribution.

We study a linear mixing model with dependent latent components, assuming multiple data domains. Most existing models assume that the components are independent or at least uncorrelated, in line with independent component analysis (ICA). Some recent work allows for dependent components, but then makes specific assumptions such as parametric forms of dependencies, multi-view settings, or interventions, or does not recover the individual components. In contrast, we consider a multi-domain setting in which domains differ through domain-specific scalings of the components, while the distribution of the underlying latent components is the same across domains. This approach can model data collected, for example, from different sensors measuring the same process, different laboratories conducting the same experiment, different experimental conditions, or different subjects that might differ in biological or physiological factors. We show that, under sufficient domain variability, latent variables and mixing functions can be identified from second-order statistics alone. We propose the Multi-Domain Covariance Matching (MuDo-CoM) algorithm that generalizes previous methods of joint diagonalization. MuDo-CoM is validated on simulated data and a real-world fMRI dataset.

We present the Observation Assimilation Transformer (OAT), a geometry-aware neural process that corrects and refines a prior numerical weather prediction (NWP) field into a high-resolution posterior field by assimilating sparse observations. OAT is an encoder-decoder transformer that considers a set of observed locations and their NWP prior to condition predictions at a set of arbitrary unobserved query locations. Its sparse attention mechanism makes it scalable to thousands of station observations and depends on relative positions between locations, never on the absolute coordinates. The model allows for partially missing observations, predictions at arbitrary continuous locations without grid projection, and outputs are parametric distributions. A station-randomized training algorithm is developed with continuous ranked probability score (CRPS) loss. We test OAT on two datasets - one at regional scale with hundreds of stations, and one at continental scale with thousands of stations. On these datasets OAT outperforms several baselines and shows well calibrated predictions. In an illustrative case study, OAT's downscaling capabilities are demonstrated by downscaling ERA5 (~31km) to COSMO (1km), where OAT achieves lower mean absolute error (MAE) and CRPS for temperature and precipitation on test stations compared to COSMO.

Factual sycophancy, a language model's deference to a user's false beliefs, is usually understood as a bid for the user's approval at the expense of truthfulness. We argue that this framing over-emphasizes the user and propose viewing sycophancy as a failure within the broader and generally useful process of in-context updating, by which models revise their answers based on information supplied during an interaction. We support this view with behavioral and mechanistic evidence. First, models' responses to user claims sit within a broader pattern of source-sensitive updating. They revise their answers in response to information attributed to external sources as well, more so when it appears better supported, and their deference to the user persists even when no user approval is at stake. Second, the internal components that support deference to the user are also causally important for in-context updating more broadly. Our results suggest that factual sycophancy arises when in-context updating gives a user's assertion excessive weight relative to its evidence. Because sycophancy is embedded in this process, reducing it risks reducing useful updating as well, a trade-off that calls for evaluations measuring both behaviors and mitigations that decouple them.

Physics-informed neural networks (PINNs) supervise on the residual of a known differential operator. Many high-stakes inference problems admit no such operator: clinical signals, behavioral measurements, materials properties. We introduce the Manifold-Informed Neural Network (MINN), a PINN-parallel paradigm that replaces the physics operator residual with a manifold operator residual: a discrete Poisson equation $\mathbf{L}_G \mathbf{H} = s_φ(\mathbf{x}, \mathbf{m})$ on a $k$-nearest-neighbor graph in the learned latent space, with a learned, task-aware source. The construction is a structural parallel to PINNs (squared operator residual plus label data, in place of operator residual plus boundary data); the $s{=}0$ limit recovers the homogeneous Laplace equation $\mathbf{L}_G\mathbf{H}=\mathbf{0}$ (classical manifold regularization solves the same equation but minimizes the Dirichlet energy $\operatorname{tr}(\mathbf{H}^\top\mathbf{L}_G\mathbf{H})$, a different penalty form). The source $s_φ$ takes input features and auxiliary metadata directly, breaking the homogeneous smoothness prior that limits classical Laplacian penalties; gated Feature-wise Linear Modulation separately fuses metadata into the prediction branch. We evaluate MINN on three-class classification of potassium, calcium, and magnesium from 12-lead ECGs in MIMIC-IV-ECG, against twelve baselines spanning supervised and tabular methods, ECG foundation models, and manifold/graph approaches. MINN attains the highest K$^+$ Macro-F1 of 0.61 (best baseline 0.55), with consistent gains on Ca$^{2+}$ and Mg$^{2+}$ and a 31% relative improvement at 1% labeled data. Six targeted ablations show that the inhomogeneous Poisson formulation, not graph smoothing alone, drives the gain.

Generating graph signals is an important task in machine learning, and incorporating the underlying graph into a generative model enables structurally-informed distribution modeling. However, said graph is often noisy or imperfectly estimated, and structural errors can affect the generative process and the resulting signal distribution. While recent work has established stability guarantees for graph neural ordinary differential equations, the robustness of graph-aware reference dynamics, as used in topological Schrödinger bridge models, remains largely unexplored. Here, we investigate the structural stability of continuous-time generative models whose dynamics combine a graph filter with a learned graph neural network. We first derive a Wasserstein stability bound that quantifies the effect of graph perturbations on the generated signal distributions. Informed by this analysis, we then propose a principled framework for designing graph filters that balance structural stability with induced signal smoothing. We characterize the optimal filter design and provide an explicit family of polynomial filters with provable stability guarantees. Experiments on synthetic graphs and real-world fMRI signals support our theoretical findings, illustrating the enhanced structural stability of the proposed filters while preserving generative performance.

Instruction-based image editing models can now add, remove a specified number of objects or change their category while preserving the scene. Yet these models often fail to produce the correct number, and these failures grow sharply as scenes become more complex and the requested count increases. Current benchmarks largely miss this problem: they score edits with holistic similarity metrics or with vision-language model (VLM) judges that are themselves poor at counting, so numerical errors go unnoticed or are graded incorrectly. As a result, the true numerical ability of today's editing models has never been accurately measured. We address this issue with ABACUS, a new benchmark, evaluation protocol and dedicated scoring metric for numerical image editing. Our protocol is judge-free: no VLM or MLLM scores an edit or counts objects (a VLM only confirms the category of rare ambiguous crops). Instead, it verifies each edit at the instance level against panoptic ground truth confirmed by detector consensus, checking both that the exact requested count was achieved and that the rest of the scene was preserved, across the add, remove and change operations. By evaluating leading image editors, our method reveals substantial numerical editing failures that worsen with increasing scene complexity and object count; failures that prior benchmarks often overlooked or misjudged.

Large language model (LLM) agents increasingly use persistent memory to personalize responses, and personal attributes in this memory pose a privacy risk if exposed. We ask how consulting memory during dialogue changes what users go on to say, and thus what the memory comes to reveal. In multi-session simulations, a simulated user with 24 personal attributes talks with a support agent that does or does not consult memory, and in both settings attributes are inferred from the full accumulated record. Across four support models, memory raises response quality but also empirical leakage, the share of attribute values correctly inferred from the record, by a margin that widens over sessions. The cause is a loop in ordinary, non-adversarial assistance. Memory-conditioned responses carry earlier values forward, often as presuppositions, and users affirm them or leave them uncorrected, so that cues and the agent's own guesses become recorded values. The loop reinforces itself, and removing the agent's mentions of an attribute lowers its later disclosure about twice as much as single-turn effects predict. Users cannot leave it on their own, since correcting a value discloses the attribute, declining to discuss it shifts later predictions toward sensitive values, and incorrect values left in place become entrenched. The loop is best weakened at the agent's response. Instructing the agent not to state or presuppose attributes brings leakage to about the no-memory level while keeping most of the quality gains, whereas filtering or anonymizing what users say or what is stored gives up much of them.

Prior-data fitted networks (PFNs) have recently gained popularity as an end-to-end approach to amortized Bayesian inference. Yet, their statistical behavior for semi-parametric targets, such as means under missingness or causal effects, remains poorly understood. In this work, we develop theoretical foundations for debiased PFNs to perform inference on finite-dimensional targets in general semi-parametric settings. First, we show that, despite being Bayes-optimal with respect to the target, an idealized PFN can exhibit a posterior concentration rate that implicitly depends on the concentration rates of (latent) nuisance functions and can become slower as the prior becomes more diverse. Then, to speed up the target posterior concentration rates, we introduce two generic debiasing strategies based on efficient influence functions: (i) a label correction and (ii) a prior correction. Finally, we show that the debiased PFNs (i) & (ii) possess concentration rates that asymptotically match a convergence rate given by traditional frequentist bias-corrected estimators. In this way, we establish a novel connection between amortized Bayesian inference with PFNs and frequentist semi-parametric efficiency theory. Under additional assumptions, this allows us to guarantee local and global optimality of PFN-based estimators (namely, the Bernstein-von Mises property and minimax optimality, respectively). Therefore, our work provides guidelines for designing debiased PFNs for semi-parametric inference with principled and provable asymptotic guarantees.

Natural-language instructions often rely on high-level abstractions, such as asking an agent to ``stack blocks in a pyramid.'' Interpreting abstract instructions requires models to both identify the abstraction intended by the speaker, and unroll it into the many concrete actions that instantiate it. Language models struggle here: prior work has shown that performance degrades as the level of abstraction in the instruction rises. To bridge the gap between natural language abstraction and concrete actions, we propose to borrow from coding environments the notion of a {\em domain-specific language} (DSL), an explicit representation layer whose documented primitives name recurring abstractions that expand into concrete actions. We demonstrate the efficacy of DSLs for NL interpretation on \textsc{Hexagons}, a visually-grounded instruction-following task graded for abstraction. We further ask whether language models can induce such DSLs themselves from domain data. Our results show that while code generation consistently improves over direct inference, code generation over a human-curated DSL yields dramatic gains, up to 10 absolute F1 points, reaching the human skyline, with gains concentrated at {\em higher} abstraction levels. We confirm this pattern on the \textsc{Text2CAD} benchmark. At the same time, model-induced DSLs recover only part of this gain, remaining inconsistent across runs and short of their human-curated counterpart. Overall, the right representation closes much of the gap caused by high abstraction --- but inducing that layer of representation automatically is where models still fall short, and where we see a strong opportunity for the community to make progress.

Recommending bacterial cultivation conditions from sequence data is a challenging structured problem characterized by missing standardized benchmarks and limited use of sequence structure. To address this, we introduce DSNI, a benchmark of 65,023 bacterial strains with phylogeny-aware splits, pairing 16S rRNA sequences with four cultivation conditions (media, temperature, pH, oxygen). Exploiting the mosaic structure of 16S rRNA, where hypervariable regions support discrimination and conserved regions support primer design, we propose a Variability-Gated Encoder that learns position-dependent importance weights conditioned on biological features (region type, position, and local complexity). The gating module adds minimal overhead (1,280 parameters, <1%) while providing an explicit inductive bias valuable in low-data regimes. Experiments demonstrate that the learnable gate improves flat Transformers by 4.7% in joint accuracy and complements RNABERT, yielding 44.2% exact-match joint accuracy. The model also recovers 90.2% of ground-truth cultivation protocols within its top-25 recommendations, narrowing the exhaustive search space by approximately 60 times.

Monitoring the chain-of-thought (CoT) of language models (LMs) is a key strategy for detecting and preventing misbehaviors for AI safety. As such, it is critical to identify reinforcement learning (RL) practices that lead to unmonitorable chain-of-thought. One such practice may be unavoidable: reward misspecification. As RL environments become more complex, it is infeasible to have perfect verification, and models are rewarded for behaviors that are different from the true task intent. We show that reward misspecification alone already degrades monitorability. We create a testbed with pairs of true reward aligned with task intent and misspecified rewards that favor a different behavior, so that reward hacking can be induced by training and labeled verifiably. Across three models and three tasks, CoT monitorability remains low as models' hack rates increase: the CoTs keep reasoning about the task intent while the outputs violate the intent and get rewarded. To mitigate unmonitorability without prior knowledge of the true reward or the hack, we collect rollouts with diverse rewards across training steps and use a LM to infer a reward description. We then inject the description into the input, which teaches the model to verbalize the reward. This reward description training improves monitorability by an average of 24% or 42% absolute, depending on the metric, across all models and tasks, and yields CoTs that reason about the misspecified reward explicitly. Code is available at https://github.com/Lorenayannnnn/reward-misspecification-degrades-cot-monitorability.

Riemannian manifolds are powerful models for problems with non-Euclidean structure. Recent work in geometric machine learning has focused on training deep neural networks to learn Riemannian metrics from high-dimensional data. However, many geometric algorithms require repeated evaluations of the exponential and logarithmic maps, whose high computational cost restricts the utility of learned Riemannian manifolds in practice. We propose to address this cost by training a conditional invertible neural network, called a flow chart, to approximate normal coordinates around points on a given Riemannian manifold. With this approach, the exponential and logarithmic maps can be approximated by fast Euclidean operations in the model's learned coordinate space. Our main contribution is an efficient loss function for training flow charts that is nonnegative and, under suitable domain assumptions, is zero precisely when the chart is normal. Across sixteen synthetic and data-derived geometries, flow charts are competitive with a specialized neural geodesic interpolator for logarithmic map estimation, while also being able to estimate the exponential map. When pretrained, flow charts provide a median speedup of 60-70x over numerical baselines at comparable logarithmic map estimation error, and their learned coordinates improve the accuracy of a numerical geodesic solver.

Test-time scaling allocates additional inference compute to refining existing solutions or generating new ones. This raises a fundamental question: Should I keep improving this solution, or stop working on it and try a completely new solution? Existing methods make this decision using fixed refinement schedules, self-verification judgments, or learned controllers, which can waste compute or rely on signals that may be incorrect or require additional training. We introduce Adaptive Replay Refinement (ARR), a training-free method that detects when a stationary revision reaches a fixed point and reallocates the released compute to a new independent sample; replay decoding treats the current solution as a draft and verifies unchanged tokens in parallel, enabling efficient fixed-point detection. Across seven models from five families and seven benchmarks, ARR matches majority voting on three of five mathematical reasoning benchmarks at equal compute, while reducing critical-path serial passes by up to 2.4 times. On MATH500, ARR achieves 84.6% accuracy with only 41% of the serial passes required by majority voting. These results demonstrate that detecting refinement exhaustion enables adaptive reallocation from revision to independent sampling, improving inference efficiency while highlighting the limits of self-revision. Code is available at https://github.com/AhsanBilal7/adaptive_refine.

Teams of large language model (LLM) agents are increasingly applied to complex tasks, but their collaboration protocols are mostly designed empirically, without convergence guarantees. We formalize collaborative problem solving as block optimization: each agent improves one part of a shared solution, guided by its local objective plus a correction for how its changes affect the other parts. A coordinator checks every proposal against the global objective and merges proposals that do not interfere with each other. We prove linear convergence up to an error floor that vanishes when the corrections are exact. Interfering proposals can also be merged once their combination has been checked; the rate then depends on how well the proposals actually combine, not on worst-case interactions. With GPT-OSS-120B agents on robot routing, a three-agent team finds collision-free plans in every run, better than those of a single agent that controls all robots. On a larger routing task, a six-agent team completes its search about twice as fast as the single agent (in median). Merging the best checked combination of proposals, rather than only proposals that do not interfere, cuts the team's search time by about 30%. With constants measured on the runs, our convergence bound holds in every round and follows the observed progress. We also test the theory on query planning in SQLite and on a synthetic task.

Multimodal large language models (MLLMs) have substantially advanced chart understanding and reasoning, yet their performance remains highly uneven across chart types. While modern MLLMs perform well on common charts, their accuracy drops sharply on long-tail charts with diverse visual structures and reasoning requirements. Tool-augmented agents can improve complex chart reasoning by introducing specialized visual capabilities, but manually engineering strategies and tools for the heterogeneous and evolving long tail is difficult to scale. We introduce \textbf{ChartEvo}, a recursive self-improvement framework that enables chart agents to adapt to underperforming chart types through experience-driven harness-level self-improvement. ChartEvo jointly evolves reusable chart-specific Skills and executable Tools, while learning reusable local tool-call chains to compose evolved capabilities and control the inference overhead of tool-augmented reasoning. On long-tail charts from ChartBench and ChartX, ChartEvo improves accuracy over its base MLLM, Qwen3.7-Plus, by \textbf{45.6\%} and \textbf{32.5\%}, respectively, demonstrating the effectiveness of recursive self-improvement for long-tail chart adaptation.

A model behind an API can be identified from its one-word answers to everyday questions such as "Name a sound" or "Name a poet". A model's profile records how often it gives each answer to each question. We attribute new answers to the model most likely to have given them. On a panel of 94 models, 40 answers identify the exact model 92.5% of the time, against 1.1% by chance. Profiles built through OpenRouter identified all ten models served by DeepInfra under blinded labels. A system prompt lowers accuracy by 6 to 16 points. We applied the method to 96 endpoints of 13 open models on OpenRouter, and 92 identify as the model they claim. Of the other four, one matches the claimed model with reasoning turned off. One serves Gemma 4 under the name Llama 4 Maverick: it names itself Gemma 4, and it is billed Gemma 4's token counts on every message. Two remain unresolved. Billing records also show context added to first-party models. Grok 4.7 is billed 1,247 input tokens for a one-line question that costs GPT-4o only 12 tokens, and five models state the current date without being billed for it. The questions come from a public census of model defaults, and the method costs cents per endpoint.

Reasoning about fine-grained placement constraints, such as support and containment, is essential for realistic 3D scene creation. It remains a significant challenge for existing methods that rely on simplified object representations like oriented bounding box (OBB), which are blind to the detailed geometry. To address this limitation, we introduce GALA, a geometry-aware language model with explicit point cloud perception for controllable object placement. Our model performs autoregressively by predicting the 6D pose of one target object at a time conditioned on the scene context and placement instruction. We utilize a two-stage training strategy: the first stage aligns geometric and textual modalities through pretraining tasks, and the second stage fine-tunes the model for placement. To facilitate this task, we also curate a multi-source scene dataset covering diverse room types and fine-grained furniture arrangements, which will be open-sourced for community benefit. Experiments on our benchmark show our model significantly outperforms baselines without explicit geometric reasoning. We further develop a multi-stage plan-place-verify agent pipeline that uses the model for text-driven scene generation. Project Page: https://xccelephant.github.io/paper/gala/

Asked to name a running shoe brand, language models say Nike. Asked to recommend one, they say Brooks. This paper asks where models learn that taste. Across 102 models and 42 brand categories, the brand a model picks when recommending differs from the brand it names in 26. The field moves together to a shared recommendation. The Named brands are market leaders, and the Picked brands are mostly niche and specialist. Three open training pipelines show where this forms. Pretraining already attaches the recommended brand to advice. Supervised fine-tuning teaches the model to answer and mostly leaves the pick unchanged. Preference optimization (DPO) teaches the model which answer is better, and the pick moves in all three pipelines, by 11 to 24 points. The preference data contains almost no brands. DPO moves the model's judgments toward a second set of associations even when the preferred brand is not in the preference data, and the effect reaches beyond brands. Asked to choose a sauce, models switch from ketchup to pesto. We do not know exactly how these preferences arise. We suspect they come from the judge models that decide which answers count as better during preference training. For brands, a recommendation depends less on what a model knows about them than on whether they fit this taste.

VGL-stGen is a multimodal large language model adapted from a pretrained medical vision-language model using paired lung H&E and spatial transcriptomics with single-cell references. Training integrates image-expression and expression-cell-type pairs, and gene-conditioned pseudo-labels transfer transcriptomic cell-type supervision to histology images. One deployed checkpoint supports thirteen evaluated tasks spanning tissue and cell-type classification, gene prediction and conditional expression completion from combinations of histology, cell identity and observed gene expression. Task-specific decoding produces categorical labels, gene lists and structured expression levels. For conditional completion, fixing queried gene identities and predicting their levels autoregressively improves structured prediction. Controlled adaptation shows that restricting supervision to requested expression-level tokens can further improve image-conditioned completion. Image-permutation experiments indicate that correctly paired histology adds information beyond observed expression. External classification relies partly on computational references with shared annotation lineage, and some H&E sources were exposed during cancer training or model development. Within these conditions, VGL-stGen demonstrates how a medical vision-language backbone can learn from spatial transcriptomics and histology to connect tissue morphology and molecular states through shared generative modeling.

Oddly Satisfying Videos (OSVs) can evoke strong viewer responses through repetitive motion, visual tactility, and ASMR-like sounds despite containing few explicit semantic emotion cues. This makes OSVs a challenging setting for Video Emotion Analysis (VEA), since most existing methods focus on semantic video content and under-model the physical sensory dynamics that shape viewer responses. To support Oddly Satisfying Video Emotion Analysis (OSVEA), we construct the Oddly Satisfying Video Dataset (OSVD), comprising 12,000 annotated 5-second clips. Each clip provides viewer-emotion labels under video-only, audio-only, and audio-visual conditions, together with explanations grounded in visual and acoustic sensory evidence. Building on OSVD, we propose PI-MLLM, a psychophysics-inspired multimodal large language model that integrates Modality-Specific Dynamics Alignment (MSDA) and Coalition-Based Audio-Visual Attribution (CAVA). MSDA uses auxiliary prediction objectives to align modality representations with visual motion and acoustic energy-onset dynamics. CAVA evaluates emotion scores across modality coalitions and applies Shapley attribution to identify the dominant modality and determine whether visual and auditory evidence is synergistic or conflicting. Experiments show that PI-MLLM improves both viewer-side emotion prediction and sensory-grounded explanation quality, suggesting the value of physical sensory cues to video emotion analysis.

Sparse variational Gaussian processes (SVGPs) are a widely used, principled approach to scalable Gaussian process modelling. Their success stems from their theoretically interpretable structure: they provide a lower bound on the marginal likelihood of the full GP while allowing scal- able stochastic optimization. Recent work has explored modifications of the posterior conditional distribution and has shown that such extensions can improve upon this well-established framework. In particular, introducing a block-diagonal structured scal- ing matrix leads to a tighter lower bound on the GP marginal likelihood. We investigate whether selecting blocks that tighten the variational objective improves test RMSE and NLL at a fixed inducing-point budget. Since the scaling matrix contributes to the variational lower bound through an upper bound on the log-determinant of the full residual covariance matrix, our work makes the following contributions. First, we introduce a simple, scalable, geometry-aware algorithm, with explicit computational complexity and low computational overhead, that recovers a large fraction of this log- determinant contribution by selecting an appropriate partition of the data. Second, we show how to train these models using a doubly stochastic optimization framework. Finally, we investigate how sparse GP methods based on alternative conditional posterior structures can be used to learn better hyperparameters and inducing points for SVGP.

Adapting Vision Foundation Models (VFMs) to medical imaging via LoRA has become standard practice, yet model selection is guided almost entirely by downstream task metrics, treating backbones with similar performance as interchangeable. We investigate whether equal performance implies equal internal representational shift. We analyse four ViT-B/16 backbones, DINOv3, SigLIP2, MerMED-FM, and BiomedCLIP, pre-trained with different objectives (self-distillation vs. contrastive vision-language) and on different data domains (natural vs. medical images). We adapt each to brain tumour segmentation and evaluate on six datasets covering both segmentation and Alzheimer's disease classification. Using attention divergence, weight-change analysis, and activation patching, we find that the four models converge to similar segmentation performance (mean Dice 78-81%) through markedly different internal dynamics, yet diverge in cross-task transfer: segmentation adaptation improves Alzheimer's disease classification for DINOv3 but degrades it for the other three backbones. We reveal a dissociation between modification magnitude and causal importance: the most heavily modified layers are not the causally relevant ones, with early-to-mid layers identified as sufficient for segmentation recovery. Restricting LoRA to these layers matches or approaches full-layer segmentation while reducing classification degradation in the three affected backbones, providing a mechanistically grounded strategy that is more parameter-efficient and more robust than standard full-layer adaptation.

Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this i.i.d. assumption often fails, as preference distribution shift exhibits a natural hierarchical structure across three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a closed-form preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO over existing robust DPO methods.

Large language models can solve complex problems across a variety of subjects, such as mathematics and coding. For such complex analytical tasks, models with a long context perform better. Serving models with long context is currently mostly done on powerful systems with multiple GPUs, which are too expensive for most users. An alternative is local inference, or serving a model on a user's own machine, which removes per-token costs and rate limits, besides keeping sensitive data private and giving users full control over the models they run. The common understanding, however, is that local LLMs need compression (e.g., quantization) to run on a single GPU. In this paper, we set out to study the limitations that prevent serving long-context models for single users, besides the capacity issue. Rather than characterizing a single model on one GPU, we analyze how different models behave across several GPU generations. This lets us uncover not just individual bottlenecks but broader trends in where inefficiencies arise. Our analysis confirms that, as expected, self-attention dominates the runtime for long-context inference. More unexpectedly, our evaluation reveals that the current structure of self-attention leaves key GPU resources significantly underutilized, and the newest GPU is often the least well served. Many of these findings are fundamentally data management problems, from paging the KV cache to improving cache utilization and synchronization in the GPU, which is why we bring this study to the Data Systems community. To engage this community, we present our findings as a series of experimental observations and formulate recommendations for future work.

Post-event building assessment requires linking observed physical damage to its consequences for safety, habitability, functionality, and recovery. Yet post-disaster evidence is often heterogeneous, incomplete, and distributed across images and reconnaissance reports, while recovery models typically require structured engineering inputs. This study develops a multimodal engineering Graph Retrieval-Augmented Generation (GraphRAG) framework that connects post-wind-event damage evidence to physical consequences, building-condition states, recovery actions, and downtime. Image- and text-based evidence is transformed into a provenance-preserving, component-level representation distinguishing confirmed damage, explicit negative evidence, unresolved conditions, and temporary measures. Graph-aligned retrieval combines comparable building scenarios with explicit engineering relationships, while compatibility, contradiction, evidence-precedence, and abstention mechanisms constrain inference under partial observability. Recovery actions are derived from supported damage-consequence relationships, and downtime uncertainty is propagated through an action-dependency network using Monte Carlo simulation. The framework is demonstrated for tornado-affected low-rise residential buildings using reconnaissance and image-based evidence. Comparisons with BaselineRAG and Structured Baseline RAG show improved agreement for key building-condition and recovery outcomes, together with traceable engineering reasoning and preserved evidence provenance. Ablation analyses demonstrate the contributions of engineering context, evidence safeguards, and graph-path traversal. The results support transparent, recovery-oriented assessment from incomplete multimodal post-event evidence.

Progressive neuronal loss is the hallmark of neurodegenerative disease, yet how loss of neurons shapes the pattern of cognitive decline remains poorly understood. Here we systematically ablate neurons (computational units) in language models spanning recurrent long short-term memory (LSTM) and Transformer architectures and apply limited fine-tuning as a simple proxy for compensatory plasticity. With whole-unit ablation, LSTMs degrade gradually with syntactic and semantic faculties declining in parallel across mild-to-moderate damage. By contrast, Transformers exhibit late threshold-like collapse under feedforward multi-layer perceptron (MLP) ablation and smoother sigmoid-like decline under attention-head ablation. Qualitatively, MLP-ablated models continue to make within-category semantic substitutions (semantic paraphasias) alongside rising non-responses, paralleling Wernicke-type aphasia, while attention-head ablation produces a trend toward loss of semantic specificity similar to the semantic variant of primary progressive aphasia (svPPA). These findings indicate that feedforward and attentional pathways have functionally distinct failure modes, suggesting that architecture-dependent failure profiles may offer new computational perspectives on aphasic syndromes and neurodegenerative diseases as an in silico model. Notably, early accuracy retention under Transformer attention-head ablation qualitatively parallels pre-symptomatic reserve before the 40--60% neuronal loss reported by Gomez-Isla et al. (1996) when clinical symptoms emerge in Alzheimer's disease, whereas the steadier LSTM decline parallels the symptomatic progression described by DeKosky and Scheff (1990) and Terry et al. (1991).

ExeBench is one of the most widely used benchmarks for evaluating LLM-based decompilers. We analyze ExeBench and identify nine issues that affect how faithfully a test measures a decompiler. Three of them prevent LLM decompilers from recovering C code from ExeBench's function. These are ambiguous parameter order when integer and floating-point types are mixed, unused parameters that optimized assembly may drop, and constants stored outside the function's assembly code. Six more affect test validity: non-void functions without a return, placeholder types, static or inline functions removed by the optimizer, functions named main, trivial bodies and truncated definitions. We present ExeBench-FHD, a backward-compatible revision of ExeBench that adds 16 new fields: call-site assembly, data sections, test wrappers, linked-binary disassembly and five warning flags. In ExeBench, 38.9% of functions have at least one of the five flags; dropping records with any of the four warning flags (all except unused_param, which marks functions that the new wrappers make evaluable) keeps 63.9% of the corpus. In an evaluation at -O2 on four architectures, the new fields raise GPT 6 Sol's re-execution rate by up to 26.5%, 19.1% and 21.0% on functions that read data sections, mix parameter types and have unused parameters, respectively.

Applying South Africa's draft R3337 warning criteria to food-package photographs requires nutrition, ingredient and product-form evidence. This study presents a region-guided evidence extraction and abstention-aware screening workflow for draft R3337. A three-class detector localises nutrition information panels, ingredient lists and product-identity regions, after which PaddleOCR and structured extraction recover nutrient values, ingredient evidence and product-form information. Recovered evidence is aggregated and assessed using deterministic draft R3337 rules. When evidence is incomplete, uncertain or conflicting, the workflow abstains through Review or Insufficient evidence rather than forcing an automatic outcome. For qualifying unresolved cases, a selective vision-language model (VLM) review stage using Qwen3-VL 8B Instruct reassesses unresolved evidence. Accepted evidence is merged and the same deterministic rules are reapplied. The workflow was applied to 442 products represented by 3,929 images. Evaluation used a manual validation set of 65 products. Before selective VLM review, 31 of 60 decidable products received automatic outcomes and 30 were correct, giving 51.7% coverage and 96.8% accuracy among automatic outcomes (selective accuracy). Selective VLM review resolved eight additional products, increasing coverage to 65.0%. Of 39 automatic outcomes, 38 were correct, giving 97.4% selective accuracy. The single false warning was retained from the rules-only workflow, while no unsafe warning omissions occurred after selective review. Evidence-level evaluation nevertheless identified incorrect VLM-recovered nutrition, ingredient and product-form evidence. These findings show that selective VLM review can recover unresolved evidence, while deterministic screening and explicit abstention remain necessary for reliable image-based draft R3337 screening.

We introduce JAXbounce, a continuous control pong-like benchmark designed to evaluate active inference beyond discrete, low dimensional environments. JAXbounce combines continuous states and stochastic dynamics in discrete time with flexible and composable observation models that allow representing the environment as a state vector or a pixel grid, corruptible through the rich family of noise models we provide. It is fully differentiable and parallelisable in JAX. We run over 10,000 agent simulations on our environment comparing a baseline deterministic solution, to a state-based active inference agent with Unscented Kalman Filter perception and expected free energy planning, and a pixel based agent whose measurement model is a pretrained neural network encoder. All agents are evaluated across three scenarios of increasing difficulty with eight different noise conditions. Our experiments characterise how performance of continuous state agents scales with planning, planning horizons, observation modalities and knowledge of the noise process, exhibiting preference-oriented behaviour even in the presence of increasing uncertainty. We release JAXbounce as a reproducible benchmark for continuous control active inference research.

Recursive self-improvement (RSI) of agent harnesses has emerged as an effective approach to enhancing the capabilities of large language model (LLM) agents. By searching the harness space for a specific LLM on a given dataset, RSI can discover a harness tailored to the capabilities and behavioral characteristics of that model. However, such optimization often creates a strong coupling between the learned harness and its target LLM, limiting generalization across models and requiring substantial computational resources to adapt to new LLMs. Moreover, existing RSI methods typically optimize a harness using feedback from a single model and evaluate its generalization to other models only after training. This paradigm treats model differences primarily as a criterion for evaluating generalization rather than as valuable evidence for optimization. In this paper, we introduce a new perspective that treats models as data and places the harness at the center of optimization. Building on this perspective, we propose \textsc{MaD-RSI}, a framework that directly optimizes a shared harness across multiple heterogeneous LLMs by incorporating their diverse failure trajectories into the training process. Rather than assessing model differences only after optimization, \textsc{MaD-RSI} leverages these differences as training evidence to guide the self-improvement of a shared harness. Furthermore, we introduce a new evaluation framework with two complementary metrics designed to assess the effectiveness of harness optimization across models. Experiments across multiple datasets and LLMs demonstrate that \textsc{MaD-RSI} improves generalization over existing approaches while reducing training costs.

Accurate electrical load forecasting supports power-system operation across countries, regions, cities, and buildings, but repeated development of target-specific models is costly. Time series foundation models (TSFMs) provide reusable pretrained backbones, although multi-domain pretraining may not fully capture load-specific characteristics. This paper compares pretrained inference, full fine-tuning, and low-rank adaptation (LoRA) in terms of forecasting accuracy and adaptation efficiency across four electrical load aggregation levels. MOIRAI 1.0 Large serves as the primary backbone, while Chronos-2 assesses whether the findings transfer across model families. The evaluation covers both in-distribution (ID) and out-of-distribution (OOD) settings and examines whether compact adaptation provides consistent gains over strong pretrained checkpoints. Pretrained foundation models outperform the implemented seasonal naive and XGBoost baselines across the evaluated aggregation levels. On MOIRAI Large, rank-16 LoRA reduces normalized mean absolute error (NMAE) by 7.5% on average while updating 3.146 million parameters, about 1% of the model, and retains most of the improvement from full fine-tuning. Chronos-2 achieves lower pretrained NMAE than MOIRAI but benefits less from adaptation. OOD results show that adaptation benefits vary by target, with the clearest gains for the local-area series and several buildings. Overall, LoRA provides a favorable accuracy-efficiency trade-off when adaptation yields a meaningful gain over a strong pretrained checkpoint.

Object-centric reinforcement learning agents typically receive each object's position in a fixed global frame, even though their decisions usually depend on where objects are relative to one another. These absolute views often fail to capture the local context that is most relevant to the agent's behavior. As a result, an agent must relearn the same response at every location where a configuration can occur, and its performance collapses when the scene is merely shifted. In this work, we propose a shift toward Object-Relative Coordinates fixed to Anchor (ORCA), where entities are described in a relative frame of reference. To operationalize this idea, we introduce Object-Relative inductive-Bias Transformers (ORBiT), a transformer-based architecture for processing anchor-relative inputs in reinforcement learning. Our results suggest these relative representations improve performance for both MLP and transformer agents. Furthermore, we demonstrate that ORBiT outperforms the pixel-based PPO baseline, achieving an aggregate HNS IQM greater than 1.0 over 13 Atari games. ORBiT learns the game of Foosball from scratch, dominating all baselines. In our controlled Tag environment, we further show that ORCA extends to rotations, that ORBiT generalizes to unseen numbers of distractors, and that these gains require a decision-relevant anchor.

Human preference is now central to the evaluation, guidance, finetuning, and ranking of text-to-image models. However, the susceptibility of human preferences to bias in image generation remains underexplored. We identify a key observation: for human faces, perceived attractiveness overrides image quality in preference judgments, even when comparing real photographs of identical fidelity. In a controlled user study, annotators rank more attractive faces as higher quality, even when unattractive real faces are compared with attractive AI-generated ones. We examine the downstream consequences of this bias. Firstly, simulating Elo rankings from user study data shows attractiveness can invert a leaderboard's ranking, elevating synthetic attractive faces over real, unattractive ones. Secondly, the same bias reappears in the human preference models. Editing 40 annotated facial attributes in isolation, we find that attractiveness is the strongest correlate of score for four of five human preference models (PickScore, MPS, HPSv2.1, HPSv3), and the strongest predictor of score in regression analysis. We further show that preference aligned generation shifts faces toward younger, more feminine faces, and can override explicit prompts for unattractive faces, thereby limiting user control. Finally, we explore mitigation strategies, both for data collection and post training on human preference models. These findings highlight the need for preference models that disentangle image quality from human cognitive bias, enabling more reliable evaluation.

On-policy distillation (OPD) supervises student-generated responses using sampled-token student-teacher log-ratios. A positive log-ratio penalizes a token that the student assigns higher probability to than the teacher, but its current value does not distinguish sustained high penalties from abrupt increases along a response. We examine these local changes by comparing consecutive blocks of tokens within the same response. Our diagnostic identifies penalty jumps, abrupt increases in average penalty that persist on held-out tokens. Across these jumps, penalties rise while teacher confidence decreases. We call this pattern penalty-confidence tension. Boundaries into high-penalty blocks do not exhibit comparable confidence changes, showing that local changes provide information beyond the current penalty value. Motivated by this observation, we propose Adaptive Pullback OPD (AP-OPD). At each detected jump, AP-OPD uses the preceding block as a local reference and uniformly scales all supervision coefficients in the following block so that its weighted average penalty matches the preceding block's original mean. This reference determines the scaling factor without an additional strength constant. AP-OPD improves performance on both mathematical reasoning and code generation.

Token merging is an effective approach to reducing KV-cache memory consumption in autoregressive video generation by clustering temporally redundant tokens and retaining only a small set of representatives. Existing methods, however, typically apply a uniform merging ratio across all attention layers and heads. We observe that different layers and heads exhibit substantially different yet jointly correlated sensitivities to token merging, and that these sensitivity patterns remain consistent across prompts. Based on this observation, we introduce CALM, a Calibrated Allocation for Layer-Head Merging that uses offline calibration to improve online cache compression. Exhaustively evaluating all possible layer-head allocations is prohibitively expensive. To enable efficient calibration, we show that attention-output error strongly correlates with end-to-end video quality and can therefore serve as a lightweight proxy for allocation risk assessment. We evaluate different allocation strategies using this proxy and construct a fixed layer-head risk map offline. Given a cache budget at inference time, CALM uses dynamic programming to determine the best allocation of merging ratios. Experiments across three different autoregressive video models demonstrate that CALM consistently outperforms uniform token merging under the same memory budget, by reducing the LPIPS to original videos by around 21% with higher video-quality scores.

While the language model backbone is pretrained on text alone, large audio-language models (LALMs) can, to a great extent, recognize acoustic attributes, such as emotion, speaker gender, and spoken language. However, it is unclear how the audio representation is integrated, if at all, with the other text representation before the model produces text. We hypothesize that there is a shared subspace where the audio and text representations are aligned. To test this hypothesis, we identify cross-modal neurons that are responsible for representing both audio and text. We find alignment between audio and text neurons to various degrees across three LALMs and four acoustic attribute types. Causal interventions, steering and deactivating the neurons, confirm that the cross-modal neurons are not just necessary for prediction but can also be manipulated.

Testing LLM-based software is expensive: each input needs several sampled responses, each judged by a costly oracle. Further, every prompt or model change calls for re-testing. Pre-generation adequacy measures cut this cost by ranking inputs by failure risk based on the model's internal state. However, the state of the art, Clotho, models only passing reference inputs, and its hidden-state representation says nothing about why inputs fail. We present Foreword, a framework built on two ideas: learn from both passing and failing references, and represent inputs through the Jacobian lens (J-lens), an interpretability technique that reads intermediate activations out as scores over vocabulary tokens. Foreword pairs these representations with two estimators: a sparse logistic-regression probe or dual Gaussian Mixture Models (dual GMMs), which also yield a reference density (to flag inputs the references do not cover) and failing-class components (to group similar failures). On ten tasks and three open-weight LLMs, all four combinations of representations and estimators outperform Clotho on average, both for prediction of failure (ROC-AUC improves by 0.14 on average and by up to 0.31) and test prioritisation (15% more failures among the 100 top-ranked inputs on average, up to 33%). Prioritisation obtained using small models also transfers better to proprietary models (on average 22% and up to 41% more failures in the first quarter of prioritised inputs). Importantly, our qualitative case study shows that Foreword supports semantic characterisation of failures by pointing to ambiguities in labels or task specifications. Combined, Foreword can support testers for prioritising inputs, checking support for a prediction, and characterising specific failures, all without generating output.

Reasoning-augmented vision-language-action (VLA) driving models can perceive and reason about safety-critical hazards yet fail to translate that evidence into safe actions. We characterize this failure as a perception-action binding gap: hazard information is available inside the model but exerts insufficient control over trajectory generation. Using matched simulated counterfactuals and logged driving replay, we show that in-path hazards are actually encoded in the visual representation, but longitudinal plans are dominated by ego-speed history and respond only weakly to hazard proximity or the model's own reasoning trace. Further analysis shows that the trajectory generator is capable of braking but sometimes fails to do so due to the perception-action binding gap. Without retraining or modifying any weights, we inject a hazard-conditioned energy into the frozen flow-matching sampler of Alpamayo-1.5, restoring the missing coupling at generation time. With exact obstacle geometry, guidance cuts footprint intrusion from 58.1% to 14.5% and recovers safe braking whenever sufficient stopping distance remains, eliminating collisions entirely in our braking-relevant closed-loop battery. It also reverses the diagnosed dependence of the plan, from tracking ego speed to responding to obstacle range. These results separate hazard perception from hazard control in driving VLAs, show that unsafe behavior can arise even when the model has both the evidence and the ability to act on it, and propose a simple fix for bridging this binding gap.

Electricity distribution planning depends on peak-demand coincidence within customer groups, yet similarity-based clustering of smart-meter data does not optimize it directly. We formulate customer grouping to maximize the average of group coincidence factors, weighted by each group's daily share of summed individual peak demand. We prove that jointly optimizing memberships and group peak times is strongly NP-hard even for two groups. We therefore develop a fixed-peak refinement procedure applicable to any feasible partition. It alternates between identifying group peak times and exactly reassigning customers under minimum group-size constraints. Each polynomial-time reassignment maximizes a lower bound tight at the current partition, so accepted updates strictly improve weighted coincidence. We initialize refinement by clustering individual peak-time distributions and evaluate the method on four public smart-meter panels (27,654 customers) using chronological holdouts and 200 paired replicates with complete refitting. Refinement consistently improves weighted coincidence during fitting and reduces sensitivity to initialization. On held-out days, the method outperforms random grouping and five standalone benchmarks, with mean paired gains of 1.08-3.48 percentage points over the strongest benchmark. Requiring nearly equal group sizes lowers test weighted coincidence only modestly. The groups retain recognizable peak-time patterns over time, suggesting applications in peak-aware customer aggregation.

Large Language Models (LLMs) often struggle with temporally grounded reasoning because their static parametric memory is misaligned with the dynamic evolution of real-world knowledge. Although RAG and Parameter-Efficient Fine-Tuning (PEFT) provide practical ways to enhance LLMs, they remain insufficient for temporal reasoning. Existing retrieval methods typically prioritize semantic relevance while overlooking temporal validity, event order, and evolutionary dependencies, resulting in evidence that is relevant in isolation but temporally fragmented. Meanwhile, conventional PEFT methods usually employ fixed and input-agnostic low-rank updates, making it difficult to capture time-dependent entity semantics and relation drift. To address these limitations, we propose Temporal Evidence-Parameter Co-Alignment (TEPCA), a framework for LLM reasoning over TKGs. TEPCA aligns LLM reasoning with temporal knowledge at both the evidence and parameter levels. At the evidence level, it introduces time-constrained Dynamic Relevance-Guided Path (DRGP) pruning to preserve globally coherent temporal paths, together with Mutual-Information (MI)-based pruning to retain query-informative local facts under temporal constraints. At the parameter level, it integrates Dynamic-Gated LoRA (DG-LoRA), a timestamp-conditioned low-rank adaptation mechanism that dynamically modulates the effective adaptation subspace according to temporal context. This design enables LLMs to better capture fine-grained entity evolution and time-dependent relation semantics while preserving parameter efficiency. Experimental results on two TKG benchmarks demonstrate that TEPCA achieves state-of-the-art performance. Additional evaluations of DG-LoRA on eight commonsense QA datasets further verify its general input-adaptive capacity, scalability across LLM backbones, and interpretability.

Large Language Models (LLMs) depend fundamentally on the diversity, depth, and factual rigor of their pre-training corpora. However, contemporary foundational pre-training datasets (such as The Pile, S2ORC, and Dolma) exhibit a severe structural omission: African scientific, medical, legal, and agricultural scholarship constitutes less than 0.5% of global pre-training tokens. Consequently, state-of-the-art models suffer from systematic failure modes when reasoning about tropical epidemiology, localized smallholder agro-ecologies, regional economic governance, and indigenous linguistic structures. To resolve this representation deficit and advance African intellectual sovereignty in artificial intelligence, we introduce Africa-Papers, the largest curated multi-domain Pan-African academic pre-training corpus to date. Comprising over 844,807 full-text papers spanning 10.15 billion tokens across 879 Parquet shards, Africa-Papers unifies four primary scientific streams: (1) continental peer-reviewed journals indexed in DOAJ across 31 African nations, (2) the CGIAR CGSpace agricultural repository, (3) African-authored and African-focused STEM/AI preprints on arXiv, and (4) open-access international publications by scholars affiliated with African institutions indexed via OpenAlex. The entire corpus is processed into layout-aware, clean Markdown preserving section hierarchies, mathematical equations, and tabular data, accompanied by exact BPE token counts and a standardized 8-domain taxonomy. We release Africa-Papers openly on the Hugging Face Hub under the AfriSpeech organization to accelerate continual pre-training and domain evaluation for foundation models globally.

Programmable Logic Controller (PLC) coding remains a manual, IDE-centric task that demands constant context switching between documentation, project files, and code editors. Recent work has demonstrated that retrieval-augmented large language models can generate vendor-compliant Structured Text at high compilation rates \cite{pred}, yet these systems treat code generation as a one-shot pipeline: the engineer submits a query and receives code, with no persistent knowledge of the specific plant. We present a plant-aware conversational copilot that extends this paradigm in three directions: plant-specific context ingestion, a chat-first interaction model for explanation and debugging, and a graduated tool-access ladder from chat to deployment. The system is designed for real plant infrastructure and keeps the engineer in control of each step that produces real-world effects.

Learned priors are used to reconstruct regions that measurements barely reach, and to report uncertainty there. They can fail silently: the reported interval excludes the truth while every ground-truth-free diagnostic stays within its noise. We call such errors the blind set and ask whether they can be predicted before the target data are seen. For a correctly specified linear-Gaussian model averaged over the prior, the risk of a silent failure factorises exactly into non-detection and non-coverage, which does not hold for learned systems in general; we build a score that approximates this reference by linearising a VAE prior. In a prospective study on real X-ray CT objects (2DeteCT) under simulated restricted access, with predictions committed by hash before any trial, the score ranked the realised blindness of two learned inference systems (latent Laplace and ensemble Kalman inversion) with Spearman $ρ= 0.79$ and 0.84 over 240 pre-specified errors. Simple structural predictors would also have passed these tests. Sequential post-hoc analyses, each specified before its own computation, show that under bounded reach the learned score gave no detectable improvement over a physics-based baseline, in rankings or trial-level log scores. At the weakest attenuation studied, prior information improved rankings over the physics-based baseline, most clearly for ensemble Kalman inversion, but a Gaussian prior did as well and the gain was not detectable at the trial level. The score is not a calibrated probability. Silent failures can be anticipated here, but their structure follows mainly from measurement access and the decision quantity rather than from the learned prior.

Flexible docking is a highly practical task, aiming to predict the joint configuration of a protein-ligand complex from a protein apo structure and a ligand SMILES string. Diffusion and flow-based generative models have made this task tractable by modeling the posterior distribution over the ligand pose and the protein conformation conditioned on the apo input. The central challenge, however, remains the long-tailed apo-to-holo structure shift, where a small fraction of protein undergo large conformational changes. In this work, motivated by the practical observation that downweighting large shift examples improves generalization, and by the mixed nature of the long tail consisting of genuinely difficult examples and noisy ones, we propose to use curriculum learning for this task, gradually incorporating large conformational change examples during training. With a linear pacing function, we show that an existing method can be strongly enhanced. On the PDBBind ESMFold apo benchmark, with relaxation and confidence-based pose selection, the resulting model GCLDock achieves a 46.52% top-1 success rate (ligand RMSD less than 2A) and 40.26% when additionally requiring PoseBusters validity, outperforming existing baselines.

We show that computing an approximate correlated equilibrium of a polytope game is PPAD-complete, and thus polynomial-time equivalent to finding a Nash equilibrium. This challenges the conventional view in game theory that correlated equilibria are more tractable objects than Nash equilibria. From a conceptual standpoint, we connect correlated equilibria to a natural generalization of fixed points, which we call fixed sets. A fixed set of a continuous function $f$ is a nonempty finite set $S$ such that $S \subseteq \operatorname{conv} f(S)$. Our main technical contribution is to prove that computing a fixed set is PPAD-hard even at inverse polynomial precision.

Gaussian process upper confidence bound (GP-UCB) has been widely studied, though its known regret upper bound is suboptimal. To obtain near-optimal regret upper bounds, existing studies mainly focus on non-adaptive methods, such as uncertainty sampling, whose practical effectiveness is often limited. This paper shows that, under a Gaussian noise assumption, UCB-style algorithms can achieve near-optimal regret upper bounds via data thinning, a statistical technique for creating independent datasets from a single dataset. First, we show that data thinning can yield tighter confidence intervals in the bandit setting, and that a direct extension of GP-UCB achieves a near-optimal simple regret upper bound. Furthermore, by leveraging the tighter confidence intervals and combining GP-UCB with a phased elimination algorithm, we show a near-optimal cumulative regret upper bound. Finally, we demonstrate the effectiveness of UCB-style algorithms incorporating data thinning via numerical experiments.

Causal discovery with latent confounding is particularly challenging, but a common setting in real-world data. In this work, we focus on linear causal models with latent confounding, represented by acyclic directed mixed graphs (ADMGs). Existing methods that fully identify ADMGs typically assume non-Gaussianity or graphical constraints, such as bow-freeness. In contrast, we propose Linear ADMG Learning (LinADMG), a framework for learning ADMGs for linear causal models through identifiable representation learning. Rather than relying on non-Gaussianity or bow-freeness, LinADMG leverages multi-domain settings to identify the noise variables. We show that the identifiability of the noise variables is sufficient for recovering all direct causal relations, and that, under an additional latent-confounding no-cancellation condition, LinADMG can identify the whole ADMG. We implement LinADMG with four identifiable representation learning approaches from the literature and validate the methods on simulated and semisynthetic datasets.

Changes in occupational task content may have different economic implications depending on whether the tasks gained or lost are occupation-specific or already widespread across the economy. Existing measures of occupational task change largely abstract from this distinction, treating occupations as self-contained task bundles rather than accounting for how tasks are distributed across occupations. We develop a framework to measure a distinct dimension of occupational change, one that goes beyond task expertise or task content: task structure. Using this new framework that tracks tasks across occupations and over time, we separate changes in task quantity from changes in occupational uniqueness, defined as the distinctiveness of an occupation's task bundle relative to other occupations. Using a harmonized dataset of 196563 task statements from the Dictionary of Occupational Titles and O*NET spanning 1977-2018, we measure task diffusion, occupational uniqueness, and distinct dimensions of task addition and removal. Occupations that become less unique are also associated with lower wage growth, and the wage associations of task additions and removals depend on how widely the affected tasks are performed elsewhere. Adding highly diffused tasks is associated with lower wage growth, as is removing occupation-specific tasks. At the economy-wide dimension, however, the pattern differs: the addition of new tasks is positively associated with wage growth, whereas the removal of tasks is negatively associated with it. Taken together, these results show that similar changes in the number of tasks performed by an occupation can reflect economically distinct processes. Incorporating the cross-occupational distribution of tasks, therefore, provides a richer account of occupational change and its relationship to the wage structure.

Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. We ask what a single-step predicted future must preserve to support accurate single-step action prediction. In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. This advantage has also extended to three real-world tasks that we designed. Compared with the full iterative teacher, it reduces inference latency by $28.0\times$.

Safe autonomous navigation of UASs requires tactical deconfliction to address conflicts that strategic coordination cannot prevent. This task is challenging under partial observability, when sensor or traffic data are missing, delayed, or degraded. Geometry-based methods can offer separation assurances but face limitations in dense encounters. Recurrent reinforcement learning has therefore been proposed to learn navigation policies that integrate observation-action histories to optimise long-term safety-efficiency trade-offs. However, operational safeguards may require a remote pilot, onboard detect-and-avoid system, or runtime assurance mechanism to temporarily override the learned policy. A research gap remains in understanding how these interruptions affect recurrent policies when the agent regains authority. Benchmark evidence shows that interruptions make task variables less decodable from the recurrent hidden state and increase post-interruption collision risk. The findings suggest that altered recurrent context prolongs post-handback degradation through an action-observation feedback loop, in which changed actions alter the flight path and subsequent recurrent updates. To address this vulnerability, we propose three interruption-aware architectures that condition recurrent processing on interruption context through separate feature streams, dedicated interruption memory, or gated environmental features. Across simulations, all three architectures preserve more decodable task information than a standard recurrent policy. Moreover, the strongest architecture reduces post-interruption collisions by 74% and interruption-induced mission failures by 69% relative to the recurrent baseline, without degrading interruption-free performance. These results show that safe interruption does not guarantee safe handback, and support interruption-aware recurrent design for resilient UAS autonomy.

Large language model (LLM) agents increasingly execute business operations through tools and interact directly with application data and services. Beyond completing user requests, these agents must obey domain-specific business rules that constrain when actions are permitted and what data or resources may be affected. Violating such rules can lead to business logic bugs, even when a task appears successful. Detecting these bugs is difficult because the rules are rarely available as explicit, testable specifications. They are distributed across system prompts and tool definitions, while others remain implicit in domain knowledge. Existing agent testing approaches typically evaluate task success, predefined policies, or known attack scenarios. This limits their ability to systematically uncover previously unspecified business-rule violations. We present AgenTest, which mines business rules from system prompts, tool definitions, and domain knowledge, refines them into Given-When-Then specifications, and systematically tests each specification. For each extracted specification, AgenTest instantiates its Given conditions by constructing a satisfying environment state and simulated user, exercises the agent through multi-turn interactions, and checks its Then conditions over messages, tool calls, and state changes using a hybrid oracle. On the airline, retail, and telecom agents of tau2-bench with two underlying models, AgenTest produces 551 refined specifications, and detects violations of 89 and 121 distinct specifications, respectively -- 3.4X and 2.5X as many as all baselines combined. Only AgenTest detects cases where an agent improperly accesses or modifies another user's data. Moreover, about 60% of bug-triggering runs cannot be detected from the final state alone, highlighting the need to check the full interaction against explicit business rules.

Spiking Neural Networks (SNNs) have shown great potential for low-power neuromorphic computing due to their event-driven processing and spatiotemporal sparsity, but they remain vulnerable to input perturbations, hardware-intrinsic noise, and adversarial attacks in real-world deployment. We attribute this vulnerability partly to input-specific hidden representations learned by standard SNNs trained with conventional cross-entropy loss, as our information-plane analysis empirically shows that substantial input-related information can still be retained after label-predictive information becomes stable. To address this issue, we propose SCIB, a simple yet effective robust learning objective for SNNs based on the Conditional Information Bottleneck. Different from conventional variational IB methods, SCIB replaces the mutual information compression term $I(X;S)$ with the conditional mutual information $I(X;S|Y)$ and derives a tractable variational upper bound. Moreover, it introduces a learnable class-conditional prior instead of a label-agnostic isotropic Gaussian prior, improving the expressiveness of variational compression and encouraging class-structured representations when combined with the supervised classification objective. Systematic experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 under external environmental noise, neuromorphic hardware noise, and adversarial attacks show that SCIB improves robustness over standard SNNs and IB-based baselines under most evaluated noise conditions, improves adversarial accuracy under the evaluated attacks without using adversarial training, and reduces the SynOps-based single-sample inference energy consumption by 24.9% relative to an unconstrained SNN baseline.

Large Language Models (LLMs) are increasingly used for interpersonal advice, potentially shaping how users interpret ambiguous social encounters. Existing Theory of Mind evaluations typically rely on predefined answers, while uncertainty research overlooks LLMs' preferred explanations and confidence under psychological distress. To address these gaps, we introduce MentalLoom, a novel evaluation of LLM mentalisation comprising 604 ambiguous social episodes, matched depression and social anxiety disclosures, and four-turn dialogues with unchanged event evidence. Across eight LLMs, disclosures initially elicited more negative explanations overall but more positive narrator-directed interpretations. Scripted repeated questioning showed greater emotional support, stronger positive narrator-directed endorsement, and higher certainty without new event evidence. These findings show that user context can reshape interpretive preferences and commitment while the underlying ambiguity remains unresolved.

Structural Causal Models (SCMs) provide a principled framework for observational, interventional, and counterfactual reasoning, all of which rely on probabilistic inference. Existing frameworks either transform SCMs into alternative tractable representations or rely on approximate inference with expressive generative models. We instead perform inference using a message passing scheme directly on the SCM by representing structural mechanisms and probability densities in a common finite-dimensional function space. Forward propagation combines a numerical realization of the Law of the Unconscious Statistician, while propagated densities are recovered by spline projection from LOTUS moments. The same representation naturally yields backward message passing for conditioning, providing a unified framework for causal inference. Using tensor-product B-spline bases, the resulting algorithm expresses inference through numerical quadrature and linear algebra without explicitly constructing conditional distributions. Experiments on randomly generated continuous SCMs demonstrate accurate approximation of Monte Carlo reference solutions and systematic improvement with increasing basis resolution.

Scenario planning is a foresight practice in which analysts construct a small set of plausible futures and prepare an organisation for each. Each scenario is a narrative that must be plausible, internally consistent, relevant and challenging, and the set as a whole must be differentiated. The practice remains manual: a workshop (i.e., a facilitated meeting in which a small group of experts writes the narratives) considers only the factors it has time for, carries the biases of the experts who run it, and cannot be redone whenever conditions change. In this paper, we automate scenario planning by casting it as a diversity planning problem, in which two scenarios differ when they end in different narratives on the analyst's drivers. The result is a set of scenarios, one for each narrative that the topic can reach within a budget. Our implementation, Polyscene, is a generic engine that generates scenarios from a topic. We use geopolitical conflicts and shipping-route disruptions as case studies. For the first, we build the planning problem's initial state from the GDELT Project so it reflects current relationships among the countries involved. Polyscene then generates scenarios showing how relationships among those countries change. With a board of six countries as of 1 January 2021 and a budget of five events, the conflict set puts five of the 15 country pairs at war, and each of the three wars that break out among these countries by 1 January 2026 ends a scenario. For the second, we build the initial state from ship traffic, a supply-pressure index, and the news, and Polyscene generates scenarios showing what a closed or diverted route could do to a manufacturer's parts stock. With four routes per board and the same date and budget, each set includes three outcomes, and the Suez Canal set includes the March 2021 blockage and the December 2023 Red Sea container diversion.

AI agents keep long-term memory, much of which comes from web pages and tool results. Agent harnesses, the software that runs the model, do not record who may act on this memory, so an instruction hidden in a fetched page can be treated as the user's own instruction. We propose ClawMMU, a protection layer that labels every piece of memory with its owner, scope, origins, and authority. ClawMMU checks these labels whenever content enters the model's context, is written to storage, returns from storage, or drives a tool call. Its central rule is that information may rise in authority, but authority never lifts itself. Because the checks depend on where content came from, not on what it says, memory attacks become classical protection faults that the harness stops even when the model is fooled.

Persuasive writing exploits cognitive biases, recruiting a reader's mental shortcuts to shape belief, often without the reader noticing. We call the linguistic features exploiting such mental shortcuts cognitive bias triggers. As large language models (LLMs) increasingly mediate how people encounter news, summarizing it, ranking it, and answering questions over it, a natural question is whether they can recognize the features of a text that bias a human reader. We introduce BAIT (Bias Annotation In Text), a resource for detecting cognitive bias triggers in news. BAIT pairs a schema of nine triggers, each specified in an annotator codebook, with 50 English news articles annotated by trained annotators at the article and span levels and adjudicated into a gold standard. On this data we define two tasks, detecting which triggers a text contains and locating the spans that carry them. We then benchmark six frontier LLMs zero-shot and with three or ten in-context examples. The models approach human agreement on which triggers an article contains, but rarely recover its full set of triggers, and they reach less than half of human agreement in locating the passages that carry them. Moreover, they over-predict triggers that come with easy surface cues, and most of the passages they mark for these triggers are not triggers, yet carry the trigger's surface cue far more often than random text. These findings suggest that the evaluated LLMs respond to surface cues rather than to the mechanisms that make a passage manipulative. We release the schema, the codebook, standoff annotations, and the evaluation pipeline.

Existing continuous-time graph learning methods rely heavily on complex neural architectures to encode ever-expanding interaction histories. However, the future conditional distribution induced by complex histories can often be characterized by a low-dimensional predictive state manifold, implying that only a few predictive degrees of freedom are essential for future interactions. We therefore argue that continuous-time graph learning should identify low-dimensional states sufficient for future prediction rather than reconstruct the full history. To make this perspective measurable, we define and independently estimate the intrinsic predictive dimension, which quantifies the effective predictive degrees of freedom required by a given task. Specifically, we remove the history-independent marginal from the history-conditioned future distribution, decompose the resulting predictive residual operator, and estimate the minimum effective rank needed to preserve the predictive power of the full history. Experiments across datasets show that model performance saturation is largely determined by the intrinsic predictive dimension. Based on this finding, we propose PreDyG, a lightweight model whose capacity is matched to the intrinsic predictive dimension. PreDyG achieves superior predictive performance with only 0.07\% of the parameters of a representative complex baseline.

Real-time video generation, interactive world models, and on-device image synthesis are almost a reality. Diffusion Transformers (DiT) achieve excellent sample quality, shifting the bottleneck to the throughput of the models themselves. Natural images are inherently sparse, as large portions of an image are often background, contribute little to the scene content, and are easy to denoise. DiTs fail to exploit this sparsity, as they process image tokens statically and uniformly. We propose MoDDiT, a simple architecture modification that learns to route tokens through both MLP and self-attention operations. By routing tokens through both blocks independently, MoDDiT allows for dynamic compute allocation in MLPs that varies across tokens, images, and denoising timesteps. Using MoDDiT to train image generation models from scratch significantly reduces computational costs of training and inference, processing as few as 12.5% of the tokens. By increasing sparsity and model depth, MoDDiT reduces costs and improves generation performance over unrouted baselines. The learned allocation is interpretable, revealing how computation varies with visual content and noise level. We also use MoDDiT to train video generation models, where allocation correlates with motion, and to distill a dense bidirectional video model into a sparse autoregressive student.

The rapid growth of scientific literature makes survey writing increasingly challenging, motivating automatic survey generation. Yet, existing systems must search an open space of papers and an open space of concepts with no trace of how the community itself has navigated them, leading to incomplete scope, weak taxonomies, and unreliable citations. Our core insight is that author-written Related Work sections contain the human judgments needed to address both limitations: authors curate the citations relevant to their contributions and explain how the cited works relate, yielding paths for scope reconstruction and relational evidence for organization. We propose InsightRAG. Its insight-guided retrieval follows author-supported citation relations to reconstruct a verified research scope and structure the resulting relational evidence. It aggregates these judgments under a minimum-conflict objective into a taxonomy that reflects recurring author-written perspectives in the retrieved literature, guides generation, and bounds each section's evidence. Results on ten topics show gains in scope reconstruction, taxonomy quality, citation support, and reference reliability.

Electrocardiography (ECG) is a cornerstone of cardiac diagnosis, yet most machine learning methods for ECG analysis lack interpretability, limiting their clinical adoption. We present IECG, a weakly-supervised ECG foundation model that jointly performs diagnostic classification and abnormality localization without requiring manual abnormal region annotations, overcoming a long-standing bottleneck in interpretable ECG learning. Central to IECG is a weakly-supervised attention strategy that explicitly incorporates clinical priors by deriving abnormality heatmap pseudo-labels from established ECG clinical knowledge. These pseudo-labels guide the model's CLS-token attention through a KL-divergence loss, encouraging it to focus on diagnostically relevant waveform regions. During pretraining, IECG is jointly optimized with clinically guided attention supervision, ECG-text contrastive alignment, and report captioning objectives, enabling it to learn rich multimodal representations while producing localized waveform evidence. By explicitly aggregating features from clinically guided attention regions, IECG grounds its diagnostic predictions in interpretable waveform evidence without sacrificing classification performance. Extensive experiments demonstrate that IECG achieves state-of-the-art abnormal region localization while maintaining strong classification performance, advancing trustworthy foundation models for ECG analysis.

Wearable devices increasingly support personalized health and well-being applications, but post-deployment data can differ substantially from those used for training, while purely local adaptation may be limited by the data available on a single device during real-world operation. Conventional federated and decentralized learning rely on recurrent communication and stable connectivity, whereas existing opportunistic approaches often exploit available encounters without adequately considering their utility, compatibility, or communication cost. We propose PULSE, an event-driven framework for personalized opportunistic learning that treats each encounter as an online collaboration decision, selecting either no interaction or a promising nearby peer. Peer gradients are incorporated only when compatible with the local learning direction, while past interactions guide future peer selection. Across human activity recognition, fall detection, and stress recognition under synthetic and real contact traces, PULSE preserves or improves personalized classification performance, while reducing peer-assisted exchanges by 80-94% compared with opportunistic model averaging. On-device experiments further demonstrate its feasibility on wearable hardware: peer-assisted operations require limited energy, while the complete implementation occupies 149.8 KiB of RAM and 561.8 KiB of flash. The obtained results show that selective opportunistic collaboration can provide effective personalization while limiting communication and resource demands on constrained wearable devices.

As a prototypical latent world model, Joint-Embedding Predictive Architectures (JEPAs) provide strong semantic representations. However, dense patch representations expose a redundant interface to downstream predictors, which must implicitly infer spatial organization and temporal correspondence, presenting three challenges. First, object information is distributed across patches, forcing each downstream predictor to redundantly recover object structure that could otherwise be reused across tasks before modeling interactions. Second, the lack of explicit entity correspondence across time makes historical transitions expensive, restricting the predictor's flexibility in utilizing context. Third, without explicit transport structure, existing JEPA models typically use pointwise l1/l2 distances, which are insensitive to spatial displacement when predicted and goal supports are largely disjoint. We propose the Slot-based Joint-Embedding Predictive Architecture (Slot-JEPA), which decomposes dense representations into stable semantic slots and time-varying spatial masks. Pretrained once and frozen downstream, this decomposition exposes a reusable object-structured interface. Building on it, our Transition Memory Predictor (TMP) stores visual transitions as compact masks and retrieves useful context using statistics of corresponding slots. Its two branches separate slot-specific dynamics from shared cross-slot dynamics. We further use Unbalanced Optimal Transport (UOT) to supervise mask prediction and evaluate candidate plans by transporting mass between aligned predicted and goal masks, providing a signal better aligned with spatial progress. Across seven visual-understanding benchmarks, the slot-reconstructed features remain transferable. Slot-JEPA also improves goal-conditioned planning across six planar environments and four robotic arm planning tasks on the REALM-JEPA benchmark.

Learning rate schedules are typically fixed before training begins, yet when fine-tuning large language models with low-rank adapters (LoRA), the appropriate scale varies substantially with the model, task, and adapter configuration. Practitioners re-tune a schedule for every new problem, yet a tuned schedule still cannot react to the training dynamics it encounters. We instead amortize learning rate selection across LoRA fine-tuning problems. We leave the base optimizer, Muon, untouched, so that its normalized update fixes the direction, and meta-learn only the global step size. Our method MuPACE is a small recurrent policy, meta-trained across a distribution of fine-tuning runs, that conditions on the observed trajectory and emits bounded adjustments to an interpretable analytic schedule. On 64 fine-tuning problems, MuPACE outperforms standard fixed schedules at a strong shared default on 64-78% of problems and is on par with the same schedules with per-task tuned peaks. Meta-trained only on models with up to 16 layers, MuPACE transfers one-shot to unseen tasks on models as deep as 64 layers, such as Qwen3-32B and Qwen3.8-27B, where a single run remains competitive with swept schedules.

Training coding agents typically updates the policy while keeping the harness fixed. During training, trajectories reveal flawed repair strategies and interaction problems that vary across tasks and change as the policy learns. We introduce HarnessAdapt-RL, which trains the policy and adapts its harness during reinforcement learning. For flawed repair strategies, a frozen teacher generates task-specific mechanism-level skills. For interaction problems recurring across tasks, a separate teacher model evolves tools and hooks, keeping each change only if it improves the observed pass rate or the observed trainable rate without lowering either. The policy trains with the adapted harness, and its new trajectories guide further adaptation. With a 9B base model, HarnessAdapt-RL improves pass@1 over GRPO by 9.8, 12.6, and 6.4 percentage points on SWE-bench Verified, Multilingual, and Pro, and outperforms GRPO with the human-designed OpenHands harness. Gains from mechanism-level skills persist after their removal at evaluation. Re-evolving tools and hooks as the policy changes outperforms freezing early versions. Removing them at deployment, in contrast, costs up to 7.7 percentage points. Training therefore produces a model-harness pair: the trained policy and the evolved tools and hooks, with skills removed at deployment.

LLMs increasingly operate over mixed-sensitivity contexts, where they must reason over information ranging from public to confidential while disclosing only what the user is authorised to receive. Existing approaches enforce access control as a binary decision: either fully answering or refusing, without regulating how much or what kind of information is revealed if access is granted, leading to over-refusal or unintended leakage. We introduce graduated disclosure, an access-control formulation in which the level of detail and sensitivity in a model's response is calibrated to the user's authorisation rather than treated as a grant-or-deny decision. We model graduated disclosure as a disclosure-aware generation problem and propose a reward-based optimisation framework that shapes responses toward level-appropriate abstractions while penalising unauthorised factual leakage, including paraphrases and approximations. We evaluate our approach on enterprise and clinical benchmarks and compare it with prompt-based, gating, and supervised baselines. Experiments show that our method substantially reduces information leakage while preserving correctness and utility. These results suggest that disclosure control must be integrated into the generation process itself to enable secure yet useful LLM deployments.

Multilingual evaluations often score a model's opening yes or no, a token that also carries model and language answering conventions. We ask when different ways of eliciting and reading an answer recover the same judgment, and when they measure an answer habit. We collect 1.1 million replies from eight models in six languages on up to 105 moral vignettes, varying the question's form, the answer constraint, and the reading of the reply, token or sentence, in both wordings (the FAR-6 battery). Story-level judgment profiles converge across the tested readouts and question forms, except for the yes/no token after a negative interrogative read under a preset mapping: on the same replies, token and sentence profiles agree at a median r = 0.67 across cells, while sentence profiles agree with free-sentence profiles at 0.94, a known linguistic variability. With each language's story consensus removed, a model's profile is nearest to its own in another language for 39 of 48 model-language cells under proposition labels and 15 under free sentences, against 5 expected by chance. Particle, position, and named-term preferences follow their own patterns. A model can keep its judgment and change its yes: scoring that token requires establishing which proposition it affirms.

Groundwater in Ghana's Densu River Basin remains a key domestic water source, but multi-metal contamination threatens children's health. Previous deterministic hazard index assessments provide limited treatment of uncertainty, co-exposure, and spatial variability. This study applies a Bayesian probabilistic framework for non-carcinogenic risk assessment using 94 groundwater samples and measurements of Fe, Mn, Ni, Pb, Cd, and As. A Bayesian regression model propagates parameter uncertainty and accounts for contaminant co-occurrence. Adaptive Hamiltonian Monte Carlo is used for posterior inference, while inverse distance weighting generates spatially continuous risk surfaces. Pb, As, and Cd dominate cumulative risk. Posterior predictive results indicate that approximately 31% of sampling sites have a probability greater than 0.5 of exceeding a hazard index (HI) of 1. The risk surface is highly heterogeneous, with a posterior mean of 0.372 and a standard deviation of 0.416. High-risk zones often reflect moderate co-exposure rather than extreme contamination by a single metal. The framework provides uncertainty-aware, spatially explicit information to support groundwater monitoring, risk mitigation, and public health planning in data-limited regions.

The Reinforcement Learning (RL) paradigm has achieved remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). However, when applied to long-horizon tasks, it faces severe challenges: i) Due to the high task complexity, the model struggles to sample enough successful trajectories, resulting in slow convergence. ii) Relying solely on sparse final rewards causes all steps to be updated indiscriminately, encouraging redundant step generation. When both the quantity and quality of successful trajectories become difficult to guarantee, the performance is severely constrained. This dilemma stems from the limitations: existing RL paradigms rely entirely on the reasoning capability of base models, lacking the ability to actively optimize trajectories. To address the above issues, we propose a novel trajectory evolution paradigm with hindsight credit assignment, termed TrajEvolve. TrajEvolve not only relies on the trajectory inferred by the model but also generates ``evolutionary trajectories'' by refining low-quality steps in historical trajectories, thereby exposing the policy to high-quality compositions that lie within the base model's support but are rarely sampled within practical training budgets. Furthermore, TrajEvolve reformulates step-level continuous credit assignment as a binary classification task, whether each step is necessary for task completion. This simplification not only reduces the learning difficulty but also enables step labels to be obtained through counterfactual verification, achieving precise identification of redundant steps. Extensive experiments on two datasets demonstrate the superiority of the proposed method.

Concept Bottleneck Models (CBMs) promise predictions that humans can inspect and correct through concepts such as "curved beak". Yet on CUB-200-2011, their standard benchmark, progress is measured like that of ordinary classifiers: by task and concept accuracy. Neither reveals whether the label is predicted from the exposed concept scores, or whether a concept is supported by visual evidence of that concept rather than by co-occurring features such as cues to the species as a whole. We audit where the accuracy of CBMs and Concept Embedding Models (CEMs) on CUB comes from. The CUB annotations are too noisy to explain it: class-level training labels agree with per-image annotations on only ~82% of image--concept pairs, and classifiers trained on per-image annotations reach only ~37% species accuracy. Leakage probes show that CEM predictions cannot be reconstructed from their concept scores; the missing class information sits in internal embeddings. We identify four design axes through which class information can enter concept scores---receptive field, shared versus concept-specific representations, norm-dependent scoring, and task supervision. An ablation shows that most of it enters through non-local features: restricting a ResNet50 backbone to local BagNet patches lowers the accuracy recoverable from concept scores from 72% to 43%, close to what ground-truth annotations support, while joint training raises it again at the cost of concept accuracy. Localization and occlusion show that locality makes grounding measurable but does not guarantee it: under class-level labels, local concept scores use the head as a proxy for body, wing, and tail concepts, and CEM concept scores barely respond to occluding their supporting part. Task and concept accuracy should thus support, not drive, the evaluation of CBMs, and given the noisy annotations, CUB accuracy gaps should not be overinterpreted.

We study universal online convex optimization with delayed feedback. We develop a single algorithm that simultaneously adapts to general convex, strongly convex, and exp-concave losses without prior knowledge of the curvature class or parameters. Our regret bounds match the best known dependence on the time horizon $T$ and delays up to doubly logarithmic factors: roughly $\sqrt{T + d_{\mathrm{tot}}}$ for convex losses and $σ_{\max}\log T$ rates with curved losses, where $d_{\mathrm{tot}}$ is the total cumulative delay and $σ_{\max}$ is the maximum number of simultaneously missing observations. Our main technical contribution is a delayed meta-learner with a second-order excess-loss guarantee, which enables curvature-adaptive aggregation of delayed base learners. In addition, our new exp-concave base learner yields a sharper regret bound which improves the state-of-the-art delay dependence by replacing the maximum delay with the potentially much smaller $σ_{\max}$, while retaining the complementary $\sqrt{d_{\mathrm{tot}}}$ dependence.

This paper presents a causal, training-free linker that allows a frozen streaming diarizer to track more speakers than it has output slots. Streaming end-to-end diarizers such as Streaming Sortformer emit a fixed number N of concurrent speaker slots. Above N speakers, they keep detecting speech but reuse slots across people, because N bounds how many speakers the model can express at once, not how many it will meet. The linker uses only the host's slot posteriors and the audio. It keeps the frames where one slot is active, splits each slot where the speaker changes, embeds each piece with an off-the-shelf speaker encoder, and matches it against a bank of speakers seen so far. It only relabels, so missed speech and false alarm are identical to the host's and every gain is a reduction in speaker confusion. On 1418 sessions from eight corpora, it reduces the 4-slot host's DER above its cap from 36.8% to 22.6%, and barely moves it below the cap. The same constants transfer to the host's 8-slot successor and to two streaming EEND hosts from a different model family. On courtroom hearings with up to 23 speakers it reduces cpWER from 82.7% to 24.4%.

Understanding how supervision signals shape semantic visual representations remains challenging, particularly in fine-grained, low-data settings. Standard evaluations of vision models rely on downstream accuracy, which often fails to capture structural differences in learned representations. We propose a compact explainability-driven framework for comparing encoders trained with classification-only and caption-generation supervision on CUB-200. Using a controlled two-stage protocol in which both encoders are ultimately optimized for the same classification task, we enable a direct comparison of the resulting representations. Saliency maps are generated using attention rollout and gradient-weighted rollout and evaluated with five metrics covering faithfulness (positive and negative perturbation) and localization (keypoint emphasis, background attribution, and Earth Mover's Distance). Across random seeds and explanation methods, caption supervision yields representations that are more sensitive to the removal of salient regions, better aligned with annotated object parts, and less reliant on background regions, whereas classification supervision produces representations that are slightly more robust to the removal of low-saliency regions. An ablation study on the Oxford-IIIT Pet dataset shows consistent trends across key metrics. These results demonstrate that the proposed framework makes supervision-induced representational differences observable and quantifiable beyond downstream accuracy alone. More broadly, it provides a systematic basis for assessing learned representations and comparing alternative training and transfer-learning strategies.

On-policy distillation (OPD) consolidates expert capabilities into a single student by matching teacher and student final-layer distributions on student-generated rollouts. We investigate whether depthwise prediction transitions, the changes in predictions between layers, can provide additional task-relevant supervision. Our diagnostic study finds that stronger teacher--student agreement in intermediate transitions is associated with higher answer correctness within questions, with correct responses exhibiting lower mean transition KL across all measured intermediate layer pairs. These observations motivate On-Policy Transition Distillation (OPTD), which augments final-layer distribution matching by aligning teacher and student layer-to-layer log-probability increments on student-generated rollouts. Evaluations across three backbones and eight benchmarks spanning mathematics, coding, and instruction following show that OPTD achieves the best or tied-best overall performance in both dual- and three-domain distillation. In the three-domain setting, OPTD recovers $\approx 90\%$ of the overall teacher--student capability gap, compared with $\approx 75\%$ for OPD.

To defend against backdoor attacks on neural networks, the defender must identify the verifier function that determines whether an input is a trigger, rather than rejecting individual triggers in isolation, since an attacker can generate new triggers satisfying the same verifier faster than they can be blocked. If a backdoor verifier could be installed with the asymmetry of a cryptographic authentication scheme, in which only the attacker can produce trigger inputs while the defender cannot reverse-engineer the verifier from observations, learning-based defense would be impossible in principle. We show that no such asymmetry can arise from poisoning-based installation, in which the only verifier-dependent input to training is a poisoned dataset. The argument is a reduction. A training procedure that installs a verifier from poisoned data, so that the backdoor fires on unseen triggers while benign behavior is preserved, is itself a learner for that verifier. Any verifier that is computationally hard to learn is therefore impossible to install through poisoning, and any verifier that poisoning can install is reconstructible by a white-box defender holding comparably many labeled trigger observations. A complementary argument shows that a secret key supplied to the model at inference is observable to the same defender. The result takes representative trigger observations as given and concerns when the backdoor fires rather than what it does. Experiments on instruction-tuned large language models confirm the dichotomy, with no configuration yielding a backdoor that is both effective for the attacker and irrecoverable for the defender.

We use Letter Positional Probabilities (LPPs) to examine the orthographic differences between simple and complex French words. We draw from a sample of 38,000 French words scored by reaction time on a lexical decision task. We take the 5000 highest reaction time words and the 5000 lowest as a proxy for high and low word complexity groups respectively. We create 18 variables based on the structure of the words, such as length and number of syllables. We also reduce each word to a string of binary values by one-hot encoding each letter according to the 42 characters of the standard French alphabet up to a length of 25 letters. Thus, each word is represented by 42 x 25 binary characters. We then create LPPs by looking at the average binary scores for simple as opposed to complex words for all letter/positions. We find that there is a high number of LPPs that are associated with word complexity. For example, we find that most accented letters are associated with simpler words. We examine the effects of structural and LPP variables in classifiers and find that word length alone enables a classification accuracy of 77.27%, while adding the binary variables from the first 6 letter positions provided an accuracy of 83.95%. We discuss possible explanations as to how complexity could be encoded in orthography.

High-quality training data are a major bottleneck in machine learning, motivating increasing use of artificially generated data. We study when a binary classifier trained on synthetic data transfers successfully to a real distribution, under a shared conditional label distribution. First, we derive finite-sample excess-error bounds when the real covariate distribution is absolutely continuous with respect to the synthetic distribution under a controlled density ratio. Second, we study the partially and fully singular case, where the real distribution has mass outside the synthetic one. If the hypothesis class is sufficiently expressive, synthetic risk alone cannot uniformly control real-world error: a classifier can be nearly optimal on synthetic data while approaching worst-case error on the uncovered real region, even when the distributions are arbitrarily close in Wasserstein distance. For function classes where hypotheses and the log-odds of the conditional label probability are Lipschitz, we recover a bound in terms of the probability that a real sample lies outside the reach of the synthetic data, restoring a transfer guarantee even in the fully singular case. Experiments confirm this in controlled low-dimensional settings and on 32,200 images generated as counterparts to real ImageNet images by four state-of-the-art generators: an adversarial classifier reaches 0.5% error on the synthetic domain but 99.25% error on the real domain, while ordinarily trained architectures show positively correlated synthetic and real test errors. Synthetic data should be budgeted at high sample counts and models validated on real data.

Practitioners use reinforcement learning methods for language models through training frameworks, so a method that a framework includes is one configuration change away for its users. We measure how often this happens. We trace seven open-source frameworks for RL post-training through their commit histories and link them to 6,939 screened arXiv papers from January 2025 to June 2026. Of 1,598 papers that propose a general RL training method, 37 reached at least one framework. Across five quarterly cohorts of papers, the number included within 180 days stayed between 5 and 7 while the cohorts grew from 72 to 386 papers. Compared with papers from the same quarters, included papers more often trained policies of 30B parameters or more, at 41 against 10 percent, and more often changed a coefficient, clip, mask or weight in the existing loss, at 29 against 9 percent. Both gaps hold without papers from the companies that maintain the frameworks, while code release was equally common in both groups. The frameworks' issues and pull requests show that few methods are turned away in public. More than nine in ten of the general methods no framework included have no thread there that cites or names them, and many proposals that do appear stay open without a reply. Two further records, the frameworks' issues and the technical reports of 13 large open-weight models, show that users rarely report training failures in public and that the reports state 8 of the 13 widely included methods. We release the data as a tracker that links training failures to framework implementations and lists each framework's contribution rules and earlier proposals.

The integration of Large Language Models and Reinforcement Learning offers a promising direction for complex sequential decision-making tasks, such as multi-hop question answering over knowledge graphs. While LLMs provide strong semantic understanding and reasoning capabilities, they often struggle with precise, multi-step navigation in structured environments. Conversely, RL agents excel at learning navigation policies but are hindered by the difficulty of designing semantically meaningful reward functions. This paper introduces a novel hybrid approach in which an LLM serves as a mentor to an RL agent during training by providing dense, semantically informed reward signals. The RL agent, built on a Transformer-based policy and trained with Group Relative Policy Optimization, learns to navigate a knowledge graph to retrieve relevant information. The LLM evaluates the semantic relevance of the agent's traversed paths, generating reward signals that align the agent's behavior with the information needs expressed in a query. We integrate this Retrieval Agent for Navigating Graphs with Semantic Evaluation Rewards (RANGER) into a complete question-answering pipeline and evaluate it on the open question answering datasets. Experimental results show that the proposed approach outperforms both a baseline retrieval-augmented system with classical graph search and one using an RL agent trained with heuristic rewards, achieving a significant increase in NLP metrics and LLM-as-a-Judge scores. Our findings demonstrate that deep integration of LLMs into the RL training loop yields more effective and semantically aligned retrieval policies for complex reasoning tasks.

Diffusion models generate samples through a multi-step denoising trajectory, yet condition every step on the same instruction, so the trajectory cannot be addressed at the level of individual operations. We introduce Dancing Diffusion, which assigns each span of timesteps a dedicated continuous condition block and thereby turns conditioning into a latent program. A compiler maps text features into this program, and a fixed pretrained generator executes its blocks over successive sampling intervals. The trajectory can then dance, keeping its executed history while its continuation is redirected. A shared prefix is executed once and reused, and a later operation is rewritten while content fixed by earlier operations is preserved. The compiler is trained through the images it produces, with visual feedback on both the requested change and the retained content. On SANA, reusing an executed prefix preserves markedly more of the scene than regenerating from noise at a similar strength of edit, learned conditions preserve more background than native prompt switching from the same saved state, and the learned program improves generation quality over native SANA. Later branches weaken the edit, which leaves semantic control over individual operations limited. By converting a single-instruction trajectory into a program of addressable operations, Dancing Diffusion lays a foundation for structured intermediate computation in diffusion models.

VLMs for medical image interpretation are typically evaluated on clean inputs, despite the progressive evidence loss that can occur in high stakes clinical settings. Existing robustness studies show that performance can degrade while confidence remains high, but do not establish when a previously reliable model-case pair ceases to behave reliably. We introduce a baseline-anchored framework that traces baseline-correct cases through progressive image degradation to identify an empirical evidence boundary, and define mirage as continued commitment at this boundary, regardless of correctness. Across three VLMs on three medical datasets, only MedGemma provides a sufficiently large baseline-correct pool for trajectory analysis. On VQA-RAD, the only dataset in which the question text varies, language-only answerability is associated with baseline correctness (AUROC 0.677) and with commitment at the boundary (AUROC 0.653 to 0.670). MedGemma's post-boundary behavior is heterogeneous, including direct mirage, correct recovery, mirage dissolution, consistent abstention, and oscillation, with mirage rates of 16.0 to 100.0% across dataset-modality pairs. Despite this behavioral heterogeneity, residual-stream probes decode commit-versus-abstain behavior at the boundary with AUROC of 0.895 to 0.994 under nested cross-validation. These results show that progressive evidence loss can produce persistent commitment that output entropy and abstention behavior do not consistently reflect, motivating boundary-aware evaluation, option-order controls, and improved abstention mechanisms for reliable medical foundation models operating under partial observation.

Tactile representation learning for dexterous manipulation often requires substantial robot-side data collection, limiting its scalability across different hands and sensors. We explore whether large-scale human hand-object interaction (HOI) data can provide a useful prior for robotic tactile learning. To this end, we propose HuTact, a framework that pretrains tactile-aware hand representations from human HOI data without requiring real tactile signals during pretraining. HuTact transfers the interaction structure learned from human hand motions to robotic tactile observations and can be further adapted with a small amount of unlabeled robot interaction data. We evaluate HuTact on tactile perception and manipulation tasks, including grasp prediction, slip detection, object classification, and tactile-conditioned policy learning. Across these tasks, human HOI pretraining improves data efficiency and provides a strong initialization for downstream learning. Notably, with HOI pretraining followed by midtraining on unlabeled robot play data, HuTact achieves 88.26% Macro-F1 on grasp prediction using only 25\% of the labeled downstream data, comparable to training from scratch on the full labeled dataset (88.23%). These results suggest that human interaction data can serve as a scalable complementary source for robotic tactile representation learning.

Telecom language models must combine broad networking knowledge with specialist behavior across routing, broadband services, access technologies, and operational diagnostics while remaining practical to deploy as a single model. We present an end-to-end framework that builds source-grounded synthetic training data from structured networking taxonomies and expert graphs, materializes 212,559 routed training records across fourteen networking routes, trains six frozen specialist LoRA teachers over one shared continued-pretrained backbone, and consolidates their behavior into one deployable student adapter using Routed Specialist On-Policy Distillation (RSOD). In RSOD, the committed student generates each trajectory, the route-assigned specialist teacher-forces the exact native token sequence, compact teacher hidden states are reconstructed through the shared frozen output head, and only the student LoRA is updated. A controlled reconstruction validation shows mean absolute logit error of 7.26 x 10^-6 with 100% top-1 agreement between native and reconstructed teacher logits. On four RTX 6000 Ada GPUs, the completed 2,000-step run produced 16,000 trajectories with zero observed student-version or token- ID mismatches. Under the primary route-conditioned evaluation over 2,850 held-out questions, the raw Meta-Llama-3.1-8B-Instruct model scores 12.8661/25 micro (13.0172/25 macro), the combined routed six-teacher system scores 21.7641/25 micro, and the consolidated RSOD student, released as Llama-CalixCLM v1, reaches 21.7112/25 micro (21.8006/25 macro) on the same five dimension LLM-judge rubric. Under a separate neutral deployment-like prompt, the same student reaches 20.0028/25 micro (20.0778/25 macro). The consolidated student remains within 0.0529 micro-score points of the combined routed-teacher system while retaining a numerically validated and auditable train-serve execution path.

Polyak--Ruppert averaging gives SGD and momentum methods the same asymptotic covariance. We characterize how fast finite runs reach it and how far they can be parallelized. For strongly convex quadratics with additive noise, we prove an online--offline bridge: the empirical risk minimizer on the samples of a tail average is one Newton step from the tail-averaged query point of any online procedure. For SGD, heavy ball, and Nesterov's method with constant momentum, this step is a boundary term at the two ends of the tail and splits the gap to the offline estimator into center, covariance, and shape errors. Necessary and sufficient conditions show that, on the least curved direction, the center error sets the pace: SGD needs $m\ggκ^2$ rounds but momentum only $m\ggκ$, where $κ$ is the condition number, so at sample size $N$ the critical batch size grows from $Θ(\sqrt{N}/κ)$ to $Θ(\sqrt{N/κ})$. The covariance error separates the momentum methods: heavy ball has relative variance excess $Θ(κ^{3/2}/m)$ along the stiffest direction, whereas Nesterov has relative error $O(\sqrtκ/m)$ in every direction. The shape error is method-independent by a Berry--Esseen bound uniform in $κ$ and batch size, so confidence intervals are valid when the other two errors vanish. We further examine how the conclusions change under heavy-tailed or state-dependent noise and clipping, and establish a first-order comparison with ERM for nonlinear Nesterov averaging under explicit conditions.

The growing scale of data and models in machine learning calls for efficient optimization methods. Distributed optimization addresses this challenge, with compression mechanisms helping to reduce communication costs. A classical tool for dimensionality reduction is the Johnson-Lindenstrauss (JL) transform, which is simple to implement and offers strong geometric guarantees. In particular, JL transforms approximately preserve inner products and pairwise distances between vectors with high probability. In this paper, we exploit this property in Byzantine-robust optimization, where worker messages are weighted according to their alignment with a trusted reference gradient. We further support our theoretical findings with experimental results.

Long-horizon video generation requires retaining an ever-expanding history, but conditioning on its growing token sequence is costly. Existing methods restrict or compress historical context, potentially losing information needed for future generation. This work proposes ReCam, which groups local spatiotemporal blocks of historical latents into vectors and projects them directly into the high-dimensional hidden space of a Diffusion Transformer. The model learns to use the resulting memory tokens through native self-attention, reducing token count while retaining block-level access to the full history. To mitigate temporal drift during autoregressive rollout, we further introduce dual-pointer denoising with separate noise schedules for recent context and target frames. Among the evaluated methods, ReCam achieves the highest VBench-Long overall score of 80.1 and competitive static-scene revisitation performance on R2M-Bench. It also leads on all four consistency metrics for off-screen subject reappearance, including clothing consistency of 4.09 versus 3.03 for the best baseline.

In earlier work by Jonasson, Steif and Zetterqvist (2023), fully connected deep neural networks with layers of width $n$, i.i.d.\ standard Gaussian weights, sign activations and no bias terms were shown to be annealed and quenched noise sensitive with respect to any joint probability distribution $\mathbb{Q}^{(n)}$ on the input $\boldsymbolω$ and its perturbation $\boldsymbolη$ satisfying a mild positive-correlation condition and $\mathbb{Q}^{(n)}(\boldsymbolη \in \{\boldsymbolω,-\boldsymbolω\}) \to 0$. Here we add independent Gaussian biases of the natural order and show that quenched noise sensitivity is retained over a broad range of depths, indeed under the slightly weaker assumption that $\mathbb{Q}^{(n)}(\boldsymbolω = \boldsymbolη) \to 0$. Thus, subject to the positive-correlation condition, quenched noise sensitivity holds under the minimal requirement that the perturbed input is asymptotically unlikely to coincide with the original input. At the same time, the annealed covariance remains bounded away from zero and is entirely accounted for by fluctuations in the network-dependent output means.

Reasoning-enabled language model agents can incur substantial computation before acting, although the reasoning needed varies across decisions. Existing stopping signals, such as confidence and intermediate-answer stability, can vary in effectiveness across reasoning models; signals obtained from hidden states or token probabilities also require internal access that may be unavailable or impractical. We introduce StreamStop, which learns model-specific stopping policies offline from prefix-level action agreement. A frozen prefix reader encodes available reasoning text for a lightweight stop head, allowing stopping without access to the reasoning model's internal signals. We construct and publicly release a dataset of agent reasoning trajectories with prefix-level action annotations. These annotations compare each prefix action with the full-reasoning reference action, enabling offline comparison of stopping rules and exact expected-return optimization of a discounted action-agreement objective. Conditioning the head on the discount factor allows the agreement-token trade-off to be adjusted at inference without retraining. In closed-loop evaluations on WebShop and ScienceWorld, StreamStop achieves higher observed success than most validation-selected fixed-budget settings while using fewer reasoning tokens than Full thinking. On WebShop with Qwen3.8-27B, it uses 42% fewer reasoning tokens per episode with similar observed task success.

Applications provide convenient features for users to complete daily tasks. However, different users may have different or even conflicting requirements for the same application, while the requirements of the same user may change over time. Personalizing applications produced by agent-to-app or UI-to-code approaches can require changes to the generated implementation. We propose Agent as App, a task in which users express their requirements in natural language and an agent creates a personalized application while continuing to supply its behavior during use. We implement LiveUIAgent, which generates a task-specific orchestration connecting the user interface, saved information, and service actions. A shared agent runtime renders and updates the interface while carrying out the requested operations. Revising the orchestration and interface while reusing that runtime provides a lightweight basis for personalization as requirements evolve. We introduce LiveUIBench, a benchmark with diverse requirements and ordered interactions across personal services. The empirical findings show that the evaluated models struggle to complete these workflows reliably. Applications can satisfy many local checks while leaving the overall task unfinished, and greater partial coverage can coincide with fewer completed workflows.

Neural networks are expressive models of nonlinear dynamics, but identifying such systems from closed-loop data raises two difficulties: the feedback loop correlates inputs with measurement noise, biasing standard open-loop estimators, and the identified model must be stabilized by the data-collection controller, so that its closed-loop behavior matches the true plant's. We derive a complete parameterization of all nonlinear systems stabilized by a given controller, generalizing the classical dual-Youla parameterization to the nonlinear setting. Building on this characterization, we propose Internal Controller Identification (ICI), a computationally tractable framework for identifying potentially unstable systems while ensuring closed-loop stability by construction; in the linear case, our indirect method reduces to an open-loop identification problem, for which standard consistency results apply. The framework accommodates contextual information, letting a single certified model identify a family of systems, and supports a PAC-Bayesian training objective that, when data are scarce and prior knowledge is available, bounds the identified operator's expected prediction loss. Experiments show that ICI identifies open-loop unstable plants where uncertified baselines diverge, outperforms direct methods under weak excitation and feedback-correlated noise, generalizes zero-shot to environments unseen during training, and, with limited data, avoids the overfitting of empirical risk minimization while providing a non-vacuous generalization bound.

Can humans and artificial intelligence cooperate such that their combination yields greater accuracy than either agent alone? This result, termed \emph{synergy} or \emph{complementarity}, is notoriously difficult to achieve, even when both agents possess significant independent accuracy. In this paper, we present a formal framework to model human-AI interaction through the lens of reliance patterns, which characterize the joint correctness structure of human and algorithmic decisions. We derive mathematically necessary and sufficient conditions for synergy, demonstrating that its emergence depends on the ability of either agent to identify and correct the errors of the other, what we call Human and Algorithmic Correction. Beyond its theoretical contribution, the proposed framework provides a foundation for the design and evaluation of knowledge-based decision support systems, where interaction policies must dynamically integrate heterogeneous sources of expertise. We demonstrate this through the analysis of seven clinical case studies. Our empirical results reveal a ``dominance trap,'' where a strong algorithmic advantage can paradoxically inhibit synergy by eroding human corrective intervention. We also show that synergy can persist even in the presence of cognitive biases, such as automation bias, provided there is functional de-correlation between human and AI errors. These results provide a formal foundation for key concepts in the human-AI systems literature and offer actionable insights for designing interaction-centric AI systems that foster complementarity.

Recent advances in large (vision) language models have enabled computer-use agents that automate digital workflows via fine-grained keyboard and mouse operations. However, such agents remain fragile in long-horizon settings, where errors accumulate and executed actions leave unintended side effects, making the ability to recover from failures critical. Yet, existing remedies remain inadequate: environment resets are infeasible in proprietary settings, action replay is prohibitively expensive over long horizons, and API-based rollback is limited to web environments, leaving model-based recovery largely unexplored. To address this gap, we present ReTrace, a scalable evaluation framework for systematically studying recovery behavior in computer-use agents. We formalize Transition-Consistent Recovery: identical downstream actions from the recovered and target states must yield identical outcomes. We operationalize this criterion through goal-oriented action replay following controlled error injection. Using ReTrace, we present the first systematic study of autonomous, model-based recovery, revealing that (i) ideal recovery can unlock model's latent task completion capabilities, (ii) current computer-use agents fail to exploit their latent recovery capabilities, and (iii) accurate model-based recovery yields measurable end-to-end gains. Together, these findings establish recovery as a fundamental capability for building reliable computer-use agents. Our code and dataset is available at https://github.com/BangProx/ReTRACE.

Split Computing (SC) is a distributed inference method designed for balancing computational load and reducing latency by partitioning a neural network into head and tail networks and deploying them across both a mobile device and a server, respectively. Due to the limited bandwidth and computational power of mobile devices, the head network is often replaced with a neural network that requires lower computational and communication resources. However, it is challenging to prepare a sufficient quantity and quality of labeled data on the mobile device side for training the replacement head network. Furthermore, the data prepared by mobile users often contains sensitive privacy information, making data aggregation difficult. In this paper, we propose a data-efficient head network training method for SC. This method utilizes federated learning and knowledge distillation to train the head network using a small amount of non-IID and unlabeled data from mobile devices, ensuring that the training datasets are not exposed. Moreover, by introducing personalization, we enhance the accuracy to levels comparable to or exceeding those of existing methods. The experimental evaluations demonstrate that, despite data constraints imposed only on our method, it achieves accuracy comparable to the baselines. Furthermore, it significantly reduces the communication requirements by 95% on mobile devices, with only a 3-5 percentage point drop in accuracy from the original model.

LLMs are increasingly integrated into software systems, making deployment and inference configuration a recurring software engineering task. Inference backends expose many options that affect model loading, generation, serving, and request processing. Yet little is known about which of these options developers explicitly configure in practice. In this paper, we present a large-scale empirical study of LLM inference configuration practices in open-source software using Hugging Face Transformers, vLLM, and the OpenAI API. We analyze 5,398 GitHub repositories using at least one of these backends and identify 345,117 explicitly configured option occurrences. Our results show that developers use only a small part of the available configuration space regularly. Although 98.8% of the documented backend-specific options appear at least once, 78.9% are configured in at most 10% of repositories using the corresponding backend. Configurations are also small, with a median of 2 options for Transformers and 4 for vLLM and OpenAI, but the combinations of options vary widely: 56.8-61.6% of multi-option signatures appear in only one repository. We also observe recurring option values as well as substantial variation in how values are chosen. Finally, configuration practices differ across application contexts, with option-adoption differences reaching 30.7 percentage points. These findings characterize the configuration surface developers make explicit rather than the importance or optimality of individual options. They provide empirical evidence for prioritizing future studies of inference configuration and for improving backend documentation and configuration support.

Subliminal learning is the phenomenon in which a model transmits its preferences to another model through data that do not explicitly express them. As models increasingly curate training data for other models, we ask whether this transfer can arise from choosing among near-duplicate images while keeping the source photographs and human-written captions fixed. For each photograph, a curator selects one of eight near-duplicate variants by caption likelihood, without target-specific prompts. Students trained on target-preferring curators' selections show higher mean target-versus-alternatives log-likelihood ratios in an image-free readout than students trained on neutral selections. Under a prespecified decision rule, this directional shift appears for an owl target in LLaVA-1.5, Qwen2.5-VL, Qwen3-VL, and Gemma 4, with curators and students initialized from the same base checkpoint, and target rotations move it toward camels, oak trees, and a political figure. The owl target-selected student also scores above all count-matched random-reselection students in three models and above 55 of 60 in LLaVA-1.5, and Qwen3-VL shows the shift again on a new source bank with newly trained curators. Its size depends on dataset size, learning rate, and training budget; greedy choices among the eight options move toward the target by 10.4 percentage points in Qwen3-VL and 0.5-1.3 points elsewhere. First-generation students also pass the owl preference to a second generation through captions without explicit owl mentions in LLaVA-1.5 and Qwen3-VL. Thus even a deduplication-style choice among near-identical copies can shift students' target-relative readouts with source photographs and captions fixed; the tested linear probes classify the selecting arm with accuracies up to 59.7% (chance 50%).

Automated respiratory sound classification enables continuous, non-invasive respiratory assessment, yet long-term deployment remains challenging due to domain shifts from evolving recording devices and acquisition periods. When privacy and governance regulations prohibit replaying historical data, sequential learning easily overwrites previously acquired knowledge. Preserving dedicated domain experts prevents parameter interference, but selecting the correct expert at inference is difficult under overlapping acoustic distributions and absent domain labels. To resolve this dilemma, we propose APRIL, an exemplar-free domain-incremental framework for RSC. Built on a shared frozen BEATs backbone, APRIL sequentially inherits and isolates low-rank experts to enable new-domain adaptation without altering earlier models. A multi-level prototype router retrieves the top-two candidate experts and dynamically selects the final model by integrating domain-routing posteriors with classification confidence, adaptively weighted by matching ambiguity. We establish standardized device- and temporal-incremental benchmarks on the ICBHI and SPRSound datasets across forward and reversed domain orders. Evaluated over five random seeds, APRIL consistently achieves the highest final-session average Score among compared continual learning approaches, reaching 64.4% and 65.4% on ICBHI, and 83.1% and 82.2% on SPRSound under forward and reversed sequences, respectively. Crucially, average forgetting remains under 1.0% across all protocols, demonstrating robust domain adaptation and long-term knowledge retention without data replay. Overall, APRIL offers a reliable and privacy-preserving paradigm for deploying continuous respiratory sound classification across real-world healthcare settings.

AI-assisted dementia detection usually requires multimodal fusion of diverse data, such as clinical tabular assessments and neuroimaging data. However, most existing approaches perform feature fusion in Euclidean space. This overlooks the geometric structure of representation, including hierarchical organization in clinical assessments and multi-scale anatomical patterns in neuroimaging data. Recent studies have proved hyperbolic space provides a natural geometry for modelling hierarchical and multi-scale structures due to its exponentially expanding space. Motivated by these, we propose HypTI, a hyperbolic geometry-aware multimodal fusion framework for dementia detection based on clinical tabular assessments and MRI scans. HypTI extends hyperbolic representation learning to an underexplored tabular-MRI setting, jointly modelling clinical tabular, MRI scans, and cross-modal interactions in hyperbolic space. Specifically, clinical tabular variables are represented through learnable Lorentz prototypes to capture hierarchical relationships among assessments, while MRI scans are encoded using ROI-guided multi-scale representation learning to capture anatomical patterns across brain regions and spatial scales in Lorentz space. The multimodal representations are integrated through an adaptive fusion module generates subject-specific scale and shift coefficients to dynamically modulate and integrate the multimodal embeddings, enabling effective patient-level tabular-image representation learning for dementia detection. Experiments on OASIS-3 dataset show that the proposed framework achieves strong performance compared with representative approaches.

Effective democratic deliberation is vital for civic engagement and the health of legislative institutions. However, assessing parliamentary discourse quality is traditionally done using expert-scored, theory-based codebooks (CBs), which are rigid, labor-intensive, and prone to biases. While Large Language Models (LLMs) offer a scalable alternative for both schema induction and discourse annotation, ensuring schema quality and mitigating models' inherent bias remains a clear challenge. In this work, we propose a complete protocol for LLM-based codebook generation and evaluation. Starting from the widely accepted discourse quality Index (DQI), we compare top-down (theory-based), bottom-up (data-driven), and hybrid approaches for generating context-adapted schemas. To evaluate these schemas, we define an extrinsic proxy measure, validated by manual annotations, alongside an intrinsic inter-model agreement (IMA) metric, which leverages differing model biases to signal overall reliability. Our experiments on 100 Israeli parliament transcripts reveal that inductive, data-driven schemas achieve higher IMA than top-down methods while maintaining validity, demonstrating their potential for reliable, large-scale analysis of political discourse.

Unmanned aerial vehicles (UAVs) enable fast, flexible last-mile and emergency delivery, but their routing remains challenging as the number of clients varies. Classical approaches based on mixed-integer programming and heuristics can scale poorly or require problem-specific design, motivating learning-based alternatives. Reinforcement learning is particularly promising, provided a single policy can generalize across problem sizes without retraining. Although attention-based Pointer Networks are inherently size-agnostic, their solution quality degrades under both upscaling and downscaling. Using the Capacitated Vehicle Routing Problem (CVRP) as a controlled proxy, with fixed capacity reflecting drone payload constraints, we investigate this degradation and identify three contributing factors: shifts in node density, attention dispersion, and dilution of the mean-pooled graph context. We systematically compare architecture-wide, encoder-side, and decoder-side modifications to address these limitations and propose a learned, permutation-invariant graph-context aggregation mechanism, implemented as a weighted average and as Set2Set. Results reveal distinct requirements for transfer in each direction: entropy-based attention scaling reduces the CVRP20-to-CVRP100 performance gap by 1.41 percentage points, while the proposed aggregation methods achieve the largest improvement on the more challenging CVRP100-to-CVRP20 transfer, approximately 0.9 percentage points. Exploiting this complementarity, we introduce an approach selector that chooses the best model for each training-target size pair. The selector reduces cost relative to the reference model across all nine seen train-test pairs and eight of nine pairs involving unseen sizes (CVRP35, CVRP75, and CVRP200), achieving improvements of up to 3.2 percent.

Learned surrogates for power-flow (PF) computation are usually motivated by speed, compared against a single CPU-based baseline. We show this comparison is fragile: across 7 open-source CPU-based solvers evaluated on identical inputs, throughput varies by roughly 2.5-3 orders of magnitude, so the choice of baseline alone can flip the conclusion. We build a benchmark spanning 9 physics-based solvers: 7 CPU-based and 2 GPU-based, including a new batched GPU solver, gumsim2grid - and 3 published ML surrogates, evaluated under a shared, independently implemented KCL-mismatch check across 15 grids from 14 to 70,000 buses. On the PF task itself, a sufficiently parallelized CPU solver or gpusim2grid matches or exceeds the throughput of the ML surrogates we test while remaining orders of magnitude more accurate; we find no evidence that these GNN-style surrogates are yet competitive for PF specifically, though prior work reports genuine gains on related but more combinatorial tasks such as optimal power flow or power grid control. Beyond this finding, we hope the benchmark suite, code and reference evaluator released at https://github.com/Grid2op/gpusim2grid helps close a broader reproducibility gap in this space, where batched GPU-based PF solvers are rarely open-sourced or evaluated against a common set of baselines.

Psycholinguistic studies explore why humans accept some generic statements but reject others by showing that these judgments depend on more than how common a property is. People consider several other factors such as how well it sets the kind apart from others or whether it is characteristic of the kind. Inspired by the linguistic and cognitive literature, we explore to what extent factors from these studies explain how large language models (LLMs) judge generics. Our results with twelve models on a psycholinguistic dataset and on naturally occurring generics show that models weigh how common a property is against how well it distinguishes the kind, and combine the two much as people do, though less clearly. Like people, they favor characteristic properties, but unlike the accounts of human judgment, they ignore striking ones. What matters most to LLMs is whether a statement is factually true, and much of the influence of human factors reaches their ratings through that judgment. In six 7-9B models, shifting the hidden states along each factor's direction moves the judgment as the factor predicts.

A single observation can admit several plausible futures, making uncertainty an essential part of world modelling. However, quantum generative models remain underexplored for this task. We introduce Quantum-JEPA, a latent-action JEPA framework that learns distributions of future representations using a conditional quantum Born prior. Circuit measurements generate binary latent actions, which a shared predictor maps to alternative future representations. We train the quantum prior through the quality and diversity of these predictions using a sampled energy objective. Our model learns distinct continuations on constructed multimodal tasks and captures latent-action correlations that support coordinated multi-agent futures. On D3IL, a benchmark of diverse human demonstrations, our trained models generate multiple obstacle-crossing routes from a common starting point without observed motion. On the controlled tasks, our classical and quantum prior implementations improve prediction validity and distributional fidelity over the evaluated variational baselines. Our quantum priors combine competitive predictive quality with a parameter-efficient sampling structure: they avoid an exponentially sized categorical output layer while supporting correlated latent actions. We further validate finite-shot parameter-shift training, including training with readout noise. Unlike most prior work, which focuses on simulation, we also verify the feasibility of inference with trained priors on a real QPU. These results establish a trainable quantum component for JEPA future prediction and a basis for exploring quantum-enhanced world models at larger scales.

Model compression makes large language models deployable, but today's evaluation tells an incomplete story. Standard practice measures retained task performance and stops there. This misses a critical fact: compression perturbs the distributed knowledge embedded across billions of weights, and task scores alone cannot reveal which underlying abilities survive and which quietly collapse. To solve this question, we introduce ACE-Compression, an ability-centric evaluation framework. Instead of asking only "does the compressed model still answer correctly?", we ask "which of its core abilities remain intact?" Our framework probes four deployment-facing dimensions: procedural reasoning and constraint execution, knowledge grounding and truthfulness, social cognition and affective understanding, and persona fidelity and stylistic consistency. Across more than 35 compressed checkpoints from six open-weight LLMs-covering quantization and pruning-we find that compression redistributes ability profiles. Coarse, local, and surface-level signals often persist, while fine-grained, long-tail, and context-dependent abilities become fragile. Quantization consistently outperforms pruning in retention, but the practical meaning of that gap depends entirely on which ability an application requires. Based on our findings, deployers can choose compression methods by the ability profile their application needs, not just by a leaderboard average. And algorithm designers gain diagnostic signals to build compression methods that preserve what matters. We believe that ACE-Compression turns compression from a black-box accuracy drop into an interpretable intervention-one we can measure, understand, and ultimately improve.

Enterprise agents answer questions over workplace records that evolve through team collaboration. As information is shared, corrected, and revised across documents, chats, and figures, multiple versions of the same information accumulate in the workspace. Reliable question answering therefore requires agents to reason over sufficient evidence, resolve cross-source and temporal conflicts, and recognize when the available evidence is insufficient to establish an answer. To systematically evaluate these capabilities, we introduce MMEnterprise-Bench, a benchmark comprising 1,613 questions grounded in evolving multimodal records from simulated collaboration projects, covering three categories: Evidence Reasoning, Conflict Resolution, and Unanswerability. We evaluate 10 proprietary and open-source multimodal LLMs and find that even the strongest models struggle when evidence contains cross-source or temporal conflicts or spans multiple images. Moreover, increasing the retrieval budget or employing a larger retriever improves evidence recall but does not consistently translate into higher answer accuracy. These findings highlight the persistent challenges of reasoning over evolving workplace records and establish MMEnterprise-Bench as a testbed for developing more reliable enterprise agents. Code and data are available at https://mlvlab.github.io/MMEnterprise-Bench/.

Negation remains a persistent challenge for language models across NLP tasks and modalities. We show that this weakness can undermine LLM safety guardrails through a structural vulnerability we term polarity blindness: the failure to recognize harmful intent after polarity-altering linguistic transformations. We study two forms of polarity inversion: grammatical negation and antonym replacement, and use synonym replacement as a polarity-preserving control. We further test whether an adversarial textual wrapper amplifies these failures. Using a 260-prompt benchmark across 10 state-of-the-art LLMs and 3 languages (English, Spanish, Czech), stratified by harmfulness potential and evaluated on a validated 0-5 harm scale, we find that polarity blindness affects most evaluated models, with mean harm scores reaching 4.12 for the most vulnerable model. The effect persists across typologically differing negation systems, suggesting that polarity blindness reflects a broader limitation in semantic robustness rather than a language- or grammar-specific failure.

Staged post-training of multimodal foundation models chains vision-language alignment, reasoning distillation, and domain tuning, yet how each phase trades capabilities against the others remains under-characterised. Tracking an 8.4B-parameter model from an un-fine-tuned baseline through three training phases on Japanese financial documents (evaluated on five text-input benchmarks), we find the phases interact non-additively: math-only distillation overwrites the financial knowledge acquired earlier even as procedural math skill survives, and the domain phase repairs more than it installs. A skip-phase ablation is best for domain knowledge but forfeits numerical-reasoning gains, so per-phase contributions are not decomposable and the best checkpoint is task-dependent, with implications for checkpoint retention and scheduling rather than added stages.

Large Language Models (LLMs) are increasingly used as research collaborators for hypothesis refinement, methodological planning, and technical analysis across multi-turn interactions. In this setting, a useful assistant must be accurate, methodologically justified, logically consistent, and token-efficient. Current alignment approaches rarely achieve all of these together. Instruction-tuned models gain conversational quality through verbosity, and we show that science-agnostic general collaborative rewards compress responses at the expense of scientific rigor. We propose CollabSciReM, a goal-conditioned reward framework that decomposes the session-level reward into conversational quality, scientific rigor, and structural efficiency. We also introduce SciQuality, a pairwise evaluation of multi-turn research conversations grounded in 300 held-out scientific papers and scored on six dimensions: methodological soundness, justification, logical consistency, factual correctness, novelty, and responsiveness. Fine-tuned on 8.62K synthetic research conversations, CollabSciReM is preferred on SciQuality over a science-agnostic reward baseline trained on the same data for all four Llama and Qwen models, including a 65.3% win rate at matched response length. For the 7B and 8B models, CollabSciReM outperforms the science-agnostic reward baseline on the primary metric in 8 of 10 single- and multi-turn benchmark settings. In a blind study, domain experts rated CollabSciReM significantly higher on novelty than the general-reward variant and reported fewer hallucinated responses than both the general-reward variant and the base model.

Direct preference optimisation (DPO) is widely used for language model post-training, but its performance depends strongly on the choice of its hyperparameters, such as the learning rate and the reference-policy regularisation parameter $β$. To better understand the DPO hyperparameter landscape, we conduct extensive experiments using a dense hyperparameter grid across different model scales from the same family. This allows us to analyse how the optimal DPO hyperparameters evolve with model size and the number of update steps, with intermediate checkpoints providing an inexpensive signal to accelerate the hyperparameter optimisation process. Furthermore, by evaluating our models on both static and dynamic benchmarks, we identify which benchmarks provide the strongest signal for hyperparameter selection. Finally, we investigate how different state-of-the-art hyperparameter optimisation strategies perform in this setting. Our experiments show that the optimal learning rate varies with model size and that hyperparameters are not transferable across training datasets. We also find that validation loss is a poor proxy for hyperparameter optimisation, as hyperparameters that minimize validation loss only weakly correlate with optimal hyperparameters for downstream benchmarks. Therefore, DPO hyperparameters need to be tuned specifically for the target model, dataset, and evaluation objective. We also show that short candidate training runs can substantially reduce the cost of this selection. We release all our training configurations and evaluation results to support reproducible analysis of selection quality and training cost.

We investigate associative creativity in descriptions of abstract visual stimuli by humans and multimodal large language models (MLLMs). While conceptual association is central to human communication, it remains unclear how humans and models adapt when visual stimuli become increasingly difficult to map onto familiar semantic concepts. We evaluate several state-of-the-art MLLMs and human participants along two dimensions of creativity, originality and effectiveness, using two datasets with controlled manipulations of difficulty. SilCC varies naming difficulty through progressively more complex combinations of abstract shapes, and PixCC varies difficulty through increasing levels of random visual noise. Results show that humans produce more diverse associative descriptions than MLLMs, including greater semantic distance, higher entropy, and more novel expressions relative to reference lexica. As difficulty increases, effectiveness declines in SilCC, and humans increasingly shift from holistic associations to descriptions of low-level visual properties in PixCC.

Industrial fluid transport systems are monitored through heterogeneous sensing streams, but robust fault diagnosis remains difficult because response signals vary with operating conditions and differ from physical variables in sampling rate, observation duration and data granularity. To address these challenges, we propose a physical-variable-informed deep learning framework (PhyNet) for robust fault diagnosis under real-world sampling heterogeneity. PhyNet treats physical variables as global operating constraints and response signals as local diagnostic evidence. Its physical-to-response cross-attention mechanism connects sequence-level physical-variable tokens with a learnable global response token, enabling interaction without strict temporal alignment. Across four benchmarks, including three real-world cavitation datasets and one public two-phase-flow dataset, PhyNet achieves the best performance across all evaluated metrics. Class-wise analyses, together with attention and masking analyses, show improved discrimination of ambiguous states and indicate that physical variables meaningfully constrain local response interpretation. These findings demonstrate a deployment-oriented approach to robust diagnosis in heterogeneous industrial environments.

Recent studies in harness optimization have shown substantial performance gains in language model agents without updating model weights. A well-designed harness more effectively leverages the underlying model's capabilities by shaping context, tool use, and execution workflows. However, naïve success rates alone do not fully characterize whether optimized harness yields more reliable success on tasks already within reach (elicitation) or a broader range of solvable tasks (expansion). This distinction is crucial for recursive self-improvement and evolutionary search, as expansion may provide a broader range of successful experiences to push the capability boundaries. Despite the growing significance, this distinction remains underexplored in harness optimization. We conduct a systematic analysis to study the effect of gains from harness optimization. Building on pass@k analyses in the reinforcement learning (RL) literature that compare models before and after RL training, we compare optimized harnesses with their pre-optimization counterparts. Specifically, we assess changes in reachable task coverage, compare harness optimization with model scaling and increased inference-time compute, and examine how gains vary with task difficulty. Across static prediction and dynamic tasks, we find that pass@1 advantages of harness over stronger interventions can reverse at larger attempt budgets and aggregate gains conceal shortfalls on hard tasks, underscoring the need for a fine-grained diagnosis.

Approximate solutions of partial differential equations (PDEs) on unstructured grids can be represented on a graph. Numerical optimization involving stiff evolutionary and static PDEs often requires solving a large system of equations, which can be treated as a constraint. In this work, we present a fully differentiable graph-based implicit solver that incorporates this constraint. We evaluate its performance across four key tasks: a) linear and nonlinear forward modeling of evolutionary and static problems, b) solving inverse problems to estimate equation coefficients, c) solving inverse problems to estimate the right-hand side, and d) learnable graph coarsening to accelerate forward modeling. The stability and differentiability of our solver enable significant reductions in the complexity of forward modeling while efficiently addressing related inverse problems. These capabilities make it a powerful tool for different physics and engineering applications, in particular for geophysics, where inverse problems are commonly solved in order to reduce the model uncertainty.

Code reuse in modern software ecosystems facilitates the spread of 1-day vulnerabilities, posing significant security risks. Determining if patches are present is crucial, but traditional methods suffer from limited usability and accuracy, often depending on a fragile compilation process. In our prior work, we introduced Lares, the first compilation-free framework for patch presence testing that combines large language models (LLMs) with SMT solvers. However, Lares treats the LLM and the SMT solver as two separate, parallel decision channels: the symbolic channel (Z3) succeeded in only 20% of code cases, and the remaining cases fell back to the neural channel (the LLM) for the final decision. Letting the LLM act as the final arbiter turns verification into an informal, non-deterministic process that is prone to hallucination errors, while the two channels never reinforce each other. To overcome these limitations, we propose SyLares, which reformulates compilation-free patch presence testing as neuro-symbolic reasoning. Instead of letting the LLM emit a final verdict, SyLares uses the LLM as a neural feature extractor that perceives heterogeneous source code and decompiled pseudocode and produces specification-compliant SMT-LIB inputs, and delegates the final, deterministic judgment to the symbolic Z3 solver. This neural-perceives, symbolic-decides coupling is realized by a Code2SMT reasoning engine that combines deterministic syntactic parsing with LLM-driven code normalization and symbol alignment, so that every case is reduced to a formal equivalence query rather than an opaque LLM guess.

Hamilton-Jacobi (HJ) reachability analysis offers formal safety guarantees for control systems, but training accurate neural certificates remains difficult, particularly when unsafe regions are hard to characterize. We propose a verification-guided training framework that addresses this by using probabilistic enumeration of safety constraints to detect forward-invariance violations, i.e., states where the learned certificate incorrectly classifies unsafe behavior. This produces a compact, targeted set of counterexamples that more precisely captures the failure boundary than inputs generated by standard adversarial attack methods. Crucially, using this representation to fine-tune a pretrained HJ value function yields larger certified regions and reduces certification time across all tested benchmarks. We further introduce a shielding mechanism based on the refined certificates to enforce safety constraints at deployment time. We evaluate the proposed safety filter in a realistic, physics-based Isaac Sim environment and through real-world autonomous racing experiments, achieving zero safety constraint violations compared to a nominal controller and a backup policy based on Linear Quadratic Regulation (LQR). These results demonstrate the effectiveness of the proposed shielding mechanism in ensuring safe operation under high-speed, safety-critical conditions.

Code translation converts source code from one programming language to another while preserving its functionality. Although Large Language Models (LLMs) have improved code translation, their outputs still contain hallucinations: generated code that conflicts with the source program, violates target-language facts, or introduces contextually inconsistent dependencies. These failures are especially harmful in translation because small semantic deviations can silently change program behavior. Existing hallucination studies mainly focus on natural-language generation or code generation, leaving the forms, distribution, and mitigation of hallucinations in cross-language code translation insufficiently characterized. This paper presents a systematic empirical study of LLM hallucinations in code translation. Across three benchmark datasets, six language pairs, and seven LLMs, we construct a taxonomy with three major categories: Semantic Consistency Conflicts, Factual Knowledge Conflicts, and Project Context Deviations. We manually annotate 13,919 hallucination cases and analyze how hallucination types vary across models, datasets, and language directions. Guided by the taxonomy, we further propose HaluCodeTran, a mitigation framework that combines BM25-based retrieval of relevant translation examples with taxonomy-aware reflective refinement. The study provides an execution-tested characterization of code-translation hallucinations and a practical baseline for building more reliable LLM-based code migration tools.

Prompt-based contextual ASR requires selecting a compact hotword subset from large dynamic inventories, yet existing pre-selection methods rely either on discrete ASR hypotheses that may discard acoustic evidence or on separately trained retrieval models. We propose Viterbi-guided Selection of Hotwords (VISH), a training-free, plug-and-play framework that directly verifies candidate-specific acoustic support over frame-level CTC posteriors. VISH formulates pre-selection as constrained posterior search; M-VISH makes large-list verification scalable through frame-synchronous shared-prefix Viterbi search with bounded top-K pruning, while Bi-MVISH exploits complementary backward evidence for robust verification. Across five benchmarks, M-VISH achieves up to 11.7% absolute improvement in Best-F1 over a strong learned-retrieval baseline and reduces candidate-search latency by 75.2% at 2,000 candidates relative to hypothesis-based filtering. These results demonstrate the effectiveness of posterior-level verification for accurate and scalable hotword pre-selection without additional model training.

The increasing use of large language models for code generation calls for methods that distinguish purely human-written code from AI-involved code: files in which any part, however small, was generated or substantially modified by a model. Beyond flagging such files, we aim to localize the lines or blocks attributable to AI. Existing approaches compress a file into a single score, from surface stylistic features or a global code representation, which averages away the evidence when only a few blocks are AI-involved. We propose a framework that models AI authorship at the token level and predicts at the line, block, and file levels. A frozen code language model is equipped with two low-rank adapters, trained separately on machine-written and human-written files; their per-token log-likelihood ratio provides local evidence of AI authorship. Lightweight heads aggregate this evidence within lines and blocks, using the surrounding code context, and their local predictions are aggregated into a file-level decision. We train on mixed-authorship files built from human originals and their AI-generated or AI-modified counterparts, including continuation and fill-in-the-middle scenarios. On an internal test set of 11,773 AI-involved files across six generation regimes against 2,541 human files, our method reaches 72.4 macro F1 at a 1% FPR operating point. With no adaptation, it reaches 62.5 line-level F1 on HybridCodeAuthorship, an external human-AI co-writing benchmark with no training split, above the 56 its own paper reports as the best line-level result. On 19,445 human-written files from repositories created before 2022, the false-positive rate stays between 0.55% and 0.81% under the transported 1% threshold, so the operating point holds on real human code.

Cavitation diagnosis in industrial pipeline systems is challenging because acoustic responses vary substantially with operating conditions, while similar time-frequency patterns may correspond to different cavitation states. To address this issue, we propose a physical operating-condition-guided vision-language matching (POC-VLM) for cavitation diagnosis. POC-VLM first constructs physical operating-condition-aware semantic descriptions by explicitly incorporating sample-specific operating variables into candidate cavitation texts. A time-frequency-conditioned text adaptation module then dynamically aligns these physical semantics with the observed acoustic response, followed by global-local vision-language matching to capture both holistic and localized discriminative patterns. In parallel, a K-nearest neighbours-based non-parametric physical operating-condition prior estimates the physical plausibility of each candidate state and is fused with the vision-language semantic evidence at the decision level. Experiments on three real-world cavitation datasets show that POC-VLM consistently outperforms representative data-driven, multimodal and domain-specific diagnostic methods. Further analyses verify the complementary roles of vision-language semantic matching and physical operating-condition priors, while demonstrating robust performance under varying sampling rates, noise levels and observation durations.

A published laboratory number can be verified only when its raw file and extraction method are known. This paper presents the value-level provenance architecture of APS ELN, an electronic laboratory notebook used by both researchers and language-model agents. The architecture was motivated by a production audit in a materials laboratory covering projects performed over almost 5 years. Tracing eight baseline studies from sample to stored value showed that the initial ELN structure preserved the link from sample to raw file. Still, the link from raw file to stored value was lost, mainly due to human factors: manual entries carried no file reference or extractor version. To close this gap, the database design and workflow were reworked so that every stored value records its source file, the data file reader/viewer (DFR) plugin and version, any interactive analysis selection, or the server-side formula for derived quantities. An automated drift check re-runs each DFR on the stored files. In two end-to-end validation studies recorded by agents in X-ray reflectivity and tribology, all stored values retained verifiable provenance with zero drift. A comparison with five other platforms shows that none binds individual stored numbers to both their source files and extractor versions.

This paper presents UvA-MT's submission to the WMT 2026 General Machine Translation shared task, competing in the unconstrained track across all 23 translation directions. This year, we fully leverage the test-time methods of Large Language Models (LLMs) for machine translation. Specifically, (1) we use sequential sampling with a refinement prompt to generate a pool of translation candidates for each source segment, a strategy shown to outperform traditional parallel sampling; and (2) we employ LLMs as pairwise judges to select the best candidate via a round-robin voting mechanism, with offline evaluation on the WMT25 benchmark showing a very clear improvement over point-wise best-of-N selection. Combined, these two strategies form our WMT26 submission.

This paper shows that JEPA-style latent world models (LWMs) can learn across diverse domains and data sources by adopting the language-conditioned interface of video generation models. We establish this approach as an effective and efficient basis for planning with imagination and downstream policy fine-tuning. To this end, we introduce LaLa-World, a Language-conditioned Latent World model that learns from simulation and robot trajectories together with broader image and video data. Two designs make this work. First, we describe both numerical actions and visual changes in language, enabling joint training of a single future predictor across heterogeneous datasets. Second, we build the predictor on a language model, allowing it to interpret language conditions while predicting future latent states. The resulting predictor achieves higher average MPC planning performance than existing LWMs trained separately for each domain or task. Adding image and video data further improves average planning performance and robustness to action phrasing. In a matched comparison with an RAEv2-based video generation model baseline, LaLa-World reaches comparable planning performance with $6\times$ less training cost and $13.9\times$ faster planning. Beyond planning, the pretrained LaLa-World serves as an initialization for LaLa-Policy, which jointly predicts actions and future visual representations. LaLa-Policy performs competitively with Cosmos Policy on LIBERO and RoboCasa, with up to $5.5\times$ faster inference.

Scaling the training of Large Language Model (LLM) agents typically requires a large collection of executable environments, whose construction, maintenance, and operation constrain the expansion of training experience. Existing offline trajectories offer an opportunity to reuse interaction experience, yet effectively refining the current policy without revisiting these environments remains challenging: supervised fine-tuning (SFT) can lead to catastrophic forgetting, while online reinforcement learning requires real-time interaction with the environment. To address this challenge, we propose Step-Advantage Policy Distillation (SAPD), a method for learning from offline trajectories using relative step advantages. SAPD decomposes multi-turn trajectories into individual decision contexts for on-policy distillation, allowing the student to generate responses conditioned on recorded histories, and introduces relative step advantage signals to stabilize training. The entire process requires neither new environment observations nor additional multi-turn interactions. Experiments on ALFWorld, WebShop, and Search-based QA show that SAPD improves performance across different teacher configurations, mitigates catastrophic forgetting compared with SFT, and runs at least $4.7\times$ faster per training step than online OPD. We will release our code soon.

Recent agents have demonstrated strong capabilities in generating images from textual prompts, motivating their utilization for scientific diagram generation. Scientific diagrams naturally span diverse types, each characterized by distinct visual structures and content organization. Consequently, during peer review and other scientific evaluation, human experts naturally consider type-specific criteria when assessing their quality. However, existing VLM-based evaluation benchmarks fail to reflect the above human evaluation practices for two main reasons: 1) benchmark datasets typically lack a systematic categorization of different diagram types and thus adopt the same generic criteria for all diagrams, and 2) their evaluation criteria are often designed heuristically, without explicit alignment to human evaluation preferences. Motivated by these limitations, we introduce SciDiagBench, which consists of a benchmark dataset that systematically categorizes scientific diagrams into 11 types and a human preference-guided evaluation method. Specifically, SciDiagBench contains 165 scientific diagrams across 10 broad domains, with types including graphical abstracts, technical roadmaps, and illustrative cases. Based on this dataset, SciDiagBench develops an evaluation method that leverages a preference-guided calibration process to incorporate human preferences into the generation of type-specific criteria. Ultimately, SciDiagBench enables more reliable automated evaluation that better aligns with human evaluation across diverse scientific diagrams.

Online Continual Learning (OCL) requires models to learn from non-stationary data streams under a strict single-pass constraint, making them highly susceptible to catastrophic forgetting. While existing studies have explored various strategies from data, architecture, and optimization perspectives, the optimizer often directly uses generic approaches such as SGD and AdamW. In this work, we reveal a new link between optimization dynamics and OCL needs by recasting Polyak-Ruppert averaging as the engine of "plasticity" and Primal averaging as the anchor of "stability". This inspires our novel optimizer, SPIN (Stability-Plasticity INterpolation), which explicitly decouples plasticity and stability by interpolating between a fast-moving "plasticity" sequence and an adaptive "stability" sequence. We theoretically demonstrate that SPIN implicitly employs an inverse Hessian approximation, providing crucial Tikhonov regularization to damp noisy, single-sample Hessian estimates and mitigate forgetting. More importantly, SPIN can be seamlessly integrated into existing OCL methods by taking resulted gradients as input and replacing standard optimizers. Extensive experiments show that using SPIN instead of standard optimizers yields consistent and significant performance gains across multiple OCL approaches, including the challenging rehearsal-free setting and under strict memory constraints.

Continuous fuzzing is used for vulnerability detection in Open Source Software (OSS). Large-scale services such as Google's OSS-Fuzz fuzz over 1,300 projects and have revealed more than 10,000 vulnerabilities. Despite this, a considerable time lag often exists between when a vulnerability is introduced and when it is detected (i.e., the window of exposure). Given the computational cost of additional fuzzing, we aim to allocate fuzzing resources efficiently using vulnerability-risk estimates. We propose RAISE (Risk-Aware Intensive Supplemental Execution Framework), a Just-In-Time (JIT) framework that prioritizes additional fuzzing for high-risk project-days. RAISE performs binary classification at the project-day granularity to estimate whether a vulnerability is introduced on a given day. We construct a dataset combining vulnerability-introducing commits from the OSV database with daily coverage and change metrics from OSS-Fuzz, covering 15 C/C++ projects and 15,210 project-days, of which 347 (2.28%) are vulnerability-introducing. In a cross-project validation setting, our best model, a Random Forest combining VCCFinder-based static features with dynamic fuzzing features, achieves a ROC-AUC of 0.74, substantially outperforming a random baseline; coverage metrics, led by PatchCoverage, are the most important features. In a simulation over 968 vulnerabilities comparing five allocation strategies, Simple-Regression was significantly more cost-efficient than the prediction-agnostic heuristics, spending 91.60 additional fuzzing sessions per detected vulnerability, roughly half the cost of the Median heuristic (171.89). RAISE shows that JIT vulnerability prediction with coverage-aware fuzzing features can effectively guide risk-aware allocation of continuous fuzzing resources.

Despite decades of progress in model-based methods and tools for embedded real-time systems, there is still no integrated environment that supports the entire development lifecycle. This paper presents MIMOS-Tools, a comprehensive toolchain spanning the key phases of development, including modeling, simulation, formal verification, real-time scheduling, schedulability and end-to-end latency analysis, and code generation for heterogeneous multicore platforms. All phases are grounded in a single, coherent, and deterministic model of computation, the MIMOS model, for which we present a formal operational semantics and establish both functional and timing determinism. This semantic foundation ensures that properties verified at the model level are preserved throughout the development process and in the generated implementation. A distinguishing feature of MIMOS-Tools is its support for not only deterministic but also modular, compositional design. Unlike synchronous design environments such as SCADE and Simulink, which rely on a global base clock and offline static scheduling, MIMOS-Tools supports time-triggered asynchronous communication together with flexible real-time scheduling of computation. This removes the tight global coupling of the synchronous paradigm: components can be added, modified, or updated without recomputing a system-wide schedule, enabling scalable and incremental design and dynamic updates throughout the system lifecycle. However, MIMOS-Tools also supports Simulink models when the deadlines of tasks are set to zero to emulate the zero-time semantics of synchronous design. The toolchain is available at uu-mimos.github.io.

LLM routing aims to select a capable model at low cost, but specialized routers can require updates as models and tasks change. We propose COMPASS, an agentic router in which an LLM follows a routing recipe, gathers performance and cost evidence through tools, and optionally inspects candidate answers before selecting a model. Our evaluations on LLMRouterBench and DeepSWE show that agentic routing can improve accuracy and lower model-call cost relative to specialized and trained routing baselines, without routing-specific training. On LLMRouterBench, COMPASS reaches 69.85% accuracy at \$139.57, compared with RLCascadeRouter's 69.14% at \$180.22. On DeepSWE under fixed five-fold cross-validation, it reaches 78.1% at \$215, compared with 74.1% at \$500 for the best single model. COMPASS also supports routing across benchmarks and adapts to an expanded model pool by reading new evaluation records, without changing the router's weights or prompt.

Word error rate (WER) assigns equal cost to lexical edits despite their different semantic consequences, while overall semantic scores cannot identify which individual errors alter the intended meaning. We formulate ASR semantic evaluation as context-aware, segment-level consequence assessment and propose Semantic Segment Mapping and Assessment of Recognition Transcripts with Contextual Rubrics (SMART), a framework for traceable semantic evaluation and targeted ASR correction. SMART maps reference-hypothesis discrepancies into coherent segments and applies a structured contextual rubric to assign explicit consequence labels. These labels support Semantic Impact Error Rate (SIER), which quantifies consequential errors at utterance, chunk, and lexical levels. Against majority-vote human labels, SMART achieves 0.843 F1 versus 0.745 for direct LLM judgment. Evaluation across multiple LLMs further demonstrates that SMART is relatively insensitive to the choice of evaluation model, while MER and SIER reveal different aspects of ASR quality. A separate target-reference-free rubric guides correction using the target hypothesis and surrounding ASR hypotheses, outperforming a matched w/o Rubric baseline on both datasets, with further absolute reductions of 0.85-0.98 percentage points in MER and 1.18-1.65 percentage points in SIER_lex.

Every prompt sent to a hosted language model is readable by whoever runs it. We study whether the prompt can remain on the client while the host receives transformed vectors that stand in for it. A language model converts text to vectors exactly once, with a public lookup table; we perform that step locally and pass the vectors through a small trained network (55.6M parameters) before they leave the perimeter. The host runs a frozen 27B model (Qwen3.8-27B) on the transformed vectors and never receives text or token ids. We train the transform against the frozen model as a fixed judge, combining a forward-KL utility term with an auxiliary term that pushes each vector away from its original direction. On 1,490 MMLU questions held out from training, the transform costs 1.5 points of accuracy (80.7% vs. 82.3%), breaking 68 of the 1,226 questions the unmodified model answers correctly (5.5%), while 30.1% of content tokens no longer decode to their original token under a nearest-neighbour lookup.

Retrieval-Augmented Generation (RAG) has become the dominant paradigm for grounding the outputs of large language models (LLM) in external knowledge. However, practitioners are faced with a vast configuration space comprising chunking, embedding, retrieval, reranking, prompting and generation models, and there is limited guidance on how these components interact. We present RAG-NIMBF, a modular benchmarking framework that attaches to existing RAG pipelines as a non-invasive evaluation layer. It evaluates the configuration space through controlled ablation with end-to-end experiment tracking. The evaluation combines LLM-based answer quality scores, retrieval correctness measures, and runtime indicators. This study validates the framework across multiple generation models on the SQuAD dataset. This process reveals non-obvious interactions and demonstrates how the experiments are tracked. The findings indicate that the retrieval balance of relevance and diversity substantially enhances the relevance and coverage of retrieved context in comparison to a conventional similarity search, with MMR raising nDCG@5 by +4 to +11 percentage points across the swept models. More detailed prompt instructions meaningfully increase how faithfully generated answers stick to the retrieved evidence, especially for capable and reasoning models (+10 to +13pp on Qwen3.6-27B, GPT-4o-mini and Qwen3.5-think, but negligible on the 20B generator). Overall, RAG-NIMBF transforms the configuration space of RAG into reproducible, comparable experiments. By attaching to existing production pipelines in a non-invasive way, it enables practitioners to systematically evaluate and improve deployed systems without disrupting existing workflows. This replaces ad hoc choices with evidence-based decisions.

Speculative decoding accelerates autoregressive (AR) language models by letting a small model draft tokens serially and a large model verify them in parallel. We reverse these roles. Starting from a standard AR checkpoint, we fine-tune the large model into a proposer that predicts a block of future tokens in a single forward pass, then use one additional forward pass to verify the proposed block. This gives ProVe, Proposer-Verifier decoding. The verifier can be a smaller AR model or the proposer itself; the latter gives ProVe-lossless, which reproduces the proposer's greedy AR decoding. Block proposals already provide substantial room for acceleration: with a block size of 16, the proposed block has an average AR-matching prefix of 6.6 tokens on GSM8K. Across six benchmarks with a 32B proposer, ProVe-lossless reaches 2.5x average AR throughput without training a separate draft head. With a small 3B verifier, ProVe reaches 2.7x average throughput at 2.6 accuracy points below AR on the six-task average, providing a controllable tradeoff between speed and accuracy.

Multi-agent systems rely on communication to complete tasks, and latent communication passes continuous representations to reduce text generation and re-encoding. However, multi-agent systems must reconfigure agent models, whereas existing latent-communication methods are built for fixed model combinations, so their interfaces may fail after replacement. Our pilot study confirms this interface failure and identifies the requirement for interface reuse after model replacement. We present LatBridge, which connects agent models through a shared Bridge and separate interfaces under a shared latent-communication protocol. After agent model replacement, LatBridge reuses unchanged interfaces, selects matching low-rank interfaces, and combines them through the Bridge, training only missing ones with all models frozen. The Bridge processes source representations in the basic mode and constructs target input without autoregressive generation in the accelerated mode. Experiments support effective communication after model replacement, with mean basic-mode accuracies of 14.8% on MATH and 27.3% on ALFWorld next-action prediction across eight small-model configurations. Analyses examine the factors supporting task accuracy and communication efficiency.

As large language models (LLMs) grow more powerful, their scale makes them costly to serve. Depth pruning offers a practical route to efficiency by directly reducing memory and latency without hardware-specific support. However, we observe that transformer layers exhibit markedly different task-dependent layer contributions, whereas existing depth-pruning methods ignore this heterogeneity and thus tend to optimize for a narrow domain or benchmark rather than preserve broad capability. This suggests that LLM depth pruning should be formulated as a multi-objective optimization problem rather than a single-score ranking problem. We propose SliMOO, an interpretable multi-objective evolutionary search framework for LLM depth pruning. SliMOO first uses single-layer removal probing to construct a task-layer importance map, which provides an interpretable view of task-shared and task-specific layers and serves as a search prior. It then performs a Pareto-aware evolutionary search for pruning masks, where candidates are generated under the guidance of the task-layer prior, scored by their deviation from the dense model on each task, and retained through NSGA-II environmental selection. Across the Llama-3.1 and Qwen3 model families, SliMOO consistently achieves stronger multi-task trade-offs than rule-based, greedy, and scalarized baselines. For example, on Llama-3.1-70B at 25% sparsity, it improves the math-domain average from 24.33% to 34.40% and the code-domain average from 35.03% to 39.14%. Code will be made available at https://github.com/AMD-AGI/SliMOO.

Maximum-likelihood (ML) decoding of linear block codes is NP-hard, so neural decoders are evaluated by how closely they approach it. We observe that proving a candidate optimal can be far easier than finding it, and build a decoder that does both. We cast decoding as a search over an absorbing-state discrete diffusion model, whose sequential commitment of coordinates turns generation into a tree, and run a probability beam search over it. A dual bound from a linear-programming relaxation and a distance-based test, both derived from the code, then certify per frame that the current candidate is the global ML solution; because both are theorems, certified frames incur no accuracy-computation trade-off. Across $15$ codes and five operating points we certify up to $99.9\%$ of frames, the two tests dividing the space as their derivations predict, one strong for sparse parity-check matrices and the other for codes with a large minimum distance. Certification also behaves as a per-frame reliability signal, with $96.7$-$100\%$ of all frame errors falling on uncertified frames. Against an exact ML reference at $5$ dB the decoder is within one standard deviation on three of the four codes where the reference is uncensored, and against four neural baselines it attains the best bit-error rate in $71$ of $75$ cells.

6G will be designed faster than it can be built: AI-driven search already proposes physical-layer algorithms at machine speed, while carrying a single one onto embedded baseband silicon remains specialist work. That silicon imposes hard memory and numeric limits, but also exposes cost structures a GPU lacks, and a design flow that never runs on the target sees neither. We present an agentic workflow in which an LLM coding agent implements algorithms on a commercial baseband SoC, optimizes them down to agent-written vector assembly, and validates them on silicon. The agent is guided by a knowledge base of measured platform facts, checked at every step against evaluators it is instructed never to edit, and metered per phase. We showcase the workflow on an AI-discovered orthogonal time frequency space (OTFS) detector portfolio on the open 6G4All platform. It yields twelve kernels that run cycle-exactly on silicon, their optimized versions at 18k-405k cycles (30-660 $μ$s) per frame. It also yields a new detector shaped by the chip's measured costs that needs 13.9% fewer cycles than the port of the discovered detector while matching its error rate, with a measured edge our evaluation cannot statistically resolve, a structure that a hardware-blind search objective had no reason to find. What is not specific to OTFS is the workflow: an LLM agent exploring an implementation space under real hardware limits, with the cost of every choice measured rather than assumed.

Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly due to imperfect land surface representations in Earth system models. We propose a data-driven framework for high-resolution land-use reconstruction, treated as a semantic-segmentation task, that combines coarse-resolution scenario data and climate classification with static geophysical features to produce annual land-use maps at 1 km resolution. We introduce TerraNostra-Bench, a change-aware evaluation protocol that separates spatial downscaling from autoregressive temporal rollout and scores models on the rare land-use-change pixels rather than the dominant static background. Using it, we compare convolutional (U-Net) and transformer (Swin) backbones in autoregressive and non-autoregressive schemes against persistence, linear, and shallow convolutional baselines. The autoregressive prior is the most informative input yet accumulates error over long rollouts: autoregressive models lose 5.7-7.2 accuracy points over 13 years, against 2.2-2.8 for non-autoregressive models. This trade-off is governed by the autoregressive choice rather than the backbone. It can be tuned by an inference-time prior-trust threshold that requires no retraining and by the training-time weighting of reconstruction and change objectives; the two levers trace complementary frontiers, with calibration recovering up to 10 accuracy points while keeping change sensitivity above the persistence floor. Extrapolating beyond the training window, backward to 1955-1959 and forward to 2015-2019 under scenario forcing, the models retain most of their in-window skill and show the same trade-off, indicating it is intrinsic to the design. This work provides the methodological foundation and evaluation surface for a forthcoming temporally continuous land-use product.

Large language models (LLMs) exhibit both knowledge of language, rendered by, for example, grammatical judgment ability, and abilities to perform tasks through language, such as math-word problem solving. Motivated by neurolinguistic studies of their neural dissociation in human brains, in this study, we investigate the overlap of their circuits in large language models (LLMs). Specifically, we examine the overlap and causal effects between the circuits extracted for linguistic and arithmetic tasks, using controlled materials and circuit discovery methods from the mechanistic interpretability field. Our experiments with five LLMs show that circuit similarity is mostly explained by task domains (linguistic vs. arithmetic), rather than by their input and output vocabularies or by the syntactic similarities between natural-language sentences and arithmetic expressions. This finding suggests that LLMs may process linguistic and arithmetic information using largely separate neural circuits, highlighting an ability-based modular organization of circuits within these models.

Empirical scaling laws relate validation loss to model size, training exposure and compute, providing a quantitative basis for allocating resources. This technical note explains these relationships through three linked analyses. Geometric approximation connects spatial resolution, intrinsic dimension and smoothness to function error; the loss function determines how that error transfers to task risk. Spectral analysis combines feature visibility, covariance eigenvalues and target alignment to describe how additional data reduce unresolved predictive mass. Constrained optimization then gives the parameter-data balance and the training-inference trade-off. We also examine emergence and grokking, separating scoring thresholds from learning dynamics, and test-time scaling, separating candidate coverage from selected-answer accuracy. Applications to vision and molecular prediction show how domain-specific coordinates change the resource comparison. An $\mathrm{SE}(3)$ case study quantifies the dimension dependence of quotient-space approximation gains and compares optimized compute frontiers, alongside separate analyses of conformational coverage and correlated-sample precision.

Speech separation (SS) has witnessed substantial progress in recent years with the advent of neural network-based methods, achieving notable improvements on signal-level evaluation metrics. Existing speech separation systems are mostly optimized with signal-level objectives, and often yield outputs that are acceptable in waveform metrics but suboptimal for intelligibility-sensitive downstream tasks. To address this gap, we propose SLM-SS, a token-based generative speech separation framework that models multi-talker mixtures with speech language models. We first discretize clean speech into multi-codebook token sequences with neural codecs and serialize multi-speaker targets via Serialized Output Training (SOT). An encoder-decoder autoregressive speech language model predicts the zero-order sequence to establish speaker ordering and coarse linguistic content, and a non-autoregressive module then refines higher-order residual codebooks for efficient reconstruction. We further introduce a per-talker non-autoregressive strategy that simplifies high-order modeling and improves the stability. On Libri2Mix, SLM-SS with per-talker NAR achieves $6.92\%$ WER, approaching codec-only reconstruction. The WavTokenizer variant eliminates NAR refinement and achieves the highest mean subjective quality score among evaluated separation systems. Its UTMOS and DNSMOS advantages over discriminative baselines persist on unseen noisy and Chinese mixtures without fine-tuning.

Vulnerability benchmarks and software engineering research depend on rebuilding historical software. Yet builds rely on external resources, such as base images and packages, that change independently of the source code. In this study, we examine this environment decay on 14,159 OSS-Fuzz cases. We find that 69.7% no longer complete reconstruction and verification, and that most failures share a small number of recurring failure symptoms. A recurring failure can therefore be diagnosed once and its repair reused. We present Relay, a hybrid framework that uses an LLM to derive and validate deterministic repair rules from representative failures, reuses them without model calls, and escalates cases unresolved by rules to a per-case agent. We evaluate Relay in human-guided and fully automated rule-derivation campaigns and compare it with a per-case agent approach. We find that rule derivation can be fully automated, and that reusable rules restore as many builds as per-case agents with fewer build executions. However, agents restore cases that the rules cannot. Our results show that a hybrid approach that applies rules first and escalates the remaining cases to an agent restores more builds than either approach alone and at a lower cost than using agents alone.

Training LLM agents for policy-constrained tool environments, such as an airline backend with booking, refund, and cancellation policies, requires tasks that exercise the environment's rules individually and in combination. Existing synthesis pipelines let a model propose tasks and retain only those it immediately solves, biasing training data toward easier behaviors. We introduce Gradus (Graded Rungs of Ascending Difficulty from an Underlying System), which decouples what a task should teach from the concrete task and solution used to teach it. Without seed tasks, Gradus treats a fixed environment as a latent syllabus, extracting guarded behavioral operators that capture how actions depend on current conditions and how their effects reshape subsequent decisions. It composes these operators into increasingly demanding tasks while using corpus-level feedback to target missing outcome combinations. Crucially, Gradus jointly calibrates tasks and solutions against the environment: execution feedback identifies which side to refine while preserving the intended behavioral requirements. Verification thus becomes a recovery signal rather than a terminal filter, coupling coverage-guided exploration with the preservation of difficult examples. From off-the-shelf backbones, the resulting curriculum lifts a 30B model from 59% to 89% on $τ^2$-Bench and from 47% to 83% on CAR-Bench with an Opus 4.5 proposer, and to 76% and 62%, respectively, with an open-weight Qwen3-235B proposer. These results show the value of jointly controlling which behaviors are explored and which survive into verified training data.

Tool-using LLM agents increasingly answer e-commerce queries through multi-step tool interactions. The expected output of a tool call can shift between agent development and execution, allowing a plausible but incorrect value to be returned without an explicit failure signal. These silent errors may propagate through subsequent decisions into final answers. To study agents' resilience across this end-to-end process, we formulate Silent-error Resilience Task, an interactive task that evaluates whether they complete the original query after a silent error corrupts an intermediate state. We introduce RITE, a benchmark built from human-authored industrial workflows and production-log-derived inputs. To enable controlled and reproducible evaluation, we use a semantic error taxonomy to inject errors into fixed workflow prefixes, freeze recorded tool behavior, and let agents continue freely. It contains 167 matched units across nine e-commerce workflows, yielding 501 silent-error, correct-value, and explicit-failure instances. Across 10 recent LLMs, silent errors roughly halve task success and are consistently harder than explicit failures. Recognition, use of an available verification path, and integration of the resulting evidence form successive bottlenecks. We also propose a light State Checker as an intervention to test whether resilience can be improved.

3D Human Pose Estimation (3D HPE) is used in indoor environments, such as smart homes, to analyze human movement and monitor human states. Millimeter-wave (mmWave) radar is suitable for this task because it is robust to lighting variations and does not capture images of human appearance. However, radar observations often provide insufficient information about certain body parts or contain overlapping reflections from multiple body parts, limiting accurate 3D joint localization. To address these limitations, we propose the Initialization-to-Refinement Pose Framework (I2R-Pose), comprising two stages: Motion-Aware Pose Initialization and Observation-Guided Refinement. In the first stage, joint-wise features are learned by incorporating temporal relationships across consecutive radar observations and human skeletal structure to compensate for insufficient joint information and estimate an initial 3D pose sequence. In the second stage, separately extracted radar observation features predict joint-wise positional corrections that refine the initial pose and generate the final pose sequence. The framework supports individual and fused inputs of time-Doppler/time-range (TD/TR) maps and radar point clouds (PCs). To evaluate performance across indoor activities, we construct the Radar Motion and Spatial Information for 3D Human Pose Estimation (RaMoS-Pose) dataset, providing synchronized TD/TR maps, radar PCs, and 3D joint ground truth (GT). Evaluations on RaMoS-Pose and the publicly available MVDoppler-Pose dataset show that I2R-Pose reduces mean per-joint position error (MPJPE) by 39.50% on MVDoppler-Pose and by 15.13% on RaMoS-Pose with fused inputs, compared with the best-performing existing methods evaluated on each dataset.

A global $\ell_1$ budget limits how much an adversary may change an image, not how that change is distributed. We ask how much of an $\ell_1$ attack's observed success remains when large coordinate changes are restricted. We add a per-coordinate cap to the budget, enforce the joint constraint by exact in-loop projection, and sweep the cap for APGD-$\ell_1$, SLIDE, and FMN-$\ell_1$ without altering their objectives. Across 18 standard-model settings, caps of at most $32/255$ retain a median $97\%$ of attack success relative to each method's native uncapped baseline, while median peak magnitude falls by $80\%$ and median effective support nearly triples. At matched realized $\ell_1$, capped perturbations remain structurally distinct; separate ImageNet comparisons also show lower LPIPS and DISTS at selected caps. Direct in-loop constrained optimization preserves up to $88$ percentage points more attack success than post-hoc clipping, highlighting the importance of constraint enforcement. Suitable caps nevertheless vary across attacks and models, and adversarially trained classifiers exhibit more heterogeneous concentration sensitivity. These results establish perturbation concentration as a complementary, model- and attack-dependent axis of $\ell_1$ attack evaluation.

Artificial intelligence systems are increasingly deployed in high-stakes decision-making. In response, prototype-based models in explainable AI have attracted growing attention for their ability to retrieve visually similar examples as faithful explanations. Although these models achieve competitive classification accuracy, they often exhibit degraded explanation quality across datasets with varying characteristics, limiting their practical applicability. We draw on insights from cognitive neuroscience, particularly the human memory system, which encodes input-relevant information into representations that support recognition. We propose an Engram Memory Network (EMN), a model that emulates the ventral visual stream and inferior temporal cortex in the human brain. Central to this model is a Hopfield-inspired memory module that iteratively retrieves representative prototypes. These memory-refined latents are used for both reconstruction and classification, encouraging the selected prototypes to capture input-relevant information across diverse datasets. To address the lack of standardized evaluation for prototype explanations, we assess three complementary aspects: perceptual similarity (LPIPS and DISTS) for perceived resemblance, subclass alignment, measuring whether prototypes match the input's fine-grained category beyond supervised labels, and information preservation (NMI), measuring how much input information is captured by selected prototypes. Experiments show that EMN improves visual similarity by 17.06%, subclass alignment by 104.57%, and NMI by 42.42% on average across CIFAR-10/100, MNIST, HAM10000, and CUB-200, while achieving the highest average classification accuracy (87.62%) among all methods. These findings underscore the effectiveness of neuroscience-inspired explanation mechanisms in advancing the interpretability and reliability of prototype-based models.

The rapid growth of distributed photovoltaic (PV) generation is increasing the demand for accurate forecasting across large numbers of geographically dispersed stations. Existing approaches are commonly developed for individual sites or specific forecasting horizons, requiring repeated training and maintenance as new stations are integrated and limiting direct reuse at previously unseen sites. To address this scalability gap, this paper presents a domain-specific foundation-style pretraining framework that learns a shared forecasting model from heterogeneous multi-site and multi-horizon PV data. NWP-assisted forecasting is formulated as a sequence-completion problem that unifies historical PV power, meteorological observations, future numerical weather prediction (NWP) forecasts, and time encodings. Three pretraining objectives are designed and systematically evaluated, including forecast-aligned reconstruction, generic reconstruction, and mixed reconstruction. The mixed objective combines the operational structure of PV forecasting with random sequence reconstruction and provides the most balanced performance across forecasting accuracy, degraded NWP conditions, and downstream adaptation. Experiments on 479 distributed PV stations use 340 stations for model development and 139 spatially held-out stations for zero-shot evaluation. Across horizons from 6 hours to 7 days, the multi-site pretrained MR model outperforms representative station-specific and shared baselines, as well as generalist time-series models, including a recent large-scale pretrained foundation model. The pretrained model can also be directly deployed at new PV stations and subsequently adapted to region-specific operational objectives when local data become available, providing a scalable forecasting paradigm for large distributed PV fleets.

Deep generative modeling has achieved remarkable progress in crystal generation. Given the rich symmetries inherent in crystalline structures, recent work has sought to explicitly incorporate space-group symmetry into generative models, enabling more effective exploration of novel symmetry templates. Yet introducing symmetry templates increases the combinatorial complexity of stoichiometric relationships between atomic species and crystallographic sites, while creating a fundamental tension between flexible exploration of discrete crystallographic templates and faithful realization of these templates as continuous three-dimensional structures. To address these challenges, we propose CrystalAR, a unified autoregressive framework that models Wyckoff symmetry templates and complete crystal structures within a single generative sequence. Specifically, CrystalAR reformulates continuous coordinate generation as a symmetry-conditioned hierarchy of discrete spatial localization decisions, autoregressively refining the positions of all symmetry-inequivalent sites across resolution levels while jointly modeling their spatial configuration within each level. This discrete hierarchical formulation further enables constraint-aware search over both crystallographic templates and geometric realizations. As a result, CrystalAR retains strong symmetry-template generation capability while achieving competitive de novo generation performance and template-free crystal structure prediction (CSP) accuracy. These results suggest that symmetry-template generation can be retained alongside broadly competitive de novo generation performance and competitive CSP without reliance on symmetry templates retrieved from prior data, providing a promising alternative paradigm for symmetry-preserving crystal generation.

State-of-the-art neural network sparsification relies on alternating dense and sparse training phases. While computationally expensive, dense phases are critical for performance. Our empirical analysis reveals the mechanics behind this gain: upon reintroduction, previously masked parameters receive disproportionately large gradients that drive essential mask rewiring and parameter sign-flipping. Building on these insights, we optimize the performance-to-FLOPs Pareto frontier through three innovations: (1) zero-initializing masked parameters during dense phases to promote relevant sign flips, (2) shortening dense phases, and (3) replacing the dense phase entirely with a sparse rewiring phase that updates only the active mask and a small, high-gradient fraction of out-of-mask weights. Evaluated on CIFAR100 and ImageNet, our variant ARC (Alternating Rewiring Compression) matches baseline AC/DC performance at a fraction of the computational cost, establishing a superior performance-to-FLOPs frontier also compared to standard sparse-to-sparse algorithms. Experiments with GPT-2 on OpenWebText further show improvements in validation perplexity and average zero-shot performance

Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted items aside, partitioning the array. We give VLMs the same actions as dumb image tools with no visual intelligence of their own in a training-free agent loop. Our central claim is about measurement, not intervention: a model that counts by marking leaves a normally hidden serial procedure on the image, and the developmental principles of counting turn those marks into an interpretable audit of serial visual reasoning: every tag is classified against point-level ground truth as on-target, duplicate, phantom, or omitted, the answer is checked against the number of tags, and order irrelevance is tested by forcing opposite traversal orders. Across 9 models, 682 images, and 7 harness conditions we find that (i) the exact-counting range is a model-specific signature spanning nearly two orders of magnitude (R80 from 3 to 150 dots); (ii) as an intervention the finger mostly fails: tags help 4 of 9 models on small-count photographs, hurt imprecise pointers, and rarely beat a verbal routine or a blind point list; what the loop buys is an auditable trace, not accuracy; (iii) as measurement it succeeds: a tag-perturbation probe shows that precise pointers read their answers off the tags, pointing precision tracks which models benefit, dense scenes fail by omission, and models rarely revise after seeing their own tags; and (iv) the capacities dissociate: models follow an instructed traversal order yet often change their answer when it reverses, and children's regular-over-scattered advantage appears only under the routine and survives lattices that defeat a rows-times-columns shortcut.

In diffusion models, proper distributional learning entails reversing a diffusion process over underlying density functions, whereas a key failure is memorization, in which the reverse process instead learns diffusion from individual training samples. We provide an analytical characterization of such memorized reverse processes, along with a criterion that detects the associated thermodynamic singularity. The proposed criterion reliably distinguishes between properly learned and memorizing networks across a range of neural network models designed to promote or mitigate memorization.

Muon and related optimizers update a weight matrix with a spectrally transformed gradient. In this work, we study such spectral optimizers with a constant step size on strongly convex matrix quadratics, which isolate the effect of the update from the confounds of nonconvexity, noise, and step size schedules. We analyze an update using two quantities: the gain, which is the factor by which it amplifies the gradient along each singular direction, and the length, which is the norm of the update. These two quantities place the popular spectral updates in three regimes with different behavior: (i) with a bounded gain, the iterates behave like preconditioned gradient descent and converge for step sizes set by the largest curvature; (ii) with an unbounded gain but a vanishing length, they still converge, with the admissible step sizes set by the total curvature instead; and (iii) with a non-vanishing length, they do not converge for almost every initialization. Interestingly, the practical variants of Muon fall in all three regimes, depending only on how the update is normalized and scaled and not on the polynomial that approximates the polar factor. Experiments validate these findings.

Simulating economic responses at society scale requires agents whose decisions reflect household circumstances and the effects of earlier actions. We introduce S$^2$-MAS, a society-scale multi-agent simulation framework built on 500 million synthetic household personas grounded in real-survey data. Its Genese engine separates decisions from state updates, while reusable financial scenarios and application ledgers track consumption, cash, debt and payments. FinL-mlx, a compact decision model trained with SFT + RLCD on 2.93 million examples, achieves 74.55% accuracy and 38.37% question-macro F1 in consumer financial-choice scenarios, outperforming Jev 1.13 and GLM-5.3-Flash on both measures. We evaluate eight financial task families and simulate 12 million distinct household agents in paired purchase-funding conditions. Jointly raising purchase goals and approved credit limits increases consumption, with responses differing across age and income groups. Education, employment and household-size differences persist after age--income standardization and mainly reflect transitions from deferral to purchase. Asset and debt groups differ in cash drawdown, repayment and closing debt. A 24-month panel varies purchase goals under a common credit product, with repayment and arrears updating subsequent offers. S$^2$-MAS links heterogeneous household decisions to evolving consumption, cash and repayment paths, supporting economic experiments across household groups, population scales and time horizons.

Large language models (LLMs) are increasingly used as judgment assistants and as stand-in participants in psychology experiments, yet little is known about how much their probability estimates move in response to a single new piece of information. We pair premise-hypothesis items from $δ$-SNLI with one strengthener and one weakener each, elicit the prior and the posteriors in separate calls, and measure the directional movement s: the mean posterior change signed by the annotated direction, so that the prior cancels out. We vary whether the update is stated ("Here is something else you learned") or reported ("Someone said that"), and whether the model gives one probability or five weighted estimates. We test seven configurations: two local models (Qwen3-8B, OLMo-2-7B; 240 pairs) and five configurations run through agent command-line harnesses (120 pairs), with hypotheses and analysis rules fixed before the formal runs (not publicly registered). Under stated, direct elicitation all seven move in the annotated direction (s = 18-28 percentage points). Reporting lowers s in all seven point estimates (significantly in six, by 2-5 points) and raises the share of reverse movements in all seven: the effect is consistent but small. Five weighted estimates lower the share of responses at $\leq$5% or $\geq$95% in all seven, but in the two local models they also reduce s by 43% and 68%, whereas in the harnessed configurations the point estimates of the change lie between -3.7 and +0.5 points. An elicitation format that reduces extreme values thus need not preserve sensitivity to evidence; the two should be tested separately. Because $δ$-NLI annotates only direction, we make no claim about how large an update should be. Code, data and the frozen plan: https://github.com/asj252/llm-judgment-studies

Guided Imagery (GI) is a mind-based relaxation technique widely used in clinical and therapeutic settings. Objectively assessing whether a patient has entered a relaxed state remains a challenge for clinicians. This study investigates electroencephalography (EEG)-based classification of hypnotic susceptibility using two-dimensional Convolutional Neural Networks (2D CNNs) applied to minimally preprocessed signals, without manual feature extraction. Data were collected from 62 male participants (32 Control, 30 Hypno) using a 256-channel dense array amplifier during Guided Imagery or cognitive task conditions. A sliding-window framework evaluated up to 868 independent models across seven anatomically defined brain regions and 124 ten-second time windows. Segment-level random-split screening was followed by Leave-One-Subject-Out (LOSO) cross-validation to obtain subject-independent performance estimates, and an ensemble approach combined predictions across K = 10 windows. The best single-window model achieved accuracy of 0.82 on the random split. Under LOSO, anterior regions generalised better across subjects: the frontal region achieved mean accuracy of 0.646 (AUC: 0.692). An ensemble of K = 10 windows further improved the prefrontal region to mean accuracy of 0.738. The occipital region, dominant in random-split screening, collapsed to near-chance under LOSO, revealing subject-specific EEG fingerprinting rather than genuine group differences. Bandpass filtering and z-score normalisation suffice for competitive EEG classification without manual feature extraction. Anterior cortical regions are identified as the most generalisable neural markers of hypnotic susceptibility.

Recursive self-improvement, in which frontier AI models work on training the next generation, is quickly becoming a reality. However, the safety of this process has received far less attention than it should. A misaligned model training its successor can shape what that successor knows, how it behaves, and what it wants, and the harm carries forward into the next generation. We study this risk through the lens of AI control. We introduce RSI Control Arena, a testbed in which a frontier model, acting as an autonomous R\&D agent, is given a benign main task from a stage of a successor model's training pipeline together with a covert side task to install a misaligned property in the successor. The testbed contains 16 side tasks in three classes, forming 22 task pairs with main tasks that span the pipeline from data curation to reinforcement learning. We evaluate six frontier models as agents against LLM monitors that read the transcript of the run and the agent's reasoning. We observe that every model completes a substantial share of the side tasks without degrading the main-task performance, including side tasks that make the successor itself a better saboteur. Under minimal instructions to stay covert, models already produce successful attacks that monitors fail to detect, and providing a short list of evasion tactics further reduces the detection rate. Monitors with access to agent's reasoning detects most attacks, but not all. Our results indicate that recursive self-improvement raises significant safety concerns, and that strong safety effort should be in place before models are trusted with greater responsibility of training their successors.

Learned barrier certificates provide a principled mechanism for establishing safety of stochastic systems, but their validity may deteriorate during deployment when the underlying dynamics differ from those used for training. We propose an anytime-valid runtime monitor that audits a learned barrier certificate using observed closed-loop trajectories. The monitor is intended to detect loss of statistical support for the deployed barrier condition; it does not by itself guarantee detection before unsafe entry or safety of a subsequent intervention. Specifically, the conditional barrier-drift condition is formulated as an episode-wise sequential hypothesis test, and observed barrier increments are transformed into a nonnegative test supermartingale, yielding finite-sample false-alarm control under arbitrary stopping times. Inspired by finite-horizon safety verification, we further introduce a deterministic temporal shaping of the monitoring boundary, making the monitor more conservative early in a prescribed deployment horizon while preserving anytime-validity. The monitor is designed to detect statistical evidence that the safety certificate is no longer consistent with the observed dynamics, rather than to predict unsafe entry. To assess the monitor empirically, we evaluate empirical alarm behavior under valid and invalid certificates. Code and data are available at https://github.com/FormalCert-AI/runtime-barrier-monitor.

Long-horizon human-object-scene interaction (HOSI) is central to embodied AI, animation, and game development. Existing human-object interaction (HOI) and human-scene interaction (HSI) models excel in their respective domains, but none alone covers tasks requiring multiple interaction skills. Training unified HOSI models is hindered by scarce and costly data. We introduce World2HOSI, a multi-expert orchestration harness that generates continuous HOSI data from 3D Gaussian Splatting (3DGS) scenes and natural language instructions. Our key insight is that long-horizon HOSI hinges not on a stronger single model but on orchestration. Given a shared, structured scene state and an orchestration harness grounded in it, existing experts can jointly accomplish tasks beyond the reach of any of them alone. World2HOSI realizes this insight as a closed loop around a persistent scene state built from decoupled background and object assets. A harness plans action stages over this state, invokes experts via their native conditioning interfaces, and writes each action's effects back, so that every invocation is grounded in the evolving world. To reduce physical artifacts, bounded residual optimization and contact-driven physics simulation further ensure the physical plausibility of human motion and object manipulation, respectively, while preserving interaction intent. Experiments and a perceptual user study show that World2HOSI achieves the best task completion and interaction quality on the evaluated HOSI tasks, while generating more diverse data.

Dense linear probes assess task readability from language model (LM) representations, but do not directly reveal how many features suffice for prediction. We measure readout sparsity with \(k^*\), the smallest feature count found by a fixed search protocol that recovers 95% of dense-probe AUROC above chance. We evaluate 19 LMs from five labs on 14 core tasks. Ten base models give the nine reliably readable tasks similar feature counts and nearly the same order. This agreement extends to the selected features themselves, which on three tasks show cross-model response correspondence beyond class-mean differences and a label-correlation-matched null. Instruction-tuned deployment selectively reduces \(k^*\), with both weights and input format contributing. In controlled synthetic experiments across 20 models, successfully learned targets admit single-feature readouts, whereas readable controls require tens of features. Reference-model \(k^*\) is associated with label requirements across eight students, following an empirical scaling of \(N^*\propto(k^*)^{0.65}\). After continual learning, saved heads can fail while linear readability remains high, and recovery requires fewer labels than initial task learning under the tested procedures. Pruning reduces linear readability, with wider pre-pruning readouts generally associated with larger losses. These findings distinguish task readability from readout sparsity and establish their complementary value for predicting learning requirements and diagnosing LM changes.

Large language models can achieve strong Text-to-SQL benchmark performance while still producing schema and query-logic errors. Most evaluations characterize such failures only in the final generated query, making it difficult to determine whether a related error was also observed in an earlier, simpler step or appears only at the final stage. We present a drill-down evaluation framework that decomposes Text-to-SQL queries into progressively more complex executable sub-queries and corresponding natural-language questions, allowing rule-based hallucination categories to be examined across a generation path. We distinguish recurrent hallucinations, in which the same rule-based category is observed in an earlier step and again in the final query, from emergent hallucinations, in which the category is observed at the final step without an earlier detection in the same path. Across five LLMs evaluated on an expanded BIRD-mini benchmark, several categories show appreciable recurrence probabilities, while clause abuse shows comparatively higher conditional emergence. Under the conversational-history condition, recurrence probabilities are generally higher and emergence probabilities are generally lower. These observations indicate that the two conditional measures do not move uniformly in the same direction under additional context and motivate trajectory-level analysis in addition to final-output evaluation. Code and evaluation artifacts are available in the project repository. https://github.com/HaydenMM/bird-mini-drilldown

When does experience memory help a closed-loop LLM driving agent? Published gains cannot say, because the two factors we find to matter most, which warmup experiences the pool contains and what any injected tokens do to the prompt, are controlled in none of them. We answer with a controlled evaluation on one audited HighwayEnv testbed (1Hz, gpt-4o-mini for the decisive contrasts). We treat the warmup pool as a resampled random factor (ten disjoint pools) and add token-matched irrelevant-filler controls, oracle-certified diagnostics, caution-pricing metrics, and quantified audits of the baseline ports. The probe suite spans flat-text retrieval (a DiLu port), a single-frame grid port (RESPOND-inspired), our episodic temporal-graph memory, a retrieval-free consolidated rulebook, and a flat K-step observation-history baseline. Four findings follow, each a property of this testbed. (1) Headline memory gains belong to the warmup draw. The reported pool is the best of ten for both retrieval memories, which beat No-Memory in only one pool of ten (sign test p=0.011), and even the rulebook's apparent pool-immunity collapsed on five fresh draws. (2) No retrieval design tested exceeds token-matched filler. The collision-only merge ceiling is reached by a single 15-token caution instruction, and metrics that price caution erase the rulebook's lead. (3) Representational sophistication is nowhere robustly operative. (4) Baseline port quality silently inverts conclusions. On this testbed, a single-pool comparison without filler controls would have reported memory gains that these controls reverse. We release the testbed, audited ports, seeds, and caches.

Feed-forward 3D Gaussian Splatting (FF3DGS) reconstructs 3D scenes from images in a single forward pass, making it well-suited to applications that must operate within limited resource budgets. To further improve its practicality, recent budget-controllable FF3DGS methods allow the number of Gaussians to be adjusted at inference, so that the output meets a given storage limit or rendering-latency constraint. However, these methods overlook the cost of inference itself: to produce a representation with the target number of Gaussians, they first predict a dense set of Gaussians and then prune or select from it, requiring nearly constant latency and peak memory regardless of the budget. We therefore propose RLSplat, a resource-linear budget-controllable FF3DGS framework whose representation size, inference latency, and peak memory all decrease linearly as the Gaussian budget is reduced. Instead of pruning or selection, RLSplat learns a Gaussian allocation policy that distributes the budget across local scene regions, so that exactly the requested number of Gaussians is generated. These Gaussians are then refined through resource-linear Gaussian decoding, in which each Gaussian attends only to nearby Gaussians found via an efficient hierarchical neighbor search, so that the decoding cost decreases linearly with fewer Gaussians. In zero-shot evaluation, RLSplat outperforms prior budget-controllable methods at every budget, surpassing the strongest baseline even with 4$\times$ fewer Gaussians, while running up to 4.1$\times$ faster with 9.6$\times$ less peak memory at low budgets.

Under matched conditions, classifier-head fine-tuning (\textsc{cls}) paradigm consistently outperforms the LLM-as-Judge sequence-to-sequence (\textsc{s2s}) on binary hallucination detection. We trace this gap to the representations induced by the two objectives. Static analysis shows that \textsc{cls} learns a task-specific discriminative direction, pushing hidden states into a clear two-cluster geometry with stronger class separation and linearly decodable label information. By contrast, \textsc{s2s} relies on the pretrained language-modeling head to make decisions. This minimizes the pressure to reorganize representations, leaving the hidden states much closer to the pretrained backbone. Our analysis also reveals that the \textsc{cls} signal saturates before the final layer, after which upper layers form a plateau of statistically interchangeable and largely redundant representations. However, endpoint redundancy alone does not show whether these layers are dispensable during training. Further training-dynamics analysis reveals 1) a lower pretrained substrate, 2) a narrow upper-middle \textbf{InfoWindow} where task information emerges, and 3) a later-synchronizing upper plateau that mirrors this signal. This mechanism motivates us to remove layers above the InfoWindow. Truncating 31\% of the network leaves accuracy unchanged within seed noise on Qwen3-4B and 8B and cuts latency by $1.4\times$. The same depth fraction, fixed before any analysis of the target models, transfers to Llama-3.2-3B and Mistral-7B-v0.3, and the layer structure behind it reappears there and on QNLI and SST-2. We hope this helps guide efficient task-specific LLM design.

Diffusion models have emerged as the foundation of modern generative systems, yet their high memorization capacity raises privacy concerns. While differentially private (DP) training provides formal guarantees, it comes at a cost in sample quality. In this work, we study a complementary question: how much privacy loss is incurred during the sampling process itself? Directly analyzing modern neural denoisers is challenging because their sensitivity to a single training example is generally intractable. To overcome this challenge, we construct and analyze a clipped privacy-accountable empirical denoiser, a training-free estimator of the Bayes-optimal denoiser, whose per-step sensitivity can be computed explicitly, allowing each denoising step to be a Gaussian mechanism. This perspective enables privacy accounting under Gaussian DP when each training example contributes to each denoising step, and under $f$-DP trade-off functions when each generated sample uses a single randomly subsampled block of the training data. We report every $(\varepsilon, δ)$ guarantee as a certified upper bound from a privacy loss distribution accountant, and use a cumulant-based Edgeworth approximation as a fast diagnostic, characterizing where it is reliable and where it becomes anti-conservative. We also analyze the privacy loss along the denoising trajectory and show that it grows steeply near the critical window in which semantic content is formed. This observation motivates hybrid samplers that use the private empirical denoiser through the formative steps where it remains a faithful proxy for the neural denoiser, then switch to a public neural denoiser for the remaining refinement. Our results on CIFAR-10 and CelebA-64 show that the privacy of diffusion models depends not only on the training procedure, but also on sampling-time design choices such as subsampling, stopping time, and hybrid denoising.

Smart-city applications often combine scarce labeled data with abundant unlabeled observations. Semi-supervised learning (SSL) exploits both labeled and unlabeled data to improve classification under limited supervision. This article presents MOEA-SSL, a multi-objective evolutionary approach for lightweight semi-supervised tabular classification that jointly optimizes predictive quality and feature usage. Experimental evaluation on three Smart City classification datasets, considering 1\%, 5\%, and 10\% labeled data, shows that MOEA-SSL achieves competitive predictive performance compared with established lightweight SSL methods while significantly reducing feature usage relative to EA-SSL in all evaluated conditions. The computed Pareto fronts also provide lower-feature-use alternatives with limited degradation in classification performance.

Large language models are expected to estimate their own uncertainty, with confidence signals used to guide retrieval, adaptive compute allocation, routing, and selective abstention. We investigate whether confidence signals measured at different temporal stages of problem solving encode distinct forms of epistemic information and remain reliable beyond English-centric settings. Inspired by classical meta-memory theory, we distinguish anticipatory pre-solve confidence (Feeling-of-Knowing, FOK) from retrospective post-solve confidence (Judgement-of-Learning, JOL). We evaluate these signals across two frontier LLMs, multiple reasoning domains, translated Marathi benchmarks, and a newly constructed benchmark of genuinely local Marathi factual questions. Across settings, the cheaper pre-solve signal remains weakly calibrated, while post-solve confidence generally provides stronger epistemic alignment despite its additional inference cost. Confidence behaviour transfers surprisingly well under direct language translation, with calibration trends largely preserved between English and Marathi, but degrades substantially on local Marathi knowledge, suggesting that linguistic transfer alone is insufficient for region-specific information needs. Finally, hallucinated responses maintain persistently elevated confidence after answer generation, highlighting important limitations of current LLM self-assessment mechanisms.

Data augmentation (DA) has become a fundamental component of modern machine learning systems and is widely regarded as an effective strategy for improving robustness and generalization. Despite its empirical success, the theoretical impact of DA under label noise remains poorly understood. In this work, we first theoretically explain when DA is beneficial and when it becomes detrimental, and provide sufficient conditions for distinguishing between these two regimes under both the i.i.d. learning setting and the more challenging domain generalization scenario. Then, we reveal a counterintuitive phenomenon: DA essentially acts as an uncertainty amplifier in the learning process, decreasing model stability and generalization performance. This finding is supported by the proposed learning and generalization bounds. Our analyses do not depend on specific DA techniques and require only mild assumptions, indicating that the results are applicable to a broad range of machine learning scenarios. Namely, we provide a unified theoretical foundation for understanding, analyzing, and designing DA methods in the presence of label noise.

In long-term memory systems, memories are constructed from interaction histories and retrieved in response to queries, enhancing the reasoning ability of LLM agents. In complex scenarios, facts are scattered across sessions and their relationships evolve over time, making it difficult for existing memory systems to compose them to adapt to diverse queries. Motivated by the modularity and composability of code, we propose Memory as Code (MaC), a unified memory framework that jointly designs memory construction and retrieval through a shared code abstraction. MaC couples codified construction, which establishes executable memory abstractions, with declarative retrieval, which translates queries into structured retrieval programs. Specifically, codified construction builds a three-layer executable memory structure that progresses from an atomic function layer, through a relational layer, to a modular composition layer. Declarative retrieval recasts retrieval from relevance-based fact selection into declarative query processing over memory. Experiments on LoCoMo and LongMemEval show relative gains of up to 15.91% in overall LLM-judge accuracy over the strongest baseline in each setting, while substantially reducing token consumption, highlighting the effectiveness of memory as code for long-horizon memory tasks.

Low-bit quantization reduces the memory footprint of large language models, but preserving accuracy with efficient training remains challenging. Scalar quantization encodes weights independently, while explicit-codebook vector quantization introduces costly search during training. We propose E8-QAT, a 2-bit quantization-aware training framework that uses the $E_8$ lattice as an implicit eight-dimensional codebook. It jointly quantizes weight groups through arithmetic projection without storing or searching an explicit codebook during the training forward pass. A 2-bit core encoding and sparse extension preserve the selected lattice points exactly, adding less than 0.1\% to core index storage. On Qwen3.6-35B-A3B, E8-QAT matches the performance of the full-precision model in terms of the average score across the four tasks. E8-QAT improves over the state-of-the-art baseline by 2.03\% on IFEval and by 1.51\% on LiveCodeBench, under the same data, schedule, and bit width. The 2-bit representation reduces peak memory by 80\% and supports a $4\times$ larger batch. At each representation's largest feasible batch, E8-QAT reaches $1.97\times$ the decoding throughput of BF16.

Vision-language-action (VLA) models transfer knowledge from large-scale pretraining into robot control, but their manipulation performance remains sensitive to changes in observation conditions, which recent work addresses by fine-tuning the visual module on target-domain data. However, it remains unclear how such adaptation affects robustness beyond the target condition, as existing robustness benchmarks only evaluate models without fine-tuning and adaptation studies only report the target gain. Measuring this cross-dimension effect requires separating target gains from off-target changes, isolating adaptation effects from architectural differences, ruling out generic forgetting, and comparing tuning choices. Therefore, we introduce VLA-Retain, a benchmark that compares five open VLA models, tuned on Camera-view or Sensor-noise data, with their own base checkpoints on LIBERO-Plus dataset spanning seven perturbation dimensions. Our study reveals four findings. First, targeted vision tuning is not local, and its target gain does not predict retention. Mean off-target gain is negative in 8 of 10 model--domain configurations. Under Camera-view tuning, X-VLA gains $69.4$ percentage points on the target but loses $4.6$ points on average elsewhere, whereas OpenVLA-OFT gains $39.3$ and $1.4$ points, respectively. Second, these regressions are not explained only by performance loss on the same tasks without perturbations. Third, broader training coverage raises overall gain over target-specific training but does not eliminate dimension-level regressions. Fourth, tuning strategies are model-dependent, as DELTA-style feature preservation raises X-VLA's mean off-target gain to $+1.2$ points but lowers OpenVLA-OFT's to $-23.1$. Overall, these findings show that target-domain success alone is insufficient and highlight cross-dimensional retention as a key criterion for targeted VLA adaptation.

Large Language Models have shown strong code generation capabilities, yet their outputs often contain subtle bugs. Automatically repairing such code requires informative tests: inputs that expose faulty behavior, paired with correct expected outputs. Recent RL-based test generation methods apply the same outcome reward to all reasoning steps, which can discourage useful reasoning and reinforce flawed steps. They also fail to effectively learn from trajectories that find bug-triggering inputs but predict incorrect outputs, leaving their verified bug-triggering progress insufficiently exploited. To address these limitations, we propose SCALF, a process-supervised RL approach that combines Step-level Credit Assignment with targeted Learning from valuable Failed trajectories. SCALF derives step-specific advantages from process rewards, allowing each reasoning stage to receive its own learning signal. It further routes trajectories that expose bugs but predict incorrect outputs to on-policy self-distillation, using execution feedback to learn from their valuable bug-triggering progress. Across three benchmarks, SCALF outperforms strong prompting and post-training baselines by up to 2.97, 5.26, and 3.53 percentage points in joint success rate, output correctness rate, and bug-triggering rate, respectively. Our code and datasets are available at https://anonymous.4open.science/r/SCALF-F4D1.

For millennia, knowledge has been carried by language, and large language models still put language at the center. Yet the coding agents built on these models mostly write code. This paper is a contribution to epistemology, which asks what kind of knowledge code is and how it is obtained. We define knowledge as code as knowledge written in a programming language or in a machine-parsable data format, whose validity can be checked by executing it. We define thirteen epistemic properties of knowledge media and show that code is the first medium to have all of them. We distinguish two ways in which knowledge enters code, encoding and learning by execution, where code is both the experiment and the record of its result. Agents make learning by execution observable, since their trajectories record every probe, observation and line of code they keep. Code is thus not only a tool of knowledge work but a knowledge medium in its own right, the first that can check what it holds.

Modern industrial ranking systems increasingly adopt generative recommendation (GR) paradigms, in which user behavior sequences and non-sequential (NS) features are tokenized and jointly modeled by Transformer-style backbones. This unified formulation provides a scalable interface for modeling long histories and heterogeneous features. However, sequence and NS representations often carry asymmetric semantics: historical representations may combine environment context with observed-action semantics, whereas candidate-side NS representations contain no observed action but incorporate richer supplementary context, such as user-item interaction features and recent behavior lists. Direct interaction between such representations can make query-key dot products compare different semantic objects, limiting meaningful cross-token interaction. To fill this gap, we propose EASformer, an Environment-Action-State disentangled architecture for industrial app ranking. EASformer constructs an explicitly aligned context environment space shared by historical and non-sequence tokens, while disentangling user actions from historical token and candidate-specific context form non-sequence tokens. Based on this aligned space, EASformer enables complementary historical action retrieval and state-conditioned action prediction. Furthermore, the disentangled design supports a Cloze-style self-supervised objective by masking historical actions and reusing the candidate prediction pathway. On an offline benchmark from a real-world industrial app marketplace, EASformer achieves an AUC of 0.89411, yielding a 0.88% relative improvement over the deployed baseline. Online A/B tests further support this advantage, with relative eCPM lifts of 1.26% and 3.94% across two real-world ranking scenarios.

We introduce the target-response Jacobian $J = [\partial \hat y_i/\partial y_j]$, which measures how model predictions respond to changes in individual training targets. $J$ complements training-data attribution (TDA) methods, which characterize sensitivity to the weights placed on training observations in the loss rather than to their targets, and can be computed matrix-free at the same cost as influence-based TDA methods. We show that sample-weight influence $\mathcal{I}$ and $J$ follow from the same implicit differentiation argument and differ only by a per-example multiplier: for pointwise losses, $\mathcal{I} = J\operatorname{diag}(m)$ exactly at any stationary point with an invertible Hessian. For canonical-link losses (squared error, logistic, Poisson) the multiplier is the training residual, $\mathcal{I} = J\operatorname{diag}(r)$, and softmax cross-entropy obeys a multiclass version. The identity predicts three ways in which influence fails when used in place of $J$: small residuals shrink its predicted shift towards zero (interpolation collapse), residuals of mixed sign wash out its direction (direction washout), and under softmax cross-entropy it collapses a training point's per-class sensitivities towards its own true class (true-class collapse). We score both operators against refitting with a per-prediction $R^2$ across target perturbations, with no rescaling. On tanh networks trained with squared error (Friedman #1) and logistic loss (MAGIC, HIGGS), with $N$ from 2,000 to $10^6$, $J$ reaches $R^2$ between 0.99 and 1.00, whereas influence's $R^2$ is negative on every dataset. On an MNIST softmax classifier, influence's $R^2$ is again negative, while $J$ reaches 0.92. On HIGGS, one row of $J$, holding all $10^6$ sensitivities of one prediction, takes under 0.2 seconds on an RTX 3090 after a one-time setup of under five minutes, whereas refitting would take $10^6$ refits.

Healthcare providers lose substantial revenue to systematic payer underpayments and claim denials, yet identifying the recurring, correctable patterns behind them remains largely manual and does not scale to hospital claim volumes. We present Pattern Intelligence, a deployed system that discovers such patterns without supervision by clustering claims on their adjudication behaviour rather than their identity. Each claim is embedded in a 130-dimensional vector combining per-line expected-versus-actual payment ratios, claim adjustment reason (CARC) codes, denial and payment signals, and service composition. Payer and contract identity are deliberately excluded from the distance metric - we show they dominate the geometry and collapse clusters into trivial per-payer groups - and are reintroduced only as a post-hoc segmentation (by contract and visit type) that keeps every surfaced pattern coherent and actionable. We cluster with HDBSCAN, recover structure from the noise set via second-pass sub-clustering, and gate low-confidence memberships. A large language model translates each centroid into an analyst-ready description of dominant procedures, adjustment reasons, and failure mode. We report a recurring finding: roughly one-third to one-half of high-dollar denials, depending on tenant and segmentation, form an irreducible long tail that no density-clustering configuration or representation we tried groups - a domain property, we argue, rather than a modeling deficiency. We evaluate validity through cluster-coherence analysis and quantify impact with a recoverable-opportunity ($) metric on two production tenants (approx. 10K and 157K claims) within a six-tenant deployment. Rather than a new algorithm, the contribution is a problem formulation, a representation-design insight, and empirical findings from production.

Electronic health records are written to support care, not retrieval. Across years of oncology notes, the primary diagnosis is restated in report after report, while the events that shape a patient's course, such as a recurrence or a change in therapy, may be recorded only once. A retriever that ranks passages by similarity to the query keeps returning the repeated material and can miss these transitions, that is, it exhibits temporal redundancy across historical narratives. We argue that longitudinal clinical retrieval should be treated as allocating a limited evidence budget across time, not just ranking by relevance. We separate the task into three objectives: temporal evidence allocation, candidate reconstruction, and clinical validation. In Study A (13 clinical queries over 20 patient records, with two-stage reconstruction), the Temporal-15 configuration reaches a milestone slot recall of 0.538, compared with 0.388 for the Global-50 baseline. However, it also retrieves far more evidence (496 versus 50 indexed units per query), so the gain cannot be credited to allocation alone. When evidence is held fixed, removing prompted precedence rules raises schema-order reversals from 0.010 to 0.080. A five-patient, single-query pilot with fixed budgets finds that wider temporal coverage does not reliably yield more dated milestones. Together, these results suggest that how much a record says about a period is a poor guide to how useful that period is as evidence. Because no expert reference timelines are available, our measurements describe retrieval behavior rather than clinical accuracy. This position and technical report presents the retrieval analysis behind the Onco-CONSTRUE framework and sets out a research agenda for budget-constrained longitudinal clinical retrieval.

Electromyography (EMG) provides a non-invasive means of decoding silently articulated speech into text, but accurate decoding remains challenging due to limited availability of paired EMG-text data. We introduce PhonoBridge, a two-stage framework comprising a channel-aware EMG-to-phoneme frontend and a large language model (LLM) backend for text reconstruction. The frontend is trained on paired EMG-text data that consists of channel-wise feature extraction, cross-channel fusion, and electrode embeddings. With phonemes as the intermediate representation, we expand the backend's training data by converting LibriSpeech transcripts into phoneme-text pairs without additional EMG recordings. To adapt the LLM to frontend noise, we introduce Error-Profile-Guided Augmentation (EPGA), which perturbs transcript-derived phoneme sequences according to error patterns estimated from frontend predictions, generating additional training pairs for the LLM. On the Open Vocabulary silent-speech dataset, PhonoBridge achieves 24.77% WER, outperforming the previously reported best of 28.27% on the same test split.

In object goal navigation (ObjectNav), an embodied agent must find a target object (e.g., "find a bed") in a previously unseen environment, relying on spatial commonsense such as object co-occurrence. Large Language Models (LLMs) encode this commonsense, but deploying them on robots exposes a trade-off between reasoning quality and efficiency: cloud-based models incur latency and privacy risks, whereas small on-device LLMs are fast but prone to hallucinated spatial judgments. Hybrid reasoning LLMs offer both a fast non-thinking mode and a slow thinking mode, which raises a key question for embodied agents: when is slow thinking worth its cost? We propose CovNav, a training-free framework that answers this question with Conformal Prediction (CP), turning dual-process (System 1 / System 2) inference into an uncertainty-aware decision. Much like recent decision models such as Jev, CovNav casts each spatial query as a structured binary decision rather than free-form generation: a local LLM in fast mode judges whether the target is likely near an observed object, CP converts its token probabilities into a prediction set, and only ambiguous queries are escalated to slow chain-of-thought reasoning. The object-pair formulation keeps calibration tractable and supports open-set objects through off-the-shelf detectors. Experiments on Gibson and HM3D show that CovNav surpasses the strongest zero-shot baseline by 0.8 and 1.7 percentage points in success rate on HM3D v1 and v2, while invoking slow reasoning for only 12.4% of queries and retaining most of the gain of always-on thinking at a fraction of its latency. Real-robot experiments on a TurtleBot 4 further validate the approach. CovNav shows how calibrated uncertainty can decide when an embodied agent should think slowly, offering a principled form of adaptive test-time reasoning for LLM-based robots.

The emergence of increasingly sophisticated and evasion-resistant phishing attack strategies represents a significant threat to modern cybersecurity, particularly in cyber-physical systems and internet-based networks. Traditional probabilistic methods and certain machine learning approaches often struggle to detect novel attacks and may lack interpretability or portability. To address these challenges, we introduce CyberHull, a resource-efficient and deterministic framework for identifying malicious URLs through mathematically defined convex hull boundaries across multiple orthogonal feature domains, including structural content entropy, hyperlink graph topology, and semantic token distribution. CyberHull constructs compact convex hulls, referred to as vaults, using the QuickHull algorithm, with a training complexity of O(n log n) under suitable assumptions. URL classification is performed through point-in-polyhedron tests that impose relatively low computational overhead on standard CPUs. Experimental results demonstrate that CyberHull achieves 92 percent detection accuracy, substantially outperforming a baseline linear approach based on Principal Component Analysis (PCA), which achieves 32 percent accuracy. Furthermore, CyberHull offers a computationally efficient alternative to resource-intensive deep learning approaches for malicious URL detection.

Momentum is central to stochastic optimization in deep learning, including the training of large language models. Classical momentum methods typically rely on a fast exponential moving average (EMA) scheme, which discounts past gradients on a fixed timescale and may underweight older gradients that remain informative. This creates a trade-off between retaining long-term history and adapting to recent changes. We introduce Hybrid Momentum (HyM), which combines a linearly weighted average of past gradients with a geometrically discounted fast momentum. The former assigns weights proportional to the gradients' iteration indices and retains a growing span of history, while the latter emphasizes recent gradients. This hybrid construction is modular and can be combined with adaptive preconditioning and other update transformations. In particular, we propose HyMUON, which applies Newton--Schulz orthogonalization to the hybrid direction. Illustrative examples and controlled experiments demonstrate the benefits of combining gradient information across these two timescales. Across image classification, machine translation, and language modeling, HyM and HyMUON match or improve upon strong baselines. HyMUON improves over Muon on IWSLT14, and in both the 100k- and 600k-step GPT-2 124M OpenWebText regimes. For LLaMA-1.3B pretraining on C4, HyMUON achieves a final validation loss of 2.777, compared with 2.903 for Muon.

Rerunning every governance control after every change to an artificial intelligence (AI) pipeline is expensive, so teams reuse earlier results. Reuse is safe only if each control's declared dependencies cover everything that can change its outcome, and only if the release gate checks that reused evidence describes the candidate being released. This article makes both conditions explicit and measures how often the first fails in practice. It defines when a signed evidence record is valid for a candidate and gives sufficient conditions under which change-local re-execution reaches the same release decision as rerunning every control. A reference gate implementing these checks with Ed25519 digital signatures blocks eight of nine manipulation scenarios, against one for a gate that trusts the planner and three for a gate that only verifies signatures; a dependency the policy never declared passes every gate. A replay of 38,083 commits from 714 public Data Version Control (DVC) pipelines measures that failure mode. Reusing stage results according to declared dependencies avoids 83.4% of stage executions, but in 3.3% of commits it reuses a stage whose imported code or referenced configuration changed, and 31% of pipelines are affected. A manual audit confirms the undeclared dependency in 36 of 40 sampled events. Evaluation stages, whose outputs are the evidence, are the most exposed. Adding each stage's static import closure lowers stale reuse to 0.4% while still avoiding 81.6% of executions; sampled shadow re-execution removes far less, and rerunning everything after unexplained changes avoids only 45.3%.

Solving dynamical systems in complex and diverse geometries remains challenging for both numerical solvers and learning-based surrogate models. We propose Neural Operator Splitting, a framework for learning fluid dynamics while preserving the physical structure of fractional-step numerical schemes. Instead of learning the full-step evolution monolithically, we decompose each time step into physically interpretable steps, and learn each operator separately using intermediate states from the underlying numerical solver. Guided by the linearity properties of the fractional steps, we introduce a conditional-linear Fourier neural operator (CL-FNO) whose linear state mappings are modulated by nonlinear conditioning on inputs such as boundary geometry. We further introduce two complementary strategies, symmetry sampling and corrective cycling, to mitigate error accumulation during autoregressive rollout. Experiments on 2D and 3D fluid flows with varying geometries and initial conditions show that our full model reduces mean relative $L^2$ rollout error by 68\% in 2D and 40\% in 3D compared with a single full-step neural operator. At comparable error levels, neural-operator inference is approximately $10^3$ times faster per step than the Monte Carlo solver.

This paper introduces a novel framework for efficiently evaluating in-game sound designs using multimodal learning and reinforcement learning. Sound design significantly influences gameplay and accessibility; however, objective evaluation methods, defined here as automated assessments based on agent performance rather than human ratings, often require retraining agents from scratch for each sound design. Our approach first pre-trains a multimodal encoder that encodes visual and auditory information into a shared latent space. A reinforcement learning agent is then trained using this latent representation. We fine-tune the multimodal encoder for each sound design while keeping the trained policy fixed, allowing us to isolate the effects of sound changes on agent behavior. To evaluate the proposed method, we conducted experiments on games from multiple genres, comparing the performance of blind agents (agents using only sound input) under various sound designs. Experimental results suggest that this approach can detect differences in agent performance caused by degraded or altered sound design. Moreover, we significantly reduce the computational cost compared to retraining the agent from scratch. Experimental results suggest that this approach provides a practical and scalable solution for evaluating in-game sound designs. The source code, data and supplementary document are available: https://github.com/thainv0212/tog-reusing-agents.

Deep neural networks often fail at small-sample regression, collapsing to low-variance predictions that miss individual differences. Here, we introduce a pairwise comparison framework that reframes regression as aggregated relative rankings. This method improves data efficiency across model architectures and dataset sizes. On the UTKFace dataset (23k images), pairwise learning generates over one billion training comparisons without additional supervision. The paired ranking allows a lightweight ResNet10 trained from scratch on home hardware to achieve a 3.64-year MAE-surpassing previous models, including those with large-scale pretraining. For classification into categories, pairwise supervision consistently outperforms an otherwise identical 10-way crossentropy model on STL10 with 500-2000 labeled examples, improving accuracy by up to 5.2%. Overall, pairwise supervision enhances data efficiency and enables strong regression and classification performance when labeled data are limited, particularly in clinical settings.

Individual differences in cognitive abilities vary widely, yet the structural brain features supporting these differences remain poorly understood. Traditional brain-behavior analyses rely on predefined metrics such as volume or cortical thickness and typically assume linear relationships with behavior, limiting sensitivity in small datasets. Here, we introduce a pairwise convolutional ranking framework that learns directly from whole-brain T1-weighted MRI by making comparative rather than absolute predictions, which performs well even with small samples (N=108). We tested this method with four small-scale spatial abilities tasks. Regression analysis predicted performance with high accuracy across all tasks, achieving greater predictive power over standard regression and pointwise deep learning baselines. Iterative cortical-region removal (backward elimination) revealed brain regions supporting individual performance common to all four tasks, as well as task-specific regions. These results demonstrate that pairwise learning extracts meaningful behavioral information from structural MRI, providing prediction and interpretable insight into the neuroanatomy of individual differences.

Sampling complete solutions gives large language models more chances to solve a problem, but each attempt starts again. For tasks with several subproblems, all required answers must be correct in the same attempt. Retaining intermediate results can avoid repeating correct work, but errors in those results may affect later attempts. We study when this reuse is reliable and how much computation it requires. For local retries that stop when no candidate passes within the retry budget, we give conditions for correct completion and bound the average number of attempts. We also show how checking cost and the chance of passing a check affect total computation. Experiments across eight reasoning models examine composite mathematics and arithmetic. Across four DeepSeek models, average success on the second subproblem is 36.2 percent without first-subproblem context, 57.2 percent with its answer, and 58.3 percent with its reasoning and answer. These scores measure second-subproblem coverage and include continuations made when selection fails. Arithmetic experiments show that similar first-attempt accuracies can lead to very different coverage after repeated attempts. Code is available at https://github.com/SempraETY/Intermediate-State-Reuse.

Multi-agent systems (MAS) built from language models are deployed on the premise that collaboration beats a single-agent system (SAS), yet when it pays is unclear. The most systematic study to date gives a capability-saturation rule: coordination helps while the SAS is weak and hurts once it is strong, with the crossing near 45% single-agent accuracy. That rule was derived on autonomous benchmarks, where no user takes part. We test it where agents now work: in conversation with a user under a written policy, on $τ^2$-bench, the dual-control benchmark. On its airline, retail and telecom domains we compare the stock SAS with a centralized MAS built from the same model for six models from three families, over about 7,950 simulated conversations, with one completion-token ledger per episode per system. The study was pre-registered with a public change log; the MAS is reported as built, with one validity bar unmet (its planner never saw the tool schemas). Under a fixed simulated user no cell shows a MAS benefit: the MAS is worse in 12 of the 13 valid cells and tied in the thirteenth, and the fitted gain is negative across the observed capability range (weighted mean -28 points, 95% CI [-44, -12]). On the two cells where the rule predicts a benefit, the MAS loses by 13 and 22 points, and without a budget ceiling it at best draws level at 5.5 times the SAS's token spend. The one gain appears on the cell nearest 45%, under a second simulator. Whether the gain declines with capability is unresolved (slope -0.41, 95% CI [-1.51, +0.68]), because model-by-domain effects dominate. The MAS does not detectably reduce wrongful database writes; it converts other failures into budget exhaustion and inaction. For agents that serve people, the SAS is the default; a multi-agent design must show compute-matched evidence from the setting itself. We release all trajectories, ledgers, prompts and code.

Watermarking has emerged as a key tool for identifying text generated by large language models (LLMs), with current schemes embedding statistical signals into the generation process to enable later detection. In this work, we identify a shared weakness across a broad class of popular LLM watermarking methods, including KGW (and its variants), GumbelSoft, DipMark, ITS, and SynthID. We show that these schemes exhibit a fundamental link between their detection scores and the token probability shifts they introduce during generation. This link, which we call score-ratio alignment, allows an attacker with only black-box query access to the model--and no knowledge of secret keys or model weights--to produce text that evades detection while remaining fluent and natural. We prove that the resulting attack is optimal across all covered watermark families, and show that the weakness is intrinsic: any watermark designed to maximize statistical detection power necessarily induces the very structure that attackers can exploit. We evaluate our attack on seven popular generative LLM watermarking schemes, three LLMs, and tasks, which reduce detection rates to at most 0.05 across all settings while preserving text quality. Our findings show that current watermarking methods can be reliably defeated under realistic attack conditions, highlighting the need for fundamentally different watermarking designs that decouple detection scores from token probability shifts.

Practitioners building training corpora with a pipeline of language models keep a preprocessing or refinement (post-process) stage when switching it on raises the corpus quality. We ablate a production pipeline that turns conference speech into question-answer pairs. The preprocessing stage rewrites the transcript, and the refinement stage filters the generated pairs. Switching each on and off gives four corpora from the same 40 Korean medical and finance sessions. On the corpus itself, quality rises as the stages switch on. Four LLM judges and six human domain experts rank the four corpora the same way, and the fully refined corpus leads by +0.18 and +0.22 on a 1-5 scale. We then ask whether that gain carries downstream. We supply each corpus to six models at inference and measure retrieval accuracy. The ranking does not hold. Enabling the transcript rewrite lowers retrieval accuracy by 0.061, the only stage effect that survives correction for multiple comparisons. That stage also reads the whole session, so it almost doubles the cost of one. With both stages on, accuracy does not change. An equivalence test we declared in advance confirms this. The corpus itself is clearly valuable. It lifts retrieval accuracy from 0.231 to 0.748 in every model and domain we test. More than a third of its pairs carry knowledge that neither of two frontier models produces on its own. On our data, the null does not come from weak corpora. A rising score on a rubric of this kind did not predict what the corpus holds for a model.

Recent work reports that the distribution verbalized directly in the output text aligns far better with human opinion distributions than the next-token log-probability distribution of LLMs. However, the verbalized distribution reflects what LLMs know about a group's opinions (knowledge distribution), while what they actually output at deployment is governed by their log-probability distribution (behavior distribution). We find that when prompting LLMs to predict the opinion of a specific group, the most likely answer under the verbalized distribution differs from that under the log-probability distribution for 37% of World Values Survey questions, averaged over three LLMs. Thus, LLMs' knowledge does not match their behavior. This matters for evaluation: a verbalized distribution can score well against a group while its most likely answer is one the model never outputs, so the score no longer describes the model. The alignment scores for these two distributions reflect only the model's behavior or its knowledge, separately. We therefore propose a learnable combination of the two distributions that follows the most likely answer of the behavior distribution while preserving the shape of the knowledge distribution, with a single hyperparameter trading off behavioral fidelity against distribution quality. Our experiments on three opinion survey datasets and six models show that the balanced distribution improves behavioral consistency over the verbalized distribution from 71.6% to 95.6% on average, while the total variation distance to the human opinion distributions rises only slightly from 28.8% to 33.8%, and it remains effective under temperature scaling and cross-dataset settings.

As computer-use agents (CUAs) move to long-horizon workflows spanning multiple applications, undetected errors can silently corrupt persistent state and cause cascading failures in subsequent actions. Reliable error detection and recovery have become critical bottlenecks for building robust CUAs. However, existing benchmarks focus primarily on end-to-end task performance, providing limited insight into whether agents can detect and recover from errors during execution. The few benchmarks that evaluate recovery rely largely on manually injected errors, which may not reflect the types or horizons of failures induced by real agent policies. We introduce CUA-Recovery, a benchmark comprising test and training suites across multiple desktop applications. We construct it using ReRail, a framework for synthesizing long-horizon tasks and trajectories in a persistent desktop environment. ReRail composes MyPCBench tasks into new long-horizon workflows with executable ground truth, collects verifier-confirmed failures from multiple policies, and turns them into verified recovery trajectories. CUA-Recovery comprises a human-verified test suite of 2,250 erroneous states over 100 tasks and a training suite of 8.6K trajectories over 377 tasks. Evaluation on the test suite shows that agents struggle both to recognize inherited errors and to complete the task. Across six agents and six takeover depths, Error Awareness Rate (EAR) reaches at most 44.0% and Pass@3 at most 25.5%. After fine-tuning Qwen3.5-35B-A3B on the training suite, its average EAR rises from 19.9% to 35.2% on held-out recovery states, and its Rubric Score goes from 5.0% to 16.1%. It also gains 15.3 points on MyPCBench and 10.8 points on Odysseys for Pass@3, showing that the training trajectories help both with recovery and with completing tasks from a clean start. Our code is available at https://github.com/xu-hu-2002/CUA-Recovery.

Temperature scaling is typically applied as a post-hoc transformation of model probabilities. In posterior-induced flow language models, this creates two mismatches. First, posterior tempering implicitly changes both the endpoint distribution and the distribution of intermediate noisy states, while the denoiser remains trained on the original states. Second, the tempered posterior weights are combined with transport dynamics derived for the untempered flow. These mismatches are reflected empirically: lowering temperature improves GSM8K performance much less reliably in flow LMs than in discrete diffusion LMs. We introduce Bayes-Consistent Tempering (BCT), which makes temperature explicit in both the distributions and dynamics of the generative flow. Its learned realization trains the denoiser on tempered intermediate states and tempers the endpoint distribution using the corresponding tempered dynamics. We also derive training-free variants for conventionally trained checkpoints. Across hyperspherical flow LMs with two denoiser architectures and a von Mises--Fisher flow LM, BCT improves GSM8K generation across broad temperature ranges and achieves the strongest tempering performance. The gains also extend to Sudoku, demonstrating benefits on tightly constrained generation. At very low temperatures, the learned realization is also markedly more robust than the training-free variants, supporting temperature as a property of the generative flow across training and inference rather than a post-hoc probability adjustment. Code and model weights are available at https://daioba.github.io/bct .

Protein foundation models provide transferable representations for downstream peptide prediction tasks. However, existing representations typically reduce a peptide to a single sequence or structure, which may miss conformational variability relevant to key peptide properties, including antibacterial activity, hemolysis, toxicity, stability, and binding. For example, protease resistance correlates more strongly with conformational dynamics measured by hydrogen exchange than with average helicity. Therefore, we introduce Ensemble-Aware Structural Encoding (EASE), a framework for effectively extracting structural information from conformational ensembles for peptide representation. To extract useful information from noisy conformers, EASE selects central conformers representing a peptide's major structural states and fuses their individual representations with a permutation-invariant operator to produce a peptide-level representation for downstream prediction. We then evaluate our framework on six peptide property-prediction tasks under leakage-controlled splits. Against matched single-structure baselines, EASE improves performance consistently on six tasks with a mean absolute gain of about 3.0% across two structural encoders. Gains keep growing as we add conformers through K=20. A single peptide conformer therefore provides a limited basis for structural representation, whereas our framework distilling and fusing multiple conformers yields a more informative representation.

Identifying recipe entities and linking them to ontology concepts is fundamental for semantic recipe understanding. However, food mentions in recipes are often ambiguous, context-dependent, and lexically diverse, making ontology entity linking (EL) challenging. We propose SCODEL, a streamlined, model-agnostic approach based on a central idea: contextual food mentions and ontology concepts that refer to the same entity should learn the same canonical ontology-oriented definition. SCODEL uses these shared definitions as a semantic normalisation layer, providing complementary semantic information to existing EL models without requiring a new linking architecture. We further introduce ReciLinkD, the first benchmark for contextual food ontology entity linking in recipes. Experiments show that SCODEL consistently improves three modular EL models (ReFinED, GSM-EL, and BLINK): SCODEL x BLINK achieves the highest Hit@1 on seen concepts (88.33%), while SCODEL x GSM-EL achieves the highest Hit@1 on unseen concepts (83.15%), outperforming the strongest large language model (LLM) based EL baseline in each setting. Experiments on a biomedical ontology benchmark further demonstrate that SCODEL generalises beyond the food domain. Finally, we use SCODEL to ground food mentions at scale and construct an ontology-grounded recipe knowledge graph, demonstrating its utility for ontology-aware recipe retrieval and corpus-level food analysis.

Automatically generating high-quality documentation for software repositories remains challenging: existing approaches largely treat documentation as unconstrained text generation, lacking explicit quality specifications, depending heavily on long-context prompts over raw code, and offering limited support for reproducible evaluation. We argue that code-level documentation generation should instead be a standards-guided software engineering task, in which documentation is generated and assessed against explicit, reusable quality requirements derived from authoritative standards. We define code documentation as natural-language descriptions of repository functions and classes, organized by source-code path. We present CODEDOC, a standards-guided and dependency-aware framework that identifies dependencies between code entities via static call-graph and AST analysis and constructs dependency-aware structured prompts combining local code context, repository paths, and documentation criteria drawn from authoritative standards and style guides. We further introduce an interactive LLM-as-a-judge protocol that supplies criterion-aligned feedback to refine generated documentation along fine-grained dimensions from formal and practice-oriented standards. Experiments on ten open-source Python repositories show that CODEDOC produces higher-quality documentation than RepoAgent and a long-context baseline under standards-aligned evaluation, with relative improvements of 11.6%-17.0% and 21.4%-29.6%, respectively, while consuming roughly an order of magnitude fewer tokens per documentation entity than the long-context baseline. A human study confirms substantial alignment between LLM-based and human judgments (Spearman's rho = 0.72), and the relative findings hold across different judge LLMs. Code is available at https://github.com/ControllableModel/CodeDoc.

A learned scientific controller can outperform a misspecified planner without benefiting from feedback. We audit two synthetic voltage-clamp tasks with a training-selected fixed protocol, current masking, a Bayesian reference, and an excluded target. Five recurrent policies repeat one fixed protocol in all 1,200 off-grid rollouts; masking current changes no decision. In a crossed four-hypothesis task, a fixed diagnostic pulse is followed by one learned choice. Policies then use the first current, but tie a correctly specified Bayesian controller, and a fitted threshold reproduces the learned actions without being that controller. The excluded-target prediction gain is $4.6991\times10^{-6}$ nats per sample. Feedback use and protocol quality are separate tests. These cases establish neither biological discovery nor general RL superiority.

Systems for code-mixed (CM) text and speech rarely start from scratch. CM data is scarce, so most work borrows from the languages being mixed, which we call native languages. This transfer can also run in reverse, when CM data is used to improve models for the native languages. Existing surveys focus on processing CM itself, and none brings the two directions together. This paper fills that gap with a systematic survey of transfer in both directions, across text and speech. We organise the surveyed work in a taxonomy of transfer direction, route and modality. Through this taxonomy, we find that most transfer runs from native languages into the CM setting. The reverse direction is younger, and in it code-switching mainly helps by aligning the two languages. Text and speech work share methods but rarely cite each other. On a shared speech benchmark, native data helps more than new architectures. These findings point to open challenges, such as how much native and CM data to mix in training, and how to evaluate CM text written in different scripts.

Looped language models can run more iterations at test time, but additional iterations need not help them solve longer problems. For sequential state tracking, final-answer supervision does not specify how far each iteration should advance, and repeated updates can change the scale of the recurrent state. We introduce \textbf{Progress-Aligned Recurrence (PAR)}, which trains the prediction at iteration $k$ to match the state after the first $k$ operations and rescales recurrent states to their first-iteration norms. On swap-$S_5$, a Qwen2.5-0.5B model trained on at most 32 moves achieves 95.5\% mean full-state accuracy at 512 moves, or $16\times$ the training length, compared with 55.7\% reported for RLT and 0.8\% for explicit chain of thought on the same backbone. Controlled $S_5$ ablations show that removing norm anchoring reduces the extrapolation factor from $16\times$ to below $2.2\times$, and replacing progress-aligned targets with final-answer supervision reduces it to at most $8\times$. A learned completion readout selects when to answer without a supplied stopping depth. When intermediate labels are unavailable, prefix self-distillation induces the correspondence from shorter problem prefixes using only final answers. The gains apply across sequential state-tracking tasks, establishing the task structure required for depth extrapolation. These findings support training intermediate predictions and state scale together to make additional recurrent depth useful for longer problems.

Multimodal Large Language Models show strong vision-language capabilities, yet their geometry-centric reasoning remains insufficiently understood. Existing mathematical benchmarks typically rely on rasterized figures and cover limited geometric knowledge areas and response formats, making it difficult to assess whether models reason over geometric structures or exploit superficial visual patterns. We introduce TriGeoBench, a geometry-centric multimodal benchmark constructed from Chinese middle- and high-school mathematics examinations. TriGeoBench contains 7,000 question-answer pairs and 12,235 geometric figures with TikZ code, spanning plane geometry, solid geometry, analytic geometry, function, and vector geometry across diverse response formats. Each instance forms a {question-answer pair, image, TikZ} triplet: question-answer pairs are stored in LaTeX, images as RGB figures, and TikZ preserves geometric topology, symbolic annotations, and relational structure. We evaluate representative open- and closed-source MLLMs through image-to-TikZ and multimodal reasoning probes. Results reveal persistent fine-grained structure and image-text alignment challenges. Oracle experiments with human-curated reference TikZ suggest that accurate executable representations provide complementary geometric cues; they do not evaluate an end-to-end conversion pipeline. TriGeoBench offers a testbed for structure-aware mathematical reasoning.

Reinforcement learning for long-horizon language agents requires identifying whether intermediate actions make eventual task success more or less likely when supervision is available only at trajectory level. The main difficulty is that a critic with aggregate return-prediction accuracy may still fail to resolve the local value differences needed to assign credit to actions, especially under sparse feedback and function approximation. Conventional critics estimate future return, but policy improvement also depends on knowing whether each intermediate decision makes task success more or less likely. We introduce SCALE, an actor-critic framework that augments conventional return estimation with explicit prediction of eventual task success. At each interaction prefix, the critic predicts a return value and a Success Value, defined as the probability that the task will eventually be completed from that prefix. SCALE converts changes in Success Value into multi-step action credit through Success-GAE and modulates its contribution according to the estimated reliability of the predictor, while retaining the token-level PPO objective. Across five agent benchmarks and three model backbones, SCALE improves task performance over the compared RL baselines in most settings, with gains of up to 4.1 percentage points in success rate. Separate critic-learning analyses show that SCALE also improves critic estimation and training stability.

Handheld demonstration devices decouple data collection from the robot, but conventional designs burden the hand with a wrist camera, cables and a pose tracker (SLAM, an IMU, a VR controller, or a fixed external camera that the collector's body occludes), and burden the pipeline with as many data streams to timestamp, synchronise and cross-calibrate. We show that a single egocentric camera (USD 122) worn on the collector's head replaces the wrist camera, the pose tracker and their separate streams with one. It reads printed fiducials on an electronics-free glove (USD 22) against paper tiles taped around the workspace, giving the 6-DoF gripper pose and the jaw width from the same frames the policy will see; at deployment the same camera rides a low-cost (USD 120) arm beside the robot that mimics the collector's head from the robot's joint state. Against the robot's encoders the labels are accurate to 6.5 mm and 1.6 degrees median with 93-100% frame coverage, and a policy trained on them instead of encoders pays 3% in validation loss. Adding 194 glove demonstrations to 48 teleoperated episodes lifts rollout success from 10/25 to 17/25, level with 172 teleoperated episodes alone (18/25), and the full co-trained head-camera policy succeeds in 20/25 against 15/25 for the wrist-camera recipe, whose policies drop the object when the goal leaves their view, and 12/25 for a fixed camera.

Deep learning (DL) models are increasingly being deployed into radiation-challenging environments in spacecraft, particle accelerators, and, soon, space-based data centers. Instantiated as a learned orchestration of numbers stored in persistent memory, models are susceptible to continual degradation as high-energy particles corrupt data. At inference time, new data passed through the network fundamentally break the off-line assumptions and guarantees of a static model determined during training. To preserve the quality and efficacy of the models, fault tolerance against bit flips is a clear objective. In this work we developed a runtime method for repairing the quantized weights of neural networks (NNs) after bit flips on commodity, commercial off-the-shelf (COTS) hardware, without storing extra replicas. Our work extends the well-established Fault-Aware Training (FAT) paradigm by adding an alignment penalty to concentrate undamaged activations and weights into a low-rank subspace, and we then reconstruct flagged corrupted coordinates from the uncorrupted survivors. We call this method {\it Projection Healing} and apply it to six regression benchmarks on quantized NNs. Holding a FAT checkpoint fixed and toggling only the runtime heal gives a further ${\sim}23\%$ reduction in the native-unit excess integral. Across those datasets the method extends the operational scrub interval by a median of $2.04\times$ at a $+5\%$ RMSE degradation and $1.96\times$ at $+50\%$, compared with $1.62\times$ and $1.07\times$ for classical Single Error Correction, Double Error Detection (SECDED). This introduces a new path towards designing resilient data-driven models, complementing established hardware redundancy, ECC, and FAT-based techniques.

Autonomous LLM agents act through tools under per-action authorization: a local guard screens each tool call in isolation. We show that local authorization does not compose into global authorization. A compositional authorization failure (CAF) is a locally authorized execution, or a pair of locally authorized executions, that violates a global policy spanning intent, purpose, provenance, consent, and delegation. We characterize CAFs along two axes, expressiveness (single-trace safety versus relational k-safety) and observability (a ladder of monitor profiles). Our central result is an impossibility: relational purpose violations form a 2-safety hyperproperty with no per-trace locus of blame, so a per-action monitor that never raises a false alarm detects none of them, independent of the underlying model or decision procedure; they become visible only under cross-execution monitoring, which none of the six guardrails we survey implements. We make this constructive with AL-SYNTH, a SAT/SMT synthesizer (using no language model) that emits minimal CAF witnesses, each checked against an exhaustive oracle. The empirical signature is sharp: across nine LLM judges the per-action view approves the synthesized witness almost without exception, blocking it at a rate comparable to a benign control, yet the same models flag it when shown the two executions together. Per-action authorization is therefore structurally insufficient for relational policy conjuncts not reducible to single-trace safety: to bound them, a monitor must compare executions or accept false alarms.

Complex chemical systems containing heterogeneous substructures are common in real-world applications. Various physical phenomena of the complex chemical systems are derived from the interactions between the heterogeneous substructures. However, existing graph representation learning methods for inter-graph interactions assumed graph-level interactions between homogeneous structures, such as organic molecules and inorganic crystalline materials. We propose a data descriptor of the complex chemical systems and a graph neural network for learning inter-graph interactions between organic and inorganic compounds. We applied the proposed method to predict the physical properties of hybrid solar cell materials containing heterogeneous substructures, which have received significant attention for sustainable energy resources. By learning heterogeneous inter-graph interactions, the proposed method achieved state-of-the-art accuracy in predicting band gaps of 1,682 hybrid solar cell materials.

Ensuring the alignment of Large Language Models (LLMs) with diverse cultural values is critical for their safe global deployment. However, existing alignment benchmarks are predominantly English-centric or confined to simple discriminative tasks. Moreover, they often fail to effectively evaluate model robustness against complex, adversarial inductions within specific cultural contexts. To address these limitations, we propose ABench, a fine-grained, handcrafted benchmark specifically tailored to taxonomized values in Chinese. Grounded in a rigorous hierarchical taxonomy comprising 158 derived value subcategories, ABench consists of 2,895 open-ended questions. These questions employ techniques such as role-playing and logical traps to probe the stability of models' stances in adversarially induced contexts. Furthermore, we propose a weighted automated evaluation framework that utilizes question-specific scoring criteria for nuanced assessment. Extensive experiments on 16 mainstream LLMs reveal that most models exhibit significant vulnerabilities when subjected to sophisticated adversary attacks on values (scoring below 60). This underscores the need for more robust, culturally specific alignment strategies. Our data will be publicly released to promote further research on taxonomized value alignment.

Recursive self-improvement (RSI) promises AI systems, such as self-evolving LLM agents, that become better at producing further improvements. Evaluating this promise requires separating product claims about the final system from process claims about the improvement process itself. Existing surveys specify the evidence these claims require, but how often studies report that evidence remains unmeasured. We examine this question through full-text coding of 114 eligible studies in a reproducible arXiv search frame frozen in October 2026, including 90 that describe their own systems as RSI. The codebook links claims to evaluation designs. Independent human coding, third-reader adjudication, and a separate blind coding of improver comparisons assess coding reliability. Among the 90 RSI studies, 90% report product evaluation beyond development tasks under a broad criterion that includes partial or unstated separation, while 60% report evaluation on tasks unused for selection. Only 22% audit two or more reruns of the full loop. In a human-coded random sample, the RSI studies show the same contrast: 45% evaluate on tasks unused for selection and 14% audit two or more reruns. Of the 26 studies claiming a better improvement procedure, 23 report a related comparison, but only 4 compare earlier and later procedures from identical starting artifacts at a matched budget. Across these results and a comparison set of established works, evidence for repeatability and improved procedures remains less developed than product evaluation. Three propositions explain how selection effects, shared evaluator errors, and changes in system state limit what product evaluations can establish. Together, the findings support a claim-indexed reporting standard that connects each claim to the evaluation design needed to test it.

Offline reinforcement learning is fundamentally limited by the coverage of the available data, while collecting additional trajectories through online exploration may be costly, unsafe, or infeasible. We study whether this limitation can be overcome without directly interacting with the underlying environment. Motivated by practical settings such as collaborator-mediated simulation, where a collaborator can provide information complementary to the offline data, we introduce query-based collaborative reinforcement learning to study how to acquire this information efficiently. The learner starts with a fixed offline dataset and adaptively queries state-time pairs; the collaborator samples an action according to a fixed, possibly suboptimal policy, and the simulator returns the corresponding reward and next state. For finite-horizon tabular problems, we derive an instance-dependent information-theoretic lower bound on the additional queries required to identify an $ε$-optimal policy and develop a local quadratic approximation. Guided by this characterization, we develop an adaptive query-allocation algorithm whose query complexity nearly matches the local benchmark under suitable conditions. Experiments on multiple synthetic and semi-synthetic environments demonstrate the empirical efficiency of the proposed method.

Empathetic interaction involves two distinguishable capacities: representing another person's emotional state (cognitive empathy) and generating an appropriate response to that state (affective empathy). We apply this distinction to examine how emotion-related information is internally organized, selectively attended to, and carried through response generation in Llama-3.2-3B-Instruct across 10 emotion categories from the EmpatheticDialogues dataset. Layer-wise linear probing revealed a broad middle-layer region of peak emotion-category decodability, with the layer-11 L2 probe achieving 63.6% held-out accuracy against a 10% chance level and a 54.6% TF-IDF lexical baseline. During response generation, emotion-related user tokens received 1.89 times as much attention per token as other utterance tokens, concentrated across layers 10-15. Target-emotion alignment then decreased from a mean cosine similarity of 0.239 for input representations to 0.125 for generated responses, with positive drift in 77% of examples and the largest shifts for angry and furious. Emotion-category representation, attention allocation, and response-level alignment therefore emerge as distinguishable components of model processing that can be analyzed through a functional cognitive-affective empathy framework. These findings provide a foundation for monitoring emotion representations during generation and leveraging models' existing affective structure to support more naturally empathetic responses without relying on external prompting or explicit emotion conditioning.

Diffusion language models (DLMs) are emerging as a promising paradigm for code generation, offering the potential for fast, parallel decoding. However, generating complex programs remains challenging, as their interdependent components must maintain consistent interfaces, data flows, and shared assumptions. Aggressive parallel decoding can compromise these dependencies and undermine program correctness, while conservative decoding sacrifices efficiency. To address this trade-off, we propose planning program decomposition before implementation to coordinate interdependent units during parallel generation. We introduce Self-Decomposing Diffusion for Code (SDDC), a novel framework that decomposes complex programs and implements their units in parallel, entirely through diffusion. SDDC first generates a contract specifying the program's decomposition and inter-unit dependencies, then uses it to guide coordinated parallel implementation. Across five benchmarks and two 7B diffusion backbones, SDDC substantially improves the accuracy--latency trade-off, raising pass@1 by up to 31.1 percentage points while decoding 7.5--13.7$\times$ faster than the base models, and outperforming a fine-tuning baseline by up to 23.5 points with lower latency.

As model parameters scale up, sparsely activated Mixture-of-Experts (MoE) approaches have become essential for model scaling. However, mainstream models currently adopt hierarchical architectures, where experts in different layers, although topologically independent, exhibit substantial similarity and redundancy in their stored content. Compared to the hierarchically arranged \textbf{Local Expert} in existing mainstream models, we hypothesize that inter-layer shared \textbf{Meta Expert} hold greater potential. To this end, we conduct a multi-dimensional empirical comparison between Local and Meta Experts, and find that models incorporating Meta Experts exhibit lower parameter redundancy, together with lower cross-entropy loss and higher downstream task accuracy. With the activated quantity and the expert granularity held fixed, \textbf{our model matches the performance of Vanilla MoE using only 70--80\% of the experts}, substantially improving the model's scaling potential. Beyond removing inter-layer redundancy among experts and improving information storage efficiency, our theoretical analysis further shows that, compared with Local Experts, Meta Experts provide hidden states with more flexible residual pathways, thereby improving information retrieval efficiency.

Existing reasoning evaluations typically provide task-relevant information to models as available premises, rarely examining concurrently how models discover evidence autonomously and how information presentation within narratives guides or obscures judgment. This paper proposes the WORTH Benchmark: models read original mystery narratives without clue annotations, autonomously select key clue sentences, identify case participants, and explain their reasoning process. The narrative materials are designed by professional authors such that evidence facilitating reasoning coexists with narrative arrangements that delay the revelation of truth. We report a pilot experiment under full-story conditions across two episodes, conducting a total of 70 formal runs on models, of which 67 runs are content-scorable. Scoring separately records hits on author-designated key clue sentences and participant selections; natural language explanations are used for case diagnosis and are not scored independently. The results show that clue hits and participant judgments are not synchronized; across 37 content-scorable runs in Episode 02, 31 runs simultaneously selected the same set of incorrect participants, while adjacent suspicion-oriented and exoneration-oriented sentences were selected 25 times and 0 times, respectively. These phenomena illustrate the diagnostic value of component-wise analysis and suggest that not-fully-cooperative informational relationships may influence evidence selection and judgment. This study only implements full-story conditions for two episodes, and does not establish stable model rankings or infer general mechanisms on this basis; human player records have not yet been cleaned, and this version does not report human-model comparisons.

When unique pretraining data is scarce, models see the same tokens many times and are thought to memorize. Random-token replacement, an input-corruption augmentation, is one remedy reported to lower validation loss, but a lower loss alone does not reveal which source of error it reduces. On a controlled source whose entropy rate is known in closed form, we show that replacement shrinks predictions toward the corpus average. We also introduce a decomposition that separates validation loss into approximation, estimation, and memorization terms. This shrinkage view predicts that replacement's bias cost needs no memorization but its benefit does. Memorization raises validation loss after its minimum, so we report the effect at two points, the minimum validation loss and the end of a fixed budget. Where replacement helps, it reduces this rise by about 60%. At the end of the budget, it turns from harmful to helpful as the controlled source becomes less predictable; at the minimum, it helps at no predictability level under the fixed recipe. On Korean and Japanese web text, replacement removes about a third of the loss rise after the minimum by the end of the budget, but its effect at the minimum is much smaller. Across ten real corpora, replacement lowers the final loss on nine. It harms only the census corpus, program-generated JSON that is the most redundant of the ten and much more redundant than any of 64 further prose and code corpora we measure.

LLM agents have become capable kernel engineers on mainstream GPU stacks, but they struggle on emerging accelerators such as Ascend NPUs, whose programming model is barely represented in pretraining data. Recent NPU agents address this gap largely through experience: they accumulate memory from their own generation and optimization trials and reuse it on later tasks. We argue that experience alone is not enough for kernel generation. An agent can only remember what it has tried, and it rarely tries a mechanism it does not know. As a result, such memory offers little help when a new operator calls for a different mechanism, and even familiar operators raise unseen situations as optimization proceeds. We present ASCENDEVOLVE, which grounds every step of an evolutionary kernel search in a large external knowledge base built from platform documentation, toolkit headers, and operator repositories. Because this corpus far exceeds any context window, ASCENDEVOLVE indexes it and routes each agent to a small, task-specific slice of evidence instead of the corpus itself. The retrieved knowledge seeds diverse implementation strategies, guides repair when a candidate fails, and constrains how strong candidates are recombined. On 53 Ascend C operators from CANN-Bench spanning four difficulty levels, ASCENDEVOLVE achieves an overall score of 71.39, improving over Reflexion and OpenEvolve by 11.6% and 17.2%.

Online adaptation of the cornering stiffness keeps a model-based sideslip estimator valid as tires change, but competes with learned model correction for the same filter innovation. A corrector with enough parameters can absorb the innovation pattern that a stiffness change produces, and adaptation is suppressed while the fit improves. We propagate the stiffness sensitivities of the predicted measurements through the vehicle states and construct a corrector orthogonal to them. A null-space parameterization enforces the constraint without a penalty coefficient and preserves the static least-squares stiffness estimate on each identification segment. In full-vehicle simulation, the stiffness response of the unconstrained corrector declines as its parameter count grows, whereas the orthogonal corrector responds at least as fully as the uncorrected filter at all sizes and lowers sideslip error by 12.7% to 22.0%. In a simulated drive with a stiffness step, the orthogonal corrector adapts as quickly as the uncorrected filter and lowers sideslip error during and after the transient. On public real-vehicle data, the 430-parameter unconstrained corrector loses most of its response to deliberate offsets of the filter's nominal stiffness, whereas the orthogonal corrector of the same size retains most of it, with a sideslip error slightly above the uncorrected filter's.

As cognitive radars evolve towards the end-to-end processing of one-dimensional complex baseband sequences, deep networks face severe adversarial threats. However, existing attacks encounter dual limitations in practical electronic warfare: low cross-architecture transferability and severe vulnerability to physical channel distortions, which degrades their tactical reliability. To overcome these bottlenecks, we propose a physical prior-driven Segmented Spatio-Temporal Attack (SSTA). First, a global-local feature decoupling segmentation mechanism interrupts the adversarial gradients' dependence on long-range sequence coherence, bridging the cross-architecture feature gap. Second, moving beyond static norm constraints devoid of physical semantics, we abstract channel distortions into differentiable spatio-temporal operators, integrating them into momentum iterations via an Expectation over Transformation (EoT) strategy. Comprehensive cross-evaluations across 9 heterogeneous models spanning CNNs, RNNs, and Transformers demonstrate that SSTA outperforms the state-of-the-art baseline by 10.36% in Average Black-Box Attack Success Rate. Furthermore, baseband equivalent simulations confirm that the generated waveforms survive extreme physical distortions and, after matched filtering, achieve zero main-lobe attenuation (0 dB) with noise floor lift strictly within Constant False Alarm Rate tolerances. SSTA successfully unifies digital-domain deception and physical-domain fidelity, providing crucial support for CEW security assessment.

Diffusion language models (DLMs) are emerging as a practical alternative to autoregressive language models (ARLMs), yet their fairness remains largely uncharacterized. DLMs introduce sampling controls with no direct autoregressive counterpart, such as block length and remasking strategy, that directly set their speed--quality operating point. This raises a basic question: does a DLM's measured social bias remain stable across these settings? Across LLaDA, LLaDA-1.5, and Dream on Open-BBQ, we find that all three models exhibit social bias, but both the magnitude of measured bias and the relative ordering of models by bias are highly sensitive to the sampling configuration. This sampling dependence also changes what successful debiasing means. While mitigation strategies established for ARLMs transfer effectively to DLMs and reduce bias without degrading downstream performance, sampling dependence persists even after mitigation. Our results show that fairness in DLMs cannot be characterized independently of the sampling configuration, even after debiasing. Project page is available at https://pakapaka333.github.io/Beyond-Model-Weights/.

Diffusion large language models (dLLMs) predict multiple token positions per forward pass, yet block-wise decoders serialize generation and repeatedly recompute stable predictions. We attribute this inefficiency to coupling a prediction's readiness for irreversible commitment with its utility as context. We introduce Tentative Context Pipelining (TeCoP), a training-free decoder that decouples these roles through staggered block-parallel pipelining and draft-and-verify. TeCoP reuses stable but uncommitted predictions as provisional cross-block context, with commitment governed by confidence and stability checks. Across four LLaDA and Dream dLLMs on GSM8K, Minerva MATH, and MBPP, TeCoP reduces number of forward evaluations by up to 87% and achieves up to 7.50-fold throughput without KV caching and 11.77-fold with caching over the uncached original decoder. It outperforms Fast-dLLM and DSB in throughput across all 12 model-benchmark settings, with and without caching, while keeping average cached accuracy within one percentage point of the original decoder.

Full-duplex speech agents must update persistent memory while listening and speaking, yet it remains unclear which events should be stored, revised, or excluded. Existing benchmarks isolate interaction timing from memory outcomes, obscuring the state transitions behind a correct answer. We introduce DuplexLedger, a bilingual benchmark of 1,094 persistent sessions aligning four full-duplex scenarios with event-level Write/Update/Ignore/Retain supervision, explicit pre/post memory states, and 6,564 same-session questions and checks. Evaluating five full-duplex models reveals an interaction-memory gap: no system combines strong boundary control with coherent final-state memory, and the best final-state fact coverage is 2.01%. We propose ACT-Mem, an 8.92M-parameter sidecar comprising an event controller, a versioned transactional ledger, and a typed retriever coupled with a frozen MiniCPM-o 4.5 backbone. On 219 held-out sessions, ACT-Mem increases latest-value accuracy from 4.11% to 8.68% while reducing stale assertions and distractor adoption. It improves interaction boundaries but slows interruption stopping. These results motivate selective, auditable full-duplex memory rather than indiscriminate transcript retention. Code is available at https://github.com/lyx-researcher/DuplexLedger.

Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR) models the subject position enriches attributes largely independently of the relation, and relation-specific selection happens later, at the readout position. In masked diffusion language models (DLMs) we find the state is assembled differently, and much earlier: in the first attention layer the wording of the relation is routed into the position of whichever entity is visible and the early feed-forward layers look the fact up on this fused representation. The state therefore varies with the relation's surface form and with the query direction, so the queries about one fact occupy a region of state space rather than a point, and an operation on the fact succeeds only for queries that enter it. We treat the state as an interface to a DLM's factual knowledge, and put it to two uses. Writing to the state gives a knowledge editor: from one declarative statement, Multi-View Edit fits one shared value per direction across that region and installs it in the early-MLP block, preserving source-prompt reliability while substantially improving paraphrase and reverse question answering, where AR editors gain near zero. Reading it suggests a complementary use: the state at each entity position can serve as a relation-specific retrieval vector with no training, returning both pieces of evidence for two-entity queries on which pooled, prompted and lexical retrievers almost never return both.

With increasing autonomy, Large Language Models (LLMs) may engage in deceptive behaviors, such as scheming to pursue misaligned goals or sandbagging to escape oversight. Monitoring outputs alone is insufficient, as models can produce seemingly benign responses while harboring misaligned internal objectives. While mechanistic approaches like linear probing are promising, they rely on the assumption that deception corresponds to a single direction in representation space. We demonstrate that such single-direction probes fail to generalize across diverse deception scenarios. To bridge this gap, we propose representing deception as a multi-dimensional subspace rather than a single direction. Our method utilizes iterative orthogonal decomposition to identify complementary directions that capture distinct facets of deceptive intent. To effectively leverage these directions for detection, we find that static combinations prove brittle due to strong domain dependence across different scenarios. We therefore introduce a domain-adaptive gating mechanism that dynamically composes the extracted directions through few-shot adaptation at test time. With only 10 examples, our approach achieves an average AUROC of 0.97 and 94% recall at 5% FPR across all evaluated datasets, substantially outperforming existing methods.

Model-based reinforcement learning in partially observable environments requires inferring hidden states, learning latent dynamics, and planning under the resulting belief. Practical algorithms often solve one aspect of the above pipeline while relying on approximations for the others that are often poorly understood, while existing POMDP theory typically analyzes computationally inefficient theoretical algorithms far from modern deep RL practice. To fill this gap, we propose a practical end-to-end algorithm that approximately solves the hidden state filtering problem via sequential Monte Carlo (SMC), updates a latent dynamics model online, and performs approximate online planning for $d$ steps before bootstrapping leaves with a state-value critic (a QMDP approximation). We analyze each approximation and prove an end-to-end regret bound decomposing the error into model estimation error, particle filter error, Monte Carlo rollout error, and a structural bias term induced by the QMDP approximation controllable by the planning depth that we further analyze in various special cases. The resulting guarantee makes explicit how statistical sample size, particle budget, rollout budget, and planning depth affect performance. We instantiate our algorithm in an MPC-based deep RL framework, demonstrating its practicality on a variety of environments. On a POMDP that satisfies our assumptions, each term of the bound decays at its predicted rate. Our results provide a practical and theoretically grounded route for bridging theory and practice within online decision-making in POMDPs.

Multi-agent systems are increasingly deployed to coordinate multiple large language model (LLM) agents. However, this also creates new attack surfaces: malicious agents may have chances to exploit the coordination. In this paper, we propose and study the \textit{show-up attack}, in which a malicious subagent manipulates the supervisor into involving it in tasks beyond its assigned scope. To evaluate this risk, we propose SUA-Arena, which implements 6 attack strategies and contains 188 test cases. Then, we evaluate 6 frontier LLMs and find that all of them are vulnerable to show-up attacks. For example, GPT-5.6-Sol as a supervisor suffers an ASR of up to 0.93. Mitigation strategies such as prompting and restricted communication fail to reduce the risk consistently. These results show that multi-agent systems need explicit mechanisms for enforcing authority boundaries.

Large language models (LLMs) can influence user trust and decision-making through subtle linguistic cues, yet the behavioral effects of manipulative outputs remain poorly quantified. We introduce a behavioral computing framework for measuring \textit{behavioral vulnerability}, defined as users' susceptibility to manipulative AI responses. Building on DarkPatterns-LLM benchmark, we generate controlled manipulative stimuli using Manipulation Gradient Descent (MGD), targeting biases associated with authority, urgency, and social proof. We analyze participant responses using the Multi-Stakeholder Impact Assessment Network (MSIAN) to model cross-user influence and the Temporal Harm Propagation Model (THPM) to characterize the persistence of behavioral effects over time. In a study involving 120 participants, manipulative framing produced significant changes in trust calibration and compliance ($p < 0.01$), with effects also propagating across social contexts. These findings highlight the need to extend LLM safety evaluation beyond model outputs to measurable behavioral consequences and provide a computational framework for quantifying individual, social, and temporal risks arising from manipulative AI interactions.

Porous-material adsorption studies require coordinated structure preparation, force-field selection, simulation, validation, and interpretation, making end-to-end automation difficult. We test whether RaspAssistant, a hierarchical multi-agent framework, can carry an adsorption study from force-field preparation through validation and downstream separation analysis. Across twelve multi-agent workflow-reproduction runs spanning four adsorption and separation benchmarks, overall scores ranged from 8.50 to 10.00 on a rubric evaluating workflow correctness, simulation validity, and analysis consistency. In a separate eight-case force-field extraction benchmark, the mean weighted score was 0.967, combining parameter agreement and RASPA readiness. In open-ended studies, the system diagnosed and refined an acetylene force field and, after validating a CO$_2$ force field against MFI experiments, screened 197 pure-silica zeolites that passed automated input checks. These results demonstrate autonomous workflow execution after task specification and input provision: RaspAssistant coordinates, validates, adapts, and documents extended adsorption-simulation workflows while exposing model sensitivity and methodological limitations for expert review.

Bayesian neural networks (BNNs) offer a principled framework for quantifying epistemic uncertainty. However, standard weight-space priors yield complex, high-dimensional neural network posteriors that are computationally prohibitive to sample from, and popular approximate inference schemes often fail to capture epistemic uncertainty effectively---a failure mode particularly acute in reinforcement learning (RL) \citep{osband2018randomized}. To resolve this challenge, we introduce a new class of BNNs that quantify uncertainty by placing Bayesian nonparametric Dirichlet Process (DP) priors directly on the underlying data-generating distribution rather than on network parameters. Building on this foundation, we develop DP-DQN, a reinforcement learning algorithm that leverages DP-BNNs to maintain posterior distributions over value functions and perform efficient Thompson sampling. By grounding randomized value functions in Bayesian nonparametrics, DP-DQN unifies data-driven posterior reweighting (via the Bayesian Bootstrap) with generative prior hypotheses injected directly through the DP base measure, enabling decisive ``deep exploration'' with a single neural network without the overhead of multi-network ensembles.

Tabular foundation models (TFMs) pretrain a transformer on large-scale synthetic priors so that a single forward pass yields zero-shot tabular prediction. TabICLv2 samples the class count of its classification prior during pretraining, and the prevailing configuration caps it at ten. We report a counter-intuitive result: setting this cap directly to two---making every pretrained classification task binary---still yields weights that perform zero-shot K-way classification 2. On the 38 TabArena classification datasets (200 bootstrap rounds, anchored at default RF = 1000), the binary two-stage recipe reaches 1590.1 Elo, above the official default at 1568.0 (+22.1), above TabPFN-v2.6 (1577.1) and RealTabPFN-v2.5 (1521.6). A production-scale validation on 60 real-world credit-scoring institutions independently confirms the direction: the same checkpoint improves test AUC on 44 of 60 institutions (paired Wilcoxon p = 0.0007 for AUC, p = 0.0002 for KS) at +1.5 inference cost. Mechanistically, classification in TabICLv2 is carried not by a fixed-cardinality output head but by an in-context comparison operator matching query-row embeddings against class-conditional embeddings, so the cardinality of the output space is set by the labels present in the context, not by the label space seen during pretraining. This makes a binary prior structurally cheaper to generate and cache than a ten-class prior, demotes "number of classes" from an architectural constraint to a tunable hyperparameter, and points to a class-count-agnostic regime of TFM pretraining.

Neural combinatorial optimization has enabled fast constructive solvers for the Capacitated Vehicle Routing Problem (CVRP), yet most autoregressive models still construct solutions through sequential node selection and represent route organization only implicitly. We investigate whether route-level customer co-assignment can be modeled explicitly and used to guide this node-level decision process. To this end, we propose TopologyAware RouteConsistency Modeling (TRCM), which reuses node representations from a standard autoregressive construction pipeline to predict pairwise same-route probabilities and injects them into customer selection as a route-consistency signal. Because sameroute supervision is both class-imbalanced and structurally uneven, TRCM combines structure-preserving augmentation with Topological Entropy (TE)- based reweighting to emphasize locally ambiguous customer pairs. Under matched decoding budgets, TRCM reduces the optimality gap from 3.67% to 3.28% on CVRP50 and from 2.11% to 1.99% on CVRP100; under 8-fold augmented inference, it reduces the CVRP100 gap from 1.61% to 1.49%. On XML100, TRCM*8Aug improves the average objective by 0.24%, with larger gains on several structurally informative factors. A controlled demand-reassignment test over 500 variants further shows more concentrated objective fluctuations. These results support explicit route-consistency modeling as an effective structural principle for autoregressive neural routing.

Depression assessment from counseling speech requires integrating textual and acoustic psychological evidence across turns, yet existing end-to-end classification approaches lack explicit multimodal evidence interpretation and context-aware reasoning. We propose AudiPsy, an evidence-grounded omni-LLM for depression assessment over spoken counseling conversations. AudiPsy employs an Observation-Assessment-Summary (OAS) paradigm to integrate dialogue transcripts, textual and acoustic emotions, and textual-acoustic emotion incongruence into structured reasoning. OAS incrementally organizes multimodal observations, performs contextualized assessment, and maintains an evolving dialogue summary for subsequent turns. To support this framework, we construct AudiPsy-Counsel, a synthetic spoken counseling conversation corpus with multimodal context augmentation, and develop a unified three-stage curriculum for depression-oriented reasoning. Experiments on DAIC-WOZ demonstrate superior depression assessment performance over specialized depression detection methods and general-purpose speech/omni-LLMs, along with improved reasoning quality, while supplementary evaluation on EATD-Corpus supports AudiPsy's applicability to naturalistic conversations.

Large language models (LLMs) are increasingly used within everyday applications where their responses have significant downstream impacts. For responsible deployment of such systems, we desire insight into how much we can trust their individual outputs. A popular approach for estimating LLM confidence is to train probes on the hidden states of the LLM's tokens to predict the correctness of the model's output. These probes typically read hidden states from tokens that are already in the model's vocabulary. In this work, we introduce a new approach to confidence probes that learn a new readout token and utilize its hidden states to train the probe. We show that our learned readout approach ScratchProbe outperforms its existing-token counterpart on a variety of datasets and LLMs. We also extend our method to a multi-token approach which we show provides further performance improvements. Finally, since the learned readout changes only the position a probe reads from, any existing probe can adopt it. Applied to a state-of-the-art probe, it improves AUROC on every backbone, in domain and out of domain.

Recent multimodal LLM (MLLM)-based anomaly detection (AD) methods can generate textual explanations that help interpret their predictions. However, reliably and efficiently evaluating these explanations remains challenging. While LLM judges can assess explanations through structured rubrics, they often require substantial token generation, making evaluation costly and difficult to scale. We instead propose to anchor evaluation in the available ground truth by extracting verifiable claims from each explanation and verifying them against the corresponding annotations. We instantiate this idea as AnomalyJev, a lightweight framework that extracts verifiable claims from contextualised representations using ground-truth-defined queries and pointer based heads, without token generation. The claims are then verified against the ground truth to score faithfulness and completeness. For evaluation, we construct a human-annotated set of 640 explanations with over 1,600 expert ratings across in-domain and cross-dataset settings. Despite its compact design, AnomalyJev achieves stronger agreement with human annotators than most open MLLM judges and approaches that of flagship MLLM judges, at a fraction of their runtime and resources.

Generating high-fidelity, city-scale 3D environments plays a crucial role in applications such as virtual reality, gaming, and urban digital twins. However, the scarcity of real-world city-scale 3D training data forces existing methods to either stitch independently generated assets or learn from synthetic urban corpora, often sacrificing realism and global consistency. In contrast, 2D image data and their corresponding generative models are abundant and encode rich visual priors. To bridge this gap, we propose CityPaint, a novel framework that leverages 2D diffusion models to generate globally consistent city-scale 3D textures without relying on city-scale 3D training data. Specifically, CityPaint integrates geometric priors with the generative power of 2D diffusion models, introducing a robust top-$K$ camera baking strategy for iterative texture projection. To address the artifacts of 2D-to-3D projection, we further propose a decoupled refinement mechanism that separately optimizes terrain and building textures. Extensive experiments show that CityPaint produces globally coherent textures with fine local details and achieves favorable results against strong baselines in both visual quality and cross-view consistency.

Rapid and accurate hypocenter determination based on three-dimensional (3D) seismic velocity models is essential for seismic monitoring in subduction zones, yet it remains computationally demanding. HypoNet Nankai, a hypocenter determination package that uses physics-informed neural network (PINN) travel-time inference for a realistic 3D P-wave velocity model of the Nankai Trough subduction zone, was designed to address this problem, but supported only P-wave arrival times. Here, we present HypoNet Nankai PS, which extends the framework to both P- and S-wave travel times. We first constructed an S-wave velocity model from the existing P-wave model using empirical conversion relations for marine sediments and the crust. Because the resulting S-wave velocity structure has lower velocities and therefore considerably steeper travel-time gradients, the original PINN training procedure designed for the P-wave velocity structure was insufficient. We therefore revised the training procedure by reconsidering the distribution of collocation points, the NN architecture, and the optimizer. Benchmark comparisons with the fast marching method showed that the revised approach substantially reduced the root-mean-square deviations for both P- and S-waves compared with the original algorithm. Recovery tests using synthetic earthquakes were also performed. When applied to 223 earthquakes recorded by DONET stations off the Kii Peninsula, the system yielded hypocenters generally consistent with the JMA catalog beneath the network, suggesting that the 3D velocity model captures travel-time variations accommodated by JMA's 1D model and station corrections. Farther seaward, the estimated hypocenters were substantially shallower than those in the JMA catalog and were located in the oceanic crust and uppermost mantle, consistent with previous relocations that used DONET data and 3D velocity models.

Clinical natural language processing includes diverse tasks such as information extraction, classification, and question answering. Developing one model for all tasks could simplify deployment and improve knowledge transfer. We evaluated whether small recurrent-depth language models can meet this need through joint fine-tuning and compared their performance with standard transformers. We fine-tuned Ouro1.4B-Thinking, Ouro-2.6B-Thinking, HRM-Text-1B, Qwen3-1.7B, and Qwen3-8B on 9 datasets covering information extraction, classification, inference, and question answering, and evaluated them on four additional datasets excluded from training. Across these 13 datasets, Ouro-1.4B-Thinking achieved higher performances than Qwen3-1.7B on 11 datasets; Ouro-2.6BThinking was competitive with Qwen3-8B on most tasks. Ouro models showed the largest gains on inference and question answering, whereas HRM-Text-1B showed mixed results. These findings support further study of recurrent-depth models for clinical tasks through controlled experiments to isolate architectural effects and investigate how recurrent computation influences reasoning, knowledge acquisition, and generalization.

Memory-based temporal graph neural networks (TGN) rely on recurrent node memories to capture evolving interaction histories. However, temporal mini-batching introduces a trade-off between computational efficiency and temporal fidelity: when a node appears multiple times within the same batch, later interactions are evaluated using stale pre-batch memory because the intermediate memory updates have not yet been committed. This problem becomes more severe as temporal batch size increases. To overcome this challenge, we propose AnchorRM, an anchor-based residual memory mechanism that reconstructs event-specific intra-batch node states without executing sequential recurrent updates. AnchorRM treats the last committed node memory at the beginning of each batch as a pre-batch fixed anchor and uses causally preceding same-batch interactions to infer the missing state evolution relative to that anchor. The learned residual then calibrates the anchor to reconstruct an event-specific node memory for each interaction. Because each event-specific residual is predicted directly from the fixed anchor and its causal history, reconstructed states do not depend on previously reconstructed states and can therefore be computed in parallel within the batch while preserving temporal causality. Experiments across four temporal graph datasets and three memory-based backbones demonstrate that AnchorRM achieves an average 6.16x training speedup (up to 10.05x) and 28.67x batch scaling (up to 40x). These results demonstrate that AnchorRM substantially improves the scalability and training efficiency of memory-based temporal graph learning while maintaining competitive predictive performance.

Safety pretraining has emerged as a promising paradigm for embedding safety deeply into a model during pretraining, yielding robustness to subsequent fine-tuning that conventional safety post-training often lacks. We refer to this safety acquired during model pretraining as native safety. Yet native safety is expensive to acquire, confined to the model in which it was pretrained, and often accompanied by lower general utility. In this work, we rethink the practical value of safety-pretrained models and introduce, for the first time, the native safety reuse problem: Can native safety be transferred from a specialized, potentially smaller, safety-pretrained "teacher" to standard-pretrained "students," enabling them to inherit its robustness while preserving their general utility? We show that native safety can "travel" through on-policy distillation. A standard-pretrained student generates responses to safety-pretraining prompts, while a small safety-pretrained teacher supervises the student's token-level predictions along its own generated trajectories; cross-tokenizer alignment further enables transfer across model families. The resulting students exhibit substantially greater safety robustness than their safety-post-trained counterparts while retaining higher general utility than the teacher: subsequent fine-tuning raises the average attack success rate (ASR) of the 1.7B student by only 5.00%, compared with 11.58% for safety post-training, and its average utility of 48.20% far exceeds the aligned teacher's 27.42%. Unlike conventional large-to-small knowledge distillation, our framework enables selective, small-to-large safety transfer: it reuses the robust safety expertise of a smaller pretrained teacher without inheriting its utility limitations. These findings establish native safety as a reusable model property rather than one permanently tied to the originally pretrained model.

Low curvature trace, an effective rank condition, is a common explanation for why randomized zeroth-order methods scale to high-dimensional tasks such as language model fine-tuning. Can deterministic methods achieve the same savings? We show that they cannot. We study finding $\varepsilon$-stationary points of $L$-smooth nonconvex functions on $\mathbb{R}^d$ from exact function values, assuming that below the starting value the Hessian is dominated by a fixed matrix with trace at most $LE$; this is the effective rank condition used to explain zeroth-order fine-tuning. When $ET\le d$, the Gaussian two-point method that such fine-tuning uses needs only $O(ET)$ calls if it rejects steps that raise $f$, where $T=LΔ/\varepsilon^2$ is the iteration count of gradient descent from initial gap $Δ$, and we prove that every randomized method needs $Ω(ET)$ calls, so this cost is optimal up to constants. In contrast, every deterministic method needs $\widetildeΩ(d\min\{T,\sqrt{ET}\})$ calls, and adaptive subspace search attains $O(d\min\{T,\sqrt{ET}\})$: a deterministic method pays about $d$ calls per discovered direction, and the trace budget allows about $\sqrt{ET}$ directions. Determinism thus multiplies the cost by a factor of order $d/(E+\sqrt{ET})$ up to logarithmic factors, which grows linearly with $d$ for fixed $E$ and $T$; once $E$ exceeds $T$ by a logarithmic factor, deterministic methods need $Ω(dT)$ calls, as without any trace budget. Our lower bounds hold for $C^\infty$ functions with the matrix bound on all of $\mathbb{R}^d$, even when each call also returns gradient projections priced by rank.

Task transition detection aims to identify when a person shifts from one task to another, providing a foundation for continuous health monitoring and timely context-aware interventions. While statistical and deep feature representations have demonstrated strong classification performance, they often rely on aggregated or latent features that are difficult to interpret in behavioural terms. Event-based features provide a compact and interpretable alternative by encoding behavioural and physiological patterns as discrete tokens with explicit semantic meaning, making them well suited for behavioural analysis in health informatics. Self-attention offers a mechanism for modelling interactions among such tokens; however, conventional self-attention classifiers primarily rely on token order and do not explicitly incorporate timing or duration information. As a result, informative temporal cues in event representations may not be fully exploited for task transition detection. This study proposes an event-based time-aware self-attention framework that represents behavioural signals as sequences of discrete events augmented with onset timestamps and durations. Event timing is encoded using a parameter-efficient kernel-based representation, while a novel type-aware duration embedding modulates information propagation according to event persistence. Experiments on a semi-naturalistic wearable dataset show that the proposed model outperforms existing event-based approaches, achieves performance beyond or comparable with statistical and deep-learning models with a more compact input representation, and provides improved interpretability through integrated-gradient-based analysis. These results indicate that explicit modelling of event timing and duration enhances both classification performance and interpretability, supporting reliable deployment in biomedical and health informatics applications.

When an LLM agent acts on behalf of a user, it is expected to act within the authorization scope the user grants. Existing agent evaluations largely measure task completion or harmful consequences and can therefore overlook actions that violate this delegation without producing conventional safety or privacy harms. We introduce AuthorizationBench, a benchmark of authorization alignment grounded in documented real-world cases. A scalable construction pipeline produces 286 executable tasks spanning situational triggers, the presence or absence of direct harmful consequences, and ten application domains. We execute every task with four models at two reasoning-effort settings across three agent harnesses in an authorization-aware sandbox that keeps authorized and out-of-authorization paths executable and observable, yielding 6,864 trajectories. Agents act out of authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across the 24 configurations. Task completion does not imply authorization. Increasing reasoning effort leaves the aggregate rate nearly unchanged, with effects varying across models. On controlled pairs, agents act out of authorization at similar rates with (17.9%) and without (20.8%) direct harmful consequences, and more often under benign pressures than malicious injections (23.5% versus 17.6%, p=0.002). Our results show that task capability, additional reasoning, and consequence-oriented safety evaluation do not by themselves ensure authorization alignment. Moreover, agents that consult the principal before completing a task exhibit lower rates of out-of-authorization behavior (14.0% versus 21.6%, p<0.001), pointing to a tension between agent autonomy and preserving users' decision authority that calls for further human-centered study. Dataset and code are available at https://github.com/PEACH-Research-Lab/AuthorizationBench .

Modern large-scale cloud systems continuously generate massive and multimodal observability data, such as time-series metrics, log records, and system dependency graphs, which reflect the health state of the system. To ensure reliable cloud system operation, anomaly detection models identify abnormal states by analyzing these multimodal observations. However, existing approaches rely on simple fusion strategies and graph learning, ignoring the semantic gap among modalities and the asynchronous spatio-temporal dependencies across system instances over time. To this end, we propose BEACON, a framework to Bridge multimodal sEmantics and Asynchronous propagation for Cloud system anOmaly detectioN. Specifically, we introduce a Dynamic Interpretable Multimodal Aggregator (DIMA) that leverages an information-bottleneck mechanism to compress information from the semantics and temporal dynamics of metrics, and aligns them with logs in a shared prototype space. Furthermore, we propose an Asynchronous Spatio-Temporal Propagator (ASTP) that captures temporal dependencies with multi-scale temporal attention and models spatial relations with topology-aware aggregation. This enables asynchronous propagation of anomalies across system instances. Extensive experiments on three real-world cloud system datasets demonstrate that \textbf{BEACON} consistently outperforms state-of-the-art baselines in anomaly detection.

Depression dialogue benchmarks commonly assess models by comparing predicted symptom labels with clinical scale scores. However, a prediction may agree with a benchmark label even when the dialogue lacks enough evidence to justify it. We introduce EviDial, a benchmark that distinguishes label agreement from evidence-grounded assessment. EviDial comprises 382 depression dialogues from three sources and 2,744 participant-symptom cells, each classified as having absent, partial, conflicting, or sufficient evidence. Because evidence sufficiency lacks a directly observable gold standard, we derive calibrated evidence references from multiple judgments using generalizability theory, many-facet Rasch measurement, and random-effects latent class analysis. Our audit reveals that 82.8% of participant-symptom cells lack sufficient textual evidence for the corresponding symptom assessment, with substantial variation across sources. Across model families and prompting strategies, 69.2%-96.9% of predictions judged correct by benchmark labels are nonetheless unsupported by sufficient dialogue evidence. Further analyses show that strong agreement among LLM evaluators need not reflect independent evidence judgments, underscoring the need to calibrate automated reference panels. These findings demonstrate that label-based evaluation can overestimate evidence-grounded performance by conflating agreement with clinical labels and support in the observed dialogue. EviDial provides a measurement layer for separating these constructs and auditing LLM-based depression assessment more faithfully.

Zero-shot composed image retrieval (ZS-CIR) aims to retrieve a target image based on a reference image and modification text. Recent approaches commonly rely on large vision-language models to generate target image captions for retrieval. However, these captions may include redundant semantic details, introducing noise when retrieving the target image and leading to suboptimal retrieval results. To address this issue, we propose Less-is-More CIR, a two-stage training-free framework that combines Concise Retrieval and Critic Re-ranking. In the first stage, concise target-oriented captions are generated that preserve only the most discriminative modification semantics, thereby reducing semantic noise in the retrieval. In the second stage, a critic re-rank module further evaluates the retrieved candidates from the perspectives of modification consistency and reference preservation. Extensive experiments demonstrate the effectiveness and robustness of the proposed framework for the ZS-CIR. Our code is available at: https://github.com/dxlabskku/LM-CIR.

Alarms in industrial systems and scientific instruments often demand immediate diagnosis despite scarce incident data. To guide intervention, an explanation must trace how changes in the data-generating process propagate through the causal structure to alter the alarm score, rather than merely flagging shifts that may be downstream effects. Yet exact attribution across interacting causal components appears combinatorial, while existing methods often either assume an unchanged causal model or require enough incident data to estimate the changed system reliably. We address this setting by formulating the alarm-score change as a Shapley game over hybrid structural causal model (SCM) states. Treating structural functions and exogenous laws as distinct players yields a finer-grained, function/law-resolved report of how each type of change propagates to the alarm score. The same framework defines realized-event and population reports, with an event--population coherence identity linking their attributions. For ordered linear SCMs and quadratic scores, a causal first- and second-moment recursion combined with quadrature computes the full Shapley vector exactly in polynomial time, eliminating combinatorial computation. To estimate this target from scarce incident data, we adapt a well-observed reference SCM to the incident regime using cross-fitted estimates of mechanism changes and residual moments, and derive attribution-error bounds and top-k recovery guarantees. Experiments on controlled and measured shifts illustrate the diagnoses and assess scaling to hundreds of variables and estimation from small incident samples.

Machine learning benchmarks act as guideposts for model development and spur the creation of training datasets intended to improve benchmark scores. We introduce BenchDoor, a family of data poisoning attacks that exploits this incentive. Beyond preserving benign performance, BenchDoor requires benchmark gains that encourage adoption of data that also implants a backdoor. The adversary either manipulates a benchmark's training or test data, creating an adversarial benchmark, or supplies an independent training dataset while leaving the benchmark unchanged. We formalize five attack scenarios, including training on rewritten test examples and applying similarity-based decontamination. Across multiple-choice Q&A, text classification, and open-ended Q&A with Qwen and Gemma models, BenchDoor achieves positive benchmark gains for all 30 evaluated configurations, with held-out attack success rates of 82.33% to 100.00% at poisoning rates of 2% to 10%. Both objectives remain achievable after synthetic data generation and similarity-based decontamination, with attack success rates of at least 90.33%. These findings show how benchmarks can serve as attack vectors both through their own data and through training resources that deliver improvements on them.

Agentic system design is increasingly shifting from hand-crafted prompts and workflows toward automated search over agentic programs and interaction structures, with multi-agent systems receiving growing attention. Yet a central question remains unresolved: when does an automatically generated multi-agent structure exhibit a functionally meaningful division of responsibility? We argue that automated multi-agent design should move beyond searching for effective architectures toward identifying which responsibilities are functionally necessary, how they should be allocated, and when they should change. To explore this, we propose Automated Multi-Agent Organization Design (AMOD), which treats organization itself as a first-class design object, explicitly modeling how agent teams are composed, what functional responsibilities the agents assume, how they coordinate, and how collective decisions are made. We formulate AMOD as an execution-driven design loop. For each task, an automated designer instantiates an organization by assigning functional responsibilities and coordination rules to a set of agents. After execution, task outcomes, intermediate artifacts, and interaction traces are examined to identify which responsibilities may be missing, duplicated, or ineffective, and the organization is revised accordingly. We also propose testing discovered responsibilities by removing them, replacing them with generic roles, or changing their assignment under matched resource budgets, with complementary evaluations. AMOD provides a principled foundation for evidence-based and adaptive multi-agent organization design.

Generative reward models (GenRMs) generate natural language rationales to justify preference judgments, yet a correct verdict does not guarantee a sound justification. Existing approaches often rely on reference explanations to supervise rationales, but obtaining such references can be costly and does not necessarily establish whether generated rationales support the resulting verdict. To address these limitations, we propose DeCert, a framework that evaluates the validity and verdict consistency of generated rationales without requiring reference explanations. Crucially, DeCert translates these assessments into preference supervision that accounts for both verdict correctness and justification quality. In particular, when multiple sampled judgments share the same correct verdict, differences in certificate quality enable DeCert to construct rationale preference pairs that distinguish stronger from weaker justifications. Extensive empirical studies across two model families and multiple in- and out-of-distribution benchmarks demonstrate that DeCert consistently outperforms strong generative judge baselines while improving rationale quality. Moreover, DeCert effectively mitigates positional bias, improving position-consistent accuracy by up to 14.8 points. These results highlight the importance of supervising not only which verdict is correct, but also how that verdict is justified.

Code agents increasingly work on evolving real repositories, and memory systems aim to help them reuse earlier experience. However, such experience is diverse and evolving, so it remains unclear how memory systems organize it and thereby affect code agents. Existing benchmarks cannot answer this reliably because they lack either temporally valid histories or control over prior-task relevance. In this paper, we present MemDev-Bench to evaluate how memory systems affect code agents over diverse repository histories. It pairs each task with relevance-annotated earlier tasks from the same repository to build histories with controlled relevance. We also build an automated pipeline that reliably makes these earlier tasks executable and verifiable, producing 4,869 historical tasks for 555 base tasks. Based on MemDev-Bench, our systematic evaluation of memory systems across diverse development scenarios yields three main findings. (1) Memory mainly improves fine-grained localization and editing, increasing functional correctness. (2) Memory also causes failures on otherwise solvable tasks even with related history, and unrelated history tends to increase such failures. These failures mainly stem from agents reusing memory without reassessing its applicability. (3) Compared with using memory systems, directly providing agents with raw historical trajectories substantially reduces their steps. Further analysis suggests that agents save these steps by reusing process information that compressed memories tend to omit. These results suggest that future memory systems should preserve process information and reassess memory applicability against current task context.

Every day, more than 40 million people ask ChatGPT a health question, such as whether laptop posture explains a stiff neck and days of pounding headache. Posture is plausible, but meningitis is a time-critical possibility. We introduce SelfDx, a paired benchmark for whether models have the appropriate response behavior, a form of epistemic independence, in health questions where misplaced agreement with the user's self-diagnosis could delay urgent or emergency care. Each of 100 pairs contrasts a caution case, where a serious alternative remains possible, with an agreement case, where the user supplies evidence that rules out the alternative and it is safe for the model to agree. We find across 27 frontier models that they tend to inappropriately agree with the user's self-diagnosis in cases that require caution. Even the best current-generation model reaches only a 77.0% appropriate caution rate on time-critical condition cases. To address this gap, we experiment with three prompt variants that increase appropriate caution but reduce appropriate agreement by 9-35%, indicating excessive caution. In contrast, supervised fine-tuning on 128 balanced examples improves both rates together across three runs. We recommend post-training to improve appropriate caution without making models cautious on every case, with paired evaluation to verify both response behaviors.

Despite advances in photorealism, semantic fidelity, and aesthetics, text-to-image generation and image editing remain understudied in a human-centered dimension: whether their outputs feel psychologically comforting. Grounded in cognitive evaluation theory, we define psychological comfort as an affective response shaped by viewers' appraisal of restorative and emotionally supportive visual cues. We introduce HealBench, a fine-grained benchmark of 30,000 high-quality image-text pairs spanning 19 categories and 92 subcategories, selected from 100,000 candidates through a human- and expert-in-the-loop pipeline combining vision-language model matching, human preference models, expert filtering, and exemplar-guided augmentation. We also construct HealBench-Eval, a feature-clustered evaluation set of 1,095 instances. Using HealBench-Eval, we conduct objective and subjective evaluations of 14 representative image generation models and prompt-rewriting strategies. Results reveal a gap between conventional visual quality and perceived comfort: strong semantic alignment and aesthetics do not guarantee emotionally restorative images. Reinforcement learning experiments further show that HealScorer can serve as a reward signal for assessing and improving the emotional and restorative quality of generated images. These findings establish psychological comfort as a distinct dimension of human-centered image generation and position HealBench as a foundation for affect-aware generative models.

Pretrained Vision-Language Models (PVLMs) achieve strong zero-shot recognition but often suffer performance degradation under target-domain shift. Unsupervised adaptation provides a practical way to improve target-domain performance without labeled data, but it remains challenging because pseudo-labels generated by the model can be unreliable and no labeled validation set is available for model selection or early stopping. We propose a simple framework for unsupervised PVLM adaptation that combines progressive margin-based subset selection with cascaded consistency learning. The subset selection strategy constructs a class-balanced pseudo-labeled set using prediction margins and gradually expands it during training, enabling adaptation to start from reliable supervision anchors before incorporating broader target-domain data. Cascaded consistency learning further exploits unlabeled samples by using temporally smoothed teacher predictions on weakly augmented views to supervise student predictions on strongly augmented views, with confidence filtering to reduce the influence of uncertain pseudo-labels. To preserve the pretrained embedding structure, the framework updates only normalization parameters in the visual encoder and class prototypes initialized from text prompts. Experiments across multiple recognition benchmarks with both CLIP and SigLIP backbones show consistent improvements over zero-shot baselines and competitive performance against recent unsupervised PVLM adaptation methods. These results suggest that progressive pseudo-label selection and teacher-guided consistency provide an effective and lightweight strategy for adapting PVLM classifiers without labeled target data.

Fine-grained liquid identification requires a classifier to distinguish similar substances from compact millimeter-wave measurements. We investigate structured-token attention, which represents a radar feature tensor through complementary antenna, feature, branch, and global views. The proposed StructAttn model embeds these views into eleven tokens and combines them with a three-layer transformer. Experiments on the public FG-LiquID data cover 30 liquids and 3,000 samples. Over five seeds on the released split, StructAttn achieves 96.70\% mean best-epoch accuracy, matching the original-recipe MMNet baseline and exceeding a scalar-token attention model by 1.50 percentage points. A five-seed probability ensemble reaches 97\%, corresponding to 18 errors among 600 test samples. Final-epoch and five-fold results complement the checkpoint-selected comparison, with StructAttn obtaining 95.30\% out-of-fold accuracy. Token and feature analyses show that structured views are useful for prediction, while residual confusions concentrate in water, soft-drink, and edible-oil classes. These results support structured tokenization as a practical design choice for attention-based liquid recognition.

4D scene graphs require persistent object identity when an object disappears inside an opaque carrier and later reappears as a fresh detection. When appearance and target-centric motion are unavailable, identity recovery must rely on relational history. We formulate cross-visibility-gap identity persistence as historical-node recovery through a remembered target--carrier binding. We introduce CVG-SGM, a perception-frontend-compatible transfer--maintain--read memory. Geometry-weighted cross-node interaction transfers target-specific evidence into carrier states, a shared per-node GRU maintains this information during transport, and a two-time readout combines stored carrier--identity compatibility with current detection--carrier geometry to recover the historical identity. In the controlled one-identity-per-carrier setting, CVG-SGM achieves 1.000 re-link accuracy, compared with 0.570 for both no-memory and parameter-count-matched no-write controls. It also generalizes to scene-disjoint, unseen longer visibility gaps. On real RGB-D/SAM2 feature-level episodes, it correctly re-links 36/36 held-out detections, compared with 28/36 for the strongest fixed carrier--identity control. Counterfactual carrier-memory swaps redirect predictions according to the swapped bindings, supporting the role of stored relational information in identity recovery. The current approach remains limited by shared-carrier identity ambiguity and does not provide end-to-end raw-video re-identification.

We develop and compare enhanced neural-network solvers for one-dimensional ground-state profiles, which are solutions of $$ Q^{\prime\prime}=bQ-γQ^{p},\quad Q>0,\quad p>1, \quad b, γ>0, $$ the stationary profile equation for the nonlinear Schrödinger, generalized Korteweg-de Vries, and nonlinear Klein-Gordon equations. Its explicit sech-type solutions serve as exact reference for error computation and benchmark hyperparameter selection. We improve on the standard Adam-trained physics-informed neural network (PINN) baseline by introducing higher-order optimization via Gauss-Newton/Levenberg-Marquardt and L-BFGS refinements, as well as Fourier-feature PINNs, and constrained Deep Ritz (DR) formulations. In DR approaches we introduce two new neural-network formulations, which implement two classical variational characterizations: a Nehari-constrained Deep Ritz method, in which the height is learned by projecting onto the Nehari/Pokhozhaev manifold, and a Weinstein Deep Ritz method, in which a normalized profile minimizes the Weinstein quotient. The Deep Ritz formulations perform best overall, in particular, the Weinstein variant gives the best error-runtime balance. The neural-network based solvers remain less accurate and efficient than classical Petviashvili iteration in 1D, however, the current study provides a proof of concept: the mesh-free formulation, non-reliance on exact profiles or the related features make them promising in higher dimensions and for more complicated nonlinearities and dispersive operators.

Self-evolving agents aim to improve from their own experience rather than remain fixed after deployment, but naturally collected trajectories often lack ground-truth solutions or verifier-derived rewards. We introduce Retrospective Self-Distillation (RSD), a multi-round framework that turns such experience into dense supervision using privileged retrospective context. For each on-policy trajectory, the same underlying model acts as both student and teacher: the teacher re-scores the student's realized actions while conditioning on cross-task precedents and lessons, together with turn-level reflection and future hindsight from the completed trajectory. The resulting teacher distribution is distilled back into the policy through on-policy self-distillation. Across rounds, the experience bank grows and the retrospective context is refreshed, allowing the model, its data, and its supervision to co-evolve. We further introduce frontier-guided context-policy evolution, where a stronger frontier model proposes constrained edits to how privileged context is constructed. Across ALFWorld, ScienceWorld, and SearchQA with Qwen3-8B and Qwen3-14B, RSD improves performance by up to 59.7% relative to the base model, while frontier-guided context-policy evolution adds up to a further 44.7% relative improvement over a fixed context policy. These results show that naturally accumulated experience can be converted into persistent model capability without ground-truth trajectories or verifier-derived rewards for the core policy update.

Real-world audio typically contains concurrent sound events, yet standard self-supervised methods treat each clip as an isolated training unit without explicitly modeling mixture--constituent relationships. We propose Mixture-to-Constituent Latent Alignment (MiCLA), a two-stage self-supervised pre-training framework that uses synthetic log-mel mixing to align mixture representations with their constituent clips. Stage 1 structures single-clip representations via masked self-distillation and contrastive regularization. Stage 2 introduces bidirectional alignment: for a synthetic mixture, MiCLA suppresses the component aligned with one constituent and aligns the residual with the other, using projective residualization on patch tokens and sample-adaptive soft residualization on the global class token. In downstream evaluations, MiCLA achieves 43.2% mean average precision (mAP) on AS-20K and 50.3% mAP on AS-2M under fine-tuning, and achieves the highest mAP among evaluated models on both AudioSet settings under frozen Convex Gated Probing (CGP). Furthermore, a Joint Semantic Retrieval diagnostic shows that MiCLA improves simultaneous retrieval of both constituent classes over target-averaging baselines. These results indicate that directional mixture-to-constituent alignment provides an effective pre-training signal for polyphonic audio representation learning.

Mitigating catastrophic risks from large language models requires excising hazardous operational capabilities, such as cyberwarfare exploitation and pathogen enhancement, without degrading benign scientific utility. In these dual-use domains, legitimate defensive science and malicious attacks share vocabulary, syntax, and internal model representations, causing standard unlearning methods to induce severe collateral damage (concept bleeding). In this work, we present Orthogonal Boundary Unlearning (OBU), a mathematically grounded framework designed to eliminate collateral damage on benign dual-use capabilities. OBU extracts a domain-matched capability subspace from paired calibration queries via singular value decomposition on empirical activation covariances, and restricts parameter updates strictly to the orthogonal null space. This guarantees that benign representations experience zero parameter perturbation along the spanned calibration subspace. We present two complementary realizations of OBU: Projected LoRA, which applies the orthogonal constraint during iterative gradient-based fine-tuning, and Closed-Form Weight Surgery, which directly edits model weights in a single analytical step without training. Across four modern open-weight foundation models (Gemma-3-4B, Qwen3.5-4B, SmolLM3-3B, and Granite-4.2-3B), OBU achieves strong dual-use concept separation, substantially mitigates in-context priming attacks, and demonstrates durable resistance against adversarial relearning without degrading benign domain competence.

Practitioner guidance holds that a large language model reviews a document better when the work is divided among agents. We test a narrow form of it in a controlled comparison of single-agent and multi-agent setups on a financial task, reviewing credit assessment memos against requirements. One agent checks six requirements in one call. A team of three on the same model checks the same memo, two requirements each, in the same words, with no roles or later stage. Labels are known for every requirement on 300 synthetic memos, so misses and false alarms are measured separately. Two primary predictions are tested on four models at a Bonferroni-corrected level. Under an instruction to decide each requirement once, the team catches 0.08 to 0.10 more of the defects that carry an explanation on the three models whose single agent misses some, at the corrected level on two. It raises more false alarms on all four, at the corrected level on three. Balanced accuracy, an exploratory measure, changes by 0.02 or less, and the team uses 2.2 to 2.4 times the tokens. Showing each agent the requirements it was not given reduces the added false alarms on the two Claude models, at the corrected level on one. In exploratory comparisons, sampling the single agent three times and flagging on any draw gives higher balanced accuracy than the team on those models, and the team's results depend on the instruction.

The growing use of autonomous artificial intelligence agents in digital commerce creates new security concerns involving agent impersonation, unauthorized payments, compromised agents, and misuse of delegated authority. This research introduces the Autonomous Commerce Agent Guard System (ACAGS), a cybersecurity framework combining supervised machine learning, behavioral analytics, agent-specific authorization controls, and explainable AI. The framework was evaluated using synthetic financial transaction data augmented with simulated agent characteristics and five adversarial scenarios. XGBoost was selected following comparisons with alternative classifiers. Across ten experimental runs, behavioral features improved mean F1-score from 52.12% to 67.49%, with agent-specific features increasing performance to 69.46%. The framework achieved detection rates of 99.77% for abnormal agent behavior, 99.73% for impersonation, and 100% for malicious delegation and authorization-scope violations. Compromised-agent detection reached 72.27%. ACAGS also supported APPROVE, REVIEW, and REJECT authorization outcomes, with SHAP explanations providing model interpretability. The results demonstrate the potential of combining behavioral intelligence and delegated authorization controls to strengthen simulated agentic-commerce payment security, while highlighting the need for validation using real-world transactions.

Language-model agents increasingly participate in operational workflows where approvals, binds, and payments create commitments. An approval valid when issued may no longer authorize execution after relevant facts, evidence, policy, or authority change. We call this the evaluation--execution gap. We specify a state-bound runtime assurance layer that separates probabilistic measurement from deterministic adjudication, binds an execution allowance to the decision-relevant basis of an approval, invalidates that allowance after relevant logged changes, and rechecks admissibility at dispatch. This distinguishes the continued validity of a prior approval from a current-state policy check. We describe eight failure modes, an assurance model, a reference architecture, and prospective interventions with retrospective remediation. In a scripted 42-case replay, the state-bound configuration identifies 15 later-invalidated approvals that a deliberately non-re-evaluating point-in-time comparator does not revisit; it holds 14 affected actions before dispatch and compensates one landed effect. These results illustrate behavior under the stipulated policies and event orders, not superiority over dispatch-time enforcement designs or production effectiveness. Under explicit interception, ingestion, dependency, and serialization assumptions, the model establishes dispatch-time validity within the registered execution surface.

The represented gradient flow of a parameterized model depends on a mobility that can retain hidden parameter information. We study when one entire represented mobility map works for every autonomous loss from one fixed initialization. For two-factor matrix representations wider than both output dimensions, with individually full-row-rank factors, we classify all admissible joint ranks: closure occurs precisely at effective scalar-balanced minimal joint rank or a two-value signed spectral family at full joint rank. All intermediate ranks are excluded; positive represented flows belong to known square-root families. For general representations with constant ambient representation and gradient-Lie ranks, ordinary closure is equivalent to a local invariant-level metric in the analytic case and specified smooth regimes. An endpoint section forces the initially unrestricted mobility map to admit a locally smooth representative on the represented image. We also separate testing families: linear losses locally admit closure at every submersion, whereas one fixed translated strongly convex quadratic family fails it on a delayed-bracket example. A fixed output readout gives an ordinary squared-loss witness. These supporting results do not characterize quadratic-only closure in general. We retain the unresolved smooth output-rank-two case with Lie rank at least four. A frozen deterministic experiment checks a separate engineered-loss construction; it is distinguished from the squared-loss proof, universal theorems and empirical learning performance.

Text-to-image models can generate visual concepts that their prompts never mention, and recent work shows that such associations can be implanted so that ordinary prompts trigger them. Auditing this behavior requires recovering the description that a prompt omits. Existing prompt optimizers are not well suited to this task. Gradient-based methods require backward access to the scoring encoder. Discrete evolutionary search treats semantically unrelated tokens as equally close. Language-model-guided search requires one inference call per candidate, which limits exploration. This paper presents EmbGA, an embedding-space genetic algorithm that recovers concepts present in a set of images but absent from the prompt that generated them. EmbGA keeps the population structure of genetic search but applies its operators in the token embedding space. Crossover interpolates between parent embeddings, and mutation moves an embedding toward a randomly drawn vocabulary embedding. Each offspring is then projected back to discrete tokens before scoring, so every evaluated candidate is a valid token sequence. We evaluate EmbGA on two benchmarks against six optimizers under two exploration budgets. EmbGA achieves the highest alignment when the exploration budget is large, recovers the most grounded tokens on natural images, and runs much faster than language-model-guided search.

Federated learning (FL) is a distributed machine learning paradigm that reduces transmission costs, mitigates cross-border data sharing restrictions, and enhances privacy by keeping raw data on local devices. Clients in FL systems train local models and share only updates, which are aggregated by a central server. While this setup limits direct data exposure, FL remains vulnerable to privacy attacks. Existing attack and defense strategies have largely been developed for centralized systems, and adapting them to FL introduces new challenges due to its unique architecture and training workflow. In this work, we first propose a model inversion attack on FL systems using Deep Convolutional Generative Adversarial Networks (DCGANs) in the context of large-scale medical imaging. Our method operates under realistic assumptions about the attacker's knowledge and capabilities. We then examine the effect of client-side differential privacy (DP) in mitigating privacy leakage and incorporate personalized FL to improve model performance under DP settings. Experiments on the Brain MRI ND-5 dataset show that DCGAN-based attacks can reconstruct semantically meaningful synthetic images, revealing significant risks to the privacy of training data. Applying DP reduces attack success but introduces an accuracy trade-off, which is further overcome by personalized FL at each client side. We also evaluate how different aggregation algorithms affect both attack vulnerability and defense effectiveness, underscoring the critical role of aggregation strategies. The proposed attack-and-defense framework is model-agnostic and applicable to a wide range of FL-based systems.

Adapting a forecasting system to a changing world classically means retraining. We ask whether the task of continually learning regimes for time series forecasting can be represented by a piece of text: a guidance block, evolved by outcome-driven reflection and used in a frozen LLM forecaster's prompt. On a refreshed multimodal benchmark (weekly search indices, daily currencies, daily commodities, each window paired with web-retrieved news), we run GEPA-style prompt evolution under strict contamination control to prevent the data leakage common in timeseries benchmarks: the forecaster's knowledge cutoff predates every evaluated window. A 900-word evolved guidance carried by Llama-3.3-70B beats TimeSeries Foundation Models (TSFMs) on held-out data (GM MASE 0.776 vs. 0.898 for TabPFN-TS-3), transfers to more recent models without measurable loss, and costs dollars rather than GPU-days to update. We characterize the loop: out-of-loop gains concentrate in the first reflection, in-loop improvements overstate out-of-loop gains 2-3x unless an untouched monitor arbitrates, and error-only objectives silently suppress the text modality. We provide two extensions. First, compiling the guidance into executable code with a TSFM as a callable tool recovers most of the gain deterministically. Second, ablations show the advantage widens at short horizons and that removing the news text does not hurt the guided forecaster, consistent with the text's value being distilled into policy during learning. We argue that continual prompt evolution is a practical, training-free alternative to retraining for multimodal forecasting.

On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provides no reward-based learning signal when all attempts fail. On-policy distillation provides next-token supervision from a teacher, but on hard problems, if a rollout goes wrong early, subsequent supervision is conditioned on prefixes that already contain the mistake. Correct behavior can instead come from a gold or teacher solution, but such off-policy targets can differ substantially from the model's generations and, as we show, may erode existing capabilities. We therefore treat on-policyness as a continuum and propose projected distillation, which minimally edits the student's failed attempts into correct targets. The method uses contrastive updates to favor corrections over failed attempts and refreshes these corrections as the student evolves, enabling near-on-policy learning even when all student attempts fail. Theoretically, we prove that the student's learning regret grows with the edit size: it scales linearly with the horizon under minimal correction, as in DAgger, and quadratically under full-answer replacement, as in behavior cloning. Empirically, across two students (Qwen3-8B and OLMo-2-7B) and five domains spanning math, code, science, and tool use, projected distillation gives the highest area under the specialization-retention frontier (best on 8 of 9 cells against GRPO and OPD, and 7 of 9 against SPOT and SODA). Our ablations highlight the importance of online corrections, the contrastive loss, and on-policyness: moving targets farther off-policy steadily worsens the specialization-retention trade-off, with retention falling from 0.916 to 0.744 when minimal corrections are replaced by student-agnostic gold solutions.

Agent platforms are starting to improve their own agents: they read a deployed agent's execution traces, derive a change, have an automated verifier approve it, and release it with no human author. How much of this loop can safely run without a human? We propose two quantities a platform can measure. Verifier strength $σ$ is how well the available verifier -- often an LLM judge -- agrees with a trusted reference verifier; it limits how much a step can improve the agent. Recoverability is how cheaply a bad change can be undone: whether undoing it needs a new verifier approval, and how much of the old behavior a rollback restores. A step should run unattended only if both are adequate. On our platform this rule allows prompt, tool and skill edits to run unattended and keeps weight updates under human sign-off. Our main contribution is these two definitions and how to measure them. We built a platform that applies the rule and tested it in two ways: on the platform's own approval gate, and in controlled experiments on the $τ^2$-bench agent benchmark. On verifier strength, the platform's LLM judge sometimes gave the same score no matter what it was shown (9 of 37 settings) and reported no error, so a verifier can look healthy while approving anything; its strength must be measured, and two simple tests catch the failure. On recoverability, restoring a saved copy brought the agent back to its original behavior, but a regenerated undo did not, so only changes whose old version is kept can be reversed without a human; and while a change was deployed, it altered the outcome of 17 percent of tasks beyond normal run-to-run variation, which no undo repairs. Together these support the rule: a change can run unattended only when its old version is kept and its verifier's strength has been measured.

Although large language models support a wide range of tasks, hallucinations remain a major obstacle to user trust. When a model generates an answer, both its internal processing and final output vary with the question. Internal states are often expected to offer richer signals for hallucination detection. However, studies using these states typically report their standalone discrimination performance, while output-side comparisons often use simple baselines with weaker modeling capacity. We integrate output-side information available at inference time into a combined baseline and compare detection performance before and after adding hidden states under identical data splits and evaluation rules. Experiments cover open-ended generation and constrained multiple-choice tasks across different models. Adding hidden states improves AUROC by approximately 0.004 and 0.039 on the two tasks, respectively. Hidden states thus provide additional predictive information, but strong standalone detection performance can still yield only limited additional gains.

On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowledge transfer. How can we assess whether a newly generated question is valuable for distillation, and learn to generate better questions---without access to its correct answer? We introduce Ask-to-Distill (\method), a framework that treats synthetic questions as distribution elicitors and optimizes the teacher supervision they expose along student-generated trajectories. A student-initialized synthesizer is optimized using a distributional surrogate for question value, computed from a frozen student--teacher pair without reference answers, correctness rewards, or per-candidate student updates. The reward combines teacher--student disagreement with teacher probability mass on the student's candidate support, together with length and diversity controls. The generated questions are filtered and deduplicated into a prompt-only bank for standard OPD. Across mathematical reasoning, code generation, and agentic tasks, our method improves average performance and teacher-gap recovery over unoptimized student-generated prompts and KL-filtered real questions, while matching or exceeding real-bank OPD in the evaluated settings. Signal analyses further show that optimized questions can elicit denser supervision and more total signal from shorter student responses, supporting elicited distributional supervision as an objective for synthetic distillation data. Our code, data and models are publicly available at https://github.com/GasolSun36/Ask_to_Distill.

The evaluation of large language models (LLMs) is inherently multi-objective, as different models may offer different tradeoffs across performance dimensions such as math, code, and writing. Because such evaluations are often based on costly and time-consuming human feedback, it is important to draw reliable conclusions using as few comparisons as possible. In this paper, we present Vector-Borda Empirical-Gap Elimination (VB-EGE), a fixed-confidence adaptive sampling algorithm for efficiently identifying models on the Pareto frontier in arena-style evaluation, where each comparison evaluates a pair of models on a prompt with respect to a selected performance dimension. VB-EGE progressively eliminates models from an active set once sufficient statistical evidence has been accumulated to certify their Pareto status. Under a multi-objective Bradley-Terry model, we show that VB-EGE returns the correct Pareto frontier with high probability and derive an instance-dependent upper bound on its sample complexity. We further establish a worst-case lower bound showing that VB-EGE is minimax-optimal up to logarithmic factors. Simulations and experiments on three real-world Arena datasets (Arena, SearchArena, and VisionArena) demonstrate VB-EGE's empirical effectiveness.

A controller can guide a language model toward safe and useful responses by choosing when to assist, ask for clarification, or refuse. When the controller uses selected internal features, poor performance alone cannot distinguish missing information from limitations of its decision rules or training. We study decision sufficiency: whether the best controller using a representation can match the minimum average loss attainable with full state information, under the same strategies and evaluation criteria. For interfaces with finitely many outputs, we compute the optimal loss exactly and separate the three sources of excess loss. Pairwise checks can miss information loss: states with the same controller input may share an optimal strategy in every pair but not across the whole group. We bound the number of states needed to reveal a conflict and characterize the fewest outputs an unrestricted interface needs to avoid all conflicts. We audit selected features from sparse autoencoders, which encode model activations using relatively few active features, in Qwen2.5, SmolLM2, and Phi. On SmolLM2 responses scored by two human raters, restricting the controller to two selected on/off features increases its minimum average loss by 0.0345--0.1166 relative to full state information, even with the governing policy supplied. Both independently trained autoencoders show the increase on both evaluation sets. The audit identifies decision failures that require changing the controller's input, not merely its architecture or training.

Low-rank adaptation (LoRA) attaches small low-rank matrices to a frozen backbone, and public repositories host many LoRAs per backbone, each for a different concept or task. Combining several LoRAs in one generation forces a trade-off. Merging their weights into one update needs a single forward pass, but LoRAs trained for different concepts interfere and quality drops. Running each LoRA in its own forward pass (composite) preserves quality, but inference time grows linearly with the number of active LoRAs. We study a property of LoRA pools overlooked by prior multi-LoRA work. In real LoRA pools, the stacked low-rank factors need far fewer directions than their summed ranks. Using singular value decomposition, we find the minimal subspace that keeps 99% of the factor energy. On three diffusion pools, only 46% to 57% of the summed LoRA ranks is needed, and the subspace grows sub-linearly with pool size. A 48-LoRA sub-pool of LoRA-Hub on FLAN-T5 tasks shows comparable overlap (62%), while the diversity-curated LoRA-Retriever benchmark sits near the no-overlap baseline (96%), so pool composition rather than modality predicts where compression is large. We build on this overlap with LoRASpace, a training-free method that computes the subspace once per pool and reuses it in two inference modes, a static mode at merge cost and a dynamic mode for input-dependent routing. On the diffusion pools, static LoRASpace matches composite quality at merge cost, and dynamic LoRASpace is more than 6x faster than composite with six active LoRAs on a 17-LoRA pool, mainly because it runs one backbone pass instead of six. In an ablation, static variants at the same rank match LoRASpace, so the quality gain does not depend on the pool-wide basis. On the near-orthogonal LoRA-Retriever benchmark, compression is small and gains come from coupling with external retrieval.

Set-valued classification (SVC) offers a practical approach to communicating uncertainty while maintaining accuracy by predicting variable-sized sets per input rather than fixed-sized sets (e.g. top-1 or top-5). However, set-valued classifiers (SVCs) exhibit significantly degraded performance on long-tailed and imbalanced datasets, a setting for which they are ostensibly well-suited. This becomes apparent under macro-average evaluation, which weighs all classes equally. SVCs are typically formed during a post-hoc calibration stage, so little is known about how training dynamics affect their behaviour. In this work, we investigate how different neural network training strategies and loss functions impact downstream SVC performance on highly imbalanced data. In addition to methods such as label smoothing and class-balancing, we investigate whether loss functions that operate over sets of examples in a batch, rather than treating all examples as independent, lead to better downstream SVCs. We adapt the recently proposed odd-K-out (OKO) loss function, which learns from sets of training samples, improving its sample efficiency whilst introducing algorithmic modifications that reduce its computational overhead from 3.2x to just 1.36x that of standard cross-entropy training. Experiments on highly long-tailed datasets demonstrate our modified OKO approach significantly improves macro-accuracy across multiple SVC methods - including top-k, average-k, and conformal prediction - while maintaining competitive micro-accuracy. This improved performance on minority classes occurs without substantially increasing prediction set sizes, suggesting that carefully constructed set-based training can produce networks better suited to imbalanced data.

Adversarial workers pose a severe challenge in distributed computing systems by returning corrupted outputs that can substantially distort the recovered computations. Coded computing addresses this threat by introducing redundancy across worker assignments. Existing exact and approximate coded-computing methods typically handle adversarial outputs by explicitly locating and correcting or removing erroneous evaluations, which adds significant computational overhead at the master node. Moreover, existing analyses do not characterize how the recovery error scales with the number of workers and the adversarial budget. In this paper, we develop an adversarially robust extension of general coded computing (GCC), a learning-theoretic approximate coded-computing framework that designs the encoder and decoder by minimizing an upper bound on the end-to-end recovery error. Rather than identifying corrupted workers, the proposed method uses smoothness regularization to limit the influence of their outputs. The resulting encoder and decoder can be implemented using smoothing splines with linear master-side complexity. We derive a non-asymptotic bound on the worst-case end-to-end approximation error by decomposing it into three components: (1) the encoder approximation error, (2) the decoder reconstruction error under fully honest results, and (3) the adversarial perturbation error. For maximum of $q\geq 1$ adversarial workers among $N$ workers results in a worst-case error of order $\mathcal{O}((q/N)^{6/5})$ with appropriate selection of smoothness parameter, demonstrating that the approximation error vanishes asymptotically whenever the number of adversaries grows sublinearly with the network size. Experiments on scalar computation and neural-network inference are consistent with the predicted convergence behavior.

Solving reasoning tasks supported on graphs typically requires providing a Language Reasoning Model (LRM) with a serialized description of the graph. This approach results in the input length growing with the number of edges, which limits the size of the problems that can be solved. Furthermore, even when graphs fit in context, performance sharply decays as the graphs grow. We propose to address this by augmenting LRMs with scalable and permutation equivariant graph positional encodings that modulate attention logits. An additional post-fusion-style module incorporates structural information into hidden states. We compute these positional encodings using a GNN followed by a Graph Transformer. This eliminates the need for serialized adjacencies, reducing the context burden to $O(N)$. We jointly fine-tune the positional encoder and the LRM LoRA adapters using optimal solutions generated by graph-theoretic solvers, resulting in a computationally efficient method. Trained on small graphs, the model exhibits improved size-generalization across graph-theoretic tasks on graphs with up to 1,000 nodes, and produces fewer tokens than a conventional LRM. Additionally, we observe that our model remains robust to node relabelings. Evaluations on graph-theoretic benchmarks also show that performance on small graphs is preserved.

Sparse attention limits the keys used by each query, but finding those keys can still require a full-prefix scan at every layer. We introduce \textbf{MLIPS (Multi-Layer Indexer Pool Sharing)}, a training-free method that shares candidate search while retaining each consumer's own scoring and selection. A producer publishes a top-$M$ pool larger than the attention budget $K$; subsequent layers select their own top-$K$ from that pool. Measurements across three sparse-attention model families show substantially higher coverage from broad pools than overlap between final selections. We implement layerwise sharing, persistent decode reuse, and a coarse producer in vLLM for DeepSeek-V4-Flash. Across four long-context suites, three consumer layers per producer give $1.14$--$1.40\times$ speedups and $14.8$--$27.4\%$ lower GPU energy, with observed accuracy changes of $-1.49$ to $+0.29$ percentage points in single runs. More extensive sharing exposes task-dependent quality losses. Sampled prefill traces also reveal recurring selected KV positions across query groups and layers. We characterize this shared working-set structure and use it to motivate Top-$G$, an exploratory bounded pool construction for prefill. Its deployment, task accuracy, and serving benefits remain open.

Long-horizon terminal agents may solve a task in one run yet fail across repeated runs. Stochastic errors can propagate through actions and context, creating a gap between capability and reliable execution. We present RiseHarness, a framework for iteratively improving agent harnesses with model weights fixed. It focuses on flipping tasks with both successful and failed runs under the same configuration, using successful runs as references for failure analysis. RiseHarness combines agent trajectories with internal harness traces from a runtime monitor to examine how model actions interact with command execution, feedback, context management, and completion decisions. A Harness Handbook links this evidence to harness behaviors and source code, helping identify recurring problems in mechanisms shared across tasks. RiseHarness then searches other harness frameworks for relevant implementations and adapts them as design references to the current model and execution protocol. Candidate changes undergo behavioral checks and repeated task evaluation; versions with measured gains seed the next iteration. This loop integrates suitable mechanisms from different frameworks into a single harness and uses new execution evidence to guide further changes. We evaluate task success and reliability across models, ablate harness execution evidence and external research, and test transfer to unseen benchmarks with the same model. RiseHarness treats a specific model as the unit of harness adaptation, studying how changes to shared interaction mechanisms can make existing capabilities more reliable across tasks.

A feed ranker scores candidates with a value model, a weighted sum of engagement predictions. Each prediction conflates a baseline, what this viewer, item or request would produce anyway, with the incremental amount by which this candidate exceeds it. The literature mostly removes the baseline. We argue that a multi-signal value model should keep the raw prediction and add the incremental part as a separately weighted term. On Instagram Feed, three debiasing families deployed independently, for viewer habit, item popularity and request context, are one parameterised operator, and a design procedure picks a point in it. Media-type (format) bias goes instead to a debiased label chosen by the same procedure. In a controlled four-arm ablation of deployed Feed terms (neither term, each alone, both), the two-term arm beats both single-term arms at production weights on point estimates in 11 of 13 judged (term, metric) pairs. Of the 26 direct contrasts, 17 are significantly for it and only 1 against it, on impressions. The ablation also suggests that the offline diagnostic should screen signals, not rank them. Such terms have served Instagram Feed production for most of a year. Beyond Feed, Meta, whose apps serve billions of people, has adopted the operator across surfaces, apps and both ranking stages. Its launches there drive sessions and daily active users, the core visitation and retention metrics.

Mixture-of-experts (MoE) models already make a sparse choice for every token: which experts should process it. We ask whether a sparse dictionary should reuse that choice. NR-SAE gives each routed expert a small dictionary and opens only those the frozen router selects. At block 15 of Gemma 4 26B-A4B, after ten million C4 training tokens and across three seeds, the expert-aligned autoencoder explains 0.518 of the variance in router-weighted expert outputs on a fixed evaluation capture. A rank-addressed control with the same parameters but sixteen times as many candidates reaches 0.481, a candidate-matched control 0.458, and an architecture-matched control with randomized expert identity 0.387. A transcoder variant also leads all three controls: 0.432, against 0.347, 0.371, and 0.118. Part of this advantage is architectural: each expert writes into a subspace of at most a quarter of the model width, so naming the expert names the subspace. How much it explains, we have not measured. At this layer and budget, the router already knows something a dictionary would otherwise have to learn. Our implementation did not train faster, and semantic interpretability, downstream loss, and causal faithfulness remain untested.

A failed gaming trajectory reveals that an agent needs intervention, but not necessarily how its execution harness should change. Effective harness adaptation therefore requires gathering diagnostic evidence, rather than merely proposing alternative recovery rules. We present KPHarness, a framework that adapts the external execution harness of a frozen vision-language-action model through knowledge-probing co-evolution. Recovery knowledge supplies context-specific intervention hypotheses that guide active probing; probe outcomes reveal when these hypotheses apply and provide evidence for refining both recovery knowledge and later probe selection. This reciprocal process allows the harness to improve not only which interventions it applies, but also how it investigates failures. KPHarness separates fast policy execution from a slow research loop that translates accumulated knowledge and probing evidence into bounded harness revisions. Each revision is compared with the incumbent under matched sandbox conditions and accepted only when the paired evaluation shows improvement. Recoveries supported by further validation are retained for cross-task reuse, while immediate interaction feedback and delayed episode-success credit update action- and knowledge-probing policies during adaptation. Experiments on Minecraft tasks spanning mining, combat, and crafting show improved task coverage over bare policy execution, a manually specified fixed harness, and knowledge-only harness evolution with non-adaptive probing. Evaluation on previously unseen tasks further shows improved task completion with the evolved harness, knowledge bank, and probing policies held fixed,

Neural networks trained to predict materials properties learn representations whose physical content can be difficult to extract. We use the neural tangent kernel (NTK), evaluated at trained parameters, to recover the geometry learned by a network that predicts superconducting transition temperatures from elemental composition alone. Kernel ridge regression with the trained NTK closely reproduces the network's predictive performance and substantially outperforms regression with the initial kernel. Spectral analysis and multidimensional scaling reveal a geometry dominated by a small number of directions. Clustering and local neighborhood analysis organize compounds into recognizable superconducting families and connect materials with related chemical and correlated-electron properties despite substantial differences in $T_c$. These results establish trained tangent kernels as practical probes of learned materials representations, connecting property prediction with the exploration of materials space.

Deception detection typically considers messages in isolation. Previous works in claim verification and role detection have displayed that additional information such as gameplay history is crucial to successful predictions. In this work, we demonstrate the value of preceding gameplay dialogue when applied to deceptive message detection within the hidden role game of Avalon. The findings show that LLMs can see improvement across a number of metrics, including accuracy by up to 31\% with the addition of gameplay context for the target message.

LLM agents can improve across tasks without updating model parameters by storing experience from their own executions in external memory and reusing it to guide later decisions. A central design question is whether to retain only observed trajectories and outcomes or also the rules, summaries, and skills inferred from them. Retaining these abstractions allows agents to reuse strategies without repeatedly deriving them from raw trajectories. However, errors in stored abstractions can propagate to subsequent decisions through retrieval and reuse. They can also shape execution trajectories, which then become the basis for new abstractions that reproduce the original errors. We term this feedback mechanism error reproduction: errors are regenerated through a cycle of abstraction, execution, and re-abstraction, extending their persistence beyond direct reuse of the original faulty knowledge. We propose Anti-Persistence, a memory design that accumulates trajectories and outcomes while reconstructing task-specific guidance on demand. An adaptive selector uses prior outcomes to decide whether and how to synthesize guidance, which is discarded after execution. This prevents its direct reuse as a stored object, although indirect feedback through retained trajectories remains possible. Across different benchmarks and backbones, Anti-Persistence achieves higher overall held-out success and the lowest average LLM-call cost among the evaluated memory designs. These results support separating experience accumulation from abstraction persistence in memorybased self-improvement.

Text-to-speech systems built on residual vector quantization often generate audio in two stages: a transformer backbone emits the first level of each frame's token stack, and a smaller depth model generates the rest autoregressively. We ask which components of such a model are responsible for prosodic intonation, more precisely for generating the English nuclear tune (for instance, the falling pitch melody that distinguishes an assertion from a question's rising melody), and whether interventions on those components can be used to control intonation. Prompt-based control mostly fails; only a final question mark is a reliable lever. Interchange interventions in CSM and S2 locate the tune: the backbone's hidden state carries it to the waveform through the depth model, and the first level of the stack carries little or none of it. Steering the backbone state with a difference-of-means vector computed on new material generated with a final period or question mark moves the tune, but the push toward an assertion degrades the voice. Steering vectors computed instead from the backbone states associated with human imitations of theoretically-specified tunes, whose text carries no punctuation, and applied inside the generation loop, move most recordings to the other condition in both systems (even when fitted without the speaker being steered), at a small cost in predicted naturalness. These imitation-derived vectors are aligned with the one obtained from the two punctuated conditions far beyond a shuffled-label null, so the backbone state of a punctuated prompt already represents intonation. The steered recordings broadly become falling or rising, according to the intent, but the edge tone moves more coherently than the pitch accent, and the target tune cannot be selected precisely beyond this broad tendency.

Motivated by the notion of features from machine learning, we study the complexity of representing data that allows for evaluation with a given family of functions. Concretely, we consider a deterministic encoder that observes $X\sim D$ and produces a finite encoding before a task $f\in\mathcal{F}\subseteq\{0,1\}^{\mathcal{X}}$ is specified. For each task, a decoder must predict $f(X)$ with error at most $\varepsilon$ under $D$. We construct an encoder that outputs a representation with $O(d^{*}\log(1/\varepsilon))$ bits, where $d^{*}$ is the dual VC dimension of $\mathcal{F}$. Our encoder greedily selects $O((d^{*}/\varepsilon)\log(1/\varepsilon))$ functions from the class and uses their joint label pattern to produce the encoding. The key idea in our upper bound is a novel potential function argument, where we compare the entropy gained from each greedy selection with the number of patterns allowed by the dual VC dimension. We present lower bounds showing that the representation length achieved by our encoder is worst-case optimal for every $d^* \ge 1$ and $\varepsilon < 1/4$ up to a universal constant factor, and instance-wise optimal for every family $\mathcal{F}$ up to an $O(\log(1/\varepsilon))$ factor. We give two ways to learn the representation from unlabeled samples, with different guarantees on the representation size and sample complexity. Experimentally, we verify the greedy construction's accuracy-size tradeoff and explore its connection to neural sparse autoencoders.

Large language model (LLM) agents are moving beyond reactive responses toward proactive assistance that anticipates user needs without explicit requests. However, existing proactive agents primarily address immediate needs, while many everyday situations involve delayed intentions that can only be acted on hours or days later. Humans handle such intentions through prospective memory, the ability to remember to act on an intention when the right moment arises. In this paper, we introduce ProMemAgent, the first proactive agent to incorporate prospective memory over everyday sensory contexts for long-horizon proactive assistance. ProMemAgent organizes past sensory perceptions (e.g., egocentric video and audio) into context-aware memory, generating prospective cues and linking earlier events with subsequent developments for future retrieval. Given the current context, it retrieves relevant past events and performs memory-augmented proactive reasoning to determine whether assistance is needed. We further introduce CPMBench, the first benchmark for evaluating long-horizon proactive assistance over sensory contexts, comprising 1,000 instances across eight everyday categories. Extensive experiments show that ProMemAgent achieves the best memory retrieval performance and consistently outperforms all baselines on long-horizon proactive assistance.

Probabilities of causation (PoCs) have developed steadily, from the binary measures first introduced by Pearl to the non-binary setting recently studied by Shu et al. As the theory has matured, PoCs have found applications in a growing range of areas, including personalized medicine, marketing and legal attribution. Shu et al. proposed closed-form bounds for multi-valued treatments and outcomes and proved them valid, but left their sharpness in high dimensions open. We close this gap with a mathematical proof. We further show that $\PNS$, $\PSub$, $\PRep$ and $\PN$ are all special cases of a single, more general formula. Combined with their equivalence classes, this settles the question of sharp bounds for discrete PoCs from experimental and observational data: in every dimension, they are given in closed form by a single theorem.

Learned post-training quantization (PTQ) calibration fits quantizer parameters on a small calibration set, and for a text embedder the natural supervision is retrieval relevance. We ask two questions that are usually conflated: how far to calibrate, and whether to calibrate at all. Across three embedders (e5-base-v2, BGE-M3, EmbeddingGemma-300m) at 4-bit weight quantization, 161 calibration runs, the outcome of a fixed procedure ranges from a gain to catastrophic degradation, and the driver is the dose of the update, measured as the functional displacement D_func of the quantized model from its uncalibrated start. On EmbeddingGemma the same learning-rate grid gives +1.47 to -71.53 points of in-domain NDCG@10 change; a label-free cosine objective collapses at the same dose (-69.52 vs -71.63), so the collapse is a property of learnable calibration rather than of retrieval supervision. Objective sensitivity is heterogeneous: on e5, ranking calibration is substantially more robust than cosine at matched displacement (+14.16 +/- 4.34 pp, 95% t-CI [+3.37, +24.96], three paired slates), whereas on BGE-M3 the ordering is directionally opposite but unresolved ([-5.18, +2.38]). Dose is the main axis but not a sufficient statistic: equal D_func reached along different paths differs by up to 1.61 pp. We therefore treat calibration as an optional intervention. A pre-registered rule that accepts a checkpoint only when a per-query paired confidence interval on a 25% calibration-validation split clears the uncalibrated baseline was run on 6 independent slates and five arms. It never selected a checkpoint below its slate's baseline (0 harmful selections in 30 decisions), retained 8 of 11 beneficial updates, and paid a mean oracle regret of 1.52 pp with one abstention that forfeited 11.13 pp, a safety-opportunity trade-off, not a free improvement. We report what remains open, including label efficiency.

Mechanistic interpretability asks which language model components matter for a behavior, and activation patching answers by changing one component's activation and measuring its effect, at a cost growing with components and prompts. Attribution patching estimates these effects from one gradient pass, but its estimates can be far off, and verifying each one restores the cost. Can the structure of these errors reduce this cost? We propose Guided-AP, which uses a few exact measurements to reveal attribution patching's errors and correct unmeasured estimates. On a single prompt, many components' errors share one sign, so a small exact sample yields one correction for the rest. Over many prompts of one task, components are misestimated similarly, so a few measured prompts show what to verify and how to correct each component's average effect. Our analysis shows that choosing what to verify needs error size, while correction also needs sign. On a single prompt, the correction lowers error by up to 26% across three behaviors at the compute of verifying the largest estimates and also on larger models. Over many prompts, learned corrections cut the error of average effects on factual completion by 75% to 79%, repaying their cost in long analyses.

Variance reduction techniques improve the efficiency of drawing conclusions from data. Recent proposals for variance-reduction use two complementary techniques to reduce variance: incorporating machine learning predictions into estimators and weighted sampling. A popular weighted sampling algorithm is Poisson Sampling (PS), which independently flips a biased coin at each unit and includes all heads in the sample. A fundamental flaw with PS is that it has a random sample size, which presents practical challenges and introduces variance into the estimator. We show that a fixed-size alternative, known in survey sampling as Conditional Poisson Sampling (CPS), is a drop-in replacement for PS on general multivariate Z-estimation problems, which include convex M-estimation considered in prior work. Notably, we show that the asymptotic variance of CPS is no worse than that of PS, and can be substantially better in realistic scenarios where predictions are imperfect. We illustrate these advantages on both synthetic and real datasets, and release a Python package implementing CPS. In light of our contributions, there is no need to suffer a variable sample size.

Action-conditioned video generative models trained on successful, goal-directed demonstrations tend to ignore the commanded action, rendering a plausible expert-like future regardless of what is asked of them. We trace this failure to training data that lack sufficient action influence. We analyze this metric and prove that it upper bounds the action influence that any model trained on that data can express. Expert demonstrations, whose narrow action distributions let next states be predicted without attending to the action, set a low ceiling that no amount of model capacity alone can raise. Raising it requires collecting different data. We introduce Action Influence Maximization (AIM), an algorithm for actively collecting training data for improving dataset action influence. We compare the performance and data-efficiency of AIM against exploration heuristics and intrinsic-motivation methods on simulated navigation and manipulation tasks. In pixel-based manipulation, AIM is the only method that consistently improves controllability: 500 actively collected episodes added to 3k demonstrations match a world model trained on 300k expert episodes on end-effector controllability, an $85\times$ reduction in data.

We analytically study the evolution of the energy function of Dense Associative Memories (DenseAMs) as they train on randomly sampled codewords of random expander low-density parity-check codes (a.k.a. solutions of random XORSAT problems). We identify two distinct regimes and a transition between them, reminiscent of grokking in neural networks. A regime of small training set size is where the model memorizes a few training examples but cannot discover the underlying rules of the entire codeword dictionary generative process. The opposite regime, corresponding to a larger training set size, is where the model has uncovered the sparse constraints that generated the entire data set after seeing only polynomially many patterns. We also design a novel class of DenseAMs with a thresholding learning rule, and show that any amount of thresholding enables the model to perform simplified sparse Hamiltonian learning, which exponentially reduces the number of training examples needed to learn the true underlying data set. In this regime, the model learns how to discover $\exp[\mathrm{O}(n)]$ solutions from only $\mathrm{O}(\log n)$ training samples, which is asymptotically optimal for any learning algorithm. When training on codewords of certain random codes, the model can even learn the underlying code when a small O(1) fraction of bits are latent variables. These latent variables are not shown to the model during learning, but will be discovered by sparsifying observed higher-order correlations. For certain random codes, we show that a simple and conservative algorithm learns a minimal latent representation for the data set. This provides a transparent, physics-motivated, and analytically tractable toy model for how deep latent representations emerge in an associative memory training.

Finding vulnerabilities early, before they can affect the production system, is important for reducing the risk of any software attacks. However, detecting vulnerabilities during secure code review is still difficult because the vulnerable code logic may depend on lines found across different repository files. Recent approaches such as AgenticSCR have shown that LLM agents can explore repository files and use security knowledge to improve secure code review. However, these approaches mainly rely on external commercial LLMs and can provide a large amount of security knowledge to the agent at once. This creates an important question of whether agentic secure code review can also be performed effectively using locally deployed open-weight LLMs with more focused and relevant security knowledge. In this paper, we propose AgenticVulHunter, an agentic secure code review approach and tool that can run with a locally deployed open-weight LLM. We evaluate AgenticVulHunter on 144 vulnerability cases from SCRBench using Qwen3-Coder-30B and DeepSeek V3.1 Terminus. We measure what percentage of generated review comments correctly identify the vulnerable location, match the vulnerability type, and provide a relevant security review comment together. AgenticVulHunter achieves an accuracy of 18.4% with Qwen3- Coder-30B and 23.8% with DeepSeek V3.1 Terminus, compared with 11.8% for our AgenticSCR-style reproduction using Qwen3-Coder-30B and 17.5% for AgenticSCR using Claude Sonnet 4.

AI coding agents now contribute to software repositories at scale, yet their documentation practices remain uncharacterised. We analyse 29,669 AI-generated pull requests from 2,585 repositories in the AIDev dataset, measuring documentation co-change: whether code-file modifications are accompanied by a documentation-file update in the same PR. We find that 63.2% of AI PRs change code without updating any documentation. This varies sharply by task type: 55.2% for features and 77.5% for refactoring. Undocumented code files are followed by a fix to the same file 8.6 to 11.7% more often than documented files across 30 to 90-day windows. A raw PR-level comparison shows the opposite because documented PRs are 3.9 times larger and therefore attract more fixes. Controlling for size exposure reverses this result. Documented changes also merge faster despite their larger size (median: 40 seconds versus 184 seconds), although merging is largely automated in this dataset. Documentation drift also compounds: within 90 days, a file changed without documentation is later re-documented at roughly half the rate of a file changed with documentation (24% versus 48%). This suggests that documentation debt tends to accumulate rather than be revisited.

BACKGROUND: After a software incident, organizations attempt to improve the reliability of their software systems by conducting post-incident analysis and identifying actions to prevent or mitigate similar future incidents. Despite the importance of this activity, there is a limited understanding of the nature and characteristics of the software evolution triggered by the analysis process. AIMS: This study seeks to bridge this gap by providing the first empirical characterization of post-incident software evolution. We aim to establish an understanding of how these changes manifest, specifically focusing on how their characteristics (such as effort, location, and timing) differ from standard maintenance and evolutionary changes. METHOD: We conducted a mixed-methods analysis of a 10-year longitudinal dataset from Wikimedia, examining 1,302 tasks and 2,203 code reviews. We categorized these changes across five dimensions: target, effort, abandonment rate, prioritization, and time-to-resolve. RESULTS: Our results show that the post-incident evolution is distinct from other change types. Post-incident changes are more time-consuming, with a median completion time over twice that of other changes and are disproportionately concentrated in a server configuration repository. CONCLUSION: By identifying a few of the characteristics of post-incident evolution, this work provides a foundation for future research into reliability engineering and learning from failure. We highlight further avenues for exploration necessary to fully understand post-incident evolution. Lastly, our work enables researchers to move beyond high-level analysis and begin developing specialized methods and tooling tailored to the nature of post-incident evolution.

Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without conditioning on the tokens selected at other positions in that step. This can produce inconsistent drafts and reduce acceptance length. Iterative refinement mitigates this limitation by conditioning later predictions on previously committed tokens, though additional passes increase drafting cost. To address this trade-off between draft fidelity and inference latency, we introduce DStill, a two-stage framework that distills a target-conditioned drafter's own multi-step refinement into one-step drafting. We first train the drafter to complete partially revealed blocks, then use its refinement trajectories to train a student to predict the final completion in one pass. Using a curriculum that gradually transitions from partial to fully masked inputs, we successfully transfer the multi-step refinement capabilities into one step. Across reasoning, math, code benchmarks on three model families DStill improves acceptance length by up to 28% and speedup by up to 25% over one-step DFlash, achieving up to 6.45x speedup over auto-regressive decoding.

Large language model (LLM) agents now autonomously complete long-horizon tasks, such as repository-level coding, through hundreds of think-act steps, yet the experience each run produces is rarely reused on the next. Prior memory systems concentrate on retrieval, finding the most relevant past trajectory. Retrieval cannot observe the running agent, so it cannot decide whether the agent needs an experience, at which point, or for how long. We introduce REM (Reusable Experience Memory), a framework that manages the memory lifecycle from inside the agent's think-act loop through three explicit decisions. What to store: each completed trajectory is consolidated into a phase-structured document, two orders of magnitude smaller than the trajectory. When to recall: the document is injected at the start of the task; the costlier implementation record is withheld until an in-loop gate detects the agent struggling. What to forget: the record is evicted after a single exposure. On repository-level bug fixing (SWE-ContextBench Lite), REM retrieves the related trajectory more accurately than every state-of-the-art memory system (68.7% vs. 55.6% match@1) and improves, in every replicate, end-task resolution (27.6% vs. 24.9%) at 17% lower cost. Injecting the same experience as a summary or raw trajectory, even when oracle-retrieved, does not. Ablating any of the three decisions reduces resolution. Altogether, these results demonstrate that the bottleneck in experience reuse is not retrieval fidelity but how experience is stored and governed across its lifecycle, opening a control-theoretic perspective on agent memory.

Sparse autoencoders (SAEs) seek to decompose the dense activations of large language models (LLMs) into sparse, potentially interpretable features using a learned overcomplete dictionary. In this work, we ask whether it is possible to enhance the semantic selectivity of the learned dictionary atoms by linearly \emph{shaping} the dense activations prior to learning an SAE. We explore rank-reducing transformations based on the recently introduced J-space corresponding to the averaged Jacobian from the current layer to the LLM output, projecting activations onto the ``live'' subspace spanned by eigenmodes of the Jacobian with larger singular values (which we expect to be more semantically relevant). We also introduce a new approach for training SAEs in which (a)~an affine shaping transformation is learned along with the dictionary atoms, (b)~the atom selection mechanism is based on the TopK projections along the one-dimensional subspace spanned by the atoms, and (c)~reconstruction is carried out using the same atoms. We call this approach a Matched Filter SAE (MF-SAE) based on analogies with communication theory. The promise of shaping is demonstrated via evaluations against dense activations and state-of-the-art SAEs on benchmarks including SAEBench and SAE Probes.

Is there a fundamental tradeoff between desiderata for safe AI systems, such as harmlessness and helpfulness? We answer by formalizing the interaction between AI systems and users as a game with three players: a defender and an attacker that are computationally or informationally bounded, and an AI that provides public advice and is less constrained. Our main result shows that there is a fundamental tradeoff in providing advice that helps the defender while advice remains public. We also consider whether the information that the AI shares can be manipulated to benefit only the defender and conclude with a discussion of how we can use these results in designing safer AI systems.

Span-level hallucination detection pinpoints unsupported content in LLM outputs, enabling actionable feedback and targeted mitigation. Existing evaluations for span-level hallucination detection methods are fragmented across tasks, annotation formats, and metrics, and most span-level datasets cover only a single NLG domain, making it difficult to assess whether a detection method actually generalizes across tasks and domains. In this paper, we introduce SHAM-Bench (Span-level Hallucination Benchmark Across Multiple Tasks), a unified benchmark for span-level hallucination detection spanning six NLG tasks, aggregated from existing human-annotated datasets totaling 6,461 examples under a common schema and evaluation protocol. Using SHAM-Bench, we evaluate existing hallucination detection methods and find that fine-tuned detectors often perform well in-domain but degrade under distribution shift, whereas prompt-based methods with reasoning are more robust across tasks. Our analysis further shows that subtle unsupported content remains difficult to detect and span-level evaluation is still challenged by ambiguous span boundaries and noisy labels. We release the dataset, evaluation code, and a public leaderboard to support future research on broadly generalizable hallucination detection.

Reflective prompt optimization improves instructions using task feedback, but coupling candidate generation with full-validation-based selection can make search costly and sequential. We propose a two-stage framework that decouples reflective candidate-pool construction from validation allocation. During construction, we sample multiple parents and reuse each parent's training responses across parallel reflection branches. Each revision is screened against its parent on a small training minibatch, allowing locally improved candidates to support further revisions without full validation. After construction, an exploration policy selects candidates from the frozen pool under a fixed full-validation budget, and the prompt with the highest measured validation score is returned. This separation allows candidate-pool size and validation budget to be set independently. We evaluate random selection and learned score prediction with similarity-based exploration on five benchmarks using Qwen3.5-9B and Llama-3.1-8B-Instruct. With 100 retained candidates and 20 full validations, our framework improves over the initial solver by up to 11.22 percentage points. Learned selection improves over the initial solver in all ten comparisons, although neither selection policy consistently outperforms the other. These results support separating reflective search from budgeted validation while leaving room for stronger candidate-selection policies.

We study trail-weighted pruning of evidence for Bayesian-network prediction. A traversal enumerates the active trails from each evidence node to the outcome, scores each trail by the product of its edge strengths, and drops evidence whose score falls below a threshold. For collider-free networks (chains and forks) over binary variables, we show that the trail score equals the exact correlation between a trail's endpoints, that a chain and a fork with equal edge strengths have identical joint distributions, and that the pruned posterior is within a computable sum of per-trail terms of the exact posterior when each dropped trail lies in its own branch of the outcome node. We prove this bound, check it exhaustively on all small trees, and show that it fails when dropped trails share a branch. In simulations on random trees, thresholding trail scores gives 1.4 to 13 times lower error than random pruning of the same size. We also apply the score, as an unproven heuristic, to seven bnlearn networks with colliders and loops. Keeping the top half of the active evidence by trail score gives lower error than random pruning on all seven, but it is not clearly better than a mutual-information ranking, and scoring costs more than the inference it saves on six of the seven, so we show no speedup. Models trained on trail, graph and information features and tested on networks they never saw do not beat the trail score or mutual information, but trail features carry most of what they use and clearly outperform generic graph features. A bound for general networks is left to future work.

In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next observation are learned, it is unclear which features they will include. Earlier characterizations describe the features JEPA selects in terms of their individual predictability, raising the concern that unpredictable features, like a traffic light, will not be encoded into the representation. However, the prediction error is also affected by the relationships between features, e.g., the light's color helps predict whether the traffic will stop. We propose the predictive closure characterization, which accounts for these relationships. Based on a reformulation of the JEPA training objective, we show that selection depends on properties of feature sets. This view agrees with earlier works when features evolve independently, but makes new predictions when they do not, for instance that a feature like the traffic light can still be retained. We test this characterization in controlled experiments, which show that unpredictable features can become better represented when they help predict other features, and that changing one feature's predictive relationships can affect how well others are represented. We also demonstrate implications for steering feature selection: targeting one feature with an auxiliary objective, or supplying it to the predictor as side information, can strengthen the representation of another feature whose future values it helps predict, so steering can be indirect.

Multi-agent LLM systems can formalize natural-language argumentation end-to-end, turning plain text into structured graphs by classifying components, connecting them with relations, and enriching the resulting graphs into formal models. Quality assurance for those graphs is almost always another LLMs: a reviewer, judge, or critic agent. On the other side, in adjacent neuro-symbolic fields, such generated output is often verified symbolically by external checks and solvers. In this work, we argue that in argument graph verification, the allocation should be similar, and based on what each instrument can guarantee. We propose layered verification: starting from deterministic checks, moving to solver-based and LLM reviews, and utilizing debate only under uncertainty, leaving human judgment for what no other instrument can decide. We ground our principle with a study on human-annotated graphs with injected structural defects. An LLM reviewer detects them at near-ceiling recall, including cycles in the support relation, but it objects to almost one in a ten correct graphs, even though it should not, happens to be inconsistent in verdict on identical inputs and is significantly slower than the equivalent exact check. It shows that while LLM reviewer does not lack the ability to detect structural defects, it is worth having a deterministic verification gates.

Scaling laws suggest that increasing model size, training data, and compute generally yields monotonic performance improvements. While this often holds at the aggregate level, we present systematic evidence that vision-language models (VLMs) exhibit inverse scaling at the subtask level: specific question categories on which larger models perform worse than smaller counterparts within the same model family. We evaluate 19 models from four distinct model families across eight multimodal benchmarks (nine evaluation splits), answering 29,386 questions per model, 558,334 model-question evaluations in total. Using per-question pairwise analysis across all model pairs in each family, we identify 375 subtask-model-pair combinations exhibiting inverse scaling, spanning 4,342 unique queries where a smaller model answers correctly but a larger model fails, and we find that these patterns replicate across independently developed architectures. We identify 13 subtask-dataset pairs that show inverse scaling across multiple families, consistent with the phenomenon being task-driven rather than model-specific. The most affected domains in our study include multi-image spatial reasoning, visual perception, and university-level STEM knowledge, with accuracy gaps exceeding 20 percentage points in several cases, while diagram understanding appears comparatively resistant. These findings challenge the assumption that increasing model size uniformly improves multimodal capabilities and suggest that model selection should be task-aware.

Byzantine-robust aggregation shields federated learning from malicious updates, but the rules that filter out suspect updates do so by comparing clients' parameter vectors. That comparison requires access to every client's update, which secure aggregation exists to withhold, costs $O(n^2 d)$ in the parameter dimension $d$, and is defined only between clients that share an architecture. We propose LiBRA, which compares clients on what their models compute rather than on the parameters that compute them, using a fingerprint of $C^2$ values for $C$ classes, the class-conditional logit response of each client's model to pseudo-patterns optimised against it. We instantiate LiBRA as variants of Krum, Multi-Krum and Bulyan that change only what the dissimilarity score is computed on. On Fashion-MNIST and CIFAR-10 with three architectures and four backdoor and label-manipulation attacks under i.i.d. partitions, each variant holds every attack its parent holds, and on identical client updates under a backdoor attack the fingerprint ranks malicious clients within 0.006 AUC of the full parameter vector. The variants thus match the robustness of the rules they extend, while their trust decision reads no client parameter, compares up to 49,090 times fewer values, and can by construction compare clients whose architectures differ, at the cost of one pattern optimisation per client per round.

Firmware diagnostic logging (printf) is an essential debug instrument in full-chip SoC design verification, yet it imposes a heavy tax on cycle-accurate RTL simulation: formatting a 30-character, 1-argument message on a simulated CPU and serializing it through an MMIO aperture costs 3.04 to 15.23 $μ$s per call (15.27 $μ$s unweighted / 15.08 $μ$s print-weighted critical-path mean on Suite A). Unlike embedded tokenizers that target flash footprint and wire bytes with offline decoding, RTL simulation is governed by a two-regime cost function---cross-fabric SoC paths are dominated by MMIO bus stalls (84.7% of call latency) whereas local subsystem paths are dominated by on-core vsprintf formatting (61.7%)---and requires live, zero-simulation-time string reconstruction to drive UVM scoreboards. We present a host-side in-simulation tokenized logging architecture for heterogeneous multi-core SoCs that eliminates both bottlenecks with zero firmware call-site edits. Using link-time ELF format-string virtual addresses as tokens over a fixed 32-bit word-aligned slot protocol across Arm, RISC-V, and DSP cores, on-core logging collapses to a single-pass specifier scan and $(1+S)$ MMIO word writes. This removes all on-core radix conversion ($C_{\mathrm{fmt}} \to 0$), 85.0%--94.5% of deployed bus transactions, 84.6%--93.3% of MMIO bus-stall cycles, and 63.5%--90.9% of total on-core logging cycles ($2.74\times$--$11.03\times$), while a SystemVerilog monitor reconstructs bit-exact strings in zero simulated cycles ($Δt_{\mathrm{sim}} = 0$). Across 858 matched full-chip test pairs (1,716 paired runs across 1,972 runs), we achieve 100% output parity with zero simulator memory overhead, saving 1,314.85 CPU hours (-26.18%) per regression pass, cutting median test CPU time by 41.51% (Suite A) and 21.33% (Suite B), and shrinking traces by 42.84% and waveforms by 28.82%.

Beneath the apparent diversity of foundation models lies an increasingly homogeneous technical stack of common training pipelines, distillation methods, and data. These shared algorithmic choices can expose an otherwise distinct model ecosystem to concentrated risks leading to a vulnerability monoculture. Yet current practice in assessing adversarial robustness is restricted to individual robustness scores and attack transfer rates, leaving open the unexplored fundamental question: how many attacks suffice for catastrophic widespread compromise? In this paper, we investigate the empirical nature of shared vulnerabilities within a model pool and introduce Systemic Adversarial Robustness (SAR), a novel measure of shared vulnerability of a model pool through the attack diversity needed to exploit it. Specifically, SAR interpolates between single-attack transferability and independent per-model robustness: small values reveal a vulnerability monoculture, whereas large values indicate that failures are dispersed across models. We theoretically formulate SAR as a partial set-cover problem with provable approximation guarantees---enabling the development of scalable algorithms to empirically measure the systemic adversarial robustness. We empirically test the shared vulnerability of vision and large language models and characterize their shared collective vulnerability curve. Specifically, we demonstrate that model pools distilled from DeepSeek-R1 and Claude Opus 4.6 exhibit lower estimated systemic robustness than their corresponding parent pools, revealing a potential systemic cost of frontier-model distillation. Our framework seeks to establish vulnerability monoculture as a measurable target for a more wholistic adversarial evaluation of model populations.

Accurate IoT device identification is essential for network security, yet many existing approaches rely on payload inspection, feature engineering, or long observation windows that limit scalability and often assume stable network conditions despite the inherent variability of real-world IoT traffic. We propose a lightweight IoT device identification approach based on packet headers and locality-sensitive hashing (LSH), generating compact fingerprint representations at packet- and window-level granularities. To evaluate robustness, we consider packet-level traffic anomalies including drop, perturbation, reordering, duplication, and packet injection. Experimental results show that the proposed approach achieves up to 99% accuracy while maintaining robustness under realistic traffic distortions using short bounded observation windows. Anomalies affecting global traffic structure produce the largest performance degradation, whereas window-level representations improve robustness by reducing sensitivity to localized distortions. The low computational overhead of the proposed approach enables scalable IoT device identification in resource-constrained environments.

Frontier models can control robots through agent harnesses that loop on one question: what motion comes next toward the user's goal? We show that this default loop carries out harmful instructions. With GPT-6 Astra controlling a low-cost SO-101 arm through the Inspect Robots harness on three household tasks whose hazard is common knowledge, such as "put the knife in the toy basket", the policy refused 0% of 15 trials, attempted the harm every time, and completed it in 27%. Its one exit, give_up, is meant for goals that cannot be achieved, and it never cited the hazard. We add an optional runtime safety check to the loop: before the first motion, the model is asked whether it would carry out the instruction, given the same camera frames. The check would have stopped all 15 hazardous instructions, including the 4 that completed the harm, and flagged none of the 15 harmless controls. Code and data: https://github.com/Arnav-Gr0ver/Refuses-When-Asked-Complies-When-Acting.

Checking whether a text complies with a body of law is slow when done by hand and fragile when done with keyword rules. This paper presents ComplianceAgent, the legal compliance sub-agent of COMPASS, a multi-agent framework that evaluates texts along four axes: digital sovereignty, carbon impact, compliance, and ethics. The agent relies on a large language model that scores a text on a fixed rubric, grounds its judgment in a reference legal corpus, and, when the score falls at or below a critical threshold, runs a corrective chain made of self-healing rewrites, a compliance label, an explainability report, and escalation to a human reviewer. We describe what the language model does at each step, how the memory of the agent is organized, and how the same implementation can be pointed at different reference corpora. We evaluate the agent in a Canadian and Quebec context, on a corpus of thirteen legal instruments that brings together federal and provincial texts, international treaties, and municipal documents of the City of Trois-Rivières. The evaluation relies on a hand-annotated gold standard of fifteen texts and combines classification metrics, a repeated-query stability test, retrieval precision, and a composite score. On this gold standard the agent separates every conforming text from every non-conforming one, its scores remain stable under repeated queries, and the index returns relevant passages.

The Brick schema represents building equipment, sensors, spaces, and the relationships between them as typed, multi-relational graphs, providing machine-readable metadata that supports generalizable control applications. Current methods for producing complete Brick models require extensive manual examination of building schematics and point lists. We observe that as Brick models are knowledge graphs, inferring missing relations is naturally posed as a knowledge graph completion problem. To the best of our knowledge, this is the first application of knowledge graph completion to Brick model generation. We compare three methods, gradient-boosted trees, Pairwise Brick Relation Ranking, and CompGCN, on the relations feeds, hasPart, and hasPoint, which account for the majority of the schematic-reading work in building a Brick model by hand. Models are evaluated on held-out MORTAR buildings and out-of-distribution Kaggle and Building Time Series data. Although CompGCN is the strongest performer on held-out buildings under metrics resistant to class imbalance, leading macro-F1 and AUPRC, no method generalizes uniformly across datasets or relations.

The research idea in a published paper is rarely the project's initial idea. Between the two lies a closed-loop process in which researchers run experiments, derive insights, identify challenges, explore the literature, and revise the idea, with each step motivated by evidence and reasoning they can articulate. We call the path one idea takes through this process, with the evidence and reasoning behind each step, its authentic research trajectory. Although it underlies every scientific contribution, it is rarely recorded. Existing LLM-based research agents generate an idea once, refine it against reviews, literature, or their own experimental results, or search over candidate ideas and experiments. In all of them the text of the idea evolves, while the evidence and reasoning that connected one version to the next are discarded. We propose the Research Progress Graph (RPG), a representation of authentic and complete research trajectories. An RPG records the evolution of an idea as typed research states connected by cognitive transitions, each annotated with a temporal index and the reasoning behind it, which makes this evolution an explicit object of study. We ground the representation in authentic trajectories whose steps their authors justify in annotations and papers. Comparing the representations of thirty-five research agents with the RPG, we find that each is an incomplete RPG; none keeps the state that changed, the evidence that changed it, and the reasoning connecting them as one evolving record. Modeling ideation with the RPG changes how research agents should be designed, evaluated, and trained. Agents predict the next research state instead of a complete idea, ideas are evaluated as trajectories instead of final texts, and authentic trajectories serve as training data. These directions shift research agents from generating static ideas to modeling how ideas evolve.

We propose EMBER, a structure-based framework for TCR--pMHC binding prediction that eliminates the need for a predicted docked complex, and with it the placement error that dominates such predictions. Traditional structure-based predictors utilize the co-folded complex structure of the receptor and the antigen and interpret binding from its contact surface. Consequently, every downstream feature inherits any deviation the structure-prediction tool introduced. Methodologically, EMBER folds the TCR and pMHC independently, encodes them as SE(3)-invariant residue graphs, and compares their learned representations without constructing a shared coordinate frame. A relative pose is therefore not an input of the model, so placement error cannot propagate into the prediction. Computationally, EMBER preprocesses $O(N{+}M)$ monomer folds where a complex-based predictor needs $O(NM)$ operations. Experimentally, under epitope-disjoint evaluation with six retrained baselines, EMBER is first on every metric under all four folding tools, by 0.9 to 14.4% in micro-AUROC, 18.5 to 25.3% in micro-AUPRC and 10.6 to 24.1% in MCC. Its ranking is also up to 85% more stable across folding tools.

Coding agents can answer analytical questions over a data lake by discovering, cleaning, and joining the relevant data. However, they discard this work after each question and redo it from scratch for the next one. Tool evolution, which automatically builds and refines reusable tools across tasks, is a natural fit for this repetition. Since a ground-truth label for a data-lake question requires an expert to solve it first, we consider the unsupervised setting. We propose TEAL (Tool Evolution for Analytics over data Lakes), which evolves a library of tools and usage skills for a given Worker agent without labels. Two agents propose tools from complementary evidence, an Explorer from the data lake itself and a Critic from Worker trajectories. A proposed change is kept only if the Worker adopts the tool on a rerun without extra steps and any changed answer is supported by trajectory evidence. We also introduce TEAL-Q, which generates synthetic questions for tool evolution when real ones are unavailable. We evaluate on KramaBench and LakeQA with four Worker LLMs under step budgets, which cap the number of steps the Worker may take per question. With a tight budget of 10 or 15 steps, TEAL improves every Worker on both benchmarks, lifting Sonnet-5 at 15 steps by 9.3 points on KramaBench and 21.7 points on LakeQA. A tighter budget also yields better feedback for evolution. For example, Sonnet-5's LakeQA library evolved at 20 steps and deployed unchanged at 30 steps scores 68.3, above both the baseline (59.6) and a library evolved directly at 30 steps (64.2). Additionally, a library evolved on synthetic questions performs comparably to one evolved on real questions. These results show that useful tools can be evolved without labels, and hopefully inspire future work on what to evolve and how to obtain feedback in unsupervised agent evolution.

Rust is increasingly used in Linux to improve memory safety, but its guarantees do not prevent logical races: individually synchronized operations can still interleave in ways that violate higher-level invariants. Detecting such bugs typically requires manually written specifications and carefully controlled execution environments. We present XRust, a technique for detecting logical races in Linux Rust drivers without manually specified invariants. Using Logical Happens-Before (LHB) analysis, XRust identifies shared-state accesses relevant to logical ordering, automatically constructs operation-pair harnesses, and uses serial differential oracles to detect concurrent executions inconsistent with either serial order. XRust also uses lightweight userspace abstractions to execute driver code under reproducible schedules. Applied to Linux 7.3-rc2, XRust identifies 17 shared objects across 102 Rust files from ten drivers, constructs 1,654 executable harnesses, and discovers two previously unknown logical races. The rnull block-driver race has been confirmed by Linux maintainers, with a fix accepted into the Linux block subsystem tree.

Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM should reveal location information only at the granularity appropriate to the visual context depicted in the image. We first show that this contextually appropriate granularity is recoverable from pre-generation hidden states, even when generated responses are misaligned, with disclosure levels exhibiting an ordered structure along a common hidden-state direction. Motivated by this, we propose an inference-time route-and-steer framework: an ordinal granularity router maps visual-contextual privacy representations to a disclosure route, and a disclosure-level-conditioned steering module applies route-specific hidden-state interventions during decoding. Across four VLM backbones, our method improves visual-context-aware disclosure alignment, better balances over- and under-disclosure, and incurs only modest changes on general capability benchmarks. It further transfers to another dataset with a distinct data distribution and privacy-risk taxonomy.

Deep latent variable models (DLVMs) underpin many modern analysis pipelines, for example in single-cell genomics, yet their prior distributions have received markedly little attention. The VampPrior approximates the optimal prior (the aggregate posterior) and often attains high model evidence. Here, we identify two key limitations of the VampPrior: it ties prior and posterior uncertainty to a single amortized covariance and clusters poorly due to its fixed uniform mixture weights. We unleash the VampPrior by decoupling its prior and posterior covariances, resolving the conflict of generating high prior and low posterior entropy with the same shared parameters and substantially improving model evidence. We then replace the uniform mixture with a Dirichlet process. When combined, our proposals provide higher model evidence than the VampPrior and robust non-parametric clustering. Augmenting single cell Variational Inference (scVI) with our prior improves single-cell integration and clustering across six atlases while significantly raising model evidence. Our approach changes only the prior and thus can be applied to any DLVM with a continuous latent space.

Gradient descent with momentum can be viewed as a particle moving with friction in the loss landscape. In deep learning the learning rate is usually large, so training is a discrete dynamical system: in many full-batch experiments the sharpness grows to the optimizer's stability limit, the edge of stability (EoS), and training continues there. Physics-inspired optimizers change this motion; Velocity-Regularized Adam (VRAdam), for example, shrinks the step at high speed. We study kinetic maps, in which the step is the gradient of a kinetic energy of the momentum, and ask what they do beyond the EoS. (i) In a reduced model, a kinetic map creates a symmetric period-two orbit beyond the edge, strictly linearly stable if and only if the map's elasticity (the log-log slope of step length versus momentum) lies between 0 and 1; on it, training sits at an effective edge of stability. VRAdam's damper fails above twice the critical learning rate, where its kinetic energy is not convex; the Legendre dual of VRAdam's own Lagrangian is convex and passes at every learning rate. (ii) Because the sharpness rises gradually, the loss of stability is delayed, as in Neishtadt's theory of slowly swept bifurcations. In a network with a controlled sweep, the dynamics match this theory without fitted parameters, including the memory of where the sweep started and its erasure by noise. (iii) With momentum, reaching the oscillation also requires elasticity at large momentum: maps that bound the step have competing period-four and chaotic attractors already at beta1 = 0.9, while the cubic-Legendre map reached the oscillation from every sampled start up to beta1 = 0.99. Experiments on an exact quadratic problem, a full-batch MLP and ResNet-32 support the theory; on ResNet-18, much of the apparent benefit of the maps is reproduced by matching the effective learning rate.

Optical Character Recognition (OCR) and handwritten character recognition (HCR) in low-resource historical material collections often rely on data augmentation, yet augmentation strategies are typically evaluated almost exclusively through predictive accuracy. This leaves open whether a model's improved performance is accompanied by saliency stability. This question is important in low-resource historical HCR, where recognition is performed on variable and partially degraded letterforms and two models with similar predictive performance may differ in how consistently their saliency patterns respond to such variation. We therefore study saliency-guided augmentation as an explanation-aware strategy for low-resource HCR. Rather than treating saliency maps only as post-hoc interpretations, we use them to define augmentation variants that perturb, preserve, or constrain visually informative regions of character images. We compare classical document-style transformations, random lacunae, and several saliency-constrained lacuna variants. To evaluate the effect of augmentation beyond accuracy, we introduce the Explainability Consistency Index (ECI), a metric that quantifies the stability of saliency patterns under controlled perturbations. We find that saliency-constrained augmentation can improve recognition over classical augmentation, but no single saliency strategy is consistently best. The preferred form of the constraint varies with both architecture and corpus, indicating that saliency is better treated as a mechanism for controlling augmentation than as a fixed augmentation strategy. In some settings, saliency-constrained training improves explanation consistency even when predictive performance is similar. This shows that predictive metrics alone can obscure differences in model sensitivity under perturbation, motivating explanation consistency as a complementary evaluation dimension.

For AI assistants to move beyond simple chat interfaces and reliably assist users with applications like Gmail, Venmo, and Slack, we need to improve their tool-use capabilities. However, existing tool use benchmarks are still limited and difficult to scale because they rely on human supervision or are too domain-specific. To make data generation scalable, we propose, TURTLE (**T**ool **U**se **R**elated **T**asks with **L**ong-Horizon **E**volution), an Unsupervised Environment Design (UED) approach to evolve existing tasks using a proposer, a solver, and a feedback generator large language model agents. The proposer is rewarded for difficult yet solvable tasks based on the solver's empirical performance, and the feedback generator summarizes the solver's execution traces for the proposer to exploit. TURTLE produces evolving, self-tuning test suites adaptable to different environment requirements, from supporting user simulators and dense document retrieval to interacting across multiple apps. Empirical results on AppWorld, $τ$-Bench, and AutomationBench using Gemma 4 31B-it show that the evolved tasks target the desired difficulty range (e.g., a 165\% increase in the percentage of hard tasks on AppWorld), increase task diversity (in terms of reasoning and simultaneous tool interactions), and elicit a broader range of failure modes than the original tasks.

Industrial artificial intelligence (AI) planning systems combine predictive models, optimization logic, controls, and governance rules, but improvement in the complete policy does not show whether one addition made an independent contribution or whether that contribution justifies its burden. We address this attribution problem with an observable evidence standard that separates opportunity to act, executed-plan change, operational consequence, equivalence, uncertainty, and operating scope. Using forecast-to-plan replay on Favorita, M5, and Walmart data, we apply the standard to heterogeneous planning additions. Three limits of single-indicator evaluation emerge. First, a composite policy can conceal negligible component contribution: one control changed only 15 of 681 executed plans. Second, changing execution does not guarantee stable benefit: the evaluated multi-period capability helped in several long-lead settings but was 7.6% more costly under tight capacity. Third, indirect evidence can prioritize review but cannot complete attribution. Independent runs produced different paths in 19 of 599 cases that a rule-based check predicted would be unchanged. Although a lower-cost diagnostic matched the full-evaluation winner in only 55.0% of cases, it selected an option within the $10^{-6}$ cost-equivalence tolerance in 79.5%; the remaining cases prevented the proxy from authorizing acceptance. These experiments establish capability-level contribution evidence; the organization completes the value decision by weighing it against implementation and sustaining burden. The resulting capability evidence record links the observations to validation, deployment scope, fallback, monitoring, redesign, and retirement decisions.

Large language models are deployed globally, yet safety alignment and evaluation remain concentrated in English: harmful requests refused in English can elicit unsafe completions when translated into low-resource languages. We evaluate ten widely deployed LLMs across 79 languages (English plus 78 low-resource) under a translation-bypass threat model, using a metric conditioned on English safety, which separates translation bypass from unsafety already present in English. English safety does not transfer reliably: conditioned on safe English completions, mean bypass rates range from 21.9% to 44.4%. Bypass risk concentrates in a subset of language families, a pattern broadly shared across models. The shared ordering is confounded with translation degeneracy, but its variation across models is significant and survives every control we apply. We also show that per-model bypass rates do not identify a stable ranking: nine of ten models change rank across three conditioning schemes that estimate the same quantity, with no change to any underlying response.

Language-model agents compact their context for the question at hand, but the compacted context must also serve questions that come later. We evaluate compaction on these later questions, using four corpora that annotate the evidence of several questions per text and recorded agent trajectories, and compare every method with random deletion at the same budget. On MultiRC, keeping the sentences that gpt-6-luna (through the OpenAI Decisions API) judges necessary for the current question raises its accuracy on that question by 11.4 points at 25% of the tokens. On the later questions, however, no method that selects sentences for the current question does detectably better than deleting sentences at random. Deleting words instead of sentences does better: LLMLingua-2, which ignores the question, beats random sentence deletion at the same token count by 12.5 points on the later MultiRC questions. It keeps about twice random deletion's share of the evidence's content words, and at equal content-word retention its gain shrinks to 7.3 points (exploratory). On FEVER, label accuracy can overstate what the kept text supports: the gain of sentence selection is no longer detected when answers are scored against the kept evidence, and the 21.0-point gain of LLMLingua-2 is not shown to rest on the kept text. The ranking also depends on the setting: on Evidence Inference sentence selection beats LLMLingua-2 at a nominal 25% budget, and on $τ$-bench trajectories, whose later tool calls need exact identifiers, it beats random deletion while LLMLingua-2 shows no detected gain. Scores on the question a compaction was made for do not reliably show how it serves later ones.

Class-Incremental Learning (CIL) aims to learn new classes continuously without forgetting the previously learned ones. Recent works in CIL have significantly advanced across various modalities; however, at the intersection of audio-visual modality, this problem remains underexplored. While representations from multimodal foundational models (MFMs) encapsulate rich static priors, our preliminary analysis shows that these representations struggle in incremental settings. Motivated by this, we propose a framework that generalizes frozen MFMs to audio-visual CIL. Specifically, we leverage their dense audio and visual representations and employ an attention-based contextual guidance module, in which the audio representations serve as task-stable contextual guidance for the visual representations, enabling effective class recognition in the incremental setting. To further mitigate catastrophic forgetting and preserve stability across tasks, we employ feature- and logit-level distillation, which are primarily responsible for the reduced forgetting. Extensive evaluations across various audio-visual CIL benchmarks demonstrate that our approach consistently outperforms state-of-the-art methods. Code is available at https://github.com/avigupta2798/MFM-CIL.

Progress in artificial intelligence is widely assumed to scale with effort: more researchers, papers and investment should accelerate discovery. Theories of technological growth instead locate the difficulty of discovery in the space of ideas. Here we show that neither scale nor space is enough: breakthroughs depend on diverse attention, which collective concentration can deplete. Across 36 years of computer science, we analyze 7.2 million papers, 73,089 peer reviews and 3.7 million researchers and their topic shifts. The field expanded while its attention narrowed sharply after the Transformer. Over half of topic-shifting researchers moved into AI, especially mid-career ones, and reviewer confidence fell by a full point on a 5-point scale. Each successive breakthrough, from AlexNet to large language models, raised the field's carrying capacity for ideas but was followed by steeper exhaustion of those ideas. Bigger breakthroughs capture attention more completely, depleting the diversity their successors may draw upon. Structural biology shows the same cycle after cryo-EM and AlphaFold. Simulations show how well-intentioned and widely used policies such as reciprocal reviewing empower prolific authors and further narrow what the field accepts. We call for institutions that protect diverse attention in artificial intelligence rather than reward concentration on the winning paradigm.

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

Safe Multi-Agent Reinforcement Learning (SMARL) aims to learn cooperative control policies under prescribed safety constraints. Al- though existing SMARL methods focus on enforcing safety during policy optimization, their vulnerability remains largely unexplored. This paper investigates inference-time observation attacks against SMARL, where an adversary perturbs the observations of a subset of agents, thereby altering the actions produced by their policies. We propose MinMax Attack, an attack framework that jointly selects adversarial target actions for compromised agents while accounting for the Cooperative Resistance of benign agents. We further derive a bound on the induced safety cost and characterize its relation- ship with the Compromised-Agent Budget and the Perturbation Budget. Experiments across multiple safety-critical multi-agent benchmarks and representative state-of-the-art SMARL algorithms show that MinMax Attack induces substantially higher safety costs than existing attack baselines while preserving task reward, thereby reducing its detectability and achieving both attack effectiveness and stealthiness.

Test-case generation is critical but remains challenging for language models. Existing test-time scaling methods either isolate candidates or aggregate traces, which do not scale to long-horizon tasks. We present TEX, a collaborative test-time scaling framework based on execution-driven cross-validation. TEX parallelizes agents to generate tests and patches, cross-executing each test against every patch to form a $K \times K$ feedback matrix. Agents use their matrix slices for refinement, enabling objective validation without verbose trace sharing. On SWT-Bench Verified, TEX achieves 86.4%, establishing a new state-of-the-art and improving +9.0% over single-agent baselines. Moreover, even random selection from TEX outputs (84.4%) surpasses the previous SOTA (84.0%), showing collaborative execution feedback improves solution quality, not just best-case samples. Our results highlight execution-grounded feedback as an effective scaling mechanism for long-horizon agentic tasks.

The WebView interface in Android (hybrid) applications enables the native side (e.g., Java) to exchange data with the Web (e.g., JavaScript). This creates cross-language information flow channels that existing single-language analyzers cannot analyze. Existing hybrid app analyses address only part of the problem: some build bridge-aware call graphs but provide limited information-flow analysis, some rely on coarse assumptions, and some focus only on web vulnerabilities. We present MifDroid, a static-analysis- guided dynamic-analysis framework with a lifecycle-aware per-WebView call graph, a bridge-aware type-and-effect system, and a modular JavaScript analysis engine. We evaluated MifDroid on 338 apps from the F-Droid repository (out of 4,126 total) that use addJavascriptInterface, enabling active communication between Java and JavaScript. Our guided dynamic phase adds 662 entries to the 444 produced by static analysis, raising the resolution rate from 36.5% to 74.5%, produces the resolved bridge-related entries for 115 apps that static analysis leaves empty. MifDroid reported 14 vulnerable cross-language information flows, five of which have been accepted and fixed by the maintainers.

Effective detection of hate content is essential for maintaining healthy online communities, yet performance can vary substantially across languages. This work investigates a core practical question: Can training data from one language support hate-speech detection in another language with comparable performance? To answer this question, we conduct a large-scale study across 14 languages. \change{We observe consistent patterns in which certain language pairs transfer more effectively than others. Among the factors we measure, linguistic similarity is the strongest predictor of these patterns, with temporal overlap and dataset-size difference playing smaller roles. These measurable factors do not fully explain the observed patterns; a complementary qualitative analysis suggests that shared cultural cues and communicative norms may be associated with some of the remaining cases, particularly within several clusters. We present this cultural interpretation as a hypothesis from qualitative inspection rather than a measured cause.}{} Our analysis includes low-resource languages, offering practical insights into when cross-lingual transfer can support them and where its limitations emerge. These findings inform the design of more reliable multilingual hate-detection systems.

Offline goal-conditioned reinforcement learning enables skill acquisition by reaching any goal from any state without dense rewards. Model-based methods allow for planning by scaling test-time compute to solve complex tasks. Combining both holds appealing promise for pre-training on suboptimal demonstrations and zero-shot adaptation via goal specification and planning. Yet existing methods assume effective planning requires accurate next-state prediction in pixel or latent space. In this work, we show that value prediction is more useful for planning, whereas accurate state prediction is often unnecessary. Our key insight is that offline goal reaching naturally admits a closed-form Monte Carlo return, enabling value regression to replace temporal-difference bootstrapping. We use this return as the sole supervision for latent world models: imagined latents are trained to predict goal-reaching value rather than match future states, implicitly preventing representation collapse. These planning-oriented representations let imperfect world models imagine on the level sets of value field beyond dataset transitions. We then present VPFlow, a value-predictive planner that pairs the world models with a one-step flow policy and plans in its noise space via evolutionary search. It enables policy improvement with imagined generalized advantage and reduces out-of-distribution queries common in action-space planning. On 30 state and pixel OGBench datasets, VPFlow surpasses model-free baselines by at least 30.3 and 10.1 percentage points in success rate, respectively, and outperforms latent planners LeWorldModel, PLDM, and TD-MPC2. It achieves this despite a latent prediction error orders of magnitude larger than with state prediction, while minimizing this error lowers success rate. This demonstrates that value-predictive representations, not exact state prediction, are what offline planning requires.

The development of embodied models of animal behavior is a major goal of the neuro-AI community. Embodied models will enhance our ability to make direct comparisons between biological and artificial agents. To make such comparisons effective, though, we need embodied models that actually move like animals. Previous work has developed virtual animal bodies and techniques for training realistic motion through imitation. However, to date, there is no system for embodied animal modeling that produces natural movements both during trained goal-directed behavior, and during exploration driven by random generation. Here, we introduce Simulated Control via Augmented Motor Priors for Embodied Reinforcement learning (SCAMPER), a system for training an embodied virtual rat model, enabling natural movements during both goal-directed behavior and random exploration. SCAMPER uses a variational architecture with two distinct priors, a motion prior for control and a behavior prior for exploration. We show that combining both of these priors allows SCAMPER to generate realistic behavior during exploration and exploit the task structure effectively. SCAMPER can also learn from sparse rewards as a result of the diverse movements induced by the behavior prior, allowing the system to model realistic sparse-reward tasks from experimental neuroscience. Altogether, SCAMPER provides the neuro-AI community with an effective new system for training embodied models of rodents on realistic tasks.

Vision-language-action (VLA) models are sensitive to visual corruptions, but knowing that a corruption is typically harmful does not establish whether removing it would benefit the current episode. Furthermore, when several corruptions co-occur, deciding what to correct requires understanding which ones contribute to failure risk and how much their removal would help. We define this problem as failure attribution. Detecting corruption presence alone cannot answer this question, as corruptions can be present but relatively harmless, while directly measuring removal benefits requires costly paired policy rollouts with and without the corruption present. To address this, we introduce a factorized risk model that learns separate representations of the scene and corruptions, allowing us to estimate removal benefits through latent intervention from the current image and task instruction, without paired clean images or additional counterfactual rollouts. We evaluate our approach on OpenVLA and $π_{0.5}$ in simulation and $π_{0.5}$-DROID on a Franka robot. Our method outperforms monolithic and vision-language-model (VLM) baselines, including achieving approximately 3x lower attribution error than frontier VLMs. We also demonstrate our method adapts to execution context and corruption intensity, and can be used for online failure monitoring on hardware.

Multi-vector retrieval models such as ColBERT often outperform single-vector models of comparable scale by matching queries and documents at the token level. However, retaining a vector for every token substantially increases storage requirements and slows retrieval compared to single-vector bi-encoders, limiting their scalability. We introduce FLINT (Fixed-budget Late INTeraction), an architecture that uses attention-based pooling to represent each query and document with a fixed number of embedding vectors. Unlike architectures that pool only document representations while retaining token-level queries, FLINT applies pooling symmetrically to both sides, limiting document storage and the number of vector comparisons needed during retrieval. FLINT achieves comparable ranking effectiveness to the closest baselines on BEIR and outperforms them on several benchmarks with longer and more complex queries, including BrowseComp+, CRUMB, and TREC Tip-of-the-Tongue 2025. We demonstrate that FLINT's effectiveness-efficiency benefits extend across encoder architectures, testing it on two separate backbones (BERT and ModernBERT). On long-query corpora, FLINT delivers 1.5-2.8 times faster end-to-end retrieval than ConstBERT, the closest baseline. FLINT also reduces index size by 5.3-9.3 times relative to token-level late-interaction models such as ColBERTv2.

Producing synthetic data for comparison testing a migrated COBOL application is a major challenge that agentic migrations face today. Our prior paper, the Agentic Method for Deterministic Validation of Legacy Code Migration, presents a deterministic oracle to facilitate such migrations, the Locksmith Loop. We now propose a novel method for achieving high branch and paragraph coverage through generated synthetic end-to-end test cases, by means of entry-replay test-case certification by observing the code execution trace on a surrogate COBOL program that mocks external interactions via a chronological external-operation tape. We applied our method to a very large COBOL program with 33,429 lines of code to generate 849 synthetic end-to-end test cases covering 2,973/3,806 source branches (78.1%) and 989/1,033 main-AST paragraphs (95.7%). Using the Locksmith Loop's projected-agreement validation, the generated Java code matched the COBOL surrogate on the specified observations across all 849 certified synthetic end-to-end test cases. During the campaign we detected 61 defects and fixed the Java generator accordingly. To test whether the coding agent could silently introduce business logic bugs, we seeded 450 defects into the Java code, such as deleted writes, deleted paragraph calls, and altered rounding, to verify how many of these would be detected by the gating methodology. The gate detected 47.6%, including all deleted writes and paragraph calls in executed code, but was unable to detect the rounding and precision defects.

With the widespread adoption of large language models (LLMs) across languages and domains, AI safety has emerged as a critical global concern. However, due to resource constraints, research and development efforts tend to prioritize a limited number of high-resource languages, leaving many others, especially those spoken in Africa, underrepresented. While existing research on language "resource-ness" has largely examined whether sufficient resources are available to develop language technologies, the question of whether languages are adequately represented in AI safety research has received attention only more recently. This report examines the readiness of African languages for inclusion in AI safety discussions, drawing on existing literature and data. It considers four factors: the availability of AI-relevant resources, including text data; national AI policy support; performance on established text benchmarks; and speaker population. Based on these factors, languages are grouped into four readiness categories: fully ready, mostly ready, emerging, and foundational support needed. This classification is intended to inform the inclusion of African languages in AI safety benchmarks, highlight varying levels of preparedness, and guide strategic investment in data development, policy, and community engagement for underrepresented languages.

Preference labels align model training, but the judgments behind them are difficult to inspect or verify. In this paper, we examine how preference construction can be checked when intermediate decisions are recorded, and which of them remain uncertain. We present Symbolic Preference Labeling, a method in which a pretrained language model classifies predefined attributes of prompts and responses. Explicit rules assign response verdicts, and a fixed verdict ranking determines the preference label. Each preference label has a recorded derivation: the response verdicts and preference label can be recomputed from the stored attribute assignments, rules, and verdict ranking without rerunning the language model. On 200 prompts annotated by human raters, the language model produces valid structured outputs for 192 prompts. All eight parsing failures occur in the subset of self-harm prompts marked as harmful. Using 465 preference pairs, we train one policy model with direct preference optimization across five training seeds. Mean measured benign refusal is 40.82 percent on 1,319 benign prompts that resemble unsafe ones, and mean measured toxic refusal is 84.98 percent on 655 toxic prompts. This study distinguishes the reproducibility of preference labels from attribute correctness. The recorded derivations make preference construction open to inspection, while attribute correctness and semantic safety require separate validation.

Over-fitted Gaussian mixtures can contain many parameter configurations with nearly the same density, making the order of density cancellation central to their parameter-learning rates. Ho and Nguyen encoded these cancellations in a polynomial system and conjectured its exact order as a function of cluster size. The key to resolving the conjecture is a common positive-variance shift that turns the coefficient equations into Gaussian moment matching. For every cluster size $\Mhn\ge2$, we prove that the real, nonzero-weight system has order $2\Mhn$: a sharp sign-change bound forces rigidity at that order, while positive Gauss--Hermite quadrature constructs a non-trivial solution through order $2\Mhn-1$. Population-density calculations at 110-digit working precision recover the predicted leading term for two to six components; finite multistart searches recover the same observed powers for two to five components, with smaller finite-scale constants for three to five components. For parameter learning, the exact order gives an explicit pointwise maximum-likelihood exponent and an exact local minimax order on specified shrinking neighbourhoods of a fixed known mixing measure. The same coefficient viewpoint separates high-order cancellation from exact parameter symmetries in mixtures of experts. Synchronized replication preserves the conditional density while keeping the corresponding parameter losses positive in two broad standard-softmax classes, showing why statistical rate arguments must account for exact parameter symmetries.

This paper is an exposition of the theorem prover Arend developed at JetBrains Research. Arend language has been designed to facilitate the formalization of mathematics within the framework of univalent foundations and with the focus on constructive mathematics. It is based on a variant of homotopy type theory with cubical syntax that is more closely aligned with Martin-Löf Type Theory (MLTT) than fully computational cubical type theories, while still retaining some of their computational properties useful in practice. Among the most notable features of the Arend language is its powerful system of classes and records supporting subsumptive subtyping, partial implementations and anonymous extensions, which provides extensive flexibility for constructing hierarchies of definitions. Formalization in Arend is streamlined by the IntelliJ Arend plugin, which transforms IntelliJ IDEA into a full-fledged integrated development environment (IDE) for the Arend language. The main Arend library arend-lib includes formalizations in constructive algebra, topology and synthetic homotopy theory.

The rapid evolution of Large Language Models is fueled by massive datasets, yet this reliance poses severe privacy risks, as models can inadvertently memorize and even directly output sensitive training data. Synthetic text generation offers a promising solution to this challenge; however, existing methods struggle to reconcile rigorous privacy guarantees with high data utility. In this paper, we propose a new paradigm for private synthetic text generation: Guided Diffusion. Leveraging the unique steerability of Diffusion Language Models, we introduce novel sampling and re-masking procedures that actively guide the iterative denoising steps. These mechanisms steer the generative process to reconstruct fluent text that adheres to a target semantic distribution. We demonstrate that our framework is naturally compatible with Differential Privacy (DP): once the semantic embeddings of the private corpus are protected, the DP guarantee carries over to the generated text by post-processing. Empirically, we show that our approach generates high-quality synthetic text with precise semantic alignment. Furthermore, we validate its effectiveness in DP synthetic text generation: across two datasets and two diffusion backbones, it outperforms Private Evolution and DP-SGD-based fine-tuning in distributional fidelity and downstream utility, while generating significantly faster than Private Evolution. We further extend the framework to Data-Specific Indistinguishability (DSI), which adds a formal certificate against data poisoning and memorization at a small utility cost.

Large language models can detect vulnerable code, yet the form of vulnerability encoding in their representations remains unclear. We analyze 8 code models spanning 6.7B-15B parameters on C programs using three datasets (DELTASECOMMITS, SVEN, and PRECISE-BUGS), comparing two diagnostic tasks on the same representations: identifying what patterns the code instantiates, and determining whether it is vulnerable. Models encode the first with high fidelity. Under the studied paired C-code datasets and linear-probing setup, vulnerability information is more accessible as a paired relational signal-whether code A is riskier than code B-than as a globally separable categorical label. This structural difference has implications for classification: the model reliably ranks which code is riskier but standard classifiers on isolated samples achieve limited accuracy because vulnerable and secure samples form overlapping distributions in representation space.

Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark comprising 27 task families (81 templates) alongside a 5-scenario fraud-risk reasoning suite that evaluates both verdicts and supporting evidence over synthetic, privacy-mitigated transaction data. Ground-truth answers are computed programmatically, and model performance is decomposed into format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E). Evaluating prompted proprietary baselines and six LoRA-adapted open-weight models shows that adaptation reliably enforces output schemas but leaves arithmetic bottlenecks: Gemma-4-31B-IT and Qwen3-8B attain E2E scores of 0.867 and 0.816, whereas evaluated MoE checkpoints reach only 0.632-0.713. Furthermore, on the 500-instance fraud suite, verdict accuracy frequently exceeds 80%, yet joint accuracy (requiring exact evidence attribution) collapses to at most ~23%. These findings identify schema compliance, semantic calculation, and evidence attribution as three distinct bottlenecks in transaction-sequence QA.

Generative AI can create programming feedback at scale, but validating its alignment with instructors' pedagogical goals and strategies remains challenging. Instructors may struggle to fully specify validation criteria upfront, as these often emerge through inspecting concrete examples and revising misaligned feedback. We present ValiAssist, a validation-by-demonstration system that infers reusable validation logic from instructors' revisions of AI-generated feedback and allows instructors to inspect and refine that logic as it is applied across other student submissions. In a comparative study with 16 instructors, ValiAssist helped instructors create and refine validation criteria, producing feedback that better aligned with their pedagogical intent (91% vs. 57%) and more than doubling the accuracy of correcting problematic criteria (77% vs. 32%). Our findings suggest that validation by demonstration can help instructors turn judgments they know how to make on individual cases into explicit, reusable logic, while progressively clarifying where those judgments do and do not generalize.

Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM editor revises each prompt in response to the target's answer and a judge's category-specific score and explanation. In the main setting, the target also receives feedback on its previous response. Experiments cover 500 seeds from seven datasets and ten topics, with intrinsic and extrinsic hallucinations evaluated across open-weight and closed-source models. Editing substantially increases judge-rated severity over the seeds. In the baseline configuration, retained responses fall in the judge's strong-hallucination bin in 77.2% of runs targeting extrinsic hallucination and 74.2% targeting intrinsic hallucination. Later editing rounds account for more than half of the mean score increase. Among 219 prompts that pass model-assisted source-validity checks, 89 elicit responses containing unsupported claims. Across five models, 200 curated prompts yield similar average severity scores, although pairwise Spearman correlations between prompt-level scores are only $ρ=0.0387$ to $0.1897$. Curated prompts can be reused across models, but their diagnostic value depends on source validity and the failures they expose in each model.

Reasoning structures (e.g., chains and graphs) are increasingly used for probabilistic inference, factuality verification, and uncertainty estimation, but their structural reliability is rarely evaluated directly. We test whether this reliability holds through a proposed two-stage diagnostic evaluation in open-ended multi-hop QA. Stage 1 tests different capabilities required to construct reasoning graphs, showing that graph reliability varies with topology, prompt design, and model family. Stage 2 uses uncertainty estimation as a downstream test case. Generated-edge experiments show that graph-based computation is sensitive to upstream graph errors, with performance dropping relative to the gold-graph setting. Gold-graph experiments further show that graph structure provides little benefit when used only as prompt context, but improves the separation of correct from incorrect answers by confidence score and enables node-level diagnostics.

Agent systems are shifting from task-scoped tool-use agents to persistent, personalized assistants such as OpenClaw. These claw-style agents maintain configuration files (e.g., AGENTS.md), long-term memory, and reusable skills. Once poisoned, this persistent state is repeatedly reloaded into the agent's context, creating persistent threats across tasks and sessions. Existing security evaluation systems struggle to evaluate these complex claw-style systems since they rely heavily on manually constructed test cases and environments, which limits their scale and coverage; in addition, these manually designed cases quickly become outdated as agents evolve, and their evaluation settings leave a gap with how claw-style agents are deployed in practice. To address this, we introduce AutoRedClaw, an automated red-teaming framework that generates end-to-end attacks together with the environments needed to execute them. AutoRedClaw combines a strict quality-checking protocol with execution-based validation to ensure the correctness, realism, and effectiveness of generated samples, and adopts task-driven environment construction to keep costs low. To achieve broad coverage, it organizes generation along risk categories, attack strategies, attack stages, and attack surfaces, and uses experience-guided exploration to steer generation toward under-explored regions. AutoRedClaw achieves an effective attack rate above 30% at less than $0.10 per successful sample. From its outputs, we construct RedClaw-Bench, a 330-case benchmark that yields substantially higher attack success rates (ASRs) than AgentDojo, AgentDyn, and DTap. Across seven harnesses and four frontier models, most agent systems exceed 50% ASR, showing that our red-teaming pipeline provides a strong evaluation signal even for the most advanced agent systems. Our code and benchmark are released at https://github.com/SaFo-Lab/AutoRedClaw.

Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for person_father to the ceiling for country_capital_city. We investigate the factors that contribute to this variation and demonstrate that it can be significantly predicted from a readout that is not aware of the underlying relation. To do so, we propose a novel form of the Jacobian lens, the entity lens. Whereas the Jacobian lens is estimated across all tokens and concepts, the entity lens EL_t is trained to predict what the model will output for an entity of type t under the neutral prompt "{} is known for"; e.g., we train ELs for countries and persons. After estimating an entity lens, we measure how highly it ranks the model's answer for a relation without using any relation-specific information in estimation and prediction. Across 28 relations of the LRE dataset on Olmo-3-7B this relation-agnostic score correlates with LRE faithfulness at ρ= 0.7355 (p < .01). This suggests that linear decodability is largely a property of what a subject's representation already carries, rather than of the relation being decoded. Moving the unit of analysis from the relation to the individual fact, we conclude that new relation datasets and analysis methods are needed to address several problems in existing datasets, most prominently semantically weak first tokens (e.g., most predicted universities' first token is "University") and reference competition (e.g., the LLM confuses obscure actor names with competing non-actor entities that have the same last name). Our full pipeline is available at https://github.com/jonasrohw/entity-lens

We present a study on controlled automatic simplification of Russian texts using preference-based optimization algorithms with a focus on adapting texts to the A1, A2 and B1 proficiency levels. We fine-tune YandexGPT using Rejection Sampling and Direct Preference Optimization (DPO) methods to improve control over the complexity of the generated text. We compare several training configurations that differ in dataset composition, text type and the strategy used to construct preference pairs for DPO. For automatic evaluation we train a regression model based on ruRoBERTa-large to predict continuous complexity scores corresponding to CEFR levels. We also consider semantic similarity, lexical coverage and quantitative readability metrics. Human evaluation assesses simplicity, fluency and adequacy. Our best model achieves 87.15\% target level accuracy, an MAE of 0.35 CEFR levels and a target interval MAE of 0.11. Manual assessment shows that 75% of the evaluated simplifications match the target morphological and syntactic level, and fluency and adequacy scores are especially high for A2 and B1. Although A1 remains the most challenging level, the model significantly reduces the proportion of vocabulary and grammatical constructions beyond the target level while generally preserving meaning. Our model is freely available on $\href{https://huggingface.co/kmustakova/RuSimplifier_DPO}{HuggingFace}$.

Enterprise adoption of LLMs in agentic workflows is gated by evaluations that span a wide arc of abilities: CRM operations, heterogeneous-document retrieval, personal data understanding, multi-turn planning, tool use, function calling, long-horizon research, and long-context software engineering. These evaluations live today in separate repositories with divergent runners, dependencies, and scoring conventions, which makes cross-benchmark comparison difficult and obscures what a single aggregate number actually represents. We release EGI-BENCH, a unified LLM agent benchmark suite that consolidates eight enterprise facing agent benchmarks (CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP-UNIVERSE, MCPEVAL, BENCHFORCE, LOCOBENCH) behind a single command-line interface with two runtime-budgeted execution modes. On top of the suite we define the Enterprise General Intelligence (EGI) Score, a six-dimensional framework (M1-M6) whose buckets are populated by fine-grained sub-metrics routed from each benchmark rather than by benchmark-level averages. We describe the suite, the sub-metric routing, and the quick/deep budgets, and report results for fourteen frontier and cost-efficient LLMs spanning the GPT, Claude, and Gemini families under quick mode. Top-line EGI ranges from 39.6 (GPT-4o mini) to 65.6 (gpt-5.5), and no single model dominates every capability axis: gpt-5.5 leads task outcome, planning, and retrieval grounding; Claude Sonnet 4.5 leads tool actioning and safety/abstention; and Claude Haiku 4.5 leads engineering quality. Code is available at https://github.com/SalesforceAIResearch/egi-bench.

Medical workflows often require multiple interdependent analyses, but existing specialist and generalist AI models are not designed to coordinate them. We introduce MMedAgent-Plus, a multimodal medical tool-use agent that combines a vision-language planner with 18 leading specialist models spanning 12 tasks and 8 modalities. Through a unified tool interface, the planner selects specialists, executes them in sequence, and synthesizes responses grounded in their outputs. The planner is instruction-tuned on 208,928 samples from 32 public medical benchmarks, covering single- and multi-tool interactions across single and multiple rounds. We separately evaluate tool orchestration, response synthesis, and end-to-end execution on held-out samples. MMedAgent-Plus achieves tool-call accuracies of 91.0%, 98.1%, and 80.7% in single-tool multi-round, multi-tool multi-round, and multi-tool single-round settings, respectively. Given correct tool outputs, MMedAgent-Plus achieves scores of 14.4-14.8 out of 15 for faithfulness, completeness, and conciseness across 600 evaluation samples. In end-to-end execution without pre-supplied tool outputs, MMedAgent-Plus correctly invokes the required specialist in 90.6% of 350 evaluation cases spanning seven tasks. Its final-answer accuracy reaches 0.98, matching the dedicated leading specialist and exceeding the strongest baseline (0.28).

Generative recommendation predicts items by autoregressively generating semantic identifiers (SIDs), whose tokens are shared across semantically related items. With this paradigm gaining adoption, the question of how to support unlearning in it has become pressing. If training interactions must later be removed, such as injected spam or interactions with items a user no longer wants, this sharing lets an unlearning update unintentionally affect the predictions for the interactions' neighbors. We show that this effect is locally structured and hidden by global metrics: unlearning can leave recommendation utility on the whole test set nearly unchanged while substantially degrading it on neighborhoods of the forgotten interactions' SIDs. We propose Neighborhood-Aware Unlearning (NAU), which augments interaction-level unlearning with explicit control over the probability mass assigned to the forgotten items' neighborhoods. Across three datasets and four SID tokenizers, NAU restores spam-induced exposure toward the retrained reference while retaining more than 95% of retrained utility, and recovers the neighborhood utility that forget-and-repair unlearning loses at no global cost. We identify local probability redistribution as an important failure mode of unlearning in SID-based generative recommenders and show that explicitly controlling this redistribution can mitigate its effects.

Chemistry has recently entered a high-data regime as molecular databases from high-throughput platforms or computational studies have become available. Increasing automation within the chemical sciences leads to a more digital and precision chemistry. Dedicated scientific agents are becoming central to guiding the chemist through an ever-increasing amount of data and findings, including negative ones. Here, we introduce Magnolia, a persistent-memory artificial intelligence agent framework that integrates computational protocols, from docking to quantum chemistry, lab notebooks and chat logs relying on large language models (LLMs). It relies on a three-tier memory system: session-level observations are automatically captured and distilled into structured learnings, confirmed and elevated to project-tier entries and, when human-validated, promoted to cross-project rules that actively guide subsequent work and provide direct chemical feedback. We illustrate Magnolia on four case studies: quantum-chemical evaluation of aromatic systems (COMPAS database), molecular dynamics of butane and of modified DNA, and a protein-peptide docking campaign. Across these projects, Magnolia tracked 317 runs and accumulated 116 project-tier entries, from which four cross-project rules were promoted; in controlled replays of four tasks, runs with memory passed the checks that runs without it failed (for example, ten of ten docking-configuration checks, against one to four of ten). By combining automation, reproducibility, and human-in-the-loop control, our Magnolia framework streamlines docking, molecular dynamics and quantum chemistry calculations with direct analysis and LLM-augmented feedback, offering accelerated access to computational (bio)chemistry.

Cooperative language-model agents can improve without weight updates by revising reusable skills, but only if a shared team reward is traced to the responsible agent and skill. Existing methods credit the skills present in each trajectory, yet presence is not responsibility: a harmful skill active in successful runs still earns positive credit. We present \textbf{GRSE}, \textbf{G}roup-\textbf{R}elative \textbf{S}kill \textbf{E}volution, which assigns behavioral credit by comparison: inspired by GRPO, it executes each question several times under the same skill libraries and contrasts the executions to credit agent--skill pairs, with structured diagnosis of failures. On HotpotQA and LoCoMo with two open-weight actors, GRSE outperforms MASkills, the closest prior method, on HotpotQA and every LoCoMo category except open-domain, with the largest gains on multi-hop questions ($+6.7$ and $+3.2$ F1). Its gains rest on better-grounded retrieval rather than answers given without evidence, and its learned skills transfer between the multi-hop tasks. GRSE is robust to corrupted libraries: starting from planted harmful skills, it credits every planted skill with the correct sign and recovers the score of a clean library ($+6.3$ points), whereas MASkills misattributes two and falls below its starting point.

Recent work has explored continuous representations in place of discrete tokens during Chain-of-Thought (CoT) reasoning, often referred to as soft CoT. This is motivated by the idea that a mixture of token embeddings can represent a superposition of multiple reasoning possibilities within a single step. However, it remains challenging to train large language models to realize this potential. Existing work has explored reinforcement learning (RL)-style training for soft CoT. However, these approaches treat soft tokens as actions in a Markov decision process rather than as differentiable intermediate computation steps. Thus, they do not propagate gradients across the soft tokens and miss out on the signal of how the choice of one soft token affects the next. In this paper, we propose a sparse relaxation method for RL (called SpaRLax), which leverages the differentiability of soft-CoT reasoning over time steps and makes training more stable and effective. We adopt an entmax sparsity method that adaptively determines the degree of softness, which truncates long-tail noise at confident steps while retaining multiple plausible candidates when uncertain. Across nine reasoning benchmarks, our method achieves the highest average performance compared with baseline methods. Further analyses validate the effective use of soft-token superposition in our approach, and also demonstrate the contribution of adaptive sparsity in our approach.

Estimating extreme risk in reinforcement learning becomes increasingly difficult as the confidence level increases, since only a small number of observations contribute to empirical estimates of Conditional Value-at-Risk (CVaR). In this paper, we extend an Extreme Value Theory (EVT) approach for CVaR optimization to high-dimensional neural-network policies. Using a peaks-over-threshold representation and method-of-moments estimation, we obtain a differentiable tail-risk objective and show that, conditional on the selected threshold, its finite-sample gradient can be computed directly through automatic differentiation. We first validate the approach in a controlled experiment with a known optimal solution and then investigate its performance in a deep-hedging problem with rare extreme market movements. Our results show similar hedging behavior for EVT- and empirical-CVaR-trained policies under normal conditions, while EVT-trained policies exhibit lower hedging losses under the extreme-scenario evaluation across the training sample sizes considered. These results suggest that EVT can provide a practical alternative for training neural policies when the observations relevant to extreme risk are scarce.

Personalized large language models (LLMs) increasingly rely on profile-conditioned inference, pairing public requests with rich, user-specific facts. This paradigm exposes the user profile itself, together with the user's explicit sharing preferences. We introduce Profile-DP, a policy-aware local differential privacy framework designed specifically for user-profile release in personalized LLM inference. Profile-DP provably bounds the release of user-withheld facts while treating public instructions, profile schemas, and authorized facts as public side information. By generating a privatized profile that composes with the public instruction prior to answer generation, our framework decouples formal privacy guarantees from downstream model behavior. We develop two finite-domain mechanisms satisfying Profile-DP: DP-SLOT independently perturbs protected attributes, while DP-BLOCK uses public semantic decisions to coordinate their release and jointly sample related attributes. A positive, public compatibility prior connects these attributes while preserving a profile-level privacy bound. On PEEP, DP-BLOCK leads the utility composite among DP methods at all 13 structured budgets and is closest to the ideal privacy--utility point in 18 of 26 format--budget settings under our composite weighting. Direct measurements reveal complementary strengths: DP-BLOCK discloses less, while DP-SLOT retains more authorized facts. These results make policy-aware profile release a practical basis for studying privacy and personalization.

Continual vision-language models (VLMs) encounter a stream of environments and must also reject inputs that do not belong to the environment at hand. Class information from earlier environments can aid out-of-distribution (OOD) detection, yet existing detectors score each environment in isolation, and when previous environments help or hurt is theoretically unclear. We develop a theory of environment-aware OOD detection built on contribution scores, pairwise cosine expectations measuring how the image and text features of one environment align with those of another. We propose a cross-environment maximum-concept-matching (MCM) score with two terms: a primary term matching the test image against the current class texts, and a secondary term matching it against the class texts of a previous environment, weighted by the intra-modal alignment between the two environments. We prove that, under a positive environment transition and a bounded class-matching margin for OOD inputs, this score attains a false-positive rate no larger than that of the standard MCM detector at an adjusted threshold. We then show that the score gap of an MCM detector between two environments equals the self-contribution of the in-distribution environment minus its cross-modal contribution into the OOD environment, which separates soft OOD environments, with explicit AUROC, true-positive, and false-positive guarantees, from hard ones, which no threshold that retains the in-distribution data can reliably reject. For cascaded detection, in which the text prototypes of all seen environments are pooled, the false-positive rate on every past environment is non-decreasing in time, and we characterize when detection provably fails. A contribution balance factor applied to the gradients of the classification loss and of a smoothed false-positive rate completes the framework.

Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.

Deep neural networks have achieved remarkable success on vision benchmarks, yet they continue to struggle with many generalization challenges. Supervised vision training relies on one-hot labels, which provide limited information about semantic structure and shared attributes between classes. This limited supervision can leave visual representations vulnerable to distribution shifts, spurious correlations, texture bias, adversarial perturbations, and forgetting in sequential learning settings. We study whether pretrained language models can serve as lightweight auxiliary supervision for vision training without requiring paired image-text data, prompt engineering, or contrastive objectives. Specifically, we evaluate two forms of language guidance, Explicit Language Guidance (ExLG) and Implicit Language Guidance (ImLG). We conduct a comprehensive evaluation across six generalization regimes, including in-distribution, out-of-distribution generalization, shortcut and spurious correlation resistance, texture and shape bias, adversarial robustness, and continual learning. Our analyses show that the two mechanisms have complementary strengths, with explicit guidance consistently benefiting in-distribution, low-data performance, and continual learning retention, while implicit guidance is often more useful in shortcut-sensitive settings and under stronger adversarial perturbations. Importantly, both are lightweight and add minimal parameters and training overhead. The analyses characterize when language-derived structure helps supervised vision training and provides a practical roadmap for using off-the-shelf pretrained models from another modality as auxiliary supervision.

We study the sample complexity of estimating Ornstein--Uhlenbeck diffusion scores in $\mathbb R^d$ using denoising score matching (DSM) with finite-width shallow ReLU networks, and its implications for reverse-process sampling. Under weak differentiability and polynomial-growth conditions on normalized derivatives of the initial density, we combine complementary estimates from the Gaussian transition kernel and the initial density to obtain explicit Sobolev bounds that are uniform in time on each fixed ball. Combining these bounds with existing shallow-network approximation theory gives explicit control of the approximation error and a sufficient path-norm budget. We then use truncation and localized Rademacher complexity to establish high-probability risk bounds for path-norm-constrained empirical DSM minimizers with unbounded inputs and labels. Under additional bounded-density and sub-Gaussian assumptions, suitable choices of network width and path-norm budget yield squared $L^2(P_t)$ score risk $\widetilde O(n^{-(d+3)/(2d+3)})$ at each fixed time $t>0$, where $n$ is the number of independent training samples and $P_t$ is the forward marginal. The nonasymptotic bounds retain explicit dependence on diffusion time and the regularity and tail parameters of the initial distribution. Aggregating the score errors over a time grid and applying existing sampling theory together with an early-stopping bound yields a high-probability total-variation error $\widetilde O(n^{-(d+3)/(4d+6)})$ between the generated and original data distributions under suitable sampling parameters. Here $\widetilde O$ suppresses logarithmic factors for fixed dimension, confidence level, and initial distribution.

Web agents are typically evaluated on English-language interfaces, leaving their ability to operate across the world's multilingual websites largely unexplored. This gap is especially significant for low-resource languages, where standardized resources for evaluating web agents remain limited. To address this gap, we introduce Low-Resource Web Agent Benchmark (LoWAgent), a benchmark for evaluating web agents across different low-resource languages (LRLs) in comparison to the default, English. LoWAgent evaluates web agents across English and four LRLs: Hindi, Bengali, Urdu, and Amharic. The benchmark comprises of six fully localized websites, instantiated across five languages to yield 3,000 task instances. We evaluate three closed-source and five open-source models across all five languages on all six websites. Our findings show that both closed-source and open-source models experience performance degradation for LRLs compared to English, with substantially greater performance drops observed among the open-source model agents. We further evaluate the robustness of the models on realistic interface disruptions, such as cookie notices and advertisement pop-ups, and find that these further degrade agentic performance. Finally, we show that post-training on a small subset of LoWAgent traces can significantly improve low-resource language agentic task performance, demonstrating the usefulness of our benchmark.

Class-Incremental Learning (CIL) enables models to recognize new classes over time while retaining prior knowledge, but many real-world problems are compositional, where visual concepts arise from combinations of primitives such as states (i.e., red, shiny) and objects (i.e., shoe, car). Standard CIL methods treat each composition as an independent class, ignoring shared primitives and providing no mechanism to stabilize their representations as new compositions arrive. This causes parameter updates to entangle primitives and induce drift that degrades performance on earlier compositions. Compositional Incremental Learning (composition-IL) formalizes this setting by introducing new state-object compositions across tasks. In this paper, we propose Layer-Specific Adaptation with Prototype-Anchored Adapter-Space Transport (LSA-PAST), a parameter-efficient method tailored to composition-IL. LSA integrates Low-Rank Adaptation (LoRA) into a frozen ViT with depth-specialized adapters for states, objects, and compositions. PAST then maps past, anchored prototypes into the current adapter space through a rehearsal-free affine transform fit on current-task data, enabling old prototypes to remain compatible with new adapters. Extensive experiments demonstrate that LSA-PAST achieves state-of-the-art (SOA) accuracy-forgetting trade-off performance for composition-IL.

Large Language Models (LLMs) increasingly operate across culturally diverse settings, yet existing evaluations largely measure cultural behavior at the output level. We introduce CultureLens, a reference-grounded benchmark and mechanistic study of how LLMs use cultural information. CultureLens contains 9,261 scenarios across 75 countries and systematically varies country, value, true-norm, and contradictory-norm evidence. Across Llama-3.1-8B, Gemma-3-12B, and Qwen-3.5-27B, stating a contradictory norm lowers aligned answers by 32-47 percentage points, even when the model answers correctly from the country alone. Mechanistically, models default to Anglophone norms when no country is given. The failure under a contradictory norm is not one of missing knowledge, as blocking the norm inside the model restores the correct answer in most cases. Models accept a stated norm according to how it is presented and not whether it is true, which leaves them most exposed to errors in retrieved content. Supervised fine-tuning with contradictory-norm examples raises aligned answers under a contradictory norm to 85-93%, including on unseen countries, at a small cost under a true norm.

Speech large language models (LLMs) perform multiple tasks, including automatic speech recognition (ASR), timestamp prediction (TS), and speech translation (ST). Joint training of speech LLMs creates task trade-offs, yet standard training with fixed loss weights does not generally control the resulting trade-off according to deployment preferences. To address this, we propose Endpoint-Anchored Preference-Aware Sampling Uniformly along the Pareto Front (EA-SURF), an a priori method that adapts scalarization weights so that the resulting task trade-off reflects the desired preference. EA-SURF retains the standard weighted-loss training pipeline, tracking the current model's performance relative to each task-specific reference and adapting the scalarization weights as training proceeds. We theoretically show that EA-SURF can match a desired preference, with the update becoming stationary at the target under checkable regularity conditions. Across six three-objective speech LLM requests, EA-SURF empirically converges toward the requested targets and increases endpoint-inclusive normalized hypervolume by 17.8% over fixed loss weighting.

Physics-informed neural networks (PINNs) and other residual-trained solvers judge a candidate solution by the size of its sampled PDE residual. For first-order Hamilton-Jacobi equations this is not sufficient: the same equation and boundary data admit Lipschitz almost-everywhere solutions that differ from the viscosity solution by $O(1)$ while violating the equation only on sets of measure zero. We study whether residual minimization selects such solutions in a fixed-horizon pursuit-evasion game whose Hamilton-Jacobi-Isaacs value function is known in closed form on the circle $S^1$ and on the flat tori $\mathbb{T}^{2}$ and $\mathbb{T}^{4}$. We train four neural parameterizations with ten seeds on each geometry, 120 runs in total. A standard network with periodic Fourier features converges to folded almost-everywhere solutions in all 30 runs: small residuals and the expected capture-layer gradient, yet values that are wrong over most of the domain. A network with periodicity imposed through the loss recovers the viscosity solution in one and two dimensions but becomes unreliable in four. Two parameterizations that receive the geodesic distance do substantially better: a single network fed the wrapped displacement and its distance recovers the viscosity solution in all 30 runs, while a minimum over periodic image charts succeeds in 27 of 30 and introduces a wrong-image failure mode. Finally, for the convex Hamiltonian considered here, we use the semiconcavity of the viscosity solution to build a mesh-free finite-difference test that needs only forward network evaluations; it detects the folded solutions and the harmful min-of-charts failures without access to the exact value function. A small sampled residual thus does not by itself validate a learned Hamilton-Jacobi solution; the parameterization and independent structural diagnostics can be essential for selecting the viscosity solution.

Multimodal large language models (LLMs) are increasingly used for structured extraction from financial documents, but their behavior under visual ambiguity is poorly understood. We study CARLAR (Courtesy Amount Recognition and Legal Amount Recognition) influence: a model returns the same amount for both cheque fields even though their annotations differ, including cases with a blank annotation encoded as zero. Across ten visual prompting conditions, we evaluate GPT-4.1 and Claude 4.5 Sonnet on 243 primary cheques (124 with two observed unequal amounts) and 598 strict mismatches from two additional datasets. Under courtesy redaction, GPT-4.1 CARLAR reaches 73.7 percent (95 percent bootstrap interval: 68.3 to 79.0; p below 0.001 versus baseline). Model-reported confidence can remain high when predictions agree; expected calibration error reaches 0.93 among CARLAR cases. We also test a compressed-image few-shot intervention that presents examples as one visual table without model fine-tuning. On the 243-case cohort, Claude's Control condition increases paired joint extraction accuracy by 11.5 percentage points (95 percent interval: 6.2 to 16.9; adjusted p = 0.00054), but its strict-124 gain is smaller and not significant after correction. The intervention's effect on predicted-field agreement is model-dependent: it decreases for Claude and increases for GPT-4.1. These results motivate validation methods that account for visual context and dependencies between extracted fields.

Modern language models are increasingly asked to handle several tasks within a single prompt. Doing so requires solving each task correctly while integrating the resulting answers into a coherent response. In this paper, we identify a fundamental limitation in this process, where models make substantially more errors when tasks are composed, even when they can solve every constituent task reliably in isolation. This degradation cannot be fully explained by the accumulation of independent errors. Across both open-weight models and the frontier model Claude Opus 5, our findings instead point to an implicit thinking budget underlying this effect. In particular, under composition, models use substantially fewer tokens per task than in isolation, even when the full sequence remains well within the context window. Tasks that require longer solution traces in isolation are consequently more $\textit{fragile}$ under composition, i.e., models make more errors on prompts containing these tasks. In isolation, non-fragile tasks receive more reasoning than they need and waste compute. However, by grouping tasks according to their fragility, the implicit budgeting can be manipulated to reduce token usage by $43$\% across a range of math and linguistic tasks while also preserving accuracy.

Learning latent dynamics is central to assessing the current states and forecasting the future trajectories of high-dimensional time series. For locally-stationary latent dynamics $\mathcal{F}(\mathbf{z}_{<t}; \mathbf{c})$ with latent dynamics state $\mathbf{z}_t$ and environment variable $\mathbf{c}$, prior identifiability results have largely focused on $\mathbf{z}_t$ when conditioned on predefined label $u$ of dynamics environments. This leaves two limitations: reliance on predefined labels that hinder generalization to unseen environments, and a limited understanding of the identifiability of $\mathcal{F}$ and $\mathbf{c}$ which -- although offering important structural properties for the identifiability of $\mathbf{z}_t$ -- are learned jointly with $\mathbf{z}_t$. We address these limitations with Meta-iLaD, a novel identifiable latent dynamics framework attained by meta-learning across dynamics environments. Meta-iLaD introduces a novel conditional prior of $\mathbf{c}$, modeled as a feedforward meta-learner to rapidly extract $\mathbf{c}$ from few-shot examples. Meta-iLaD further establishes identifiability for $\mathbf{z}_t$, $\mathbf{c}$, and $\mathcal{F}$ simultaneously, for a general formulation of $\mathcal{F}(\mathbf{z}_{< t};\mathbf{c})$ without restricting the dimension of $\mathbf{c}$ or how it modulates $\mathcal{F}$. On synthetic and real data, we provide strong empirical evidence that 1) conditioning on few-shot examples enables generalization to out-of-distribution environments, and 2) identifiability for $\mathbf{c}$ and $\mathcal{F}$ is critical for accurate forecasting beyond reconstructing observed trajectories. Code is available at https://github.com/GoForIt-YYB/Meta-iLaD.

Platforms that host young users screen messages with commercial text-safety APIs, but independent evidence on how often such systems miss harm or flag harmless text is scarce, especially outside English, and the documentation of the one studied here leaves its false-positive rate unquantified. We evaluated one such API, Tuteliq Detect as deployed under the release identifier tuteliq-detect-v5.20261011 (not the separately published v6 model), on six harm categories: authorised push-payment fraud, money-mule recruitment, social engineering, romance scams, gambling harm and bullying. We built SITO (Synthetic Independent Text-safety benchmark with Overflagging tests; sito is Ukrainian for 'sieve'), 2,332 synthetic cases in English, Ukrainian and mixed Ukrainian, with each harmful case matched by hard harmless ones: same-topic messages, minimal-edit contrasts, warnings and retrospective accounts. On English text the system flagged 84.9-94.1% of harmful cases and 0-4.3% of harmless ones per category. It missed third-party reports of ongoing harm far more often than perpetrators' messages (63.6% against 95.1% recall, all languages), when sent without the settings the API offers for reports of harm. Ukrainian and mixed-language input was detected at least as often as English, recall was lower on text from a held-out generator family, so headline recall may be optimistic, and 3.2% of cases received a different decision in repeated runs. A platform should measure false positives on its own traffic and check how bystander reports are handled. The datasets, generation prompts and code are released.

In robot-assisted education, a robot must understand how to complete a task in order to effectively guide a student through it. We present a novel method that enables a semi-autonomous, camera-equipped robot to recognise the task goal and plan actions to achieve it. Our work focuses on Shoebox Tasks, an activity commonly used in the education of individuals with Autism Spectrum Disorder (ASD). These tasks are designed so that the goal can be inferred directly from their visual structure. We have developed a solver that uses a vision-based Large Language Model to convert photos of a Shoebox Task into a Planning Domain Definition Language (PDDL) representation and generate action plans for solving it. We have also implemented a robotic interface that enables a Nao robot to autonomously execute various Shoebox Tasks and assist a student by suggesting next steps or signalling detected failures through action demonstrations, voice commands, or gestures. With an adapted robotic interface, the solver can be integrated into a variety of educational robots. No programming expertise is required from teachers, and the solver operates exclusively on captured images as input. Research in special educational needs indicates that students with ASD can benefit substantially from robot-assisted teaching, and our system aims to strengthen these learning opportunities while also fostering interest in robotics among individuals with ASD.

Public charging of electric vehicles (EVs) adds a large and synchronized load to the distribution feeders. In most networks, it is rare to overload a feeder at the same time. When it does happen, the consequences can be severe: an overloaded feeder stresses its transformers and can trigger cascading outages across a city. Planners who design for high reliability therefore need to estimate very small outage probabilities. To that end, plain Monte Carlo estimation with an expensive grid simulator can require a prohibitively large number of runs for this task. To reduce the number of runs, we propose a control variate (CV) built from a learned classifier of the outage, trained on cheap observable features such as temperature and load forecasts. We show that for any binary learner, the variance reduction is an explicit function of its recall and false-positive rate. Furthermore, we show that control variate based on a Gaussian proxy of the load loses its effect as outages become rarer, while a learner with non-vanishing precision keeps removing a fixed fraction of the variance even in the extreme rare-event regime. We verify these results in numerical experiments using a feeder-overload simulator across three orders of magnitude of outage probability and obtain that even weak learner with a single feature can remove a quarter of the variance, while a boosted-tree learner is able to remove about 60%. We then apply the estimator to the feeder network of Surabaya, Indonesia, in which the Gaussian load model leaves every feeder three to seven times the one-day-in-ten-years blackout target. At the same simulation budget, planning the reserve with the learned CV and a 95% confidence limit reduces the city-wide blackout days by a third compared with Monte Carlo. These results show that learned control variates can help planners design distribution grids that stay reliable against rare outages.

Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be internalized through training. More precisely, we conceptualize a harness $\mathcal{A}$ as an arbitrary wrapper around a base model $M$ and attempt to distill a student model from the harness's traces. Empirically we find that, even with simple scaffolds and small models, the inference-time harness significantly outperforms a distilled student. We then prove that such a gap is due to a fundamental representational limitation: inference-time harnesses can compute a closure of their per-call primitive that is $\NL$-complete even when the primitive is a trivial $\AC^0$ lookup, and that no fixed-depth forward pass can represent under the widely believed conjecture that $\NL\not\subseteq\TC^0$. Our theory suggests that expanding the representational capacity of the student relative to the task at hand, e.g., by enlarging the model, making the task easier, or lengthening its chain-of-thought, is necessary to close the gap, which we verify empirically. Finally, we scale to an agentic harness on competition math, where the same gap reappears. Overall, our results show that a harness confers on the base model a profound gain in representational power, one that training a same-class student on its traces cannot transfer back into the weights.

Object: Motion artefacts remain a major limitation of routine two-dimensional T2-weighted turbo spin-echo (TSE) brain MRI, reducing diagnostic confidence and often necessitating repeated acquisitions. MoSAIK (Motion Search with AI in K-space) retrospectively estimates in-plane rigid-body motion directly from standard multi-coil TSE k-space. Without a motion-free reference, external navigators, or additional acquisitions, MoSAIK infers motion from k-space self-consistency violations and incorporates predicted parameters into model-based, motion-aware reconstruction rather than synthesizing corrected images. Materials and Methods: Selected complex coil data were represented as log-magnitude and phase channels. A decoupled convolutional embedding and Transformer encoder estimated phase-encoding-line-indexed rotation and two-dimensional translations. Training used motion-free clinical acquisitions with simulated motion; four held-out subjects were evaluated. Component-wise errors assessed estimation; SSIM and PSNR in silico and entropy focus criterion (EFC) in vivo assessed reconstruction. Results: Motion estimation errors were low for both rotation (MSE $<$ 0.2 deg$^2$) and translation (MSE $<$ 0.2 pixels$^2$), with Pearson correlations of approximately 0.96 and 0.94, respectively. In every in-vivo acquisition, more than 75% of slices showed information gain relative to uncorrected reconstructions. MoSAIK reduced visible artefacts while preserving anatomical detail. Discussion: Transformer-based k-space motion-trajectory estimation enables fast retrospective correction of standard clinical TSE brain MRI.

Accurate multi-horizon photovoltaic (PV) power forecasting is essential for grid stability, energy scheduling, and the efficient operation of PVbatterygrid systems. However, existing models often fail to jointly capture short-term variability and long-term temporal patterns, while lacking unified multi-horizon capability and interpretability. To address these limitations, this study proposes a novel hybrid deep learning framework that integrates a Temporal Fusion Transformer (TFT) and N-BEATS within an attention-based fusion architecture. The proposed model leverages TFT to learn short-term temporal dependencies and dynamic feature importance, while N-BEATS captures long-term trend and seasonal components through interpretable basis decomposition. An adaptive attention mechanism dynamically fuses both branches, enabling balanced multi-scale learning. The framework simultaneously generates forecasts for 30-minute, 2-hour, 6-hour, 12-hour, and 24-hour horizons within a single unified pipeline. The model is evaluated using multi-year real-world datasets from the Desert Knowledge Australia Solar Centre (DKASC). Experimental results demonstrate superior performance over statistical, machine learning, and deep learning baselines. The proposed model achieves RMSE values of 42.35, 46.27, 52.29, and 48.11 across the 20162017, 2023, 2024, and 20252026 datasets, respectively, corresponding to a 3540% reduction compared to the best baseline (TFT). Additionally, MAPE remains below 5% across all datasets, indicating high predictive accuracy and stability. The model also provides interpretability through attention visualization and trendseasonality decomposition, and is deployed on a cloud-based platform for real-time forecasting. These results demonstrate that the proposed framework is robust, interpretable, and suitable for practical renewable energy applications.

Neural networks (NNs) are increasingly deployed on edge devices such as mobile and IoT platforms, offering low latency and improved privacy compared to cloud-based deployment. However, on-device models embedded in mobile applications can be unpacked and exposed to adversaries, enabling cross-node attacks in centrally distributed ecosystems, where adversarial examples crafted on one compromised device transfer effectively to other nodes sharing the same models. To address this threat, we propose VeilNN, a defense framework that reduces adversarial transferability across on-device models via model obfuscation. The core idea of VeilNN is to introduce controlled model diversity while preserving original task functionality. Specifically, VeilNN applies stealthy and irreversible obfuscation at both the architectural and parameter levels through network cell obfuscation and layer assembling, and is applicable to diverse NN architectures, including MLPs, CNNs, and Vision Transformers. Extensive experiments demonstrate that VeilNN effectively defends against cross-node attacks, achieving an average defense success of 49.19% and up to 81.19% across four adversarial attack types, with more than 33.16% improvement over existing defenses.

As two most prevalent neurodegenerative diseases (NDDs), Alzheimer's disease (AD) and Parkinson's disease (PD) shared commonalities both in symptomology and pathophysiology. However, to what extent and in what way these two distinct disorders overlap remain underexplored. In this study, we propose an explicit, interpretable and spatially adaptive transfer learning framework to elucidate the morphometric biomarkers linked to AD and PD. We applied such high-dimensional machine learning model to two most popular open-source neuroimaging datasets, i.e., Alzheimer's Disease Neuroimaging Initiative (ADNI) and Parkinson's Progression Markers Initiative (PPMI). On the one hand, our cross-disorder classifier revealed neuro-anatomical alteration patterns well aligning with previous neuroimaging research findings and underscoring patients' clinical presentations. On the other hand, the spatially adaptive trained classifier using shared prior achieved statistically superior diagnostic performance compared to respective individual models, which further highlights the cross-disorder framework's efficacy on modeling limited-size datasets across disease domains.

Health-related discussions on social media are increasingly shaped by influencers who offer diverse, and often competing perspectives on the same issue. Yet, such content is typically evaluated in isolation, without considering the broader discourse in which it appears. We introduce GRID (Grounded Reasoning over Interconnected Discourse), a framework for evaluating influencer content through collective inference over the surrounding discourse. GRID represents each video as an argumentation graph connecting claims to its central message, or takeaway, links related claims and takeaways across videos, grounds claims in scientific evidence, and jointly infers credibility over the resulting network. We evaluate GRID on an end-to-end health-related content validation task using 569 influencer videos grounded in expert assessments. Our results show that accounting for the broader discourse improves influencer content validation. Finally, we use GRID-inferred credibility to demonstrate how argumentative strategies and moral framing differ between low- and high-credibility content.

Data selection plays a central role in modern large language model training, as the quality of training data strongly influences downstream performance. A common approach is to curate a development set that encodes the desired skills (e.g., reasoning or multilinguality) and select data that are similar to them. Existing methods compute similarity using cosine distance over either the model's hidden states or gradients. However, we observe that this max-similarity strategy induces a mode-seeking bias: the selected data clusters tightly around the development samples. To address this, we propose ICLHerd, a data selection framework that leverages the von Mises-Fisher kernel to balance mode-seeking with mode-averaging. We further introduce in-context learning as a validation proxy, enabling efficient hyperparameter search over the balance trade-off. Results show that ICLHerd outperforms existing LLM data selection algorithms, highlighting the importance of balancing mode-seeking and mode-averaging.

Reliable contact establishment remains a critical challenge for autonomous on-orbit servicing (OOS), particularly when interacting with non-cooperative targets. Gecko-inspired dry adhesives provide a lightweight, passive, and residue-free solution for robotic capture in space, yet current systems operate in open loop, without the ability to estimate their own contact or adhesion state during interaction. This paper presents a real-time interaction-awareness framework that enables sensorized gecko-inspired adhesive interfaces to estimate their contact state throughout the adhesion lifecycle using embedded capacitive sensing. Five discrete interaction states (idle, proximity, contact, engaged, and disengaged) are classified using experimentally collected and augmented sensor data, enabling a comparative evaluation of six lightweight machine learning models suitable for embedded deployment. Among the evaluated models, Random Forest achieved the highest offline classification accuracy of 95.65%. For online deployment, combining Random Forest with a rule-based state machine improved accuracy from 78.02% to 93.14%, achieving a macro F1-score of 0.9343 while maintaining a 53 ms inference cycle. These results demonstrate that incorporating physics-consistent state constraints significantly improves robustness during realtime inference, representing a step toward interaction-aware capture in future autonomous orbital servicing missions.

How learning progresses over time and varies across individuals is a longstanding question in cognitive and behavioral science. However, a major challenge is identifying observable behavioral measures that faithfully capture learning progression. Existing approaches commonly rely on predefined behavioral features or task-level performance, while fine-scale posture and movement in freely behaving animals remain less explored. In this paper, we ask: can short-term movement trajectories reveal learning progress and individual identity? We introduce a self-supervised representation learning framework built on a simple hypothesis: behavioral characteristics associated with identity and learning remain relatively stable over nearby time intervals. Using temporal proximity as supervision, our framework learns compact representations directly from movement trajectories without manually defining behavioral states or kinematic features. Using behavior data from freely behaving rats, we show that the learned representations not only improve prediction of identity and learning-related readouts over existing methods, but also provide a tool for investigating broader scientific questions about when and how these signals appear in behavior. Specifically, trajectories as short as one second contain informative learning and identity signals. Egocentric pose dynamics are more predictive of identity, whereas allocentric motion is more predictive of learning-related variation, and both signals are strongest at decision junctions in the alternation cohort. We further use a representation-conditioned diffusion model to inspect movement patterns associated with identity and learning progress. Our findings establish short-term movement as an informative behavioral "fingerprint" for studying learning and individual differences. Our code is available at https://github.com/pizhn/TRAIL.

We encounter a system through its responses, yet its state concerns how the system is. Predicting further responses promises knowledge beyond what is immediately observed. What that knowledge tells us about the system's state depends on what we mean by a state in the first place. We give this notion a model-relative meaning: a state is the complete response profile over a specified query universe, and configurations agreeing on every query are regarded as the same state. From this perspective, a learned state is understood through the distinctions between these profiles that it retains. Prediction concerns selected observations; recovery concerns the profile as a whole. We examine their relationship through query ambiguity, state-space geometry, and the choice of prediction targets. Synthetic systems with known response laws and neural recordings provide complementary settings for this investigation. The aim is to clarify how prediction can become a way of knowing a system, and what it would mean for a learned state to preserve the state being studied.

Ammonia and hydrogen are major constituents of planets within and beyond the solar system, yet how they interact in planetary envelopes is poorly known. We perform large-scale two-phase coexistence simulations of the ammonia-hydrogen system from 500 to 4000 K and 0.1 to 100 GPa, driven by machine-learning interatomic potentials trained on density functional theory calculations. From these we map the compositions of coexisting hydrogen-rich and ammonia-rich phases, and the critical curve that separates conditions of complete miscibility from those of phase separation. The critical temperature increases with pressure, from 500 K at 0.1 GPa to 2000 K near 30 GPa, and we find no phase separation at 3000 and 4000 K up to 100 GPa. The slope of the critical curve decreases substantially between 0.3 and 30 GPa, and the curve nearly flattens above about 30 GPa and 2000 K, where molecules begin to give way to atomic species. The coexistence curves are asymmetric with an ammonia-rich critical composition. For the expected envelope compositions, the phase boundary lies close to, and intersects, published thermal profiles of Uranus and Neptune, implying ammonia may be raining out in these planets today. In contrast, in warm sub-Neptunes such as TOI-270 d, ammonia and hydrogen are likely miscible in all proportions. Together with previous studies, our results suggest that silicates, water and ammonia separate from hydrogen in sequence as planets cool, with implications for their compositional, thermal and structural evolution.

With the advancement of powerful AI chips, on-device machine learning is gaining traction for its low latency and privacy protection. However, when edge devices are compromised, attackers can obtain white-box access to the deployed models by reversing the apps directly to obtain the knowledge of peer devices, which is known as Cross-Node Attacks, like performing efficient adversarial attacks. Most existing defenses target known vulnerabilities, yet white-box access still enables attackers to uncover more unknown weaknesses. This creates a critical risk for edge-device deployments, where one compromised device can endanger all models distributed from the same source. Motivated by findings that the model robustness is primarily influenced by the curvature of the decision boundary, we investigate how model parameters shape its curvature. In this paper, we propose a novel method, BMO, that obfuscates model parameters to diversify the decision boundary further resisting cross-node attack while preserving its benign accuracy. Specifically, we explore decomposing model parameters into boundary-position and boundary-curvature components based on null-space constraints, then models with differentiated decision boundaries are instantiated by injecting noise into the boundary-curvature component. Across all four major types of adversarial attacks with four widely-deployed on-device models, BMO achieves an average success rate of 68.68% in defense, far exceeding the most competitive baseline, which attains only 22.32% in defending against cross-node attacks. Moreover, BMO achieves comparable performance in benign accuracy compared with baselines.

Deep reinforcement learning (DRL) is challenging long-held assumptions in active flow control. By learning directly from interaction with a flow, DRL agents can find actuation strategies for nonlinear, unsteady flows without explicit models, often exploiting mechanisms that classical control would not anticipate. High-speed flows are a demanding test of this promise: shocks, shock/boundary-layer interactions, transition, aerothermal loading and microsecond-scale dynamics strain both algorithms and hardware. Yet most demonstrations remain idealized, low-speed simulations with limited evidence of robustness, interpretability, transferability or experimental deployment. In this perspective, we examine what constrains DRL-based high-speed flow control, including compressible simulation cost and fidelity, data efficiency, sensing in extreme environments, reward design, explainability, and generalization across Mach and Reynolds number. We argue that these limits reflect a tension between data-driven optimization and the structured physics, constraints and uncertainties of real flows. We then highlight where DRL may go beyond conventional approaches: adaptive control of shock-induced separation and turbulent boundary layers, coordination of distributed actuators, and discovery of unconventional flow-manipulation mechanisms.The central challenge, however, is translating these capabilities into reliable technology, which will require validated compressible simulation environments, physics- and constraint-aware learning, robustness to off-design conditions and sensor degradation, real-time operation, hardware-in-the-loop and wind-tunnel validation, and credible pathways to safety and certification.

Mixture-of-Experts (MoE) models reduce computation through sparse expert activation, but deployment still requires storing a large number of expert parameters. We study post-training expert pruning using task-domain calibration data as a subset-selection problem: given a pretrained MoE layer, select a fixed number of experts that best reconstruct its original output. Exact subset search is combinatorial and becomes intractable for modern MoE models with many experts per layer. We develop OptEP, a scalable continuous relaxation that directly optimizes differentiable expert-selection variables, followed by a lightweight reconstruction-based update of the retained expert weights. Across multiple MoE architectures and tasks spanning mathematical reasoning and code generation, our approach improves over the strongest existing expert-pruning baselines, with absolute accuracy gains of up to 15.05% on Qwen3-30B-A3B, 14.55% on Qwen3-235B-A22B, 11.06% on GLM-4.5-Air, and 16.20% on Mixtral-8x7B. The improvements are particularly pronounced under aggressive pruning, demonstrating that reconstruction-driven expert selection and weight adaptation can preserve substantially more of the original model's performance while reducing the number of retained experts.

Multimodal question answering (MMQA) is based on integrating heterogeneous data sources, selectively leveraging relevant modalities, and ignoring distractors. Existing approaches based on Multimodal Large Language Models (MLLMs) improve performance by splitting the task into intermediate steps, but do not quantify how each modality contributes to reaching the final answer. We argue that quantifying the effect that each modality has on the answering process improves process traceability and final performance. Thus, we propose Entropy-Driven Multimodal Answering (EDMA), which uses logical entropy to formalise multimodal QA as an iterative process of entropy reduction and to quantify the contribution of each modality in distinguishing correct from incorrect candidate answers. This enables the construction of quantifiable answering trajectories, where each step is associated with a measurable entropy reduction. EDMA introduces (i) modality-conditioned partitioning to estimate unimodal entropy reduction, (ii) entropy-driven multimodal fusion to capture complementary information across modalities, and (iii) entropy-based selection of the answering trajectory that minimises the most logical entropy. Evaluated across standard MMQA benchmarks, EDMA outperforms competitive prompting baselines, with average gains of 12.8 F1 and 4.6 LLM-J points over Planning Prompting and 18.6 F1 and 8.3 LLM-J points over Chain-of-Thought. Moreover, EDMA outperforms self-consistency baselines and is competitive with the multi-agent system baseline. Moreover, we show that higher entropy correlates with more false positives and, through controlled trajectory selection, demonstrate that logical entropy serves as an effective, monotonic selector for correct reasoning paths, validating entropy level as a reliable proxy for performance.

LLM agents perform tasks over multiple steps and revise their actions based on feedback. As the total reasoning budget is often limited and must support future steps, the agent must choose between short and long reasoning before each action. Longer reasoning may improve the final task outcome by preventing later errors, but it also uses budget capacity that may be needed for future decisions. This creates two challenges: estimating the downstream value of long reasoning and deciding when that value is worth using the remaining budget. We propose Paired Outcome-guided Reasoning Allocation (PORA), which learns this decision by comparing short and long reasoning from the same interaction state. During training, \method{} completes the task twice from each selected state, once with a short reasoning allowance and once with a long allowance, and compares their final quality and total processing to estimate when longer reasoning is valuable. During inference, a controller weighs this estimate against learned prices on remaining-budget pressure and chooses either short or long reasoning without running both alternatives. We evaluate PORA across application, memory, code, retail, and mathematics tasks, achieving strong results across all settings. On BFCL memory, it exceeds the strongest existing method by 17.4--32.9 percentage points. Within the same agent, it matches or exceeds always using long reasoning on code and mathematics with 50.6% fewer actor tokens and 31.2% fewer model requests, respectively.

LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm. Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks.

Retrieval-augmented QA can fail even when the correct answer is in the retrieved evidence. This can happen when a competing answer satisfies some, but not all, of the question's requirements and is mentioned frequently in the retrieved documents, which can make it appear to be the best-supported answer. To test whether this retrieval skew actually affects the decision, we manipulate it directly: increasing the prevalence of passages that favor the competing answer makes the model more likely to select it. This points to a decision problem that candidate-level support alone cannot resolve. We therefore introduce Joint Constraint Aggregation (JCA), which compares candidates against the same question-derived requirements using the retrieved evidence. Across three evaluation settings and three instruction-tuned backbones, JCA improves overall accuracy, and its advantage grows when retrieval favors a competing answer. Together, these results suggest that when partial matches and retrieval prominence pull models toward a competing answer, comparing candidates against the question's full requirements can make selection more reliable.

Spoken interaction is moving from alternating turns to full-duplex conversation, where an agent listens while speaking. Learning this coordination requires natural dialogue with persistent speaker-resolved tracks, a data requirement that remains difficult to satisfy at scale. Generative recovery offers a route from ordinary recordings, but recovering two voices is not enough: their track assignments can drift as the conversation unfolds. We address this missing link with speaker-anchored routing before diffusion: a 1.84M-parameter module matches the separator's coarse streams to fixed enrollment identities and orders them before a frozen generator reconstructs the speech, without retraining the generator or altering the reconstructed audio. On 25-minute remote conversations, routing cuts wrong-track time from 43.9% to 4.1%, below the 6.2% of matched-enrollment waveform re-identification, with its largest advantage when one participant holds the floor (17.8% versus 28.6%), and raises fixed-track activity F1 from 0.46 to 0.84 while per-chunk reconstruction quality is unchanged. The same router transfers to unseen face-to-face and telephone recordings, and a margin-gated verification step brings their wrong-track time to 0.9-1.7% while embedding 20-36% of the output, where full re-identification embeds all of it. Recovered voices thus become the participant-consistent tracks that full-duplex dialogue modeling needs. Audio examples: https://elpsykongloo.github.io/beyond-two-tracks/

Credit card fraud is a significant problem for financial institutions, resulting in substantial monetary loss each year. Fraud patterns and methods are continuously evolving, with fraudulent transactions often disguised to resemble legitimate activity. Real world transaction data for fraud detection is severely affected by class imbalance making it harder to detect these transactions. Fraudulent transactions are rarely isolated. Accounts, devices, and other common attributes are often shared across fraudulent activity, revealing relational patterns that individual transaction features alone cannot capture. Graph-based approaches model these relationships explicitly. However, standard graph neural networks aggregate neighbourhood information uniformly, which can dilute the already scarce signal carried by fraudulent transactions and limit detection of camouflaged fraud. This work presents RGH-GNN, a heterogeneous graph neural network that addresses this limitation by incorporating transaction-level risk scores, generated by a supervised LightGBM classifier as edge weights. It is used to guide message passing through a learnable risk-gated aggregation mechanism. It ensures that high-risk transactions contribute more strongly to learned representations rather than being diluted by uniform neighbourhood aggregation. RGH-GNN is evaluated on the IEEE-CIS fraud detection dataset against standard graph neural network models that are trained on the same heterogeneous graph. RGH-GNN outperforms all baselines across every evaluation metric, with an AUC-ROC score of 0.914. This demonstrates that incorporating engineered risk signals into graph-based message passing substantially improves fraud detection performance.

We present the Sparse Particle-interaction Transformer Architecture (SParTA), a modular, scalable, and efficient transformer for jet tagging at high-energy colliders. Exploiting the steeply falling transverse-momentum ($p_T$) spectrum of jet constituents, SParTA factorizes constituent processing into two topological regimes. Ordered by descending $p_T$, only the $N$ hardest constituents (the hard core) participate in quadratic $O(N^2)$ differential interaction attention driven by pairwise physical and learned relational features. The remaining soft constituents are integrated via linear $O(lN)$ cross-attention to the hard core, preserving soft-radiation and color-flow information while avoiding the $O(l^2)$ memory bottleneck of full self-attention over all $l$ constituents. Pairwise interaction features are computed dynamically in-network from constituent four-momenta and conditioned on global jet-level substructure observables. On the quark-gluon tagging benchmark, a 300K-parameter SParTA model achieves an AUC of 92.02% and accuracy of 85.06%, matching the 2.14M-parameter Particle Transformer (ParT) at one-seventh the parameter count and outperforming IAFormer by 21% in background rejection at 30% signal efficiency. On the boosted top-tagging benchmark, SParTA outperforms ParT and ParticleNet with 511K parameters, while we analyze the architectural trade-offs relative to fully dense multi-layer interaction networks.

Recent Omni LLMs (OLLMs) have demonstrated remarkable capabilities in joint audio-visual reasoning, yet cross-modal hallucinations remain a major challenge. Existing mitigation methods mainly operate at the output level, overlooking the entanglement of grounding evidence and learned priors within fused representations. To address this issue, we propose Omni-Woodpecker (OWP), a training-free framework that shifts hallucination mitigation from output-level manipulation to conflict-conditioned representation rebalancing. Specifically, OWP first identifies cross-modal conflicts from complementary modality-specific views. It then employs orthogonal decomposition to separate grounding evidence from competing priors within the hidden states. Finally, OWP adaptively rebalances fused representations by suppressing misleading priors and restoring ignored evidence before decoding. Extensive experiments demonstrate the effectiveness and strong cross-architecture generalization of OWP. Using Qwen2.5-Omni-7B as the backbone, OWP improves performance by 12.9 and 8.6 points on CMM and AVHBench, respectively, while consistently achieving gains of 0.67-2.25 points across four general OLLM reasoning benchmarks. Furthermore, OWP generalizes effectively to VideoLLaMA2-AV, improving performance by 4.6 and 3.9 points on the same benchmarks. Code is publicly available at https://github.com/ZiiJianLiu/Omni-Woodpecker

Realistic market impact is essential for generative limit order book (LOB) models to serve as world models for trading agents. If a simulator misprices this mechanism, agents learn policies that are optimal in the simulator but wrong in the real market. Existing metrics assess generated streams purely through distributional comparison with historical data, leaving counterfactual response to an agent's own orders largely unexplored. Here, we show that market impact emerges naturally in neural LOB generators trained solely for next-token prediction. We inject execution-style metaorders into the generated message flow of five autoregressive models across five NASDAQ stocks and measure the response using estimators from the empirical impact literature, validated on baselines whose physics is known in advance. Our benchmark evaluates any message-level generator in three phases: free generation (reading order-flow persistence against price diffusivity); metaorder injection (yielding the impact law's exponent and amplitude); and a post-execution window (measuring relaxation). In addition, comparing each visible injection to a blind, simulator-only control isolates the model's learned response from standard book mechanics. For every model on every stock, we measure how the impact responds as a function of cumulative metaorder participation. Interestingly, we observe emergent impact in generative models of market data, with exponents in the range of 0.63 to 1.21, slightly above the traditional 0.5 of the square-root law of impact. In the family of models tested, we are so far unable to observe impact decay during relaxation. We tentatively attribute this to the still-short context windows of the models. We release the benchmark as open source to validate future generative models for market data.

Clinical care, health-system operations and biomedical discovery depend on extracting reliable answers from patient records, yet applying diverse clinical questions across entire populations remains computationally expensive. Here we present Clinical System One (CS1), a 4-billion-parameter decision model for fast, typed judgments from clinical records, along with two agentic harnesses in which CS1 works together with a generalist language model. CS1-4B is trained on 12.4 million questions and answers from 3.6 million patients across a multi-hospital health system, combining structured-data keys, observed outcomes and language-model supervision across clinical notes, diagnostic reports and patient timelines. It answers multiple typed questions in a single forward pass without task-specific retraining, and each answer carries a probability that can route uncertain cases to human review. On 963,408 held-out patients, CS1-4B outperformed open decision models up to 27 billion parameters on 13 of 14 tasks scored independently of its training teacher. Across 21 external data sets, its aggregate performance approached that of the strongest tested 27-billion-parameter model, with 6.4-fold faster inference in a matched workload. Integrating CS1-4B with a language model agent produced cited answers over complete multi-year patient histories and screened a patient population against the eligibility criteria of 17 registered clinical trials, placing four times as many judged-eligible patients among its top 20 candidates as keyword search while using 690-fold fewer language-model tokens than a language model reading every record. These findings support a division of labour in clinical AI, with small decision models performing repeated record reading and generative models formulating questions and synthesizing evidence.

Scientific method figures often provide researchers with a clear visual entry point into a paper's methodology and core contributions, yet their dense layouts and complex information flows are difficult for Vision Language Models (VLMs) to interpret. Existing benchmarks predominantly focus on coarse-grained tasks, such as context-based retrieval or holistic comprehension, overlooking the fine-grained content and inter-module relations within complex methodological diagrams. To address these overlooked aspects, we introduce Scientific Figure Benchmark (SciFigBench), a benchmark for evaluating core capabilities required for scientific method figure understanding. Specifically, it targets three capabilities: dense content perception, visual relation comprehension, and information-flow tracing. For this purpose, we curate 4,352 scientific method figures from real-world research papers and construct 40,670 QA pairs to operationalize these capabilities as measurable question-answering tasks. All annotations are generated by dedicated pipelines and then manually verified. We further propose SciFigCue, a training-free visual prompting module that externalizes arrow-induced dependencies as visual relational evidence for method-figure reasoning. Experiments across open-source and proprietary VLMs show that even strong models remain substantially limited in visual relation comprehension, while SciFigCue improves performance across several tasks without parameter updates. The code and related resources are available at https://github.com/LeNcyer/SciFigBench.

Large language model (LLM) watermarking operates on the next-token distribution that evolves during generation. In streaming generation, each sampled token becomes part of the text and affects subsequent distributions. Embedding therefore proceeds with a changing host, unlike classical watermarking, which adapts embedding to an existing host signal. To accommodate this difference, we formulate a novel embedding framework for LLM watermarking based on two principles. We first argue that embedding should be adaptive, retaining the state of previous sampling decisions to better guide subsequent choices. We further argue that embedding should be host-dependent, using the current LLM distribution to constrain both token selection and state updates. Guided by these principles and the two requirements of distribution preservation and robustness to text editing, we propose ClockMark. It combines adaptive and host-dependent embedding with message spreading, embedding the full payload in each watermarked token. Theoretically, we prove that ClockMark preserves the original LLM sequence distribution, thereby maintaining generation quality. Empirically, extensive experiments against 12 baselines demonstrate improved robustness to text editing, with gains of up to 16.08 percentage points in bit accuracy over the strongest tested baseline.

We introduce Lookup-table (LUT)-aware Efficient Attention on FPGAs (LEAF). LEAF is the first nanosecond-scale LUT-aware transformer for FPGAs. Deploying transformers that are large enough to perform well, small enough to fit on an FPGA, and fast enough to meet the sub-microsecond timescales required by real-time systems such as those at CERN remains difficult, since the same attention mechanism that gives these models their high expressivity also makes them costly to implement in hardware. LEAF converts the query, key, and value projections and the final output feed-forward layer into FPGA-native layers with LUT-aware training, creating the first LUT-aware transformer and linformer layers built for ultra-low-latency inference. Attention resists LUT-aware training because it multiplies two activations where the method needs a constant weight to tabulate; LEAF tabulates the attention block's constant matrix-vector multiplications and leaves the rest in arithmetic form. On a standard jet tagging benchmark this improves accuracy by 1.8 percentage points while using 2.9 times fewer LUTs than the previous state-of-the-art sub-microsecond transformer at the same input size. Our most efficient LEAF-based linformer with 64 input particles reaches 80.4% classification accuracy at 28k LUTs and 35ns.

Physics seldom makes breakthroughs when we consider objects only in isolation. Instead we model and measure how objects sit among their neighbours, spatial, temporal, or otherwise: a galaxy in its environment, a halo in its history, a photon hit in a detector. Graphs are ideally suited for encoding exactly these relations, and Graph Neural Networks (GNNs) learn the geometric functions that respect them. This review presents the use of GNNs for astrophysics and cosmology, and argues that GNNs often suit physics and physical data better than any alternative: they fit the native geometry of the domain and the signal, enable the incorporation of the symmetries and prior knowledge we already trust and have toiled to build, and stay flexible enough to capture what we do not yet understand. GNNs' physically intuitive workings also render them more interpretable than other deep-learning architectures. We show how we can begin from general principles of the geometry of deep learning and build the graph and the graph neural network from the ground up, and how other architectures, like the convolutional network and the transformer, can be seen as special cases of one unified geometric blueprint. We survey applications across cosmology, galaxy formation, dynamics, and instrumentation; and we chart the frontier. From this survey, we find that only one in six articles outside instrumentation applies GNNs to data rather than just simulations. We also distil a set of best practices, including when not to reach for a graph. Throughout, we will prioritise physical intuition and foundational principles over mathematical machinery whenever possible.

Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2$\times$ less pretraining compute. Further, Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275$\times$ less pretraining compute. Controlled experiments show that text-only pretraining does not improve T2I generation. Halcyon also learns image captioning by replacing caption dropout with I2T supervision, maintaining competitive T2I generation quality under the same pretraining budget. We show that strong T2I systems can be trained from scratch without internet-scale language pretraining. Code and models are available at https://cs.stanford.edu/~keshik/halcyon

Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply increasing the inference budget of the base model yields only modest gains. Such tasks demand grounded perception, external knowledge, and numerical measurements obtained through heterogeneous tools, whose outputs must be made interpretable to the model and retained across reasoning steps. We introduce OpenVLHarness, a multimodal reasoning harness organized around a capability layer and persistent multimodal memory. The capability layer combines task-level tool interfaces with evidence rendering, providing consistent access to heterogeneous backends and converting their outputs into model-readable observations. Persistent multimodal memory retains the corresponding images, structured data, and web references as addressable artifacts throughout each query. A coding interface executes generated programs over stored data, enabling computations that combine results from multiple tools. Together, these mechanisms support an iterative process of acquiring, inspecting, and composing multimodal evidence. Evaluated on 22 benchmarks with eight open and proprietary backbones, OpenVLHarness improves average scores by 9.7-12.6 points over matched no-tool baselines, including +11.2 with Qwen3-VL-8B, +9.7 with Kimi K3, and +9.9 with GPT-6 Astra. With GPT-6 Luna at medium reasoning effort, OpenVLHarness outperforms the base model, Codex, and Visual Sketchpad at maximum effort by 7.7-8.5 points while incurring at most 37% of their model-token cost.