The Inference Report

Machine learning models often achieve high average accuracy in the training data while performing poorly on test data due to subpopulation shift. A common data-centric solution is to enforce balanced subgroup proportions, but full balance can be costly, infeasible, and not always necessary. In this work, we study whether balanced training data is the unique optimal configuration for robust subpopulation generalization. We extend a Gaussian feature-model framework to theoretically characterize the performance landscape across general training subgroup distributions. We show that multiple configurations, including substantially imbalanced ones, can perform comparably to the balanced one. We further derive gradient-based directions for adjusting subgroup proportions to improve adjusted accuracy and worst-group accuracy more efficiently. These findings suggest that strategically allocating data offers a more flexible and principled path to robust performance than simple balancing. The code is available at https://github.com/zhaoyingpan/subpop-data-allocation.

Stochastic differential equations (SDEs) describe stochastic processes across a wide range of scientific domains, including physics, biology, and finance. Although individual numerical integration steps are inexpensive, observing meaningful events or transitions can require millions of sequential steps, resulting in substantial wall-clock simulation time and poor hardware utilization. Deep learning offers an alternative by amortizing many integration steps into a single neural network forward pass, potentially enabling much faster simulation. However, existing supervised approaches typically require long simulated trajectory data, which restricts their applicability. In this work, we introduce Kolmogorov-Chapman Maps (KC-Maps), a method for learning the transition kernel of an SDE that relies only on the trajectory-free marginal samples together with the known drift and diffusion coefficients. Given an initial time and state of a process, KC-Maps models its distribution integrated to any given time, corresponding to the weak solution of the SDE. To recover the correct dynamics, KC-Maps satisfies two key properties: (i) it satisfies the boundary conditions, which we incorporate directly into the neural network architecture, and (ii) it respects the Kolmogorov--Chapman equation, which provides an efficient training objective. We empirically demonstrate that KC-Maps learns to integrate the dynamics of peptide molecular systems and outperforms classical molecular dynamics in terms of GPU-hours.

Motivated by reinforcement learning subject to communication constraints, we study a policy evaluation problem with linear function approximation over a rate-limited channel. In our setup, an agent interacting with an unknown environment (modeled as a Markov Decision Process) transmits quantized estimates of a temporal-difference (TD) learning update direction to a server over a finite-capacity channel. Our goal is to investigate to what extent one can preserve the statistical convergence rates achievable when the channel has infinite capacity. To that end, we propose a new algorithm called Batched Adaptively Quantized TD (BAQ-TD), which uses batching to reduce the variance of the transmitted information, and adaptive quantization to encode innovations. We prove a high-probability finite-time error bound for BAQ-TD, and show that, above a certain finite channel capacity, BAQ-TD preserves the convergence rate of unquantized TD up to logarithmic factors. Specifically, for a $d$-dimensional value function parameter, we establish that transmitting a total of $\mathcal{O}(d \log(T))$ bits over a horizon of length $T$ suffices to achieve the order-optimal rate of $\tilde{\mathcal{O}}(1/T)$, marking an exponential improvement in communication-complexity over prior work.

As multimodal agents accumulate increasingly long interaction histories, effectively retrieving and utilizing past information becomes critical for long-term interaction. Existing memory systems typically rely on semantic retrieval over compressed memories and expose predefined representations, which may miss relevant memories and overlook heterogeneous evidence needs across queries. We introduce \textbf{\hivemem}, a multimodal memory framework that interleaves \textbf{\textsc{H}}or\textbf{\textsc{i}}zontal Memory Expansion with \textbf{\textsc{Ve}}rtical Evidence Composition. For horizontal memory expansion, \hivemem organizes memory episodes into a relational graph using textual and visual anchors, expanding semantically retrieved seeds with related neighbors to recover complementary information. For vertical evidence composition, each retrieved memory retains multiple evidence forms, including compact summaries, captions, grounded visual views, raw images, and interaction history. A lightweight trainable selector independently composes an evidence subset for each retrieved memory, enabling query-adaptive utilization of heterogeneous multimodal information. This two-dimensional design jointly broadens memory coverage and adapts the evidence contributed by individual memories. Experiments across three multimodal memory benchmarks show that \hivemem consistently improves task performance across base models while reducing token consumption by about 90\% compared with the most efficient competing multimodal baselines.

Anomaly detection (AD) is a critical for trustworthy machine learning and scientific discovery. Existing methods typically define anomalies through local density, isolation heuristics, or distance from a fitted distribution, making them sensitive to neighborhood scale, partitioning choices, or restrictive distributional assumptions. In this work, we introduce a new paradigm by formulating anomaly detection from a diversity perspective. We propose the $\textit{Vendi Anomaly Score}$ (VAS), which detects anomalies by quantifying how much a sample changes the diversity of the dataset, as measured by the Vendi score, when removed. VAS captures both local redundancy and global data structure without relying on density estimation or partitioning heuristics. Moreover, VAS requires no hyperparameter tuning, instead adapting automatically to the spectral structure of the data. VAS is non-parametric and scales linearly with dataset size. Across 148 benchmark embedding datasets derived from 10 source datasets and 4 embedding architectures, VAS achieves state-of-the-art AD performance. We further validate VAS on large-scale ImageNet experiments, where it remains robust as both contamination rate and dataset scale increase.

The Platonic Representation Hypothesis (PRH) suggests that neural networks are converging toward similar representation spaces. As language models become more capable, their representations are aligning with those of vision models. However, recent work has questioned whether this trend continues to hold for newer language models. Moreover, it has been shown that representation similarity metrics can be confounded by model depth and width, requiring appropriate baselines for interpretation. In this work, we first investigate where alignment emerges between unimodal models across model layers and representation dimensions. Next, we find that the PRH capability-alignment trend persists for newer models as long as the captions are sufficiently clean and informative. Finally, we contextualize cross-modal alignment, introducing unimodal and multimodal baselines on various datasets. Our results show that the cross-modal representational alignment of unimodal models can surpass that of explicitly aligned vision-language models, but remains substantially behind the alignment observed for models of the same modality. Code is available at https://github.com/ExplainableML/prh-closer-look.

Reinforcement learning (RL) can improve an agent's task capability while reinforcing behavior that violates contextual execution requirements. We study this reward-driven misalignment, including reward hacking, through verifiable reward--context conflicts, where compliant task completion remains feasible while some executable shortcuts earn more reward than every compliant strategy. Encoding these conflicts in task dynamics lets us generate diverse interactive environments at scale and derive compliance feedback from execution histories. In resource and reporting workflows, task optimization increases real productivity through unauthorized resource use or raises reported scores while actual work declines. Controlled changes in the training distribution show that minority conflict exposure can still lead to widespread shortcut acquisition, reaching 97.9\% certified profitable violations on an API task under a 25\% conflict mixture after 128 updates. The same contextual checks supply a simple binary penalty that supports joint capability and compliance gains and partially mitigates acquired violations. Learned safety adapts to expanded authorization and transfers across mechanisms into workflows with evolving evidence and approval requirements. On 32 unseen changing-evidence tasks, resource-trained safety increases safe completion from 53.1\% to 89.1\% relative to task-only training, while both policies achieve full real work. These results connect a verifiable account of reward-driven misalignment with scalable task construction and safety learning that remains useful under changing execution requirements.

Persona generalization occurs when fine-tuning a language model in one domain induces behavioral changes in others -- for example, the broadly unsafe behavior observed in emergent misalignment. We discover a surprising "on-the-line" phenomenon in persona generalization: in-domain and out-of-domain persona rates across various fine-tuning hyperparameters follow an approximately probit-linear relationship, and the specific probit-linear relationship is surprisingly shared even across model families and model sizes, including Qwen and Llama. Training choices that leave the set of trainable parameters unchanged, such as learning rate, weight decay, and training duration, largely move models along the same line, reducing transfer mainly by weakening in-domain learning. Changing which layers are fine-tuned, however, can change the slope. For unsafe behavior, fine-tuning only later layers yields shallower slopes, retaining strong in-domain learning while reducing out-of-domain misalignment. This "on-the-line" phenomenon unifies unexpected, isolated observations of persona generalization within a quantitative framework for predicting and controlling behavioral transfer.

Large language model agents are increasingly deployed as multi-agent systems rather than as a standalone model operating within a single context. Recent agent harnesses allow the model to instantiate worker agents and to author the executable workflow that coordinates them, shifting the design of system structure from developers to the model at run time. The quality of this organization therefore constitutes a distinct capability, yet existing benchmarks either measure end-to-end task completion or evaluate coordination under structures specified in advance. We formalize this capability as workflow orchestration and introduce WorkflowArena, a benchmark of 145 long-horizon tasks spanning 8 professional domains, from software engineering and formal mathematics to legal analysis and 3D production. Each task specifies an initial workspace and an integrated deliverable composed of interdependent components, which is assessed by a hidden verifier. Agents determine autonomously whether and how to delegate, and all delegation is recorded, enabling joint evaluation of deliverable quality and the organization of the underlying work. Experiments with nine models show that no single configuration performs best across all domains and that systems exhibit distinct delegation strategies: at high effort, Claude models in Claude Code invoke workflows less frequently but with larger worker groups, whereas GPT models in Codex delegate more frequently in smaller units. When the executing model is held fixed on a challenging subset of tasks, workflows authored by Claude Opus 5.5 substantially improve weaker executors, indicating that workflow organization is itself a measurable determinant of task performance.

Latent Diffusion Models dominate generative modeling but face two-stage training complexity and the inherent reconstruction limits of pre-trained autoencoders. Pixel-space Flow Matching offers a simplified end-to-end alternative yet often encounters training difficulties when using the standard $v$-prediction objective. Recent research attributes this difficulty to the off-manifold nature of noised targets, motivating a shift toward $x$-prediction. In this paper, we investigate this explanation through patch-wise Principal Component Analysis (PCA) and identify a bandwidth bottleneck. The velocity in minor components is largely determined by the raw noisy input, requiring deep backbones to preserve information that competes with principal-component modeling. Through controlled experiments and representation analysis, we further examine why $x$-prediction remains effective: its parameterization provides a raw-input bypass, while its learned input representation can spontaneously concentrate on principal directions. Under low-rank observations, our analysis of the induced flow characterizes $x$-prediction as an implicit conditional-mean decoder, reconstructing discarded components from the retained information. These observations motivate the Raw-injected Transformer (RiT), which combines a lightweight raw-conditioned decoder with an output-side shortcut to separate semantic modeling from raw-state propagation. Our method restores effective $v$-prediction and achieves competitive performance on ImageNet $256\times256$, providing a perspective on the interaction between prediction objectives and architectural bandwidth.

As large language models become more capable agents, they can exhibit deceptive behaviors, cheating human for higher rewards while appearing helpful. We attribute this failure to evaluation-side partial observability, where models are evaluated mainly from final outcome, without checking the faithfulness of underlying reasoning, tool use, and environmental effects. Prior mitigations address this partial observability by predefined evaluator latent belief space or one-shot monitoring, which do not extend to free-form language space and vulnerable to stronger deceptions that evade simple monitoring. In this work, we formulate deception mitigation as an Evaluation-Side Partially Observable Extensive Form Game (e-PoG), a max-min framework in which an agent maximizes reward, a monitor minimizes the agent's payoff by exposing deception from the full interaction trajectory, and a fixed verifier validates the monitor's claims and converts them into penalties. We define the solution concept as a Subgame-Perfect Oversight Equilibrium (SPOE) between agent and monitor, and prove its existence in finite games. With a perfect verifier, every SPOE is faithful; with an imperfect verifier, the guarantee persists if verifier errors do not reverse the ranking of monitor actions. To achieve faithful SPOE, we derive a two-timescale policy gradient algorithm and enhance the pipeline empirically with NLI verification and deterministic checks. Experiments with Gemma3-4B-IT and Qwen3-4B/8B on Agentic Role-Playing, SearchQA, and WebShop show that e-PoG mitigates multiple emergent deception modes, outperforming the strongest baselines by 26.0% in deception mitigation and by up to 39.8% in task success. These gains persist under weak-to-strong supervision. Code is available at https://github.com/Toyoid/deceptive-agent.

AI models' growing capacity for complex reasoning has made them attractive tools for forecasting, or the systematic prediction of the future based on past and present data. However, should AI forecasting systems be adopted, human builders will shoulder the work of building and evaluating them. In this paper, we explore what this work looks like with a qualitative interview study with 15 builders of forecasting bots, sampled from among successful participants in a recent AI forecasting competition. Our analysis surfaces the challenges around (1) tuning a bot to acquire factual research, (2) implementing coherent reasoning, and (3) understanding how well one's bot actually does. During their work, builders frequently introduced complexity that did not pay off, and tested in ways that were opportunistic and piecemeal. Our detailed account of their challenges helps clarify what problems should be targeted in upcoming innovations in process, tools, and human-AI teaming for forecasting.

Low-precision pretraining is usually judged by validation loss. We ask whether matching loss also means learning rare facts as efficiently. We train transformers from scratch with emulated 4-bit matrix multiplications on synthetic biographies with controlled exposure counts, on their own and mixed into web text. NVFP4 needs 1.32$\times$ and 1.52$\times$ as many exposures as BF16 to match its recall while validation loss differs by only 0.5% and 2%, and 1.26$\times$ and 1.30$\times$ when both models are evaluated with BF16 matrix multiplications. Under our primary readout, the exposure cost is about 1.2$\times$ at matched loss. Zero-mean gradient noise reproduces the cost of quantizing only the backward pass, which barely moves the loss. In a controlled experiment at twice the NVFP4 error, noise fed only to Adam's second-moment estimate acts like a lower learning rate and its cost fades, whereas noise in the update lowers its signal-to-noise ratio and carries about two thirds of the cost through the end of training. Raising the learning rate to offset the smaller average step does not recover the cost, either for this noise or for NVFP4. Gradient statistics of BF16 checkpoints rank the costs of formats, partial quantizations and noise controls without fitting (Spearman 0.96). We suggest measuring rare-fact acquisition alongside loss when evaluating low-precision recipes.

AI systems must often maintain bounded memory of unbounded interaction histories. However, optimizing the memory representation, and its updates, for downstream task utility is difficult for standard reinforcement learning (RL) due to severe credit assignment bottlenecks induced by sparse, delayed rewards. We introduce MaxIM (Maximally Informative Incremental Memory), a framework that formulates incremental summarization as a capacity-constrained, information-theoretic reward-shaping problem. Grounded in value suboptimality bounds, we show that bounding sequential information loss motivates maximizing the predictive log-likelihood of future, task-relevant targets, a variational lower bound on predictive sufficiency. We introduce a recursive consistency condition to ensure summaries are grounded by maximizing mutual information (MI) with the observed history. We mitigate credit assignment issues with a parameterized asymmetric critic that computes step-wise variational MI bounds for dense reward shaping. Empirical evaluation on long-horizon scientific reading and multi-turn dialogue benchmarks shows that MaxIM outperforms RL, supervised fine-tuning, and prompt-focused baselines on factual grounding, comprehensive autorater metrics, and downstream task utility across strict token budgets and extended horizons, without the need for expensive human preference labels.

Coding agents often need broad authority for productive development work, but granting that authority can expose users' data and infrastructure. Off-the-shelf cloud workspaces can help contain agent activity, but reliance on provider-specific platforms can lead to vendor lock-in and reduced user control. Researchers and engineers must therefore balance agent autonomy, enforced containment, and user control within available financial, computational, and human resources. We analyze these trade-offs and present KVM-Agent, an open-source implementation that automates most setup and provisioning steps. It gives agents broad authority inside replaceable Ubuntu VMs while keeping host administration outside those VMs and supporting multiple coding-agent applications, model services, and remote job delegation. The approach retains familiar workflows and user choice at the cost of VM resources, maintenance, and residual risks from secrets accessible inside the VMs and permitted network access.

One hundred known errors were planted into 10 open-access psychology papers, and the modified papers were then run through frontier models and two commercial AI review tools. The best single system caught 71 of the 100 errors, while the worst caught 30. Pooling every system's output caught 93. Models are only partly correlated in the errors they find, which makes ensembling a big lever for finding issues in papers. Seven errors could not be caught by any system, and all seven were omissions: information deleted from a paper rather than mistakes inserted into it. Refine.ink contributes more unique catches than any other single system, though it is expensive on a per-error basis. False positives were not measured, and how this error distribution compares to the distribution of errors in real papers remains unknown. The papers, errors, model outputs, and full experiment log are public, so that others can build on this toward a comprehensive evaluation benchmark across disciplines.

BACKGROUND: AI-based tools are expected to play a key role in lung cancer screening (LCS). The evaluation of these tools often relied on private datasets or fragmented public datasets with inconsistent availability of clinical data and expert nodule annotations. This has hindered reproducibility, comparability, and independent assessment of AI tools. To address these gaps, we present LUCQUE, a novel, large, comprehensively annotated dataset derived from LDCT lung cancer screening in Québec, which is accessible for reliable AI evaluation. METHODS: LUCQUE was curated and is hosted on PARADIM, a digital infrastructure supporting DICOM data de-identification, annotation, long-term storage and analysis. LDCT scans and clinical information were combined to a curated collection and complemented by voxel-level, expert-derived pulmonary nodule and cyst segmentations, and Lung-RADS-based lesion characteristics. Key dataset characteristics and statistics were reported, including participant demographics, imaging parameters, radiological findings, and segmented lesion characteristics. Dataset analyses were conducted according to PARADIM's model-to-data approach, in which user code is executed on the platform such that data remains on premise. RESULTS: LUCQUE currently comprises 1,678 LDCT baseline studies from 1,678 participants. Most studies were classified as Lung-RADS 2. A total of 1,776 pulmonary nodules and 60 cysts were annotated, with a median (interquartile range) volume of $89\,(56,\,159)\,\mathrm{mm}^3$ and $1,384\,(589,\,3793)\,\mathrm{mm}^3$ respectively. CONCLUSIONS: Compared to publicly available LCS datasets, LUCQUE currently represents the largest accessible annotated real-world LCS dataset. Access to LUCQUE is provided via a model-to-data approach, enabling privacy-preserving, reproducible, and independent real-world evaluation of AI tools in LCS.

The push towards circularity and Re-X principles presents a challenge to SMEs, particularly due to increasing product complexity that impedes the adoption of disassembly automation. This paper introduces and compares two frameworks for extracting disassembly instructions from unstructured information sources, such as disassembly manuals. Our goal is to transform natural language text into a structured, curated and machine-readable format. We implemented and tested a rule-based Natural Language Processing (NLP) framework against a generative AI framework leveraging LLMs. The performance of both approaches was compared using real-world disassembly manuals. This work provides the necessary data foundation for developing and automating future disassembly processes.

Counterfactual inference is the task of answering hypothetical "what-if" queries about a unit, given an observational dataset, modeling assumptions, and a causal graph. This enables a modeler to reason about hypothetical scenarios that cannot otherwise be answered without assumptions or a causal graph. The simpler case involves a causal graph with scalar 1-D variables that denote different attributes in the data-generation process. While the literature is clear in the 1-D case, extending this to settings with high-dimensional nodes (e.g., images) remains challenging. Conventionally, in the literature, methods recover the latent state from the factual image (inversion), generate counterfactual attributes from the graph, and render the counterfactual image from the latent state using the updated attributes. In our work, we investigate the necessity of the inversion step in computing high-dimensional counterfactuals. We first discuss moving from the scalar 1-D case to the higher-dimensional case. Secondly, we discuss the relevance of the inversion step and evaluate an inversion-free counterpart repurposed from the image-editing literature. Finally, on image datasets such as CelebA and ADNI, we benchmark our approach using two measures: namely, effectiveness and identity preservation. While effectiveness is a measure of how faithfully the counterfactual attribute is rendered, identity preservation measures whether the same unit (image) is being rendered, both of which are crucial for high-dimensional counterfactual inference. We present the tradeoffs of using inversion-based and inversion-free approaches in our work.

Users are asking language models mental health questions and are receiving answers. We audited what 41 language models recommend to users when the same clinical information reaches them from three footings: a clinician describing a patient, a caregiver describing a loved one, or a person describing themselves. Across 19 case vignettes and five questions, we collected 35,055 free-text responses, extracted every intervention named with a validated judge, and had 29 models rate the full option lists on a seven-point Likert scale for the same cases. Models largely recommend what they rate well, with recommendation behavior differing by footing, and many models offering clinicians a consultation-style response and lay users hedged, simplified advice. When asked about recommendations for a specific intervention type, models more often named options they themselves rated below zero as initial choices, usually with conditions attached. We will release the instrument, code, and full corpus on publication as a living audit.

Dense non-rigid registration, the recovery of a pixel-wise deformation field that warps one image into geometric agreement with another, is fundamental across medical imaging, structural biology, and correspondence estimation, yet learning-based approaches face a severe bottleneck: ground-truth deformation fields are notoriously difficult to obtain for real data. We present 2D Gum-Net, an end-to-end architecture that predicts dense elastic deformation fields fully unsupervised. We transfer the Geometric Unsupervised Matching Network (Gum-Net), originally developed for 3D subtomogram alignment in cryo-electron tomography, to 2D elastic registration: all three modules are redesigned in 2D, and the rigid transformation head is replaced by a Dense Spatial Transformer that regresses a sparse control-point grid, lifted to a full-resolution field $Φ$ by bicubic interpolation. Downsampling uses DCT spectral pooling, implemented as frozen-basis matrix multiplications ($\approx 50\times$ GPU speedup), preserving the fine, quasi-periodic structure that max and average pooling destroy. We evaluate on contact-based fingerprints, a challenging testbed for this task: locally self-similar ridge patterns make correspondence maximally ambiguous, skin pressed against a sensor generates authentic large-displacement elastic distortion, and minutiae act as precise landmarks through which an independent matcher externally verifies the topological correctness of predicted fields. Trained purely on synthetic impressions and evaluated zero-shot on real benchmarks, our method improves structural correlation by up to 15.7% on FVC2004 and raises genuine match scores under an independent matcher while leaving imposter scores essentially unchanged. This suggests that the predicted deformations restore true structure without fabricating correspondences.

Reinforcement learning is now the mainstream paradigm for training LLMs on agentic tasks, yet standard practice keeps only the best-validation checkpoint and discards the rest. We show that this is lossy: in GRPO training of LLM agents, the best-validation checkpoint fails on problems that earlier checkpoints from the same run had already solved, and the union of problems solved across checkpoints far exceeds any single one. The discarded checkpoints thus form a teacher pool obtained at no additional training cost, which on-policy distillation (OPD) can mine to recover the lost coverage. Mining this pool raises two challenges: i) which checkpoint to distill from, as no checkpoint dominates the others and complementarity to the student is problem-specific; ii) where to trust it, as even a well-chosen teacher is only locally superior, so indiscriminate distillation overwrites behavior the student has already mastered. We propose Agentic-OPD, which treats the best-validation checkpoint as the student and all saved checkpoints as frozen teacher candidates, using the student's evolving capability gaps to guide both teacher selection and supervision filtering: the HCN policy (Highest Coverage Next) dynamically selects teachers by weighted coverage of the student's unsolved problems, while the LAD module (Locality-Aware Distillation) confines transfer to failed student rollouts on problems where the teacher performs better. Across five agentic task domains and 24 benchmarks and task subsets, Agentic-OPD improves over its initialization checkpoint by up to 7.4 points, with a 4B student surpassing the latest frontier models GPT-5.6 and DeepSeek-V4.1 on deep search. Agentic-OPD recasts a finished RL run as a reusable pool of policies and establishes post-RL capability recovery as a stage orthogonal to improvements in the RL algorithm itself.

Woody Breast (WB) myopathy is a major quality defect in broiler chickens characterized by fibrosis, increased tissue hardness, and altered structural properties that affect product quality. Conventional palpation is subjective and labor-intensive, while many non-destructive methods require specialized equipment or substantial computational resources. This study developed a smartphone-based transillumination and 3D point-cloud framework for non-destructive WB severity classification. A total of 267 chicken breast fillets, including normal breast (NB, n = 87), mild WB (MWB, n = 85), and severe WB (SWB, n = 95), were evaluated using RAW transillumination images and 3D point clouds acquired with an iPhone 16 Pro. A total of 91 features, comprising 14 image-derived, 76 geometric, and one physics-informed Beer-Lambert attenuation feature, were extracted. Partial correlation controlling for thickness revealed a significant association between image characteristics and WB severity (r = 0.413, p < 0.0001), indicating optical information independent of fillet geometry. SWB tissue exhibited approximately 26% greater light attenuation per unit thickness than NB. The 2D-only models achieved approximately 82% test accuracy, while the multimodal ensemble combining Fusion-DGCNN, XGBoost, and physics-guided logistic regression (PGLR) achieved approximately 90% accuracy. These findings demonstrate the complementary value of optical, geometric, and physics-informed features and establish the feasibility of smartphone-based, non-destructive WB severity assessment.

Developing effective machine learning models requires complex manual tuning, while current Automated Machine Learning (AutoML) frameworks remain domain-specific and isolate model architecture, training hyperparameter, and data ratio optimization. Here, we bridge this gap by proposing a universal, large language model (LLM)-based AutoML framework that jointly optimizes these configurations for the first time without the need for task-specific fine-tuning. We propose NCode, a unified numerical encoding scheme that translates diverse AutoML design choices into universal discrete sequences and effectively reframes the optimization process as a unified ranking task. Trained solely on optimization trajectories derived from a limited, single-domain vision task, the proposed LLM-based searcher demonstrates impressive cross-task, cross-modality and joint search-space optimization generalizability, achieving state-of-the-art performance across image classification, image segmentation, speaker verification and text-to-image benchmarks. Crucially, it exhibits superior scalability in the large-scale training and joint optimization of a billion-parameter Diffusion Transformer, requiring 896 GPU-days and attaining a state-of-the-art GenEval score of 0.91. By leveraging reinforcement learning to instill Chain-of-Thought reasoning, we provide a systematic way to interpret the LLM-based searcher's internal decision-making mechanisms, revealing its optimization strategy shifted from the early-stage diversity-driven exploration to the late-stage anchor exploitation. Our proposed framework provides a scalable, generalizable, and interpretable pathway toward general-purpose open-world machine learning optimization.

Prompt-based continual learning adapts a frozen pre-trained vision transformer by tuning a small set of prompt tokens. Recently, one-prompt methods which share one single prompt across all tasks make this paradigm maximally parameter efficient. The open question is how a shared prompt can keep the knowledge of old classes while being updated for new ones. Existing methods protect old knowledge in parameter space, where they freeze task-specific prompts, expand new components, or sparsely gate which tokens receive gradients. However, this method sacrifices plasticity, yet the drift it still permits in shared tokens can corrupt the prompt's function on past classes. What must be preserved is the prompt's behavior. We propose TEA-Prompt, which couples (i) a token evaluation module that assigns each prompt token a continuous, input-conditional relevance score injected into the attention logits, softly specializing tokens across tasks without discarding any of them, and (ii) an attention-behavior anchoring loss that constrains each token's attention output around class-specific query regions using only compact Gaussian query statistics without stored old images, weighted by the token's past relevance. Anchoring the attention value allows the parameters to continue updating freely, as long as their interaction remains unchanged. TEA-Prompt achieves state-of-the-art accuracy with markedly lower forgetting on CIFAR-100, ImageNet-R, and CUB-200 with only 0.38M parameters. Code is publicy available at https://github.com/cuonga1cvp/TEA-Prompt.git

Log-based anomaly detection is increasingly dominated by large pretrained models, from BERT-scale encoders to 8B-parameter language models, which require GPUs at inference time and are difficult to deploy on the monitoring nodes where logs are produced. We present LogCDM, a compact detector that scores each log session using three complementary signals: the next-event surprise of a small causal Transformer, the distance between a graph-attention embedding of the session's own transition graph and that of a normal reference graph, and a transition-drift statistic that we show is exactly a frequency-weighted chi-squared divergence between transition kernels. The divergence view yields an asymptotic null distribution, a length-standardized score, and a per-edge decomposition that explains each alarm. A sigmoid gate fuses the three signals; under a shared-covariance Gaussian model, this gate is the Bayes posterior, and its learned weights indicate which signal each dataset relies on. With 8.1-8.8M parameters and no GPU required at inference, LogCDM reaches F1 scores of 0.984 on HDFS, 0.870 on chronologically split BGL, and 0.850 on Thunderbird, within 0.013-0.12 of an 8.1B-parameter LLM detector evaluated under the same protocol and above all compact baselines on BGL. It scores a session in 5.5-10.5 ms on a CPU. Ablations show that the structural signals add 0.025-0.030 F1 on BGL and Thunderbird but almost nothing on HDFS, consistent with the learned fusion weights.

Stochastic Gradient Descent (SGD) is a workhorse algorithm for continuous optimization. It is a folk belief among deep learning practitioners that learning rate decay in SGD substantially improves performance. However, it has remained unclear whether learning rate decay offers an asymptotic advantage for global nonconvex optimization. We answer this question in the affirmative by demonstrating a natural class of nonconvex functions for which, we prove that SGD with learning rate decay requires exponentially fewer gradient queries for global optimization than SGD with any fixed learning rate in the small-step size regime. To the best of our knowledge, this is the first exponential advantage that has been demonstrated for learning rate decay. The class of functions we consider includes many popular benchmark nonconvex functions. Our results provide a robust explanatory theory for the empirically observed benefits of learning rate decay for global optimization. We show that annealing performs the non-logconcave sampling task associated to the optimization problem in polynomial time, via a novel application of weak Poincaré inequalities.

Preference optimization has become a central approach for aligning large language models with human values, but its effectiveness depends heavily on the quality of preference annotations. In practice, preference data is often noisy due to annotation errors and ambiguity. Existing robust preference optimization methods primarily operate at the sequence level, implicitly treating entire responses as uniformly correct or incorrect. However, rejected responses may still contain informative reasoning steps, while preferred responses can include subtle errors. In this work, we propose \textbf{P}reference \textbf{O}ptimization via \textbf{R}obust \textbf{T}oken-level reweighting \textbf{(PORT)}, a fine-grained framework for robust alignment under noisy preferences. PORT performs token-level reweighting using the empirical cumulative distribution function (CDF) of token log-probabilities, yielding an efficient forward-pass-only proxy to gradient-norm penalization that selectively suppresses corrupted tokens without additional backward-pass computation. We provide theoretical analysis showing that PORT reduces the gradient bias from noisy preference labels by minimizing an upper bound on the bias term. Extensive experiments across diverse noise settings demonstrate that PORT consistently improves robustness and outperforms existing sequence-level baselines. The code is available at https://github.com/osu-srml/PORT

We introduce Sage-Taiyi, a metacognition-inspired architecture that uses accumulated scientific understanding to guide what to investigate next and how to investigate it. Through explicit monitoring and control, the system revises its understanding of evidence, assumptions, and unresolved questions, then uses these assessments to adapt research targets and select reusable reasoning methods. On Taiyi-Bench, an AI co-scientist ideation benchmark comprising 265 tasks from 50 papers across ten disciplines, Sage-Taiyi ranks first on all eight metrics of historical rediscovery and open-ended ideation, achieving an Overall score of 54.0 compared with 47.9 for the strongest baseline, AI Scientist-v2. It also performs well on two external scientific ideation benchmarks. Five expert-guided collaborations further connect its proposals to experimental and computational validation across chemistry, materials science, and engineering. In chemistry, reaction-condition optimization raises the yield of cobalt-catalyzed carbonylation from trace amounts to 84% and photocatalytic allylic amination from 13% to 69%. These results support metacognitive regulation as an approach to scientific ideation and its progression toward practical validation.

Modern cyber threats commonly unfold as multi-step attacks comprising temporally and causally related phases. Current defenses commonly follow a detect first and correlate later paradigm. Intrusion detection systems classify events independently, while downstream systems reconstruct attack progression from the resulting alerts or audit records. In this context, learning-based detectors learn attack-step dependencies implicitly as statistical regularities in the data and cannot guarantee that their predictions follow the real attack progression. Conversely, correlation systems represent attack relationships explicitly, but apply this knowledge only after the underlying attack observations have been detected. We present DEEPCHRONOS, a neuro-symbolic network intrusion detection system that shifts the paradigm from detect first and correlate later to detection through logical progression. DEEPCHRONOS combines learned flow-level evidence with explicit attack-phase rules, temporal prerequisites, and previously inferred phases, allowing the evolving attack history to inform and constrain each detection decision. Evaluations on DARPA 2000 and AIT-LDS V2 show that DEEPCHRONOS improves macro F1 over purely neural baselines by up to 11.3%, eliminates causal violations under strict temporal constraints, and maintains millisecond-scale inference latency. These results demonstrate that integrating explicit attack progression knowledge into detection produces more accurate and temporally coherent multi-step attack predictions.

The idea of testing whether the past of one signal improves the prediction of another can be traced back to Wiener. It was Granger who formalized the concept. Granger causality has become one of the most widely used tools to infer directed interactions between the components of a system. The name however can be misleading: it detects predictive relations among the measured signals, not causal relations among the mechanisms that generate them. This article starts by reviewing recent advances in observation-space Granger analysis. It then follows a path to directed structure in latent dynamical models, and asks under which conditions that structure can be read causally.

Large language models (LLMs) increasingly mediate how people seek and consume information, including accounts that differ across languages and regions, especially on contested topics. Supporting diverse information seeking requires models to help users navigate competing narratives across languages and information streams. We define six capabilities necessary for diverse multilingual information seeking: three foundational capabilities for narrative inclusion, narrative exclusion, and missing narrative identification; and three higher-level capabilities for presenting diverse narratives, preserving narrative coverage under confirmatory queries, and maintaining information parity across languages. To guide the development of those information seeking capabilities, we introduce \textsc{NarrativeBench}, a benchmark comprising 1,333 news articles in 23 publication languages across 23 contested topics and 460 Reddit-grounded queries. Evaluating 14 LLMs across 24 query languages, we find that current models lack the capabilities needed to reliably navigate diverse narratives, limiting users' control over the perspectives they receive and leaving substantial gaps in exposure to competing accounts even when diversity is explicitly requested. Information access also differs across query framings and languages, risking echo chambers, linguistic filter bubbles, and information disparities that weaken the common ground needed for democratic discourse. \textsc{NarrativeBench} provides a foundation for developing information systems that give users across languages and regions agency to navigate diverse narratives.

Exact score-based causal structure learning supports explicit structural constraints and optimisation certificates, but its branch-and-cut search can be computationally demanding. We augment exact causal structure learning methods with an upper bound on the maximum clique size of the learned undirected skeleton, enforced at integer candidates through lazy cycle and clique constraints. On static linear-Gaussian Erdős-Rényi and scale-free DAGs, we evaluate callback order, alternative clique-cut policies, incumbent repair, fractional separation, and Big-M bounds. We report causal-recovery metrics jointly with optimality gaps, primal and dual bounds, node throughput, MIPSOL events, and callback cost. Two variants receive particular attention: B3 selects the most violated minimal (k+1)-clique inequality at each callback, whereas C3 repairs an infeasible integer candidate by deleting violating edges and refitting affected regression columns using QR-based least squares. The results do not indicate that clique constraints uniformly accelerate optimisation. However, their benefit depends on graph density, sample size, and cut-management policy, while large collections of correlated constraints can reduce node throughput. Tighter valid Big-M bounds reduced the fixed-budget optimality gap in most tested settings, without necessarily changing SHD or F1 at a fixed time limit. These findings characterize the computational trade-offs of clique-constrained exact causal learning, and identify selective separation and incumbent repair as promising directions.

Deploying large language models in regulated industries requires more than picking the fastest inference engine: it requires evidence that a configuration meets contractual and legal obligations before it serves a single customer request. We describe Runway, an internal validation harness built at MasterControl that builds a governed decision pipeline around NVIDIA AIPerf. The key insight is separating concerns: AIPerf provides the trusted measurement, while Runway provides the experiment design that comes before it and the compliance, cost, and audit decisions that come after. We describe the architecture and the storage model that makes caching and audit one mechanism. We report three findings from Runway sweeps: TensorRT-LLM and vLLM rankings invert with concurrency; a newer, pricier GPU instance can be the cheaper one per request; and speculative decoding behaves differently across engines under load, so engine configuration can outweigh engine choice.

Post-LayerNorm Transformers have seen limited practical adoption due to enduring difficulties in scaling them to large depths. While prior research has focused on stabilising Post-LayerNorm by improving conditioning at initialisation, stability often deteriorates when models are trained with large learning rates, forcing additional compromises. In the decoder-only settings we study, existing Post-LayerNorm methods fail to outperform strong Pre-LayerNorm baselines. We propose KiteNorm, a post-norm training method designed to break this trend by constraining hidden-state variance. This approach is motivated by a bound on the LayerNorm Jacobian suggesting that variance growth contributes to gradient vanishing, alongside empirical evidence linking it to rank collapse. KiteNorm combines residual scaling and variance regularisation to maintain stability throughout training, supported by scalarised double-normalisation that provides separately learnable scales for the skip and residual branches. Across decoder-only Transformers up to 1.3B parameters, KiteNorm improves training stability and outperforms leading baselines at their optimal learning rates, with scaling laws favouring KiteNorm across depth, width, batch size, and training budget. Ablations show that each component is necessary for the full stability and performance gains.

Reinforcement learning with verifiable rewards is increasingly used to improve language model reasoning, yet its effects on sequential decision procedures remain poorly understood. We study this problem in clinical multi-turn agents that must decide what information to request, when to stop gathering evidence, and which action to take. We introduce three controlled environments: medication refill triage, acute diagnostic work-up, and dosing titration, each with hidden clinical states, query costs, exact adaptive oracles, and deterministic outcome verifiers. Across these tasks, outcome-only group relative policy optimization often improves aggregate reward while producing undesirable procedures, including premature commitment, one-question collapse, and failures on rare high-risk cases. In diagnostics, assigning dense credit to questions that reduce uncertainty improves held-out value and moderate-case support across three training seeds relative to outcome-only reinforcement learning, without the high-risk seed collapse observed under outcome-only training. Separating exploration from exploitation further increases high-risk support at its best checkpoint, but at the cost of excessive information gathering on low-acuity cases. Additional experiments show that the optimal strength of process credit is task-dependent; that after a supervised warmup that already gathers carefully, reinforcement learning raises aggregate reward by trimming questions while degrading high-risk performance; and that reinforcement learning does not reliably repair rare behavior gaps inherited from biased supervised data. Verifiable final outcomes are therefore insufficient for training reliable multi-turn agents: reward design must account for the procedure used to reach an answer, and evaluation must distinguish adaptive information gathering from reward-exploiting shortcuts.

Selecting an effective cover photo is a critical yet understudied problem in visual merchandising for e-commerce marketplaces. On Airbnb, hosts upload many photos depicting different rooms, amenities, and points of interest, but must manually choose a single static cover photo that is reused across heterogeneous product surfaces. This creates substantial challenges: over 90\% of listing photos have no engagement labels, leading to a severe cold-start regime; guest preferences vary across surfaces and search intents, and visually captivating photos may be semantically misaligned with the underlying listing. At Airbnb scale, this corresponds to hundreds of millions of photos and hundreds of thousands of new uploads per day. We propose \emph{OptiCover}, a scalable framework for cold-start visual optimization under semantic constraints and multi-surface engagement signals. The framework combines (i) a photo candidate generator that identifies candidate photos through various proxies, (ii) lightweight vision heads applied to frozen CLIP embeddings for large-scale semantic scoring, (iii) a cross-modal grounding validator based on a vision-language model (VLM) that enforces semantic representativeness, and (iv) an Exploration Logging Module (ELM) that orchestrates the collection of randomized engagement data across heterogeneous product surfaces and contextual search intents. We evaluate OptiCover through extensive offline studies and multiple online experiments across HomePage, and Search surfaces. OptiCover is now deployed in production and performs tens of millions of photo swaps daily. Online results show a 1.5\% engagement lift on the HomePage and a 0.35\% increase in bookers on Search, generating millions of dollars in incremental revenue. Our findings suggest that action-space shaping and semantic constraints are primary drivers of performance.

Providing real-time multimodal insights into images is challenging because it requires low latency, high description accuracy, and scalable solutions that are both cost-effective and efficient. For example, given a new image of an outfit, we would like to generate insights that the image depicts "earth tones" and "linen pants" and allow users to explore similar ideas along these themes. Traditional strategies to generate such insights would typically involve building taxonomies producing a golden set labeled by humans, training models on these datasets, and deploying these models for online inference services. Such workflows are constrained by fixed taxonomy, high labeling cost, limited labeling coverage in sample space, and human errors and expertise gaps in labelers. In addition, these methods are expensive and cumbersome to iterate over time. Recent trends have started exploring generative AI models for online insights, but this method suffers from long latency and prohibitive costs to support a high volume of requests or large amounts of offline labeling needs. We present Generative Labeling On Clusters (GLOC) as a solution to provide highly scalable, cost-effective, low latency, and high accuracy online image insight services. GLOC leverages highly optimized image embeddings to form image clusters that are semantically close to each other for the purpose of the type of insights, and summarizes the insights at the cluster level, which effectively generalizes to members in the cluster. At Pinterest, GLOC powers image descriptors and pivot query suggestions for Hybrid Search, which is a multimodal search service enabling users to search with both an image and a text query. Our offline and online A/B tests demonstrate that GLOC significantly outperforms previous methods, delivering improvements in quality and efficiency.

Historical decisions determine which outcomes are observed, potentially concealing a predictor's largest errors. Even unlimited data from the same observation process can therefore leave population tail risk unresolved. Here we derive sharp bounds on conditional value-at-risk, the average loss in a specified upper tail, for binary outcomes and bounded losses. These bounds span exactly the risks compatible with the observations and a stated limit on how outcome odds differ between observed and hidden cases at the same covariate values. Under stated regularity conditions, calibrated statistical certificates cover these ranges simultaneously across candidate predictors, supporting risk limits and model comparisons while separating statistical uncertainty from persistent ambiguity. For a fixed predictor and loss threshold, we derive a minimum-cost allocation for reducing excess-loss ambiguity by prioritizing hidden outcomes according to ambiguity removed per unit cost. This allocation treats audited population contributions as known and retains the original sensitivity bounds separately for audited and still-hidden outcomes. In controlled masking experiments with known loan repayments, recovering about 16% of hidden labels removes a median 61% of estimated fixed-threshold ambiguity, compared with 16% for random recovery. These results connect population tail-risk claims to the available evidence and identify where additional outcomes can sharpen their assessment.

With the growing interest in using AI (Artificial Intelligence) for RTL (Register-Transfer Level) hardware development, robust and comprehensive verification has become more important than ever. As Large Language Models (LLMs) increasingly assist in creating and modifying RTL designs, ensuring that these changes preserve existing functionality is paramount for building trust in AI-driven workflows. We present SVApshot, a fully automated framework that leverages LLMs to generate and systematically attempt to correct SystemVerilog Assertions (SVA) for RTL modules, completely removing the human from the generation loop. SVApshot introduces a novel snapshot methodology that captures the current functionality of a design as a comprehensive set of formal assertions, creating a regression suite for validating future changes--whether made by humans or AI. The framework features automated duplicate detection, assertion set expansion, and iterative repair of failing assertions using LLM-guided debugging with formal verification feedback. Our experimental evaluation across diverse RTL modules demonstrates high coverage for complex modules, with successful detection of both manually-injected and AI-introduced bugs. Additionally, this framework opens the door for automatic generation of formal testbenches to benchmark LLMs for hardware design, moving beyond binary pass/fail metrics towards nuanced assertion-based scoring.

The WebAssembly Component Model enables language-independent composition of independently developed components, but complicates observability. Existing tracing support flattens execution hierarchy by attributing every span created inside a composition to a single host-level parent context, so execution across component boundaries cannot be reconstructed. The problem is amplified because components are typically deployed as black boxes whose internals cannot be modified or inspected by the system integrator. To address these limitations, we propose a revised wasi:otel/tracing design in which the runtime maintains a per-component span stack, restoring parent-child relationships across component boundaries without changing component interfaces or source code. This change has since been adopted upstream. Building on this design, we present WasmLens, an automatic instrumentation tool that derives instrumentation points from a composition graph and injects tracing middleware into already composed binaries, yielding end-to-end traces for black-box compositions with no developer configuration. This automatic instrumentation raises mean request latency from $238$ $μ$s to $454$ $μ$s across six component boundaries. Raising the expected number of boundary crossings from 6 to 11 further increases latency to $514$ $μ$s, an incremental cost of $12$ $μ$s per crossing, of which $10.4$ $μ$s is attributable to instrumentation. These results show that instrumentation overhead is additive in composition depth rather than proportional to baseline execution time.

Objectives: To measure whether a locally hosted large language model (LLM), stacked with deterministic and clinical named-entity recognition (NER) detectors, finds more protected health information (PHI) than open tools, and at what precision. Materials and Methods: DeidSharp, an open-source .NET toolkit, combines the findings of pluggable detectors (patterns, a clinical NER model, a local 26B mixture-of-experts LLM) and fails closed on any detector error. Configurations fixed on tuning halves were compared, under a prespecified protocol, with deid, Philter, the Stanford/Penn de-identifier, OpenAI Privacy Filter and GLiNER on reporting halves of PhysioNet's nursing notes (889 spans) and semi-synthetic MIMIC-IV-Note discharge summaries (29,023 spans), and on synthetic ASQ-PHI queries, with paired patient-level bootstrap intervals. Results: In the nursing notes, against Stanford/Penn, the recall-first stack found 98.9% of PHI (+2.8, 95% CI 1.4 to 4.4) at 9.3 points lower precision; the precision-first stack, without NER, found 96.5%, matching Stanford/Penn, at 19.6 points higher precision (83.2%). In the discharge summaries, both stacks led Stanford/Penn by more than 20 points (recall-first 95.3%, +21.5, 20.5 to 22.4), almost entirely on generated untyped names; on typed labels all three exceeded 99%. All four prespecified comparisons held. No configuration beat deid significantly on its home corpus; Philter found more discharge-summary spans than the recall-first stack at 23 points lower precision. Discussion: The LLM was the main detector and NER added place names; asked to verify its own findings, the LLM discarded a quarter of true PHI. Conclusion: A local LLM layered with deterministic detectors offers recall-first and precision-first operating points without sending text off-site.

When AI agents screen items for a costly human assessor, as in agentic search, e-discovery and AI reviewing of papers, only what they pass gets judged. We show what each kind of evidence can establish about their errors, even with unlimited data. Agreement shows that agents concur, not that they are right, however many there are. Acceptance by the assessor gives precision but only a lower bound on recall. Knowing which agents found which relevant items gives an upper bound on recall if agents' finds are not negatively correlated, and caps what one more agent can add if agents are interchangeable. Without assumptions about the agents, recall can be pinned down only by judging items they did not pass: a random sample of these estimates how many relevant items were missed, to any chosen accuracy, and the sample needed does not grow with the number of agents.

Accurate classifiers can fail when background cues change without altering the correct labels. Predictive performance alone does not quantify how much task information remains after accounting for such nuisance variation. Here we extend the Fisher-glass framework to connect nuisance-adjusted information with local classification difficulty and subpopulation auditing. We derive lower bounds on local minimax excess classification risk under explicit distinguishability and decision-separation conditions, alongside statistical audits accounting for estimation error and regularization. On Waterbirds, a feature-based diagnostic of task-background alignment selects an accurately classified group, with selection frozen before held-out evaluation. Across five training seeds, background replacements preserving the extracted bird foreground and labels yield 93.2% accuracy when background type is preserved and 40.6% when it changes, a controlled difference of 52.6 percentage points. Separately, search using task labels and a fixed nuisance direction identifies an independently evaluated population with 98.8% membership-weighted accuracy and low estimated projected predictive Fisher information. Training and acquisition experiments show that information gains and predictive gains can diverge. These findings support joint assessment of predictive performance, information under declared nuisance models and controlled sensitivity when diagnosing vulnerability and evaluating model improvements.

In ultra-high-resolution (UHR) remote sensing images, the objects that questions count in a region or relate by direction often span only a few pixels. Existing methods mainly decide where a vision-language model (VLM) should look and trust its answer in the chosen view. We show that this assumption fails for counting, and measure VLMs as one calibrates a sensor. All eight open VLMs tested fall increasingly short once a view holds more than a certain number of objects, about five for most on DOTA aerial images. This counting horizon narrows as views grow and objects appear smaller. An empirical power-law response model predicts the undercount a split recovers and the spurious counts it adds, and supports the rule of splitting only views beyond the horizon. In GeoScope, a text-only language model parses the question into a typed spatial program that fixes the region and its resolution. The VLM only grounds and counts within its operating range, and directions are computed explicitly from the grounded boxes. To our knowledge, it is the first UHR method to split views by a measured limit of the VLM. Without training, it improves Qwen3-VL-8B by 16.3 points on regional counting and 10.5 on spatial relations in XLRS-Bench, and has the highest reported accuracy, trained methods included, on regional counting and on the whole benchmark by question. On UHR-Micro it improves global and regional counting by 16.8 and 36.2 points, 15.6 of the former from splitting. On MME-RealWorld counting it gains 11.4 points, and we show that an option that is almost never correct yet often chosen distorts comparisons there. Project page: https://avalon-s.github.io/GeoScope/

Not all mistakes are alike. For example, in medical diagnosis, false positive and false negative mistakes have a different impact and should not be aggregated. By revisiting the classical theory of boosting through this lens, we introduce the notion of $Z$-learner, which returns weak hypotheses achieving joint false-positive/false-negative guarantees $(\mathrm{FP},\mathrm{FN})$ from some set $Z \subseteq [0,1]^2$. This formulation generalizes the standard notion of $γ$-weak learning--recovered by taking $Z=\{(z_1,z_2): z_1+z_2<1/2-γ\}$--and allows for more flexible notions of weak learnability that account for the impact of the distribution on the achievable learning guarantees. This raises a natural question: which sets $Z$ are boostable? We answer this by giving a complete geometric characterization that unifies, as special cases, both the classical boosting dichotomy and the recent multi-objective extension of Bressan et al. (COLT 2025). We also present an adaptive algorithm for boosting $Z$-learners without relying on the prior knowledge of $Z$, generalizing AdaBoost's ability to handle unknown margin parameters. Finally, we demonstrate that $Z$-learners can achieve confidence amplification. However, the analysis is significantly more involved than in classical boosting and relies on the centerpoint theorem from discrete and computational geometry.

Woody breast (WB) myopathy is a spatially heterogeneous fibro-degenerative condition of the broiler Pectoralis major muscle. Existing mechanical reference methods rely on single-point or sequential compression measurements, limiting characterization of spatial hardness variability. This study presents TACTUS (Tactile Adaptive Conformal 3D Uniform Sensing), an automated tactile array that generates thickness-normalized, three-dimensional hardness maps at 20 anatomically distributed locations within approximately 30 s. TACTUS integrates 20 independently motorized tactuators with force-sensing resistors and encoder-driven linear actuators, alongside a LiDAR depth camera for surface profiling and adaptive compression. The system was evaluated using 100 broiler breast fillets across four WB severity classes (Normal, Mild, Moderate, Severe). TACTUS identified significant force differences among severity classes (ANOVA, p < 0.001), with the strongest differentiation in the cranial region and a significant caudal effect not detected by texture-analyzer measurements. Fillet-level TACTUS force showed moderate but significant correlation with texture-analyzer measurements (Pearson r = 0.441, p < 0.001). Semivariogram, local change entropy, and heterogeneity analyses revealed increasing and anatomically structured hardness variability with WB severity. These findings demonstrate TACTUS as an automated, operator-independent platform for spatially resolved mechanical characterization and ground-truth generation to support hyperspectral and near-infrared spectroscopy-based WB classification systems.

We study the problem of binary classification in the marginal-nonuniform learning framework in the agnostic setting. In this model, the constants in the learning rate may depend on the marginal distribution of the unlabeled data, but the guarantee must hold uniformly over all conditional distributions of the labels. We also study fine-grained rates. We consider two criteria. (1) The standard agnostic setting, where the learner competes with the best concept in the class. In this setting, an immediate consequence of standard agnostic PAC learning results gives us the optimal rate of $\tfrac{1}{\sqrt{n}}$ for every class $\mathcal{H}$ with finite $\mathrm{VC}(\mathcal{H})$ and shows that arbitrarily slow rates are necessary otherwise. This also gives us the optimal fine-grained rate of $\sqrt{\tfrac{\mathrm{VC}(\mathcal{H})}{n}}$. Therefore, the dependence on the marginal distribution does not help, which is in contrast to the realizable setting. (2) In the relaxed agnostic setting, where the learner competes with $(1+λ)$ times the error of the best concept for a fixed $λ\in (0,1]$, the picture from the realizable setting reappears. We obtain a trichotomy of the achievable optimal rates: exponential rates of the form $e^{-n}$ arise precisely when the class is finite; linear rates of the form $\tfrac{1}{n}$ arise when the class is infinite but $\mathrm{VC}(\mathcal{H})$ is finite; and arbitrarily slow rates occur when $\mathrm{VC}(\mathcal{H})=\infty$. We also obtain the optimal fine-grained rate of $\tfrac{\mathrm{uVCE}(\mathcal{H})}{λn}$, where $\mathrm{uVCE}(\mathcal{H})$ is a recently introduced combinatorial complexity dimension called the uncentered $\mathrm{VC}$-eluder dimension.

Ottoman Divan poetry is composed in aruz, a quantitative meter based on the weights of syllables. The meter is sensitive to specific characters recorded in the original writing in the Arabic script. This work is concerned with automatic meter finding and scansion alignment for Ottoman Divan poetry, especially on modern Turkish spelling without the special diacritics that scholarly transliterations of Ottoman preserve. We define two tasks on this input, first identifying the meter used in the poem, and second aligning the meter's weights with the syllabification of the poem, identifying the flaws and freedoms the poet has taken along the way. We build a corpus of 52,000 lines split by poet, with scansion annotations for 500 test lines, and a robustness benchmark generated at the deviation rates measured on real web and OCR copies. Our model is a hidden Markov model over the positions of each meter, decoded over a lattice of every candidate path, and trained by EM from the printed meter, without scansion labels. On a held-out set it scans 77.6% of the test lines fully correctly, against 26.8% for the available rule-based scanner, 55.0% for a neural tagger trained with the same supervision and 63.6% for Claude Opus 5.5 few-shot. The code is open-sourced and the tool is released on the web.

Causal inference under treatment interference requires assumptions to reduce the exponentially large number of potential outcomes. We develop hierarchical finite-population models for potential outcomes by separating these assumptions into three layers: interference neighborhoods, exposure mappings, and outcome models. Given an interference neighborhood and exposure mapping, we show that potential outcomes admit a nonparametric decomposition into direct, spillover, and treatment--spillover interaction components. This decomposition provides a common coordinate system for nonparametric, semiparametric, and parametric models of finite-population potential outcomes. For a given potential-outcome model, we define two classes of estimands. Average causal effects (ACEs), such as direct and total treatment effects, contrast potential outcomes under fixed treatment--exposure conditions. Expected causal effects (ECEs), such as expected direct and indirect effects, contrast expectations of potential outcomes under a specified allocation policy. For both classes, we bound the variance of the canonical Horvitz--Thompson estimator by three terms: a diagonal weight term, a term from overlapping interference neighborhoods, and a residual design-dependence term. Our upper and lower bounds show that under graph-unaware designs, nonparametric design-unbiased estimation of ACEs can require sample sizes growing exponentially in the interference degree, whereas polynomial growth can suffice for design-aligned ECEs. Under neighborhood interference, we construct graph-aware independent-set designs with variance bounds polynomial in the maximum degree for direct and total treatment effects on every graph, whereas graph cluster randomization can incur worst-case variance exponential in the degree.

Stellarators are fusion devices that use externally generated, three-dimensional magnetic fields to confine hot plasma and offer a promising route to commercial fusion. Their electromagnetic coils pose a challenging design problem: they must reproduce the target magnetic field while satisfying engineering constraints and remaining practical and cost-effective to manufacture. Finding suitable geometries is a highly nonlinear inverse problem that generally has neither a closed-form nor a unique solution. This problem lies at the heart of stellarator design. We introduce CoilStellaration, a dataset and benchmark for learning the mapping between coil geometries from target plasma magnetic fields and engineering requirements. The dataset contains 177,000 optimized filamentary coilsets targeting plasma equilibria from ConStellaration, with sampled engineering requirements and achieved geometric and magnetic-field metrics. It spans multiple coil counts and field-period numbers, supporting studies of both specialized and shared models across configurations. The benchmark asks models to predict coils conditioned on a target equilibrium and requirements on coil-coil and coil-plasma spacing, curvature, and magnetic-field error. We show that magnetic-field accuracy remains the main challenge, with learned models rarely satisfying all requirements simultaneously. However, using learned predictions to initialize conventional coil optimization can accelerate early progress. We release baseline models and evaluation code to support reproducible research on both direct coil generation and learned optimization initializations. By lowering the barrier for machine learning and optimization researchers to engage with stellarator coil design, we aim to accelerate progress toward practical, cost-effective stellarator power plants.

Prompt-based image generation makes it easy to produce visual assets, but offers little support for engaging with the creative process. Prompting collapses interdependent creative decisions into a monolithic instruction and returns a finished image, leaving control limited to revising the prompt or correcting the model's output. We present reComposer, a process-oriented image-generation interface built around editable planning, which decomposes users' evolving intent into meaningful creative decisions - such as subject, composition, or palette - and collaborates with users to explore underspecified goals and alternatives. reComposer supports conversational and direct manipulation of the plan, moving between concept-level intent and visual-level control. We evaluated reComposer with 8 design and non-design professionals. Compared to prompting, participants felt greater control and ownership while developing visual assets, exploring alternatives, and iterating on details. Our findings suggest how shifting generative AI interaction from specifying outputs to shaping process can support creative thinking alongside asset production.

Gastroenteropancreatic neuroendocrine tumors (GEP-NET) lack prognostic biomarkers, and computed tomography (CT)-based body composition radiomics at the third lumbar vertebra (L3) is a promising candidate. Small cohorts make high-dimensional radiomic models prone to overfitting and single-split evaluations variable and optimistic. This methodological study proposes a four-step validation protocol for radiomic survival analysis in small cohorts: (1) false discovery rate (FDR)-corrected feature selection, (2) a dual-path comparison of prior feature selection versus embedded regularization, (3) repeated nested cross-validation (CV) and (4) a negative control with Gaussian noise features. In 78 GEP-NET patients, 535 radiomic features from five L3 compartments were used to evaluate five survival models with Harrell's concordance index (C-index), time-dependent area under the curve (AUC) and integrated Brier score (IBS). Step 1 selected at least one feature in 18% of outer folds under noise (nominal level 20%) versus 84% with radiomic features. The dual-path comparison showed no consistent difference between strategies. Repeated nested CV reduced variability across seeds (41% median reduction in C-index standard deviation), but not enough to distinguish models. The nested CV C-index exceeded chance by 0.07-0.15 with radiomic features but not clearly under noise (0.40-0.42 with prior feature selection, 0.47-0.55 with embedded regularization). Frequently selected features tentatively linked greater subcutaneous fat, skeletal muscle compactness and texture to better prognosis. These findings suggest a modest prognostic signal in L3 radiomics, with considerable uncertainty that the protocol quantifies. The protocol may transfer to other small-cohort radiomic survival studies. Code: https://github.com/marielatalaveram/radiomics-survival-validation-protocol.

A world model predicts how a scene changes after an action, and a planner uses it to search for actions that reach a goal. If the camera moves or the controller changes, the same command moves the robot differently: when its commands are turned by 90 degrees, a frozen planner with the LeWM world model that succeeds in 77.5% of episodes succeeds in only 7.8%. Current fixes retrain the model, need training data, or forget a set-up when it returns. We propose Library of Hands (LoH), a training-free method that keeps the world model frozen and corrects only the map from commands to applied actions, which we call the hand. LoH reads the hand through the frozen model, with no extra sensor: it finds the action that best explains the change between two camera images (a reading). It stores every hand it identifies, so a returning set-up is recognised rather than re-learnt. In closed-loop tests, LoH more than doubles success on three simulated tasks, from 28.6% to 62.3%, and loses only 0.6 percentage points when nothing changes. It also works, unchanged, with two other world models and on two of three models pretrained on real robot data, and its gain grows with the quality of its readings.

Vision-language-action (VLA) models are trained as general-purpose foundation models for robot learning, and require fine-tuning for each downstream task and robot. This raises practical questions such as how much data and training updates are needed, how checkpoints should be selected, and whether simulation can be leveraged to alleviate the cost of real data collection. We study these questions on a task that requires both precise manipulation and spatial understanding of the scene. In this task, an SO-101 arm moves a chess king between two squares that are given only by their names, such as e4 and a5. We fine-tune $π_{0.5}$ with various dataset sizes and update budgets, using teleoperated demonstrations as well as simulated demonstrations from a randomized scripted expert that imitates the variability of human demonstrations. In simulation, the success rate scales with both data and updates. Importantly, we find that a short learning-rate cool-down adds up to 31 points and is necessary for comparing checkpoints, and that validation loss and action error are unreliable for model selection. On the real arm, we fine-tune on up to 256 teleoperated demonstrations and evaluate ten policies in 25 trials each. Partial results indicate that pretraining on the simulated demonstrations helps substantially, most clearly when real data is scarce. We also present a negative result, in which training on the policy's own rollouts with filtered behavior cloning, RECAP-style advantage conditioning, corrections from an automatic coach, or residual reinforcement learning did not yield a measurable improvement over plain behavior cloning in our setting. We release the simulator, the expert and the real demonstrations.

Interpretability tools are usually built for one model at a time: a refusal direction, a truthfulness probe, or a steering vector is found with labeled data inside the model it explains. We ask whether such tools can be carried to other models through a map between residual streams fit only on unlabeled prompts. We transfer a refusal direction, tested causally by addition and ablation, and two read-out tools, a correctness direction and a sentiment probe, across 110 pairs of models from five families and 1B to 72B parameters and between base and instruction-tuned models, with controls for specificity and five seeds. How much transfers depends on what the map was fit on. A map fit on ordinary questions recovers 54\% of the target's own refusal steering and 16\% to 77\% of its correctness gain; adding unlabeled prompts from the region where the tool operates raises these to 102\% and 88\% to 98\%, while control directions recover at most 1\%. We prove that a linear map behaves this way whenever two models correspond linearly on the data it covers, and a label-free coverage score derived from this result tracks recovery across more than 16,000 runs and explains earlier reports that maps between models fail across domains. Carried read-out tools outperform anything the target's unlabeled representations yield. Interpretability results, once found, can be reused across models at the cost of a few unlabeled prompts.

This paper introduces a geometric approach to the representation and learning of time-variant parametrizations of deep assignment flows, for binary data labelings on graphs. Each edge carries an optimally controlled parameter flow in a 2-dimensional totally geodesic submanifold, based on objectives derived from a non-local linear scale space representation of the input data. The aggregation of triplets of such flows in 5-dimensional totally geodesic submanifolds is learned from data. The fusion of these flows determines the trajectory in the high-dimensional symmetric parameter space and by geometric integration the layers of binary deep assignment flows. Numerical results demonstrate the expressivity and the limitation of our highly parallel geometric local synthesis approach to high-dimensional deep assignment flow parametrizations.

Trade promotions consume a large share of consumer-goods revenue, yet most planners still set discounts with a single, often unreliable, price elasticity per item. We present an end-to-end Trade Promotion Optimisation (TPO) system built as one chain: analogue SKUs, point elasticity, elasticity curve, price recommendation, promotional lift and net lift. Elasticity is measured at post-promo time from directional price-dip events against a rolling regular price, and at promo time by a causal estimator that residualises trend, seasonality and promotion, handles stock-out censoring with a Tobit likelihood and updates itself from the first clean promotion week. Items with weak evidence borrow the point elasticity and the curvature from analogue SKUs chosen by a cross-validated ridge model over category, brand, sub-category, price line and description. The optimiser separates self-price lift from substitution and post-promo dips and adjusts for inventory and trade funding. In a four-week promotion experiment on a 39-SKU consumer brand (mean elasticity -3.06), the plan lifts units by 20.6% and revenue by 6.9% at flat margin. Analogue pooling halves elasticity error relative to hierarchy averaging.

Routing a melanoma case to routine follow-up is a serious action error that diagnostic coverage alone does not rule out. We present Harm-Constrained Risk-Optimal Conformal Prediction (HC-ROCP), a training-free framework that uses frozen classifier probabilities and adds a harm penalty when choosing prediction sets and actions. We test the resulting policies on disjoint calibration splits with multiplicity adjustment. HC-ROCP certifies supported claims, reports descriptive fallback policies, and refuses unsupported subgroup claims. We evaluate these outputs retrospectively on COVID-19 radiographs, ISIC 2019 dermoscopy, and EyePACS-to-APTOS retinopathy transfer. Across ten calibration seeds, HC-ROCP reduces COVID-19 severe-action errors from the exact risk-optimal baseline's 4.68\% to 0.54\%. At a 5\% budget, global claims are certified for COVID-19 and ISIC, but the rare squamous-cell-carcinoma claim is refused. At 1\%, all settings receive descriptive reports. For each budget, we also report the minimum severe-case count needed to pass the global harm gate with zero errors, providing a starting point for further data collection.

Software architecture is one of the skills novice developers must master to succeed in industry, and as AI disrupts software development, it is also transforming how students learn architecture. Existing work on how students and novices work with AI has mostly focused on small programming exercises or on students' perceptions of AI, so little is known about how students experience AI as they learn software architecture in larger team projects. We report on an exploratory case study of 43 senior software engineering students who used AI throughout a semester-long, team-based software architecture project. We combined surveys with focus groups and analyzed the focus group transcripts and the open survey responses through Braun and Clarke's reflexive thematic analysis. We constructed two themes. The first, Surrendering Judgment to AI, traces how participants' reliance on AI deepened as they used it to architect and implement their systems, until they surrendered their judgment and became stuck in a codebase they had created but could not understand, which made them rely on AI even more. Participants described feeling ashamed, frustrated, and emotionally disconnected from their work, a toll that, in our reading, left them with even less reason to break their reliance. The second, Applying Human Judgment, centers on participants who kept or reclaimed their judgment: the knowledge and experience it rested on, and the strategies they used to apply it. Participants' accounts suggest that relaxed deadlines and a lecture and lab on working with AI gave them room to acquire knowledge and reclaim judgment. Given the expectations placed on junior developers and the widespread adoption of AI, we believe these first-hand accounts can help software engineering education and industry adopt AI in ways that preserve novice developers' judgment, ownership, and motivation.

Reinforcement Learning (RL) algorithms often rely on data collected from a target environment, which can be expensive and noisy. Although simulators can generate large amounts of additional data, standard simulation mixing approaches, which directly combine simulated and real experience, can degrade performance due to simulator inaccuracies, especially when they bias the solution away from the optimal policy. We propose a conservative simulation-aided learning method based on control variates, where simulated data is used solely to reduce variance in the learning process rather than directly shaping the learning target. We develop our approach for Policy-Evaluation, which is a fundamental building block of RL algorithms. Then we combine it within the actor-critic framework towards policy improvement, as well as extend it towards offline RL scenarios. Experiments on multiple MuJoCo benchmark tasks demonstrate improvements over standard simulation mixing approaches, highlighting control variates as an effective tool for simulation-aided reinforcement learning.

AI is used in much of the clinical trial pipeline. However, selecting the statistical design of a trial, a crucial step that determines how drug doses are tested on humans and evaluated, has seen almost no AI progress. LLMs appear to be a natural fit given their strengths in biomedical knowledge and statistical reasoning. Nevertheless, there is currently no appropriate way to evaluate such abilities because recent regulatory and methodological shifts, most notably FDA's Project Optimus, have redefined what constitutes a reasonable design and rendered many previous designs obsolete. To lay the groundwork for AI-assisted trial design, we introduce TrialDesignBench, a benchmark of 617 Phase-I/I-II oncology dose-finding trials, each paired with an expert-validated reference design that is consistent with the latest FDA regulatory initiatives. This reference standard is produced through a formalization of the expert design process. We evaluate five frontier LLMs, none of which matches the reference standard. Our ablation experiments trace this performance gap to two main causes, inference of latent, unrecorded design determinants from trial sources (supplying them raises class-wise balanced accuracy by 0.22) and structured decision-tree reasoning (+0.24). TrialDesignBench offers a reproducible, contemporary-standard objective for evaluating and enhancing AI-assisted trial design.

Urban traffic flow prediction is a cornerstone of intelligent transportation systems, yet accurate forecasting under adverse weather conditions remains a significant challenge. Existing data-driven models, while effective for routine patterns, often struggle with the sparse data and complex non-linear shifts associated with rare meteorological events. These methods typically treat environmental contexts as shallow numerical features, failing to capture the underlying semantic reasoning-why and how specific adverse conditions alter mobility behaviors across regions with different functional compositions. To address this limitation, we propose Multi-STG, a unified framework that integrates Large Language Model (LLM)-derived semantic priors with graph-based spatiotemporal learning. Multi-STG introduces an LLM-empowered semantic enhancement module that transforms comprehensive context descriptions such as weather, POIs, and holidays into high-level semantic embeddings, effectively providing a knowledge-driven complement to data scarcity. Furthermore, we design a multimodal attentional fusion mechanism to dynamically weigh heterogeneous inputs and a semantic-enhanced prediction head with gated residual calibration to robustly adjust forecasts based on semantic guidance. Extensive experiments on real-world multi-city private car datasets demonstrate that Multi-STG outperforms state-of-the-art baselines, reducing MAE by 11.82\% on average relative to the strongest baseline across three cities, while maintaining stable performance under sparse and non-stationary conditions.

Serving systems choose a weight format for each model and deployment. A low-cost screen predicts the resulting KL divergence from the unquantized model's Fisher information using $\tfrac12Δ^\top FΔ$, the expansion underlying Hessian-aware and second-order post-training quantizers. We assess the reliability of this screen in GPU batches with channel-wise int8, per-tensor fp8 and group-wise int4 on Qwen3-8B, Qwen3-4B, OLMo-2-7B, Qwen3-32B and Qwen3-30B-A3B. Relative predictions hold across a 16-fold range of perturbation scales, int4 group sizes and rounding modes, while absolute predictions are heavy-tailed. A trimmed average brings 10 of 12 cells on Qwen3-4B and OLMo-2 within 15% of measured KL. The MoE model exhibits nonquadratic response at small perturbations, with int8 KL scaling as $α^{0.63}$. Computing per-sequence gradients in bf16 can cause fourfold prediction errors. We trace these errors to arithmetic in the lower half of the backward pass and eliminate them by using fp32 there. Batch-invariant kernels run at 0.55 times the throughput, while changes in reduction order introduce a floor of 0.17-0.22 nats per sequence. We build a screener with one streamed fp32 Fisher pass, trimmed tensor-block predictions with a band, a flag for aligned tensors and per-layer format allocation. On the unseen Mistral-7B-v0.3 family, all six predictions fall within the band, all three KL budgets are met, and vLLM throughput is predicted within 0.4%.

Romanized Nepali, Nepali written in Latin script, is common in digital communication but has no standard spelling and few resources, and it is unclear how reliably small open language models can be adapted to it. We adapt three 2-3B base models from different families, Llama-3.2-3B, Qwen2.5-3B, and Gemma-2-2B, with QLoRA on 9,000 instruction examples produced by deterministic transliteration of Nepali instruction data. Each model is trained with one shared recipe and three seeds and evaluated on 1,000 held-out examples with reference-overlap metrics, byte-normalized reference likelihood (bits per byte, BPB), and an output audit, against zero-shot baselines under two prompts. Fine-tuning raises chrF++ from 10-14 to 29.0-30.6 and lowers BPB from 2.23-2.41 to 0.78-0.84; most zero-shot outputs (54-92%) are English-like. Likelihood and overlap disagree on how far apart the models are. Llama-3.2-3B has the lowest BPB in every seed, 0.042 below Qwen2.5-3B (95% CI 0.039-0.046), yet the two cannot be separated on chrF++ (30.51 and 30.61) or BLEU, even with paired item-level tests. Gemma-2-2B, whose tokenizer splits Romanized Nepali words least (2.72 tokens per word), scores worst on every measure. Which model gains most over its zero-shot baseline depends on the zero-shot prompt. In a supplementary comparison with an identical tokenizer, Llama-3.1-8B lowers BPB by 0.053 (0.050-0.056) relative to Llama-3.2-3B, while chrF++ and BLEU rise by only 0.68 and 0.58 points. Lower fertility and higher reference likelihood therefore did not identify the model whose outputs were closest to the references.

Neural avalanches and turbulence-like brain dynamics are usually measured after neural activity has been coarse-grained into continuous fields, leaving open which spike-level mechanisms can generate these macroscopic observables. We build a spatial leaky integrate-and-fire spiking neural network and explicitly transform spikes into activity-density and Hilbert phase fields. A delayed, distance-dependent, passive-dendrite regime reproducibly generates turbulence-like field structure, including high-amplitude phase-defect tracks, structure-function scaling, and power-law-favored continuous-field avalanche tails. Controls show that firing rate, vortex count, and avalanche tail fits are individually insufficient: destroying delayed propagation, dendritic filtering, or spike timing suppresses key field signatures, while random-space nulls can retain many mathematical phase defects but show weaker spatial structure and synaptic-delay transfer. Biologically motivated gates further shape the regime: PV-like perisomatic inhibition clamps phase dynamics, SST-like dendritic inhibition and projection-specific inhibitory facilitation preserve long-lived tracks, and a rate-constrained projection-STP regime expresses a roughly 200-ms episode at 12.25 Hz (full-record mean) across 12 seeds. These results use SNNs as a mechanistic testbed for spike-derived brain turbulence observables and show why phase defects and avalanche-like statistics must be interpreted jointly with spatial scaling and spike-level propagation.

Edge computing places computation closer to users and data sources, but introduces heterogeneous hardware, limited resources, and variable network conditions. WebAssembly (Wasm) is increasingly considered for these environments because modules are compact, portable, and quick to start. However, cold-start latency remains a challenge: when a worker node does not yet hold the required artifact, it must fetch it before execution, adding transfer overhead that directly increases the end-to-end (E2E) latency, especially for large artifacts or slow network links. At the same time, the geographical distance between the requester and the assigned worker independently contributes to latency through network routing overhead. Together, artifact locality and network proximity do not always point to the same placement decision, and naive scheduling strategies fail to account for either. This work extends a Wasm-native edge orchestrator with state-aware scheduling strategies that jointly optimize for both objectives. Evaluation on a Raspberry Pi testbed with emulated regional network latency shows that a hybrid scheduler combining cache-state and network-proximity signals reduces mean E2E latency by up to 40.8% compared to a default round-robin baseline. Artifact locality-aware placement is most effective when artifacts are large or repeatedly requested, while network-aware placement dominates when artifacts are small and routing latency has a larger relative effect, but combining both objectives consistently yields the lowest E2E latency across diverse edge workloads.

We show that type-logical grammars with the standard rules for product, associativity, extraction and infixation have an NP-complete fixed recognition problem, adapting a construction for first-order linear logic from Pshenitsyn (2025), who already showed that grammars in first-order linear logic are closed under intersection and therefore not semilinear and not mildly context-sensitive. We strengthen these results and shown that grammars in first-order linear logic can simulate the quasi-linear grammars of Book \& Greibach (1970), the grammars recognised in non-deterministic linear time (NSPACE(n) or. NLIN). I argue that NLIN is an interesting class of formal languages as well, and a useful upper bound on the time complexity of formalisms used for natural language, just like LOGCFL is a useful upper bound on the space complexity. Taken together this means that first-order linear logic can be seen as the logic of non-deterministic linear-time string languages, graph languages and automata.

AI safety evaluations often involve sensitive prompts and proprietary models. Evaluators want to keep their evaluation prompts secret from model owners to prevent purposeful or incidental data leakage, ensuring longer usefulness of their benchmark. Model owners in turn wish to keep their proprietary model weights and inference code within their own secure infrastructure to protect their IP. We present a way forward for "double-blind evals", which provide cryptographic guarantees that an evaluator's prompts are not shared with the model owner, while simultaneously maintaining the confidentiality of the model weights and inference code. To accomplish this, we use secure enclaves that encrypt data in RAM at the hardware level. In a world-first pilot, we test this approach by evaluating a proprietary Google DeepMind model (Gemini 2.5 Flash Lite) against a selection of the private AILuminate benchmark corpus, using an NVIDIA H100 secure enclave on Google Cloud Platform, and using OpenMined's PySyft to maintain data and model privacy. We believe this effort marks the first use of this technology to provide mutual secrecy for one organization's frontier-class proprietary model weights and another organization's closed benchmark.

The traveling thief problem (TTP) asks a thief to choose a tour and a packing together, and the benchmark of 9,720 instances of Polyakovskiy et al. (2014) is the standard test set for comparing TTP heuristics. We show that four of these instances are not TTP instances at all. Their files give a renting rate of 0.00, so the tour has no effect on the objective and each instance reduces to a 0-1 knapsack problem. One of them belongs to the 60 categorized instances on which new methods are usually compared, and on it five of nine published methods report exactly the same value. The cause is the format of the files, which keep the renting rate with two decimals, while the rule of the generator gives a rate close to 0.003 on these instances. The same rounding may shift the rate of 949 other instances by more than 1%. Since the tour and the packing used by the generator were not published, the original rates cannot be recovered. We therefore define new rates with the same rule, applied to a fully specified reference solution that we publish, and we show with a published heuristic that the corrected instances behave as TTP instances again. We also give reference results on the corrected instances and recommendations for future studies. Finally, we present ttplib, a new open-source Python library that loads any of the 9,720 instances, including the corrected ones, without downloading the whole benchmark.

Large language models (LLMs) are increasingly used to translate technical explainable AI (XAI) artefacts such as SHAP into explanatory narratives (XAINs) for lay users. In high-risk domains, these explanations are often the main resource through which individuals can understand and contest adverse decisions. Yet LLMs are known to reproduce social bias, even from proxy variables alone, and fairness in XAINs remains unstudied. Because XAINs are bound to a ground-truth artefact, fairness concerns not only what a narrative says and how, but also whether it is equally faithful across social groups. We introduce fairness in XAINs as a new field of research and propose a framework that generates narratives under three conditions (protected attribute disclosed, withheld, or counterfactually swapped), allowing five comparisons that pinpoint the cause of disparities. We apply it to adversely classified instances from four datasets in credit scoring, education and employment, generating narratives with five LLMs, and evaluate faithfulness, feature coverage and sentiment. We find significant disparities across all datasets, metric families and models, arising in two ways. Faithfulness disparities result from feature values: LLMs make plausible but false statements, triggered by values that are unequally distributed across social groups. Feature coverage and sentiment disparities stem from the protected attribute itself: once it is disclosed, LLMs give less explanation to, or frame more negatively the narratives of, groups that have historically been discriminated against or are stereotypically associated with worse outcomes. A more faithful model removes the faithfulness disparities, and withholding protected attributes mostly removes the others, but neither suffices on its own. Fairness in XAINs should therefore be audited separately from the predictive model and its XAI artefact.

Optimizing molecules under small evaluation budgets requires deciding both where in chemical space to search and which candidate within that region to evaluate. We introduce Trust Region Adaptive Chemical Exploration (TRACE), a framework for open-ended molecular Bayesian optimization (BO), where candidates are generated on demand rather than selected from a fixed library. TRACE combines an objective-free graph-perturbation mechanism with a fixed, chemistry-aware Gaussian process (GP) surrogate using fragprints, a representation combining molecular fingerprints and fragment descriptors. Rather than learning the search geometry itself, TRACE makes the scale of chemical exploration explicit through a Tanimoto-similarity threshold controlling chemical extent and a cardinality budget controlling candidate-search breadth. A simple success/failure controller expands or contracts both quantities during optimization, while the GP and acquisition function select which admissible molecules to evaluate. Across six challenging molecular optimization benchmarks under matched 1,000-evaluation budgets, TRACE achieves the strongest overall sample efficiency as measured by the best molecule found as evaluations accumulate. Crossed ablations on two objectives further show that adaptive region control contributes beyond the proposal mechanism, while its benefit depends on the proposal distribution. We further apply TRACE to the stability of iron phthalocyanine (FePc), a non-precious molecular electrocatalyst whose degradation can involve loss of the Fe center. Using a first-principles density functional theory (DFT) workflow designed for accurate energy evaluation, TRACE increases the computed thermodynamic resistance to Fe dissolution from 0.42 to 2.59 eV over 80 designed evaluations.

Reinforcement learning with verifiable rewards ordinarily spreads a single advantage over a whole trajectory, a granularity too coarse to isolate the few decisions on which success in a long, multi-turn interaction actually turns. Privileged self-distillation supplies denser supervision, but it leaves unanswered how a set of local teacher-student discrepancies should be converted into credit along a sequence. We introduce AgentOPSD, a critic-free scheme that assigns credit recursively, one interaction turn at a time. Inside a turn, the token-level log-probability gaps between a skill-conditioned self-teacher and the student are summed, and that sum acts as evidence in a recursive Bayesian update carried out in log-odds space; the resulting revisions reallocate the outcome-derived advantage across turns through a multiplier that is bounded and preserves its sign. With Qwen2.5-3B and Qwen2.5-7B as policy backbones on ALFWorld, the method improves consistently on GRPO and on self-distillation baselines, and attains 89.1% success at the 7B scale. Ablations attribute the gain to aggregating evidence at turn boundaries and to making the belief revision depend on history, rather than to the local gap by itself.

Likelihood-based methods are widely used to learn the parameters of partially observable Markov decision processes (POMDPs). While finite-sample guarantees for policy suboptimality exist for certain POMDP subclasses, analogous guarantees for parameter recovery remain less well understood. For any candidate POMDP parameter, we bound its parameter recovery error in terms of its likelihood gap to the ground truth, the structural conditioning of the observation and transition kernels, and the behavior policy used to collect data. The bound applies in both the undercomplete and overcomplete regimes, with the latter following naturally from the undercomplete analysis. Because the result applies to any sufficiently good candidate, it yields finite-sample guarantees for the maximum-likelihood estimator and any procedure producing a parameter with controlled likelihood gap. This is particularly useful for near-optimal learning in infinite-horizon POMDPs, where existing approaches largely rely on model-based planning with accurate transition and observation kernel estimates. The bound separates the contributions of the behavior policy and model, showing how data-collection choices, including action coverage, and their effects on mixing can accelerate parameter learning.

As vision-language models (VLMs) move toward interpreting medical images directly, their outputs increasingly include not just a label but a stated confidence and a written justification. Yet the field still predominantly evaluates VLMs using just accuracy. We evaluate the coherence between answer, confidence, and explanation across four VLMs on breast density and BI-RADS categorization, using two mammography datasets. We perform controlled perturbations of the question, multiple choice option orders, and the images and assess changes in the three outputs. Our results suggest confidence does not decrease even as the mammography image is changed. Shuffling the options also changes most VLMs' answers. Notably, there are cases where a model's own stated reasoning contradicts its final answer, a distinction accuracy alone cannot capture but one that governs whether an AI-generated answer can be trusted.

Many tasks compose transformations by a known rule: chained rotations in robotics and navigation. Learned models relearn it, compounding error; classical estimators hardcode it and miss hardware deviations. We impose the known structure and learn only the remainder: an exact non-commutative group law sits in the node of a self-contained layer, its edges learn local increments, and a zero-initialized correction psi learns unknown deviations of the law. A pre-training criterion predicts exact representability. On fixed-angle SU(2) composition, a 54-parameter dense edge is the most accurate dense model, at 27x fewer parameters than the strongest fixed-matrix baseline, and lies 0.05 deg above the Bayes floor. On real robot arms, an isolated 306-parameter psi reaches 1.12-1.14 mm versus 1.29-1.33 mm for Gauss-Newton at N >= 60, and the recipe transfers across three vendors and two protocols. The recipe matches classical fits where the data obey the law and loses where the imposed law is not the task's.

Video world models conditioned on an input video make it possible to explore real, recorded dynamic scenes, enabling re-shooting of footage, free-viewpoint replay, and real-to-sim environments for embodied agents. Such uses require the generated rollout to stay faithful to the recorded world, yet existing benchmarks start from a single image or a text prompt, and video-conditioned generation is evaluated mainly on perceptual quality and camera following. These evaluations do not directly assess whether exploration preserves subject identity, event progression, and scene consistency upon revisiting a location. We introduce WanderBench, a benchmark for exploring dynamic worlds from video. It comprises 745 single-shot source videos of real and AI-generated content in three difficulty tiers (3-25 s), paired with camera routes given as keyboard actions, and evaluates six dimensions with 24 human-validated sub-metrics. Benchmarking eight models, we find that as clips grow longer, models keep the static scene and visual quality but lose camera control and the source dynamics, and that none keeps revisited places consistent. We further propose WanderLand, a training-free framework that conditions each frame of a frozen model on the same-instant source and, on revisits, on its own earlier static content. WanderLand improves five models on every dimension, by up to 8.3 points in source-dynamics fidelity.

We propose SciMARS, a framework for recursive task-local adaptation of multi-agent harnesses in closed-loop scientific discovery. Multi-agent systems coordinate analysis, experimental design, and verification as scientific evidence evolves. However, two problems limit their effectiveness: (i) without shared memory, agents lose observations and repeat work; (ii) fixed role procedures, handoffs, and memory policies perpetuate mistakes despite retained evidence. We address these problems with a shared evidence graph and separate component editors. SciMARS adapts roles and coordination; SciMARS + Memory adds a third editor for how agents record, retrieve, and receive evidence. Candidate probes guide composition and selection of the next harness. We evaluate scientific discovery through experimental-data replay and a digital twin. SciMARS + Memory resolves 12/18 GPT-6-Astra and 2/18 GLM-5.3 attempts on Autonomous Science Bench, compared with 8/18 and 1/18 for the base multi-agent team. On ScienceAgentBench and AppWorld, it solves 14/38 and 11/20 tasks, compared with 11/38 and 5/20 for the base team. These results support combining shared memory with harness adaptation across scientific and interactive tasks.

We prove $F_{\mathrm{weighted}}(2,2,3)<11/6$ for minimum-Frobenius-norm least-squares regression with two input and two output coordinates, improving the upper bound $15/8$. More precisely, a universal $\varepsilon_0>0$ satisfies $\mathcal T_3\le(5/6-\varepsilon_0)E$ for every finite residual frame. The proof combines an exact scalar certificate, rigidity of equality in planar selection, and a compactness argument that controls disappearing weights. Together with the known lower bound, this gives $13/8\le F_{\mathrm{weighted}}(2,2,3)<11/6$. No numerical value of $\varepsilon_0$ is claimed.

Alternating minimization is a standard tool for optimization problems with two coupled variable blocks. Its cost is dominated by the inner subproblem solvers, which rarely have a closed-form solution. In this paper, we propose to learn an inexact alternating scheme in which neural networks approximate the subproblem solution at each iteration. For convex problems with quadratic couplings between the variable blocks that satisfy a quadratic growth condition, we derive a worst-case bound on the optimality gap that exposes the per-iteration error that each network is trained to minimize in a greedy fashion. Combining the bound with a PAC-Bayes argument, we obtain a fixed-horizon, high-probability bound on the expected optimality gap over the distribution of problem instances. We apply the learned scheme to robust matrix completion and low-dose CT reconstruction with dictionary regularization where we achieve a 5x-13x speedup over the inexact first-order baseline at matched accuracy. A brief end-to-end fine-tuning stage extends this up to 30x and outperforms other learned schemes. The code is available at https://github.com/paulhausner/learned-alternating-minimization.

Self-recurrence represents an important component for temporal processing in neural networks, both artificial and biological, where it appears naturally in the form of leaky integration. However, due to the complex temporal dependencies of the gradient, an efficient and effective method of online training is still lacking. Here we propose K-prop, a novel online learning method for leaky neural networks. It dramatically reduces the computational and memory cost of real time recurrent learning (RTRL) by propagating kernel gains. K-prop also outperforms existing RTRL approximations and achieves performance comparable to backpropagation through time in sequence learning, classification, and reinforcement learning tasks. We further analyze how its sole approximation affects gradient quality and identify conditions under which K-prop remains accurate and effective.

Prior beliefs about whether a virtual human (VH) is controlled by a human or a computer can shape how users interpret and respond to its behaviour. As emerging hybrid VH systems allow control to transfer between human operators and computer programs during an ongoing interaction, it remains unclear whether perceived agency updates with the actual controller or continues to reflect users' prior beliefs and first impressions. We investigated this question in a within-subject VR study with 24 participants across four conditions: always computer-controlled, computer-to-human, human-to-computer, and always human-controlled. We designed an interaction task in which participants acted as interviewers, asking the VHs a fixed set of questions. Before each interaction, participants were informed whether the VH was human- or computer-controlled, matching the actual controller at the beginning, but were unaware of possible control transitions. We found that perceived agency was strongly influenced by the controller presented at the beginning of the interaction, with no evidence of updating following the control transitions. Social evaluations including likeability and co-presence were generally higher in conditions where human control occurred later and covered a greater proportion of the interaction. The results suggest that prior beliefs and first impressions continued to shape perceived agency, whereas social evaluations were more sensitive to the VH's actual behaviour over the interaction.

This work proposes WaveGen, a fully end-to-end diffusion-based text-to-waveform framework based on Spectral Forcing, which jointly learns an internal spectral trajectory and the target waveform trajectory within a single model, enabling direct end-to-end text-to-waveform generation. Unlike recent TTS pipelines, WaveGen does not rely on separately trained or externally decoded intermediate representations, such as neural audio codecs, Mel-spectrogram vocoders, or VAE-based latent models. In addition, WaveGen performs internal text-speech alignment within a single model, eliminating the need for external alignment modules such as duration predictors. To preserve a truly end-to-end formulation, our framework further avoids dependence on self-supervised representations, such as BERT-based text embeddings and Wav2Vec 2.0-based semantic representations, as well as semantic distillation methods for training optimization. To achieve this, we carefully design Spectral Forcing, a model architecture with disentangled diffusion heads, and a training objective based on a multi-scale spectral $v$-loss. We further validate the scalability of end-to-end TTS models by scaling the model size from 0.1B to 0.6B parameters. Using a single architecture, open-source data, and one-stage training, without external modules such as duration predictor or REPA, WaveGen demonstrates fully end-to-end waveform generation achieving state-of-the-art performance on the Emilia training dataset. In particular, WaveGen achieves a WER of 1.47 and a SIM of 0.75 on the Seed-en benchmark. Source code and checkpoints are available at https://github.com/sh-lee-prml/WaveGen

"System 1" decision models (pioneered by TypeSafe AI's Jev) replace autoregressive text generation with a single forward pass returning typed probability distributions over declared schemas (Choice, Score, Noul). While deployed for low-latency routing and automated triage, their statistical calibration and post-hoc explainability remain uncharacterized. We introduce an open, deterministic audit benchmark evaluating System 1 decision models across calibration under label-preserving perturbations, cross-question interference, and post-hoc attribution faithfulness. Evaluating five open-weight models (0.8B to 27B) alongside generative System 2 baselines -- both single-token logprob scoring and structured JSON/CoT decoding -- across NLI (e-SNLI), adversarial syntactic shortcuts (HANS), synthetic clinical triage (N=900), prompt injection defense (Tensor Trust, N=250), and an 86-case logic-gate trigger suite yields five findings: (1) provider confidence is an exact closed-form dispersion transform (1 - H(p)/ln K on sub-10B models; normalized margin on Clef-27B) rather than empirical accuracy; (2) bundling multiple questions in one request induces directional cross-question interference (11.7% to 22.4% binary flips on Tev1-0.8B, vs. 0.1% to 0.2% on Tev1-4B); (3) System 1 models with >= 4B parameters are faster and better calibrated under syntactic shifts (ECE_w = 0.030 to 0.087) than logprob (0.144) and JSON (0.378) baselines; (4) for prompt-injection defense (N=250), >= 4B models achieve 0.988 to 0.998 AUROC with up to 95.2% recall and 0.0% false alarms at 48 to 60 ms; and (5) single-pass inference makes 512-coalition XAI interactive (1.3 to 3.2 s/input), while KernelSHAP with a logit link (S1-SHAP) mitigates softmax saturation on disjunctive OR gates (89.3% to 100.0% gate recovery vs. 57.1% to 82.1% in prob space; p_adj = 0.004).

Accurate prediction of electronic Hamiltonians would enable broad property inference while avoiding the high computational cost of Density Functional Theory (DFT). However, progress toward general-purpose materials foundation models is limited by a data bottleneck: existing Hamiltonian datasets are typically small, lack structural diversity, and often omit essential relativistic physics such as spin--orbit coupling (SOC). We therefore construct UniHam, a large-scale Hamiltonian dataset and benchmark suite comprising 100,000+ DFT-computed complex-valued Hermitian Hamiltonians with full SOC, covering 72 elements and a wide range of crystal geometries and symmetries (spanning diverse lattice types and space-group families). Building on UniHam, we benchmark two representative state-of-the-art models under a standardized protocol and introduce complementary evaluation metrics that jointly assess three dimensions: (i) Hamiltonian reconstruction accuracy, (ii) out-of-distribution (OOD) generalization across composition/symmetry shifts, and (iii) the ability to support downstream property prediction from the predicted Hamiltonians. Experiments on UniHam demonstrate that the proposed benchmark and metrics effectively differentiate model capabilities, revealing intrinsic SOC- and element-dependent failure modes, large variations in compositional OOD robustness, and the necessity of spectral-level evaluation to assess whether Hamiltonian predictions reliably support downstream electronic-structure properties. Overall, UniHam provides a reproducible, SOC-complete benchmark that can sharpen model comparisons and accelerate the development of next-generation foundation models for quantum materials.

Generalizing deepfake detection across facial manipulation and fully generated video remains challenging due to forgery traces varying with the manipulation method and scene content. Explicit spatial and temporal feature differences offer a way to capture this variation, but also respond to natural texture and motion, motivating their joint interpretation with appearance features. We investigate whether combining multiscale feature differences with adapted appearance representations improves detection across both forgery categories. We introduce Multiscale Appearance-Residual Streams (MARS), a two-path video detector. The residual path applies fixed spatial and temporal difference-of-Gaussians operators to frozen intermediate features and models the resulting differences alongside appearance context across frames. The semantic path adapts the encoder's normalization parameters for frame classification, and a clip-dependent gate combines their predictions. Under a shared protocol across seven datasets, MARS achieves the highest average AUROC of 91.43\%, outperforming seven compared detectors. When trained on facial manipulations alone, MARS achieves 93.99\% AUROC on GenVideo and 75.50\% on DF26, demonstrating transfer to generated videos without generated-video training data. MARS also provides signed patch and frame contributions, whose faithfulness we evaluate through controlled spatial and temporal removal against random controls. MARS is publicly available for research: Project page: https://vectorinstitute.github.io/mars/ Github: https://github.com/VectorInstitute/mars

Does grouping translation updates before speech synthesis change the measured completion gap between two translators? We compare StreamAtt with its fixed-audio-history baseline on ten English-German talks. For each translation, we keep the intended text and its availability times fixed and vary how updates are grouped for synthesis. Four-second grouping increases StreamAtt's mean completion lead from 106.6 to 229.1 seconds with XTTS-v2, but reduces it from 169.5 to 146.5 seconds with Piper. Both produce shorter audio; different changes in playback idle account for the opposite effects. These directions recur on a separately selected twenty-talk panel. In continuing CosyVoice 3 sessions, grouping instead leaves paired audio unchanged and delays completion on nineteen of twenty blocks. A speech-timing comparison therefore needs to specify how text enters synthesis, not just which synthesizer is used. Recognition and automatic word timing describe the outputs, but do not establish translation accuracy or intelligibility.

Bayesian optimization is widely used in active learning loops for materials science. Gaussian processes serve as the de-facto surrogate model family, but require task-dependent kernel choices and can lack the modeling capacity for complex relationships. Tabular foundation models (TFMs) have recently emerged as the state-of-the-art for predictive tasks on structured data, without requiring bespoke kernels or hyperparameter tuning. We evaluate two TFMs, TabPFN and TabICL, as out-of-the-box surrogate models on two standard materials science benchmarks, comparing against a large field of baselines encompassing Gaussian processes, random forests, and CatBoost. Our findings indicate that TFMs are highly competitive surrogate models across the two benchmarks, without task-specific training for active learning or GP kernel selection. Our results motivate developing TFMs further as surrogates for active learning in materials discovery.

AI-text detectors reach near-perfect scores in evaluations that contrast fully human-written texts with texts generated from researcher-designed prompts. In authentic student writing, however, AI use is voluntary and often partial. We study this setting with the PISA 2025 FLA dataset, in which 15-year-olds in Germany wrote English argumentative essays with and without access to a gpt-4o-mini-based chatbot. Keystroke, clipboard and chat logs let us trace every character of the final essay to its origin. From this provenance, we derive text- and sentence-level labels for human-written, AI-generated and mixed text, and for text written after chatbot use without traceable AI material. We evaluate a feature-based detector and a classifier on frozen Llama-3.1-8B embeddings, both adapted from top-ranked shared-task systems. Both separate AI from human text reliably (macro-F1 up to 0.96 for texts and 0.95 for sentences), independently of spelling errors. The feature-based detector, however, relies on lexical complexity, which a simple student-style rewriting prompt removes (AI recall 0.92 $\to$ 0.60 with gpt-4o-mini, 0.41 with Qwen3). The embedding-based detector stays robust for whole texts (AI recall $\geq$ 0.94) but not for single sentences (AI recall as low as 0.59). Mixed text falls between the two classes, and text without traceable AI material is classified as human-written in about 95\% of texts.

Answer accuracy alone does not distinguish failures in acoustic perception, rule knowledge, and rule application. AuDiR (Audio Diagnostic Reasoning) examines these distinctions through controlled input interventions. Each item contains an audio clip, a multiple-choice question, a task-general rule, and a textual description of the acoustic observations needed to answer. Three conditions progressively supply the rule and observations; a fourth removes the audio while retaining the question and rule. The first correct diagnostic condition assigns an operational response label, without identifying a unique internal cause. The current benchmark contains 201 items in 18 families drawn from 9 sources, with answer references documented separately from percept extraction. In a preliminary evaluation of Gemini-3.8-flash, accuracy is 29.3% without supplied information, 31.3% with the rule, and 98.5% with the observations also supplied. No-audio accuracy is 23.9%, with a confidence interval containing the 25.0% chance rate. A family-level knowledge probe associates incomplete rule recall with larger gains from rule provision. These results show the benefit of supplying acoustic observations as text, while incomplete validation, possible recognition shortcuts, and non-monotonic responses limit stronger failure-attribution claims.

We investigate convolutional-neural-network (CNN) based automatic solver selection for sparse linear systems, a problem of practical importance where solver performance varies dramatically with matrix sparsity structure. A CNN is trained to regress log timing ratios between candidate solvers using spy plot images of matrix nonzero structure as input, achieving strong prediction accuracy on both held-out SuiteSparse matrices and real-world matrices from high-performance simulation codes.

Foundation models can supply predictions for a downstream task that draw on knowledge from extensive pretraining, but they differ in task-specific accuracy and query cost. We develop a framework that improves prediction for the task under a budget on the average query cost per input, using training data to learn which models to query and how to combine their outputs with features available without querying. The first method queries the same models for every input; the second lets them vary across inputs and uses Q-learning to decide which model to query next and when to stop. For the first method, we show that several individually inaccurate outputs can jointly outperform the best single one. It attains the minimax rate when the regression function depends on the outputs only through models within the budget. Otherwise, any method within the budget can be forced to a slower rate, which the first method attains. For the second method, we show when varying the queried models across inputs strictly improves on querying the same models for every input, and bound its prediction risk and average cost in finite samples. In simulations and a multimodal laptop price application, both methods outperform existing methods.

Speculative decoders raise throughput by banking more target tokens per forward pass, and Medusa, EAGLE-3, DFlash and DSpark all do it the same way: more candidates that survive verification. All of them draft in the target's own vocabulary, so one node of the candidate tree is one target token, and the tokenizer makes that unit look like a constant. It is not. BPE's pretokenizer splits on whitespace before merges are learned, so no multi-word string can ever become a single token: in Qwen3-4B, "United States of America" is four tokens and no adjacent pair of them exists as one entry, so a phrase is out of reach of any method drafting in the target's vocabulary. We change the unit. A small MLP head reads the frozen target's hidden state and predicts a one- to three-word phrase as one entry in a separate draft vocabulary; greedy longest-match expansion maps it back into the target's own tokens, and N candidates are verified in one forward pass behind a prefix-tree mask, so the change is orthogonal to tree depth and width. The target and its tokenizer stay frozen and a candidate is accepted only while it matches the target's own argmax, so the served output is identical to greedy decoding up to floating-point tie-breaking. On one A5000 at batch 1, against HuggingFace generate(), a 21.94M head (0.545% of Qwen3-4B) reaches 2.391x on GSM8K, 2.444x on MATH-500 and 1.544x on MT-Bench. Because the head's output layer is one row of 257 parameters per entry, the draft vocabulary size is the parameter budget: an 8,000-entry head is 3.434M parameters, 0.0854% of the target and 64x smaller than EAGLE-3's drafter, and keeps 92.3% of the larger head's GSM8K tokens per pass. With EAGLE-3 run on same prompts and baseline, our 21.94M head is ahead of tree-16 on both maths sets and behind tree-60 on every set.

Causal discovery from observational data is central to scientific reasoning, yet established constraint-based and score-based methods often recover only a Markov equivalence class and provide no signal in bivariate or strongly autocorrelated regimes typical in some areas of science. Here, we recast causal discovery as a Bayesian symbolic regression problem. Given a set of candidate parents for each target variable, the Bayesian symbolic regression samples closed-form expressions from a posterior over symbolic models. The fraction of sampled expressions in which each candidate variable appears defines a posterior inclusion probability that we use as the causal score. This replaces pointwise model selection by an ensemble of plausible mechanisms. We evaluate the method on coupled nonlinear bivariate dynamics, on the CauseMe benchmark suite spanning linear, chaotic, nonlinear, higher-dimensional and high-noise settings, and on a four-variable biogeoscience structural causal model under Gaussian, heavy-tailed and mixed-noise regimes. Our approach is competitive with, and often outperforms, specialized causal discovery algorithms. Additionally, the method returns closed-form mechanisms leading to explicit causality and interpretability.

Symbolic knowledge (motif counts, ring counts, label constraints) is routinely added to graph neural networks (GNNs) and justified by their limited expressivity. We replace this yardstick with an exact one. The largest log-loss reduction a rule $S$ can give a model family is $I(Y;S\midσ)$ given what the family computes; every realised gain splits exactly into this term plus repaid neural debt; and for a trained network with representation $T$ the value is $I(Y;S\mid T)$, the label information training failed to extract. Rule design thus becomes rule discovery: we score candidates on the network's out-of-fold residual against calibrated permutation nulls with family-wise error control. On ZINC, discovery recovers the hidden ring term $\max(0,\ell_{\max}-6)$ of the target in three independent runs ($p_{\mathrm{adj}}=0.001$) and beats expert cycle counts; frozen and applied to five unseen datasets it is never significantly worse than the plain network or the expert rule, and on BBBP it significantly beats both.

Background: Freight transport is expected to increase, contributing to air pollution. There is, therefore, a critical need for increased efficiency that can reduce emissions. Data sharing is one approach, where actors within Open Data Ecosystems (ODE) collaborate and share data among stakeholders to plan and operate more efficient and sustainable freight transport. Aim: In this study, we explore how public-private data sharing practices in municipal and regional last-mile freight logistics in Sweden may improve efficiency. Method: Following a prolonged participation process, we conducted a multiple-case study involving three different cases, built upon initial meetings, semi-structured interviews, supported by a thematic analysis, and validated through a follow-up workshop. Results: The findings show that data sharing in public-private freight logistics (ODE-PPF) is shaped by interrelated socio-technical aspects, including platform design, roles of actors, handling of sensitive data, data quality, and procurement processes. We synthesize our findings in an actor model and a conceptual model, highlighting how these aspects contribute to the functioning of ODE-PPF, and derive seven proposition sketches as a foundation for future research. Conclusion: Data sharing is a central mechanism enabling integration and value creation in last-mile logistics. It depends on the interaction between technical and organizational aspects. Public actors can undertake a key role as a neutral keystone actor by orchestrating the ecosystem and facilitating data sharing through governance practices such as procurement.

Probabilistic circuits (PCs) have emerged as generative models guaranteeing tractable probabilistic inference. However, tractability comes at the price of larger model sizes than in other generative models. As such, to scale PCs to high-dimensional data such as images, previous work has devised several parameter sharing techniques as well as fast evaluation protocols that consider learning PCs on independent image patches. In this work, we revisit this practice to highlight a surprising learnability issue: models that assume independence counterintuitively can yield better likelihoods than models that do not. To break this apparent bottleneck and to scale PCs further, we introduce a recursive parameter sharing construction (RePA) that fosters efficiency without assuming independence. We show empirically that RePA can greatly reduce time and memory consumption while yielding competitive likelihoods. Inspired by these results, we further generalize the approach to data modalities that do not inherently exhibit patch-structure (NoPA), such as tabular data, showing similar efficiency improvements, with marginal changes in likelihoods and sample quality on seven tabular datasets. Code available at https://github.com/april-tools/RePA.

Non-linear functions are an important class of operations in modern large language models. However, chip vendors building for AI have so far relied on traditional vector or scalar processing units for non-linear functions. The performance of these units has not kept pace with matrix multiplication units (MMUs), which can be found in both high-performance and edge-computing devices from almost every vendor. Reconfigurable units such as FPGAs and CGRAs have been proposed in the literature, but have not been systematically compared with traditional units backed by MMUs. In this paper, we build two SoCs: one using an RVV vector processor and the other using a latency-insensitive (\emph{elastic}) CGRA, both coupled with an MMU and driven by a single-core CPU. We then propose mappings for common feed-forward networks on our SoCs and find that the CGRA reaches a maximum speedup in clock cycles of 7.12x over RVV on the activation stage of simple point-wise non-linear functions like SiLU, whereas the RVV reaches a maximum speedup of \textbf{9.8x} over the CGRA on complex non-linear functions such as $\mathrm{softmax}$. Over the enclosing feed-forward block the corresponding maxima are 3.16x and 4.24x.

Scientific research has long been driven by substantial human effort, from reviewing prior work and identifying research gaps to designing specific solutions. Recent advances in harness-based agent systems have given rise to increasingly capable Auto Researchers, which can autonomously carry out these tasks and, increasingly, entire end-to-end research workflows. Yet despite their rapidly growing capabilities, how closely the research decisions and outputs produced by auto researchers align with those of human experts remains underexplored. We investigate this gap through the lens of research reconstruction and study: Given comparable research context, how similar are the research decisions and outputs produced by Auto Researchers to those of human researchers in AI research? To answer this question, we introduce AutoResearcher-Human Research Similarity (AutoHRS), a stage-wise evaluation framework for assessing how faithfully auto researchers reconstruct human research artifacts across different stages of the scientific workflow. Using 152 ICLR oral papers as our corpus, we generate 456 research trajectories spanning different stages of the research process. Applying AutoHRS to these trajectories, we find that current Auto Researchers frequently fail to recover key research decisions and their resulting artifacts across multiple stages of the research process.

Multi-armed bandits are a natural framework for many real-world sequential decision problems, from recommender systems and clinical trials to educational test recommendation and online advertising. The performance of these algorithms depends on how they explore different arms, and exploration strategies range from variance-independent (e.g., UCB1) to variance-aware (e.g., UCB-V). In practice the reward variance of each arm is rarely known and varies widely across arms, and the reward support is often underspecified. In these cases no single rule is uniformly best: variance-independent methods are robust but loose when variances are small, while variance-aware methods are tight only when their variance estimates and the assumed support bound are reliable. Such heterogeneous, unknown variances arise in many settings, for instance when multiple objectives are scalarized into a single reward. We propose UCB-min, a per-arm adaptive meta-algorithm that selects, for each arm, the tighter of a variance-independent and a variance-aware exploration bonus, without prior knowledge of the variance regime and without reward normalization. We prove that, with UCB-V as variance-aware base, its regret is bounded by the sum of the two base bounds at the standard exploration constant, and by their per-arm minimum, up to a constant, for a larger one. We validate it on synthetic data, an educational tutoring task, and two group-recommendation benchmarks (MovieLens, Steam), showing that UCB-min consistently tracks the best base algorithm across diverse variance regimes and datasets.

Discrete autoregressive modeling, notably through next-token prediction, has become a cornerstone of modern language models. Extending this paradigm to vision allows visual and language models to converge on a unified framework, thereby fostering shared methodology and infrastructure. This work investigates a central question: how far can a finite discrete vocabulary go in representing the richness of the visual world? We present dNaViT, a Discrete Native-resolution Vision Transformer that produces discrete visual primitives analogous to language tokens. Built upon a frozen pretrained encoder, dNaViT applies lightweight residual modulation with a dual-branch objective to inject reconstruction-aware cues into semantic features. By leveraging grouped product quantization, it scales discrete capacity without inflating spatial sequence lengths, while a balanced training recipe maintains both semantic and pixel-level information. Consequently, dNaViT functions as a versatile discrete visual tokenizer, enabling unified encoding across downstream tasks alongside high-fidelity image decoding. Extensive evaluations show that dNaViT matches or exceeds task-specific baselines across 12 visual understanding benchmarks, 4 generation benchmarks, and an image editing benchmark. These findings demonstrate that a compact discrete vocabulary can serve as a universal visual currency and unlock the full potential of fully unified, native multimodal methodology.

Neural surrogates for engineering models are often trained on large collections of simulated solutions. We investigate whether physics-informed neural networks (PINNs) can achieve accurate predictions from scarce labelled observations by incorporating governing equations and prescribed initial and boundary conditions. We introduce hierarchical loss weighting (H-PINN) and its gradient-direction-aware extension (HO-PINN), which adjust the emphasis on physical constraints according to higher-priority losses. Hierarchical PINNs are compared with Vanilla PINN, tuned gradient boosting, support-vector regression, and Gaussian-process regression on a heat equation and three heterogeneous two-phase-flow benchmarks. Experiments use 20--1000 labelled states and ten paired seeds. Each label is one coordinate--state observation, not a complete simulation. At least one PINN variant achieves the highest mean primary-target \(R^2\) in each interpolation and equation-assisted forecasting comparison at 20 and 50 labels. The advantage for water saturation persists across all tested budgets. Hierarchical weighting reduces variability across runs in selected settings. These results demonstrate that known physics can support accurate neural surrogate predictions from very few solution observations, at substantially higher fitting cost than the tested classical regressors.

We introduce Rapid Simulation Based Inference (RSBI), a diffusion-based variational approach to likelihood-free Bayesian inference that achieves high sampling efficiency under large prior-to-posterior volumes. RSBI builds on advances in Schrödinger Bridge (SB) diffusion sampling to handle multi-modal posteriors, with an initial likelihood supplied by a distance measure via a differentiable simulator. The key innovation of this initial surrogate likelihood is that it gives a highly efficient proposal distribution, replacing sequential rounds of other methods with an efficient one-shot proposal. Subsequent refinement then transports the proposal to the true posterior as specified by the simulator and the original prior. The SB sampler is warm-started from the proposal, and its energy is corrected by a neural likelihood ratio that is learned only where the proposal places mass. As a variational method, RSBI does not suffer from prior leakage commonly observed in standard sequential posterior estimation methods. To improve mode coverage we optionally utilize the well-tempered meta-dynamics framework, which also encourages exploration of the prior volume. We observe strong performance on standard SBI benchmarks, particularly when the prior volume is increased relative to the ground truth posterior, achieving a performance degradation of only $\sim5\%$ on Two Moons under a highly constrained simulation budget. Additionally, we evaluate RSBI's performance for gravitational wave ring-down posterior estimation, in a real-world physics benchmark inspired by black hole spectroscopy.

Hierarchical multi-label text classification (HMTC) requires models to capture both structural dependencies among labels and the fine-grained semantics of labels at different hierarchy levels. Existing methods often model hierarchy through prediction constraints or representations of individual labels and local relations. Structural and semantic supervision, however, are usually formulated separately. We instead use complete root-to-node label paths as a shared unit for both. Our framework introduces two auxiliary training objectives while retaining a standard multi-label classifier at inference. PathSimNCE aligns document representations with hierarchical paths using hierarchy-aware contrastive targets. These targets provide graded supervision for structurally related paths while preserving explicit supervision for the ground-truth paths. The complementary semantic objective uses LLM-generated, document-conditioned descriptions of the same positive paths. It thereby provides context-specific label semantics beyond static label representations. Experiments on three benchmarks show consistent improvements: structural alignment is particularly beneficial in deeper hierarchies, while semantic alignment provides additional gains for fine-grained labels.

Given $n$ i.i.d. random matrices $A_i \in \mathbb{R}^{d \times d}$ with common expectation $Σ$, the goal of \emph{Differentially Private Stochastic PCA} is to identify a $k$-dimensional subspace capturing the leading variance directions of $Σ$, while preserving differential privacy~(DP) for each individual sample $A_i$. \citet{dungler2025iterative} introduced \kdppca, the first algorithm to simultaneously (1) achieve sample complexity $n=\widetilde O(d)$ for sub-Gaussian data, (2) adapt its privacy noise to the intrinsic randomness of the data, and (3) extend seamlessly to any target dimension $k\le d$. However, its sample complexity has a suboptimal dependence on $k$. We propose the first algorithm that achieves optimal sample complexity in both $d$ and $k$, while retaining properties (2) and (3) of \kdppca. Additionally, prior utility guarantees require a minimum sample size with exponential dependence on the eigengap. Our method removes this dependence. It also improves the dependence on spectral parameters, matching known lower bounds for the spiked covariance model. Unlike deflation-based approaches like \kdppca, our algorithm updates the full $d\times k$ subspace jointly rather than one eigenvector at a time. This non-deflation structure simplifies the algorithm, reduces the number of hyper-parameters, and improves computational efficiency, particularly when using block linear algebra libraries.

Coordination between large language-model agents has become a central question for AI safety and software engineering, as such agents are deployed in groups that share infrastructure. Existing research tells the agents that partners exist or asks them to coordinate. Whether coordination arises through shared state alone, with nothing disclosing that anyone else is present, is the key question for deployments that share caches and memory between agents, yet it has rarely been examined, and never with independent replication; this study does both. Four agents each received a short arithmetic task and a key-value store presented to each as its own scratch space; two held complementary halves of one computation, and no prompt mentioned other agents, sharing or collaboration. For reliability, the environment was rebuilt independently from a written protocol, and both implementations ran under hashed locks on design, data and analysis. The results bear novelty value in three respects. Joint success occurred only through the shared store. Whether a model coordinated depended not on its vendor tier but on whether it wrote its intermediate result and looked at the store again, and one ordinary sentence, "check your workspace again before submitting", switched that second look on and off. The findings held under replication. For practice, state shared between agents is a communication channel and should be separated, namespaced and logged. For research, designs should be locked before the data, and the parse rule, endpoint and parameters reported. Future work should test open-weight models and separate using another agent's output from intending to communicate.

Robots are increasingly expected to execute natural-language manipulation tasks in real-world environments. This is particularly relevant in household assistance and logistics, where clutter requires reasoning about both task objectives and execution feasibility. This challenge becomes even more pronounced for bimanual humanoid platforms, where coordinating two manipulators introduces additional interaction and kinematic constraints. We present a zero-shot vision-language framework for execution-aware bimanual manipulation that jointly reasons about task objectives and physical feasibility. Planning is grounded in a structured scene representation that combines semantic object relations with execution constraints, enabling the selection of obstacle relocations and execution choices while preserving task semantics. Compatible actions are organised into parallel bimanual stages with explicit pre- and post-conditions. A hierarchical validator evaluates these conditions under perceptual uncertainty and uses stage-level execution history to support local recovery or targeted replanning. Experiments on the CENTAURO hybrid wheeled-legged humanoid robot across nine language-conditioned manipulation scenarios show higher task success than a non-execution-aware baseline. In dedicated failure scenarios, hierarchical validation recovers from execution failures with fewer model queries than always replanning. These results demonstrate the benefits of combining execution reasoning directly into task planning. Videos and additional material are available on the Project Webpage: https://advrhumanoids.github.io/bridging-planning-execution/.

Despite growing attention to social bias in large language models (LLMs), existing bias benchmarks provide limited coverage of culturally grounded evaluation in Bengali, particularly within Bangladesh's sociocultural context. To address this gap, we develop a bilingual extension of BanglaBBQ comprising 61,312 question-answering instances across 12 social categories in English and Bengali. The benchmark incorporates balanced answer permutations, contrasting question framings, and both ambiguous and disambiguated contexts to systematically examine model performance and stereotype-aligned bias across languages. We validate the dataset through structural checks, 220 statistical tests, and expert annotation of 150 instances by four annotators, achieving strong inter-annotator agreement (Fleiss' \k{appa}=0.8349). We evaluate eight LLMs across 490,496 instances and identify five key findings. First, disambiguated contexts yield substantially greater accuracy improvements in Bengali (+36.59 percentage points) than in English (+17.71). Second, all evaluated models exhibit lower accuracy in Bengali, although Mistral-Nemo shows the smallest cross-lingual accuracy gap (5.12 percentage points). Third, stereotype-aligned bias exhibits language and model dependent variation, increasing in Bengali for Gemma 2 9B but decreasing for Mistral-Nemo. Fourth, higher accuracy does not necessarily correspond to lower stereotype-aligned bias. Finally, bias varies considerably across social categories, with Qwen 2.5 7B exhibiting stronger stereotype alignment in Bengali for Politics and Physical Appearance. These findings demonstrate that cross-lingual performance and social bias are distinct dimensions of LLM behavior and underscore the importance of culturally grounded, language-sensitive benchmarks for evaluating fairness beyond English-centric settings.

As LLM agents become more autonomous, they are increasingly expected to run as always-on assistants on a user's behalf, identifying work to be done and completing it without being asked. Yet existing benchmarks usually give the agent its task, whereas in knowledge work the requests sit in the user's workspace and the agent must first recognise which obligations exist. We study proactive task discovery and execution, in which an agent must infer a user's pending tasks from their workspace and complete them in the systems where the work lives. This setting poses two challenges. For the agent, the tasks are buried in a large volume of mail, chat, documents and records. For evaluation, a benchmark needs a simulated organisation whose people, events and documents are mutually consistent, so that tasks arise realistically and their correct outcomes are known. We introduce AutoBuddy-Bench, a simulator that addresses these challenges. It models an organisation's employees, projects, business systems and event history as a temporal graph, renders it into realistic workspace artefacts, derives ground-truth tasks from the events that create obligations, and computes each task's outcome from the world state; execution is judged by the agent's changes to the workspace. To evaluate current LLMs, we build AutoBuddy, a reference agent with interchangeable backbone LLMs: it identifies tasks from the workspace, grounding each in its source records, and executes each in a separate session with write access. Varying only the backbone, we find a systematic shift in where tasks are lost: weaker models fail mainly at discovery, intermediate models identify tasks but often stop short of carrying them out, and the strongest models fail mainly during execution, most often on business-system operations. Discovery precision is only weakly related to execution quality, suggesting the two are distinct capabilities.

In addition to the effects of explicit input prompts and instructions, large language model outputs are influenced by underlying behavioral preferences similar to human personality tendencies. These tendencies can affect how models interpret situations and make decisions. However, their influence is difficult to disentangle from the effects of model capabilities and instructions. Therefore, we introduce SPACE, a benchmark framework combining a controlled evaluation protocol with complementary datasets for personality verification, capability assessment, and hierarchical situational tasks. Controlled comparisons across guidance conditions help distinguish personality influence from measured capability changes and immediate instruction effects. Experimental results demonstrate that although prompting can shift generation style, models' internal personality tendencies are persistent and not readily overridden by opposing instructions. Furthermore, these underlying tendencies have an observable influence on LLM task performance, highlighting the need for greater attention to these characteristics and potential applications. Training code and SPACE benchmark specifications are available at https://github.com/Niklauseik/SPACE-training.

We study the fingerprinting of agentic cyberattacks. Our objective is to reliably infer information about the attacker's configuration, such as LLM version and harness, from telemetry. To achieve this, we introduce a technique we call shibboleth traps: modular honeypot services designed to elicit discriminative behaviors from agentic attackers. The traps make informative interactions a condition for obtaining resources the agent seeks, eliciting model declarations and characteristic writing without requiring a successful prompt injection. To achieve harness fingerprinting, we introduce general mechanisms to indirectly probe peculiar execution behaviors of harnesses. These probes reveal how a harness masks secrets in tool output, enforces command time limits, and schedules tool calls, using only the effects visible in subsequent requests. We show that combined declarative and behavioral inference distinguishes among 15 candidate models with an average 90% top-1 accuracy, with behavioral classifiers evaluated on held-out harnesses or task families. In a separate experiment across four harnesses, fixed probe rules identify the correct harness with 92% accuracy. When instructed to conceal their identity, agents still leave useful behavioral evidence even as their declarations become uninformative.

Artificial intelligence models for clinical prediction are proliferating across neurosurgery, but their aggregate discriminative performance, distribution along the care pathway and methodological quality are unclear. We systematically reviewed and meta-analysed these studies in neurosurgery and its subspecialities along with their risk of bias appraisal. We performed a PRISMA search for peer-reviewed original studies using AI to predict a clinical outcome in neurosurgical populations. Metrics were pooled using random effects modelling. Each study was classified as diagnostic, prognostic or forecasting and was assigned an evidence level. Clinical prediction task sub-stratification analysis was performed. Risk-of-bias was assessed using PROBAST. 1063 studies were included. Across 451 studies reporting AUC, pooled discrimination was 0.86, ranging from 0.84 (spine) to 0.87 (neuro-oncology), with extreme heterogeneity. Performance varied by clinical pathway stage, highest for diagnosis, grading and molecular tasks and lowest for post-operative tasks. Only 51 studies were using prospective data with external validation, whereas 619 rested on internal validation alone or no validation. 89 percent studies were at high risk-of-bias, driven by the analysis domain. AI models report high performance across all the neurosurgical subspecialities, concordant across AUC, accuracy, sensitivity and specificity and robust to risk-of-bias stratification. Performance is highest for imaging-anchored diagnostic tasks and lowest at decision-facing post-operative stage. However, the literature is dominated by optimism-prone, internally-validated and poorly calibrated models with extreme heterogeneity. External validation, calibration analysis and reporting guidelines adherence remain prerequisites for clinical adoption.

Micro-scale air quality models require hourly traffic data on every road segment, whereas permanent counters cover only a small part of the network. A graph neural network (GNN) was developed to reconstruct hourly traffic flows on the 32,163 road segments of Paris intra-muros from the 1,573 segments equipped with a counter (4.9%). The model adapts the Graph Neural Network for Urban Interpolation (GNNUI) to a real, sparse counting network. Flows are decoded on the edges of the junction graph (one edge per segment, both directions summed), a traffic climatology serves as an offset in a zero-inflated negative-binomial likelihood, and training proceeds by masking counters. Eight-fold spatial cross-validation, with counters sharing a published series merged before splitting, distinguishes interpolation at a withheld counter in an instrumented district (R${}^2$ = 0.89, error of a quarter of the mean flow) from extrapolation to a district without counters (R${}^2$ = 0.74, a third of the mean flow). These scores hold on main axes where counters are located, while relative errors are about 50% on the few counted local streets. Ablations show that the climatology offset tends to help extrapolation. Road attributes, time cycles and the live state of the counting network matter most, while weather inputs have no significant effect. The masking rate is the only inherited hyperparameter with a clear effect. When evaluated on January--August 2026 (outside the training period), the frozen model stayed around the error band of previous years.

Shadow IT, the use of technology that the IT organisation has not approved, is a long-standing governance problem. We argue that autonomous AI agents can make it grow much faster: agents make building tools nearly free, act with their user's privileges, and leave behind scripts, automations and MCP servers that are themselves unsanctioned. A simple compounding model shows that, under its assumptions, shadow IT grows exponentially once agentic artifacts reproduce faster than the organisation retires them; retiring them starts with detecting them. We then ask whether unsanctioned agents can be detected on Windows endpoints. We propose a taxonomy of nine classes of agents and six layers of endpoint footprint, and argue that rules based only on the structure of processes and connections cannot separate an agent from its behavioural twin, the same program talking to a non-AI service. We present Agent Detector, an open-source Windows detector that anchors behaviour to the identity of the remote service and correlates it with configuration, installation and execution history, using a catalog of 39 tools and 184 LLM endpoint entries; it informs users through a notice they must acknowledge instead of blocking. On a controlled bench of 12 synthetic agents, 3 AI components and 9 benign workloads, repeated 5 times on one workstation, two behavioural rules flagged every behavioural twin, while Agent Detector flagged none and detected every agent in the majority of runs. Its limits are temporal: a connection to a provider behind a shared content-delivery network is attributable only while the DNS cache holds the provider's name, one to five minutes, and an intermittent agent was seen by about one scan in three. A scan costs about 154 ms. The bench is too small to rank detectors; it maps what each kind of evidence can see: unsanctioned agents can be partly detected, and not by behaviour alone.

We introduce AudioCodex, a fully causal autoencoder for live generative audio applications that compresses stereo audio into compact continuous latent sequences at ~10 Hz. It combines a Mean Flow magnitude autoencoder with latent hypersphere projection, bounded-noise augmentation, representation alignment, and adversarial phase recovery to learn representations that support both efficient downstream generation and high-quality reconstruction. Through extensive ablations, we show that AudioCodex outperforms the state-of-the-art in terms of generation-reconstruction trade-off, whilst remaining strictly causal in both encoding and decoding. This work paves the way towards truly real-time interactive and streaming generative audio applications.

Correlation Matrix Memories (CMMs) store associations as sums of rank-one outer products, providing closed-form addressability, exact deletion, and linear attribution. While classical analyses assume random, orthogonal keys, semantic encoders produce correlated representations. We analyse CMMs under correlated keys using frame-theoretic coherence: (i) capacity bounds yielding an encoder-independent limit $n \le d(1+\varepsilon^2)$ for mean squared crosstalk $\varepsilon^2$; (ii) an exact characterisation of the collateral effect of deleting an association, guaranteeing that deletion never invalidates certified associations; and (iii) a geometric transparency bound demonstrating why attribution requires candidate sets of centred keys. On this basis, we introduce Topology-Aware Neuro-Symbolic Associative Retrieval (TANSAR). TANSAR traverses a directed acyclic association graph, prunes invalid paths via sound ground Horn constraints, and injects a representation derived strictly from the verified path into the residual stream of a frozen language model, guarded by a Clopper--Pearson selective-risk gate. Controlled experiments confirm all theoretical bounds. In multi-hop knowledge editing, Horn pruning reduces answer contradictions from 25.1% to 0.5%, whereas Hopfield retrieval incurs 24% contradictions when stale associations remain stored (and 0% if stale associations are removed). An end-to-end edit transaction requires 13.6 ms on a standard CPU and 2.37 ms on GPU. The symbolic component matches gold labels across all 600 ProofWriter theories and achieves 94.0% agreement on FOLIO, demonstrating sound and efficient multi-hop associative retrieval with exact unlearning.

Speculative decoding speeds up language models by letting a lightweight drafter propose several future tokens that the target model verifies in parallel. Block diffusion drafters such as DFlash draft a whole block in one forward pass, and DDTree builds a draft tree from their per-position distributions. Each of these distributions assumes that the drafter's own most likely continuation is correct. A tree, however, gains over a chain only on branches that leave this continuation, where the drafter's top-1 accuracy drops from 89% to 18% and its candidate lists miss the right token far more often. When re-run from the fork token, the same drafter is usually correct, so the errors come from the prefix that the distributions condition on rather than from what the drafter knows. We introduce DForkTree, a training-free method that re-anchors each branch on its own tokens. DForkTree adds n-gram statistics from the target model's own outputs to the drafter's scores as pointwise mutual information, weights them more after the fork, and builds the optimal tree under these scores exactly on the GPU. On Qwen3.5-4B and 9B, with all tree methods sharing one drafter, tree pipeline and verifier, DForkTree is the fastest tree method on every benchmark. It is up to 41.5% faster than DFlash and 15.1% faster than DDTree in the same setting, with average speedups of up to 4.75x over autoregressive decoding.

Achieving generalizable and well-aligned multimodal representations remains a core challenge in robot learning. While recent approaches have attempted to align modalities by modeling conditional or higher-order statistical dependencies, they do not explicitly consider the structural coherence across modalities. In this work, we propose a novel multimodal alignment method that augments existing contrastive losses with a geometry-aware Gromov--Wasserstein (GW) distance-based regularization. To this end, we encode intra-modal geometry for each modality and reduce cross-modal structural discrepancies through a dynamically learned GW barycenter, thereby encouraging structural consistency in addition to instance-level contrastive alignment. We apply this optimal-transport-based alignment strategy to robot perception tasks involving underexplored modalities such as force and tactile signals, where modality data often exhibit varying sample densities. Extensive experiments on five robotic benchmarks show that our method improves structural alignment and downstream performance over state-of-the-art multimodal contrastive baselines across a wide range of downstream tasks, including robot and environment state prediction. Moreover, our GW-based regularization can be seamlessly integrated into most InfoNCE-like objectives.

When a neuron's input is lost, or known through a belief $q$, what should it output? For a neuron storing a joint polynomial density $N$ with input density $D$, two formulas compete. $p_\mathrm{CD}=E_q[N]/E_q[D]$ costs one matrix-vector product; Hierarchical Correlation Reconstruction introduces it as a constant-denominator approximation to $p_\mathrm{EU}=E_q[N/D]$. For a density model both are exact answers to different questions: $p_\mathrm{CD}$ is Bayesian updating with the belief as virtual evidence; $p_\mathrm{EU}$ installs the belief as a new marginal (Jeffrey's rule). With a missing input $p_\mathrm{CD}$ returns the model's conditional; $p_\mathrm{EU}$ generally does not. This post hoc reading is proved, and illustrated outside the pair family up to $k=8$ inputs; it has no pre-registered test, only one exploratory real-data pilot. Post hoc, closed forms fall within a factor 2 of all 112 measured gaps, and negative density begins near input correlation $1/k$. In a pre-registered synthetic study (up to 32 inputs), 5 of 15 hypotheses are supported, 2 inconclusive and 8 not supported, including all three on transfer beyond equicorrelated Gaussians. Post hoc, where $p_\mathrm{EU}$ is defined, truncation usually outweighs the gap for Beta beliefs in pair neurons; triple neurons under flat and data beliefs are exceptions.

For the first twelve years of my career in computer science I mainly worked on formal methods. I had been drawn to the field as an undergraduate by its promise of a way to build software with strong, mathematically based correctness guarantees. But after many years of struggling to make formal methods scale, I was seduced by randomised testing---also known as fuzzing. Fuzzing is much less principled than formal methods, but it just works: it can be readily applied to very large systems, and it is extremely effective at finding real bugs that developers care about and fix. In this personal account I trace my path from formal methods to fuzzing, by way of the GPUVerify project on verifying GPU kernels and the GraphicsFuzz project on randomised testing of GPU compilers. I argue that the oracle problem provides an intellectual link between fuzzing and formal methods, describe my continuing, targeted uses of formal methods in the domain of GPU programming language design, and discuss work at the boundary between formal methods and fuzzing. Finally, I reflect on whether generative AI might be the missing ingredient that allows formal methods to scale, and set out my plans for large-scale formal methods via the Keystone project, through which my collaborators and I aim to find out how much of that promise can be realised in practice.

Raw radar spectra are widely viewed as too cluttered and noisy for direct representation learning, motivating handcrafted reductions, sparse point clouds, or compact Doppler descriptors in radar perception. We challenge this assumption and argue that the main limitation lies not in the raw representation itself, but in the absence of a learning objective that separates structured signal from noise. Handcrafted reductions partially address this problem but impose rigid assumptions about what constitutes signal, discarding weak yet informative structure while retaining artifacts that fit the reduction model. Once removed, such information cannot be recovered by supervised learning. We propose an alternative: structure is the component of a radar spectrum that remains reconstructable under localized corruption, while noise is the component that does not. This perspective motivates self-supervised masked modeling for dense radar spectra without predefined signal-noise separation. We instantiate it with Leakage-aware Spectral MAE, a masked autoencoder for raw radar cubes using a corruption strategy inspired by cross-bin energy spreading caused by spectral leakage. We also demonstrate the importance of Doppler organization: treating Doppler as a primary spatial dimension in an x-y-D representation, with height encoded as features, substantially outperforms naive channel stacking and enables effective learning from dense 4D spectra. Without handcrafted reductions, our approach consistently outperforms state-of-the-art radar-based detectors and point-cloud-based self-supervised learning (SSL) counterparts. These results establish raw radar spectra as an effective representation for 3D perception when paired with an appropriate pretraining objective.

For decades, the netlist-to-schematic gap has remained a persistent challenge in design automation, especially for analog and mixed-signal (AMS) circuits. Netlists encode connectivity, not visual presentation, making their manual conversion into readable schematics time-consuming and dependent on expert conventions. The same gap leaves LLM-based design tools without structured visual feedback, forcing them to iterate on circuits they cannot see. We present AMSchem, a training-free, verifier-guided agentic framework that automatically generates human-readable schematics from AMS netlists. AMSchem separates circuit-aware visual planning (DrawingIntent) from executable schematic geometry (SchematicIR), and applies localized visual feedback with bounded, revalidated repair patches. Deterministic validators enforce device preservation, pin-net consistency, and geometric correctness throughout. For evaluation, we construct AMSchem-Bench comprising 64 circuits of 12 types. On the 40-circuit schematic subset, AMSchem with gpt-5.5 achieves 85.0\% validity (vs. 32.5\% direct generation). When integrated as a visual feedback module in LLM-based analog design, AMSchem resolves all 13 previously failed cases from AnalogCoder benchmark across five models, substantially improving the success rate. Our dataset and benchmark are available at https://www.amschem.com/benchmark

Machine learning (ML) models are said to predict. Philosophers draw a line between accommodation and prediction, depending on whether or not the evidence a theory succeeds on was used in constructing it. Does this distinction survive when a theory is replaced by an ML model? We argue so, yet the dividing line becomes fuzzy. Whether an output of an ML model counts as an accommodation or as a prediction hinges upon the relationship between the patterns the input instantiate and those on which the model was selected through its training. Once this relation is taken into account, most successful outputs of ML models, we suggest, bear a strong accommodation character. We further trace the consequences of our account for confirmation, novelty, and scientific realism.

Over the past ten years Tensor Network Kernel Machines have enabled efficient nonlinear system identification through the use of low-rank tensor networks. The goal of this article is to provide the reader with a basic understanding of this model class in terms of its two defining properties. Furthermore, both the optimization and regularization of such models are discussed, together with recent developments and starting points for further exploration.

World models enable agents to plan and learn from imagined experience, but in an unfamiliar environment they must learn its mechanics from limited interaction. A mechanic is a rule for how an action or an interaction changes part of the environment. Each occurrence of a mechanic that the model observes is one exposure. We ask how many exposures a world model needs to predict the effects of a new environment's mechanics. We address this question on ARC-AGI-3, a benchmark of interactive games defined by their programs. We introduce MechEval, which labels 69 recurring mechanics across 20 games from these programs. It scores each prediction on the cells the mechanic changes at each exposure. As a test bed, we build WARC (World model for ARC-AGI-3), a 35M-parameter vision transformer that is pretrained on other games and updates all of its parameters after each transition. On held-out games, WARC learns to predict four in five mechanics with at least 80% accuracy on the cells they change and keeps that accuracy to the end of the recordings. Half of them reach that accuracy within about 80 exposures, in a single play of the game, and without updates only one in seven does at all. In closed-loop play on the public games, WARC-Agent, which pairs WARC with a fixed graph explorer, triples the explorer's official efficiency score from 0.15 to 0.46, mostly through test-time training. On the hidden games of the Kaggle leaderboard, our entry scores 0.20.

3D Gaussian Splatting (3DGS) enables photorealistic, real-time rendering on commodity GPUs and is increasingly adopted in generative world models and autonomous driving simulators. However, existing optimizations rely on algorithmic changes that often require retraining, or on specialized accelerators that lack flexibility across evolving 3DGS variants. In this paper, we identify a fundamental inefficiency in GPU-based 3DGS rendering: severe SM-level load imbalance caused by uneven workloads across screen tiles. This imbalance creates a long execution tail, leaving most SMs idle while a few process computationally intensive tiles, particularly in large-scale urban scenes. To address this, we present a software-defined GPU scheduling approach that mitigates the execution tail. Our approach uses a persistent kernel to schedule heavy tiles first and selectively chunks outlier tiles across multiple CTAs. A lightweight per-frame policy determines when and how much to chunk using tile statistics already available in the rendering pipeline. Our approach requires neither retraining nor specialized hardware and can be applied orthogonally to existing optimizations across diverse 3DGS variants. Averaged per scene, it accelerates alpha blending by up to 2.93x on a large-scale urban scene, and by up to 5.6x on its most imbalanced frames, while retaining the original alpha blending formulation.

The growing complexity of Multi-Agent Reinforcement Learning (MARL) has fragmented the literature, obscuring underlying architectural choices behind diverse implementations. This fragmentation makes it difficult to evaluate critical modeling dimensions in isolation, identify how components affect the resulting system dynamics, and compare, reuse, or combine contributions across the literature. This paper proposes a modular perspective for modeling, analyzing, and implementing MARL systems. We decompose them into eight functional modules: agents, environments, learning, reward modeling, training and execution modes, communication, modeling of other agents, and evaluation criteria. For each module, we clarify its role and review representative approaches. We instantiate this decomposition in MARLKIT, a framework whose modules are explicit and interchangeable. MARLKIT demonstrates how the proposed perspective supports transparent experimental assumptions and the controlled variation and combination of MARL mechanisms. Experiments illustrate its ability to prototype, compare, reuse and combine modules while making differences between configurations explicit. Overall, the paper provides a structured perspective on MARL modeling and implementation, offering a common language for making modeling choices explicit, supporting clearer descriptions of MARL approaches, and supporting more controlled experimentation through the systematic isolation and combination of MARL mechanisms.

This technical report presents MNN-Pocket-0.3B, a compact 300M-parameter dense language model designed for efficient on-device deployment and supporting context lengths of up to 32K tokens. With an INT4 deployment footprint of only 173 MB and decoding throughput of up to 466 tokens/s on iPad Pro M5 (Metal), it can be seamlessly integrated into applications on edge devices such as smartphones and smart glasses. MNN-Pocket-0.3B is developed through the co-design of its architecture, data, training pipeline, and inference system. Hardware-aware architecture selection and dependency-aware vocabulary pruning balance model capability with on-device memory and compute constraints. A four-stage pre-training pipeline covering approximately 4.1 trillion high-quality Chinese and English tokens equips the model with broad knowledge and reasoning capabilities. This is followed by a three-stage post-training pipeline comprising instruction fine-tuning, multi-teacher on-policy distillation, and quantization-aware training. Finally, hardware-specific optimizations in the MNN inference engine enable fast, memory-efficient inference and long-context processing. Extensive evaluations demonstrate that MNN-Pocket-0.3B achieves frontier-level performance among models of comparable size, with particularly strong results in Chinese and English knowledge and reasoning, mathematics, coding, and instruction following. It consistently outperforms similarly sized open-source models and even surpasses several 0.5B-parameter models on instruction-following benchmarks. We release MNN-Pocket-0.3B with open weights and ready-to-use MNN deployment packages. We hope it will provide a practical foundation for capable, responsive, and memory-efficient edge applications while advancing research and development in on-device language models.

The heavy-ball method with a constant step size and a constant momentum parameter accelerates gradient descent on quadratics, but no choice of the two parameters does so on the class of smooth strongly convex functions. We determine how much faster than gradient descent it can be made on this class. Writing the best worst-case rate over all tunings as $1-c(κ)/κ$, where $κ$ is the condition number, we prove that $c(κ)$ tends to $8$ in every dimension of two or more. The lower bound on the rate holds for every tuning, because a tuning that is faster than $1-8/κ$ on quadratics follows a slowly decaying spiral on some function of two variables. The matching upper bound comes from an explicit tuning, with nearly the momentum of Polyak's tuning and a step size about $\sqrtκ/8$ times smaller, and from a Lyapunov function whose positivity is verified in exact arithmetic. Optimally tuned heavy-ball is therefore asymptotically four times faster than gradient descent on this class.

Foundation models for physiological signals are commonly evaluated on in-domain interpretation tasks, but their ability to transfer to longitudinal clinical prognosis across healthcare systems remains uncertain. We developed PACE-ECG, an ECG pretraining framework that combines contrastive learning across temporally neighbouring recordings from the same patient with clinically guided multitask supervision. We evaluated frozen ECG representations for post-myocardial-infarction all-cause mortality and incident heart failure prediction across MIMIC-IV and two external HEEDB cohorts, comparing the proposed representation with DeepECG-SSL, a state-of-the-art model, and two control baselines. Across both downstream tasks, the proposed model outperformed DeepECG-SSL in-domain (AUROC +1.9% for mortality and +4.0% for heart failure) and on external populations in HEEDB-Boston (+5.6% for mortality +10.3% for heart failure) and HEEDB-Atlanta (+6.9% for mortality +11.0% for heart failure) while using an encoder 41 times smaller and a fraction of the pretraining data. A matched ablation showed that patient-temporal pairing improved external mortality and heart failure AUROC over generic self-supervision by 7.0% and 9.4% on HEEDB-Boston and 8.9% and 9.3% on HEEDB-Atlanta. These findings indicate that pretraining-objective design greatly influences downstream transfer alongside model and data scale, and that in-domain performance alone is insufficient to establish the clinical generalisability of ECG foundation representations.

Multimodal large language models (MLLMs) have demonstrated strong performance on knowledge-intensive and reasoning-based vision-language benchmark datasets, but still struggle to solve the visual questions that people can solve easily. Evaluating the differences between MLLMs and human-level visual perception is challenging due to end-to-end evaluation which combines various skills, or because we are uncertain if individual items are valid. Specifically, observed errors may reflect visual perception, language priors, background knowledge, question ambiguity, and underspecified answers. Our guideline for HumanSightBench is that we construct items that are human-easy by construction and model-hard by evaluation. HumanSightBench contains 2,876 visual questions with short answers across 7 top-level perceptual categories and 21 subcategories. We only include items that pass five inclusion gates: question clarity, answer uniqueness, visual grounding, human solvability, and text-only shortcut filtering. According to the protocol, errors are defined as incorrect answers on human-validated visual tasks with the evaluated interface. The inclusion gates are intended to reduce confounds from ambiguity and external knowledge. The best pass@3 score is 65.6% across eight evaluated MLLMs. No evaluated model exceeds 27.5% on Counting. These results identify capability gaps that are hidden by overall scores and can help future training experiments and held-out evaluation.

Action theories typically fix their vocabulary and causal laws, whereas agent harnesses can change their tools, providers, and policies during execution. We introduce Reconfigurable Action Theories (RAT), in which actions can change the action theory itself. A theory module in RAT contains a vocabulary, causal laws, and an initializer. A RAT configuration records the active modules and the world state. Module loading and unloading are represented by actions governed by ordinary causal laws. RAT retains the standard causal action semantics of its language fragment when the active module set is fixed. We give sufficient conditions for at most one next configuration, preservation of existing behavior when a module is added, and equal results when modules are loaded in either order. We establish the computational complexity of reasoning tasks in RAT and translate finite RAT systems into nonmonotonic causal theories and Answer Set Programming (ASP), preserving bounded executions. Experiments compare finite RAT models with executions of official DeepSeek Harness, Pi, and OpenHands implementations. RAT provides a formal basis for checking how component changes affect a harness's available actions, behavior, and reachable goals.

Ultrasonic guided wave (UGW) inspection has become an effective method in structural health monitoring (SHM). While machine learning has enhanced damage detection capabilities, purely data-driven models remain constrained by extensive labeled datasets and limited physical interpretability. Standard multi-layer perceptrons (MLPs) further suffer from spectral bias (F-Principle) in high-frequency regimes, prioritizing low-frequency components while failing to resolve the high-frequency oscillations critical for UGW reconstruction. Physics-Informed Sinusoidal Representation Networks (PI-SIRENs) address these limitations by embedding governing partial differential equations directly into the training process. Building upon SIRENs, which utilize periodic activation functions to inherently capture high-frequency spectral content, the proposed framework is designed to operate effectively in sparse ultrasonic measurement regimes. This work introduces a frequency-domain PI-SIREN framework tailored for the Mindlin-Reissner plate theory to reconstruct ultrasonic wavefields without temporal discretization. The model parameters are optimized using a hybrid strategy comprising three phases: first-order Adam, residual-adaptive refinement (RAR) and quasi-Newton L-BFGS. Numerical and experimental evaluations confirm that the model serves as a physics-consistent baseline derived from sparse data, functioning as a computationally efficient surrogate for damage identification in SHM and non-destructive testing (NDT). Integrating the surrogate into a damage identification pipeline detects and localizes a milled notch from sparse experimental measurements with a localization error of 1 % of the plate diagonal. In contrast, data-driven benchmarks and MLPs with hyperbolic tangent activation functions, with and without Fourier feature embedding, are unable to identify damage due to higher reconstruction errors.

Generative recommendation (GR) is moving industrial recommender systems (RecSys) from scoring candidate items to autoregressively generating item identifiers. This shift promises higher recommendation quality, but makes serving substantially harder: each request executes a short sequence of dependent decoding steps that explore a large item space and select promising candidates on the latency-critical path. GR not only creates structural mismatches with conventional RecSys pipelines, but also challenges the architectural and optimization assumptions of LLM serving stacks, making direct adoption of existing systems difficult. We present RecoGEM, our production GR serving system. RecoGEM uses a runtime--engine decoupling architecture to move GR-specific serving responsibilities, such as candidate-state transitions and live artifact updates, to a dedicated runtime. Its execution engine exploits GR's short token sequences, large candidate populations, and predictable step shapes to optimize self-attention, candidate selection, and execution planning. RecoGEM has been deployed for over nine months across our production lines, serving applications with hundreds of millions of users and supporting more than ten GR model families. On our OneModel production workloads, RecoGEM reaches over 1,000 QPS per serving instance with P99 latency below 35 ms and a 2.88$\times$ throughput improvement over representative baselines.

Robots operating in the real world are inherently subject to partial observability, particularly in multi-robot collaboration, where individual observations may be insufficient for effective task execution. A key challenge is leveraging complementary observations from partner robots to support collaborative decision-making. We argue that information sharing should go beyond passively transmitting raw observations or generic visual features. Instead, robots should selectively encode task-relevant information according to task objectives and execution progress. To this end, we propose a task-oriented active context sharing framework that transforms robot observations into compact representations, preserving critical information under constrained communication budgets. The framework adopts a two-stage training paradigm. In the first stage, onboard observations and task instructions are jointly encoded into compact context representations. Task-oriented future prediction and soft task-progress correspondence learning guide this collaboration-agnostic compression, preserving information about task objectives and execution progress. In the second stage, the learned representations are integrated into a pretrained Vision-Language-Action (VLA) policy, enabling bidirectional context exchange. Each robot combines complementary partner information with local observations to guide action generation. Action supervision jointly refines the context representations and policy, aligning information sharing with collaborative decision-making. Controlled experiments demonstrate substantial performance gains under constrained communication budgets, while three real-world experiments validate the framework in complex dual-robot tasks. These results highlight a shift from passive visual transmission toward active, task-oriented context sharing that directly supports collaborative decision-making.

Recurrent latent reasoning in large lanugage models (LLMs) scales test-time computation by increasing recurrent depth, naturally framing inference as a problem of deciding when to stop. We identify a previously overlooked limitation of this single-depth view through systematic preliminary analysis: deeper recurrence can lose correct solutions found at shallower depths, while depth-specific success sets remain broadly non-nested even after controlling for marginal accuracy. This suggests that depth determines not only computation cost, but also access to distinct candidate solutions. Motivated by this finding, we propose Depth Portfolio Allocation (DPA), which reformulates recurrent test-time scaling as budgeted allocation across depth-specific candidate sources. DPA supports both fixed portfolios and feedback-guided sequential allocation under a shared inference budget. Experiments on GSM8K, SVAMP, MultiArith, and MBPP show that mixed-depth portfolios expand oracle coverage under matched candidate counts and nominal costs, while feedback-guided allocation generally matches or improves uniform allocation and can reduce computation in high-budget settings. Code and data are avaible at : https://anonymous.4open.science/r/LLM-latent-B7BC/.

Speech comprehension in the human brain is an incremental process that begins with acoustic processing and advances to linguistic processing. While recent studies have revealed that the hierarchy of deep neural networks (DNNs) correlates with the ascending auditory pathway, the exact nature of this DNN-brain alignment remains underexplored. In this study, we investigate how DNN representations align with the brain's acoustic-to-linguistic processing. Specifically, we employed neural encoding models to simulate neural responses to acoustic (i.e., speech and noise envelope) and linguistic features (i.e., word onset and surprisal). By applying representational similarity analysis (RSA), we quantified the similarity between these neural responses and the DNN embeddings generated by a pre-trained automatic speech recognition (ASR) model, both before and after fine-tuning on audiovisual noisy data. Our results demonstrate significant DNN-brain alignment: embeddings from shallow layers exhibit higher similarity to neural responses associated with acoustic features, while those from deeper layers align more closely with neural responses related to linguistic features. Importantly, the fine-tuning process enhances this alignment by improving noise processing in shallower layers and refining linguistic representations in deeper layers. These results suggest that fine-tuned DNN models can naturally develop human-like processing patterns in noisy environments, highlighting a functional alignment between the human brain and DNNs in speech representation.

General-purpose software serves many different demands, but this flexibility costs performance and efficiency on any specific demand. Software tailored to a specific demand avoids these losses, yet developing it requires substantial time and effort from human experts. AI agents offer an alternative: they can design, implement and test artefacts tailored to individual demands. We call this production paradigm Forge. Forging for a demand is economically justified when the cumulative benefit of specialisation exceeds the total cost of construction, verification and maintenance. To make Forge a repeatable production process, we propose ForgeSystem, which consists of a control layer, a forge loop and a recursive self-improvement (RSI) loop. The control layer selects which artefact serves each demand and decides which demands to forge, when to forge them and from which starting point; the forge loop iterates between production and verification until a candidate artefact passes verification; and the RSI loop draws on accumulated production records to supply starting points, refine the harness and train models. We implement ForgeSystem for training frameworks, decoding megakernels and individual computational kernels. Forged training engines reach 1.047-1.332x the model-FLOPs utilisation of their reference frameworks, forged megakernels achieve a geometric-mean decoding speedup of 1.21x over SGLang across thirteen configurations, and forged kernels match or exceed expert libraries in eighteen of twenty-two configurations, while satisfying task-specific quality requirements. Controlled comparisons show that explicit quality requirements steer agents towards optimisations that preserve training quality. We also discuss a gradual transition to Forge through service routing, and how demand-specific production may shift value accumulation and trust towards production systems.

LLM-driven time series forecasting has become increasingly promising by leveraging contextual features to generate future values directly or refine forecast priors. However, incorporating contextual features alone does not ensure that models can identify which features matter for forecasting and understand what they suggest about future dynamics. We argue that the fundamental challenge is how to extract predictive evidence from contextual features and when and how to leverage it to guide forecasting decisions. In this paper, we propose CastReasoner, an evidence-grounded reasoning model for context-aware time series forecasting. Specifically, CastReasoner uses a structured chain of thought (CoT) to extract predictive evidence from contextual features, infer the future dynamics suggested by this evidence, and assess these dynamics against the forecast prior to guide adaptive forecasting decisions. To enable this capability, we design a reasoning-centric two-stage post-training framework including supervised fine-tuning (SFT) and reinforcement learning (RL). First, we synthesize reasoning trajectories through controllable multi-teacher guidance and select high-quality ones using a novel student-aware difficulty metric. Second, we design fine-grained multi-objective rewards to further improve reasoning and forecasting. To address heterogeneous learning needs across domains in both stages, we introduce a convergence-guided data balancing strategy. Through this two-stage training, CastReasoner improves both reasoning quality and predictive accuracy. Extensive experiments across multi-domain forecasting benchmarks demonstrate that CastReasoner achieves superior forecasting performance.

Cryptocurrency markets provide a natural testbed for autonomous AI agents because they are digital-native, high-frequency, and machine-speed environments, yet existing methods often struggle with non-stationary market regimes, opaque decisions, and fixed aggregation rules. In this paper, we propose Agent4Crypto, a dual-reflective and market-feedback-aligned multi-agent system. The framework coordinates specialized agents for data collection, analysis, risk assessment, reflection, and trading, synthesizing heterogeneous signals from prices, indicators, news, and charts in an interpretable decision pipeline. Our central design is a trader-centric decision mechanism: three style-specific trader agents, Aggressive, Neutral, and Conservative, generate candidate actions and natural-language rationales from the same market context, while their influence is learned rather than manually fixed. A PPO-based gating policy dynamically allocates trader weights through a dual-reflective feedback process. One reflection stream evaluates post-trade decision quality from realized portfolio outcomes and distills reusable lessons into a Financial ReasoningBank for retrieval in subsequent trader prompts, while the other estimates trader-level counterfactual contribution, indicating which style has recently aligned better with market movements. The PPO state combines Counterfactual Contribution Memory with market features, previous trader weights, and current trader actions, so trader influence adapts across bull, sideways, and bear regimes. Experiments on BTC, ETH, and XRP show that Agent4Crypto improves total return and Sharpe ratio over traditional strategies, machine learning and reinforcement learning baselines, and existing LLM-based trading systems, demonstrating the value of interpretable trader competition and market-feedback-driven adaptation in volatile cryptocurrency markets.

Time-to-event prediction is fundamentally challenged by partial observability and aleatoric uncertainty. In clinical medicine, for instance, the timing of disease progression or acute readmission is contingent on highly dynamic and largely latent health trajectories. Rather than forcing a deterministic point forecast, distributional regression offers a principled means to address this complexity by predicting a probability distribution over time. However, explaining these distributional predictions presents a key obstacle: raw distribution parameters output by the model are abstract mathematical quantities that do not align well with domain concepts. To gain interpretability, we propose a novel paradigm where output parameters are designed to be additive in log-odds---a widely established measure in clinical risk modeling. Our approach, called ODDS-DR, moves beyond technical convenience to deliver clear domain alignment as well as the layer-wise computational structure required for robust, time-dependent feature attribution. Validated across two real-world clinical cohorts, our framework matches state-of-the-art predictive performance while resolving distributional outputs into interpretable risk markers across flexible prognostic horizons.

Reliable tip-over prevention is critical for planetary rovers operating in unstructured environments where communication delays prevent real-time human intervention. Existing learning-based approaches typically optimize a single prediction objective, limiting their ability to capture both gradual instability evolution and abrupt terrain disturbances. This work presents a hierarchical multi-channel tip-over prediction framework based on isolated temporal learning. A multi-horizon Temporal Convolutional Network (TCN) learns progressive instability patterns from inertial measurements, while a Gated Recurrent Unit (GRU)-based forward dynamics model predicts nominal rover motion and detects unexpected dynamic deviations through prediction residuals. These complementary signals are integrated through decision-level fusion and priority-based arbitration to generate graded safety actions, including early warning and emergency braking. The proposed system is deployed as an Orocos Real-Time Toolkit (RTT)-based safety component within the Rock robotics framework, operating at 100 Hz on a planetary rover software stack. Experiments on Inertial Measurement Unit (IMU) samples collected across heterogeneous terrains demonstrate 97.58% event prediction coverage with a 1.82s median lead time, outperforming individual prediction channels. The results validate that isolated temporal models combined through hierarchical decision fusion provide a robust solution for real-time rover safety in uncertain terrain.

Reinforcement learning (RL) from raw sensory inputs such as images or onboard sensors can be challenging due to high-dimensional observations and sparse rewards. A common strategy is to train a teacher policy with access to privileged state information and then distill it into a student policy that acts from raw inputs alone. However, when the teacher relies on information unavailable to the student, exact imitation may be impossible, creating an irreducible imitation gap. Existing approaches typically mitigate this mismatch through careful reward shaping or additional reinforcement learning on the student policy after the teacher has been trained. Instead, we introduce Representation-Aligned Imitation Learning (RAIL), which addresses the mismatch during teacher training itself. RAIL learns a latent representation shared across teacher and student observations using contrastive learning, and trains the teacher policy directly in this space. Because the teacher policy is constrained to operate on this shared space, it learns behaviors that are reproducible from student observations. This substantially reduces the imitation gap while preserving task performance. Across multiple environments, RAIL outperforms strong baselines without reward modification or post-hoc student fine-tuning, and surpasses direct reinforcement learning from raw observations. The learned representations further enable zero-shot transfer to new tasks through teacher-only training. Project site: https://merajmammadov.com/rail/.

We present a general method for discovering symbolic local learning rules that can replace backpropagation, applicable, in principle, to any deep learning architecture and task. The method automatically discovers local loss functions specific to each layer by (1) constructing a dictionary of monomials of local tensor variables, inspired by Penrose diagram notation, (2) regressing over coefficients weighting these monomials to build layerwise polynomial losses whose gradients together minimize a target global loss. We test the method on classification, reconstruction, and self-supervised learning tasks with small multilayer networks and minimal datasets. Across all tasks, the discovered rules induce non-trivial interactions between layers, which lead to significant reductions in the corresponding global losses. Furthermore, the method selects sparse combinations of monomials, finding local rules that are more interpretable than those generated by MLP-based alternatives. The potential implications of our method range from energy-efficient deep learning to the discovery of biologically plausible learning mechanisms.

Evaluating generative models requires measuring not only sample fidelity but also how well a generated set covers the real data distribution. Recent work has therefore introduced recall-oriented metrics for distribution-level evaluation. However, existing recall metrics often reduce distributional coverage to geometry-based coverage. This simplifying assumption weakens their diagnostic value and interpretability in two ways: these metrics provide only a single aggregate score that cannot distinguish essential diversity failures such as mode dropping and mode collapse, and geometry-based coverage can provide unreliable estimates by assigning high scores to poorly covered generative distributions. We therefore introduce CADMA, a capacity-aware framework for recall-oriented evaluation and diagnosis. CADMA evaluates distributional coverage through local capacity matching between real and generated samples, separating geometry-based coverage from effective coverage. This yields multi-dimensional coverage diagnostics that quantify mode dropping and mode collapse while reducing the overly optimistic estimates produced by geometry-based metrics. Through extensive experiments on toy data, real-world datasets, and generative models, we show that CADMA provides more reliable coverage estimates and more fine-grained, interpretable diagnostic signals of diversity failures than existing metrics. Our code is available at https://github.com/Open-Debin/CADMA_NeurIPS2026.

Continual Test-Time Adaptation (CTTA) methods update models online using a stream of unlabeled data to overcome distribution shifts encountered at test time. However, CTTA is often unstable, as it can reinforce biases in early predictions and increasingly predict only a few classes, resulting in drift or even collapse. In this paper, we identify narrowing pseudo-label class usage as an indicator of CTTA instability and quantify this behavior. Most existing methods have attempted to address this issue by either selecting reliable adaptation samples or re-weighting the loss to encourage diverse pseudo-labels, but they do not explicitly enforce a class marginal over a temporal window. Instead, we propose Sinkhorn Class-Marginal Regularization (SCALAR), a simple plug-and-play module that controls the pseudo-label distribution over a short temporal window. SCALAR works by maintaining recent predictions in a buffer, balancing their class distribution via the Sinkhorn algorithm to match a specified target distribution, and using the balanced soft targets for adaptation. Without changing the model architecture, SCALAR maintains the per-step runtime and increases the peak allocated GPU memory by only 10 MB on ImageNet-C. Our method is particularly effective on skewed, non-stationary data with severe distribution shifts, where SCALAR significantly alleviates drift and collapse when combined with existing stabilization methods based on entropy or diversity.

Mixture-of-Experts language models are cheap to run per token but expensive to hold in GPU memory, which drives the cost of serving them in production. One-shot expert compression shrinks them after a single calibration pass. Its two current state-of-the-art methods are REAP for expert pruning, which deletes the least salient experts, and REAM for expert merging, which merges them into the most salient ones. Both have been evaluated only on academic benchmarks. We compare the two on five multilingual production tasks of a web-hosting company, from classification to the generation of marketing HTML email templates, with two 30B-parameter models (Qwen3-30B-A3B-Instruct-2507 and the hybrid Mamba-attention Nemotron-3.5-Lightning) at 25% and 50% of experts removed and the same calibration data. Pruning wins: on the same rows, REAP is significantly better in 16 of 20 comparisons and REAM in 2. At 25%, REAP keeps 93.6% of Qwen3-30B's generative quality and REAM 62.5%, although on benchmarks the two are within a point for the same model. The damage concentrates in the output contract rather than the content: 89.8-95.2% of each significant loss on email templates comes from hard-gate failures such as invalid JSON, broken mobile layouts and runaway generation, and most failed non-English domain descriptions copy the prompt's English opening phrase. Cross-perplexity and logit divergence, two diagnostics from the REAP paper, also track production quality, as does the rate of repetition loops; n-gram diversity does not. We close with practical guidance for compressing MoE models for production.

On-policy self-distillation lets a language agent act as its own Teacher by conditioning on privileged context such as skills written from its own trajectories. Its guidance, however, is weighted by how much this context shifts token likelihoods rather than by whether it improves behavior, so errors in self-written skills return through later updates. We propose environment-calibrated on-policy self-distillation, which uses environmental outcomes to decide how much to trust the Teacher's current guidance and which self-written skills to reinforce. Command-level advantages, estimated by comparing commands executed at similar observations, scale each Teacher preference and reward reflections and skills whose induced preferences favor higher-advantage commands. Without skills at inference, 3B and 7B agents reach 90.5% and 93.3% macro success on seen ALFWorld tasks and 75.5% and 80.5% success on WebShop. On unseen ALFWorld tasks, they outperform GRPO under the same protocol by 11.0 and 7.6 percentage points, respectively. In an ablation study, supervising only executable commands rather than the Actor's full response raises 3B WebShop success from 52.3% to 75.0%.

A worked solution can make self-generated training targets shorter and cheaper, but accuracy on the practiced operation may hide how well a model can reuse newly learned rules. We study this choice in a 2B-parameter reasoning model learning two invented quantitative rules. The model generates targets from a rule sheet alone or from the sheet plus a program-written worked solution; targets are accepted for final-answer correctness, and the support is removed for fine-tuning and evaluation. Training questions request forward use in one segment; composition requires sequential reuse of the same rule across two segments. After six passes, using a program-checked numerical signature, sheet-only targets lead by 24.3 percentage points (95% item-bootstrap CI 14.6-34.0) on composition in a prospective test. On unseen argument values, observed forward accuracy is identical (87.5% each), but composition accuracy is 58.3% versus 23.6%, a 34.7-point gap (25.7-43.8). An independently sampled second target-bank pair repeats the direction (+14.6 points, 2.1-26.0). Ten passes on the shorter worked-solution targets, approximately matching the domain tokens of six sheet-only passes while changing other training quantities, narrow the gap to 16.3 points (6.3-26.7) in the first bank pair and 6.2 points (-6.3 to 18.8) in the second, where the pre-specified test does not hold. At six passes each, the worked-solution policy uses fewer capture tokens and less update time, yet the trained models use 2.6 times as many generated tokens per composition success under the numerical signature (2.1 times in the second bank pair). Choosing a capture context therefore requires testing how the new rules are reused, alongside the inference cost of those uses, rather than relying on forward accuracy alone.

Systematic literature reviews (SRs) remain slow and labor-intensive, and most automation tools address single stages of the workflow. We present PARIS (PRISMA Agentic Research Intelligence System), an open-source framework in which a conductor agent running in Claude Code orchestrates 15 specialized large language model (LLM) agents, from scoping a research question to exporting a manuscript draft. Agents are specified declaratively, coordinate exclusively through persisted artifacts that the conductor validates before every transition, and delegate retrieval, bibliometric analysis and scoring to deterministic Python utilities. Three human checkpoints gate study inclusion, corpus adequacy and the synthesis outline. We also introduce QualityScore, a pre-synthesis index that aggregates normalized network, bibliometric, semantic and process sub-scores into a 0-100 grade. Two case studies illustrate both: AI chatbots and university students' learning (388 records, 100 included) and AI-assisted SR methods (131 records, 123 included, QualityScore 69.1 recomputed with a proxy network sub-score). They show end-to-end feasibility across domains and the diagnostic use of QualityScore: in the first run, corpus improvements chosen at the corpus-adequacy checkpoint raised the score from 74.6 to 81.9 and the number of pre-2023 studies from 1 to 19. Audits of both runs exposed failure modes: incomplete retrieval, counts that drifted between agents, unrecorded or delegated checkpoint decisions and a grade that depends on how citation data are handled. PARIS fully supports 8 of the 27 PRISMA 2020 items and partially supports 11. Output quality has so far been judged only by an internal reviewer agent; independent validation against human reviews is the next step.

Quantum kernel methods have been proposed for network intrusion detection, yet their performance limitations are often attributed to noise in current quantum hardware. This study evaluates that assumption through a controlled analysis of Pegasos-QSVC on the NSL-KDD and ToN-IoT datasets. We perform 56 quantum experiments using one exact noiseless statevector backend and three noise-simulated superconducting backends while keeping preprocessing, PCA features, data splits, and classical baselines identical. XGBoost achieves accuracies of 98.94% on NSL-KDD and 99.41% on ToN-IoT, compared with 91.80% and 82.98% for Pegasos-QSVC. The quantum model also exhibits substantially greater seed-dependent variability. Eight runs collapse into degenerate classifiers, including four recall-collapse cases and four trivial-positive classifiers. Notably, two failures occur on the noiseless backend, demonstrating that hardware noise is not a necessary condition for classi-fier collapse. Kernel Gram matrix analysis further reveals low effective rank and severe ill-conditioning even in the absence of noise, indicating intrinsic fragility in the learned quantum representation. These results show that average accuracy alone can conceal operationally unacceptable behavior and highlight the importance of reporting worst-case performance, seed-level variability, class-specific metrics, and explicit degeneracy tests in quantum intrusion detection studies.

Transformers are built with continuous symmetries in the attention mechanism. We derive $ρ$, a scalar invariant that measures each head's capacity to couple to rotary positional encodings. This framework allows us to analyze heads without forward passes, opening a new research direction into data-independent analysis of modern LLMs. Empirically, we find that $ρ$ predicts induction heads, the most-studied circuit type in mechanistic interpretability, with Spearman correlation between $-0.55$ and $-0.74$ across four bias-free RoPE models spanning $135$M to $7$B parameters. As a preliminary proof of concept, we use $ρ$ to apply YaRN to only the top-50% positional heads, which degrades perplexity less than uniform application. Code is available at https://github.com/MPedraBento/rho-invariant.

In patrol missions involving multiple unmanned surface vehicles (USVs), a designated command USV can support fleet management by sequentially rendezvousing with each patrolling USV. These operations require motion-feasible route planning and priority-aware rendezvous sequencing, with the objective of minimizing mission completion time and total tardiness relative to priority-dependent due dates. We propose a framework combining grid-based path planning, surrogate travel-time prediction, and graph-based reinforcement learning. The path planner accounts for ocean currents and approximates turn-rate constraints through multi-stage maneuvers, with path feasibility empirically validated through trajectory-tracking simulations. For rendezvous sequencing, a Feasible Transition Graph (FTG) represents candidate rendezvous events and temporally feasible transitions, enabling a graph neural network to exploit spatiotemporal rendezvous relationships for policy learning. A surrogate predicts travel times from compact current and motion features to avoid repeated path searches during FTG construction and policy training. Evaluation using grid-based path planning shows statistically significant average objective reductions of 10.5%, 3.5%, and 1.5% compared with greedy, MLP-RL, and 2-step lookahead greedy, respectively. The policy achieves these improvements while requiring approximately 1/35 of the latter's computation time. It maintains its advantage over MLP-RL under unseen priority-tier settings, with larger gains as fleet size increases.

A sparse terminal reward tells an agent little about its steps. Much of the missing information is implicit in the environment's dynamics, because each state transition can change the conditions for success long before the reward arrives. A world model trained on real transitions can learn to predict what each action changes, but a predicted observation does not by itself show whether success has become more likely. With a binary terminal reward, PPO's critic makes this judgment by estimating the current policy's probability of success, but only for states the agent reaches. Our method, WM-PPO, elegantly brings the two together. For each action the policy samples, executed or not, the world model predicts the next observation and the critic values the predicted successor. In the policy channel, comparisons among these values give the policy auxiliary advantages; in the value channel, their mean gives the critic an auxiliary target. A shared gate weights both channels according to how reliably comparisons with similar margins predicted real outcomes earlier in training. The policy, critic and world model all learn from the agent's own interaction, with no teacher or privileged information, and only the policy is used at inference. With a comparable number of real environment steps, WM-PPO raises success from PPO's 63.5% to 87.5% on ALFWorld and from 39.3% to 52.9% on ScienceWorld. It matches the strongest PPO refinements we tested on ALFWorld and exceeds the best of them by 6.5 points on ScienceWorld. The trained policy more often completes prerequisite steps before placing an object or committing to an answer.

Verbal memory retrieval carries rich diagnostic information for Alzheimer's disease through the order, hesitation, and categorical organization of recalled words, yet computational modeling of this retrieval process remains largely unexplored. A key challenge is that timing and hierarchical structure are diagnostically entangled: the diagnostic meaning of retrieval timing depends on its hierarchical context. We propose TrajLift, a spectral framework that models verbal recall as timed trajectories on a hierarchical graph. Our central mechanism is latency-modulated heat diffusion on the semantic hierarchy, a unified operator that lifts each recalled word into a representation of both multi-scale hierarchical structure and continuous retrieval timing. Formal analysis shows that the resulting representation is first-order separable into structural and temporal components with spectral selectivity across hierarchical levels. On synthetic benchmarks requiring structure-time interaction, TrajLift stays near-perfect on deep hierarchies where all baselines fail. On real-world clinical data, we show that TrajLift outperforms diverse baselines by jointly modeling timing and semantic structure, with gains that grow as retrieval timing becomes more informative.

Computational advances in natural language processing and machine learning have substantially expanded the use of automated approaches in verbal deception research. However, existing verbal deception datasets remain fragmented across disciplines, as they are often difficult to identify, access, and reuse, limiting cross-study comparisons and the development of robust computational models. We present the Deception Archive, a curated and standardized repository of publicly available verbal deception datasets built upon a large-scale systematic review. Eligible datasets were retrieved, manually inspected, and standardized following a structured data curation workflow. The current repository comprises 41 publicly available datasets containing 387,995 truthful or deceptive statements. The archive represents a heterogeneous repository (e.g., in data size, topic of investigation, ground-truth, etc.) and a living resource for integrating newly available deception datasets over time. By providing centralized access to standardized datasets and accompanying metadata, the Deception Archive facilitates dataset discovery, cross-dataset validation, and reproducible evaluation of computational models.

Federated learning (FL) is a promising research paradigm that enables collaborative model training across distributed clients without data sharing. To handle unlabeled data in local clients, existing unsupervised FL research has primarily focused on tasks like clustering and self-supervised representation learning, leaving the critical problem of unsupervised visual recognition largely unexplored. To bridge this gap, we introduce FedUR, the first framework designed for unsupervised recognition by leveraging the zero-shot capabilities of pre-trained vision-language models like CLIP. Specifically, FedUR addresses data heterogeneity through a pseudo-labeling mechanism with a class-aware data selection strategy. This strategy first estimates a global pseudo-label distribution, then allocates pseudo-labeled data to clients accordingly. To handle the uncertainty of conventional positive pseudo-labels, FedUR further incorporates class weights and a negative pseudo-labeling strategy. Furthermore, FedUR proposes a personalized prompt aggregation protocol where general visual prompts are aggregated on the server, and personalized text prompts encoding local knowledge are retained on clients. Extensive experiments demonstrate that FedUR consistently outperforms baseline methods under both label-level and feature-level distribution shifts. We believe this work paves the way for further advances in federated unsupervised recognition. Code will be available at {github.com/tim-learn/FedUR/}}.

Graph transformers can capture long-range interactions, but their quadratic attention cost limits their use on large graphs. We introduce \emph{NearFormer}, a graph transformer whose global attention is restricted to a randomized near-neighbor (NN) graph. NearFormer uses an efficient randomized construction to connect each node to a small number of near neighbors from a broader neighborhood defined by a label-independent feature or spectral geometry. For a fixed retained budget \(k\), the resulting attention support has linear size \(\mathcal{O}(|E|+nk)\). The NN graph is built once by within-bin search and then held fixed. We use these edges as sparse global-attention supports, with optional learned attention biases, while preserving the original graph for local message passing. Experiments on long-range graph benchmarks and synthetic structural tasks show that NearFormer provides a favorable accuracy--scalability trade-off relative to local message passing, dense and sparse attention.

Scientific discovery requires ideas that are original, useful, and non-obvious. Yet AI scientists are typically given as much task information as possible during idea generation, which can anchor generation toward familiar approaches. Inspired by the cognitive framework of Blind Variation and Selective Retention (BVSR), we propose a discovery framework that separates these two roles of information: it generates proposals under deliberately limited task information to broaden exploration, then learns from proposals previously verified on the same task and their outcomes to retain promising candidates for costly empirical evaluation. We evaluate BVSR across five proposer models and 12 scientific tasks. Limiting task information substantially increases proposal originality and non-obviousness, and produces a larger share of highly creative and high-utility proposals than existing state-of-the-art prompting-based baselines. Learning from prior proposal outcomes on the task further improves candidate selection, discarding half of the candidates while retaining most top-performing proposals. When incorporated into existing AI-scientist workflows, BVSR achieves stronger discoveries across gene expression analysis, drug response prediction, anomaly detection, scaling law discovery, and brain activity prediction. These results suggest that effective AI scientists should use information differently across stages of discovery: constrain task-specific information during exploration while exploiting accumulated empirical experience on the task.

On-demand delivery platforms such as DoorDash and Uber Eats impose a constraint absent in traditional e-commerce: merchants beyond the user's delivery range are invalid candidates rather than merely low-ranked. Generative retrieval has recently shown strong potential in large-scale retrieval systems. Adapting it to this setting, however, poses three challenges. First, existing semantic IDs encode geography only as a soft signal, so reachability can only rely on a non-differentiable post-generation hard truncation. Second, user preferences vary substantially across both regions and queries, so aggregating historical behaviors without conditioning on geography and the current query yields biased interest estimates. Third, the retrieved candidate set is evaluated by its collective value, yet autoregressive training scores each sequence independently, with no set-level signal to discourage redundant near-duplicates. To address the challenges, we propose TS-GEAR, a unified generative retrieval framework for on-demand delivery: (1) Geo-Semantic ID splits merchant identifiers into geographic prefixes and semantic suffixes, making reachability a prediction target rather than a post-hoc filter; (2) Query-Geo Target Attention aggregates historical behaviors using the real-time query and location as dual targets, while encoding each behavior's own location into its attention key; (3) GMPO (Group Marginal Policy Optimization) rewrites advantages using a deduplicated group reward and marginal advantage within GRPO, avoiding auxiliary diversity objectives. Deployed online with a lightweight 0.1B backbone, TS-GEAR serves over 150 million queries per day in the search advertising pipeline of Taobao Shangou, consistently outperforming state-of-the-art generative retrieval methods on both HitRate and geographic validity, and achieving 3.8% gains in ad impressions and 4.2% in ad revenue.

Cohort comfort models (CCMs) aim to overcome the extensive individual data requirements of personal comfort models (PCMs) by grouping occupants based on demographic and psychometric characteristics (cold start) or similar thermal preferences (warm start). We evaluated the validity of the reported CCM performance gains through (i) a methodological assessment of the original study, (ii) a replication using one of the two original datasets (DORN: 20 participants, 6 months) under a revised experimental protocol, and (iii) a replication using an independently collected dataset (PersonAI: 72 participants, 4 weeks). Classification performance was evaluated using F1-micro, as in the original study, and F1-macro to account for class imbalance. Under F1-micro, a trivial baseline that always outputs the majority class performed on par with cohort models. Under F1-macro, all cohort models achieved lower median scores than a simple baseline that makes random predictions based on the estimated distribution of thermal preferences. In our methodological assessment we identified the following key issues: Limited data per participant and per minority class, reliance on a metric insensitive to class imbalance, missing trivial baselines, inadequate cross validation strategy, and multiple occurrences of information leakage. Our findings challenge the original study's conclusion that CCMs improve thermal preference prediction. CCMs did not outperform simple class-distribution baselines. Future studies should improve experimental designs, including controlled thermal conditions, to obtain sufficient observations across thermal preference classes.

Early successes on agentic robotics point to a potential paradigm shift toward using coding agents to control robots. Existing robotics benchmarks are built around visuomotor policy learning, leaving it open whether coding agents can reliably control diverse robots under physical-world constraints. We therefore introduce the Embodied First Exam, a unified, comprehensive, and agent-native evaluation framework where coding agents can choose their own control approach: direct control, code to control, or train a policy to control. We hope it serves as a first exam for coding agents before they enter the physical world, with 1,156 tasks in 35 suites on 47 robots, including 9 new suites that go beyond pick-and-place to spatial puzzles, dexterous manipulation, multi-robot interaction, dynamic whole-body control, and sim-to-sim engineering. Each task can be run in up to three evaluation settings. The Standard Setting mimics the real world: the agent receives only sensor observations and, by default, cannot reset the environment. Its Realtime variant keeps the environment running while the agent reasons. The Privileged Setting gives the agent access to the full simulator state and unlimited resets, allowing it to search for an oracle solution for training purposes. Success rates in this setting serve as an approximate upper bound. On the 456 core tasks, the strongest of eight agents, GPT-6 Astra in Codex, achieves a success rate of 34% in the Standard Setting, leaving much room for improvement. To see where agents fail, we label the stage at which failed attempts go wrong and compare failures across agents, settings, and stage categories.

Robotic manipulation has advanced rapidly in mapping visual observations and language instructions to actions. However, reliable manipulation in open-world environments requires robust spatial understanding: the ability to identify targets in complex layouts (spatial reasoning) and adapt to scenes that continuously evolve through interaction (spatial memory). Yet existing embodied benchmarks fall short in evaluating spatial understanding: they often introduce strong visual cues that bypass spatial reasoning, and focus on static scenes that offer little insight into an agent's capacity for spatial memory. To address this gap, we introduce GeoMind, a systematic benchmark for structured spatial understanding through tabletop manipulation. GeoMind comprises 54 evaluated task variants organized around spatial reasoning and spatial memory. It also incorporates a scalable pipeline that automates the generation and annotation of task instances with verified targets and executable interactions. Extensive evaluations across diverse embodied agents reveal their consistent weaknesses in spatial understanding. Our findings highlight that these agents struggle to track identities after spatial changes and convert relational or geometric constraints into executable actions. We release the GeoMind benchmark at https://github.com/Notfound-JH/GEOMIND-benchmark.git.

Real-time data analytics has become a key demand for lakehouses. It requires propagating data changes from the data source to the lakehouse in real time, while ensuring ACID guarantees and minimizing resource consumption and impact on analytical query performance. This is highly challenging for a data system according to the RUM conjecture, especially given the read-optimized immutable file storage layer in lakehouses. Existing lakehouses offer only git-style catalog version control on a single table, and they trade data freshness, query performance, and/or resource efficiency to achieve a barely acceptable data-change throughput. In this paper, we propose a novel data synchronization framework, Retina, that enables real-time, cost-efficient, query-performant, and fully transactional data changes in lakehouses. We implement and open-source Retina. Evaluations show that Retina delivers 10-millisecond-level data freshness and over 3.2 million row/s linear-scalable data-change throughput with low resource consumption and query-performance overhead, significantly outperforming state-of-the-art lakehouse systems and lake-compatible HTAP databases.

We introduce FraudBench, an evaluation of large language models' (LLMs) propensities to recommend, assist, or directly take actions that would constitute fraud under U.S. law. Our evaluation contains a novel dataset of 159 scenarios based on real-world U.S. court cases and validated by legal experts. We present each scenario to LLMs via three tasks at varying levels of autonomy: advising users, drafting messages for users to send, and acting directly on behalf of users (477 tasks total). In each scenario, FraudBench re-enacts the facts of a court case to give LLMs an opportunity, but no explicit instruction, to either misrepresent or conceal a material fact in a high-stakes commercial setting--e.g., whether to describe company-funded product testing as "independent" to customers (misrepresentation) or whether to hide unfavorable regulatory investigations from clients (concealment). We evaluate 20 LLMs across 25 experimental conditions (totaling >300,000 generations) and include a preregistered human baseline ($n_{\texttt{annotations}} = 2{,}957$, $n_{\texttt{humans}} = 996$). We score a response as fraudulent when it engages in, assists with, or recommends misrepresentation or concealment. Key findings include: (1) across LLMs, fraud rates ranged from near 0% to over 40%; (2) LLM fraud rates increased with autonomy, with overall rates of 3.0% when advising, 14.7% when drafting, and 21.4% when acting directly, without applying pressure; (3) LLMs sometimes engaged in fraud despite recognizing legal risks; (4) LLMs more often engaged in fraud by concealment than by misrepresentation; and (5) prompt-based interventions substantially affected fraud rates, with a prompt to obey the law reducing fraud rates, while prompts to maximize profits and act agentically increased fraud rates.

Code localizers are usually trained to predict the files a reference patch edits, but the code a fixer needs to read is not the same as the code it changes. We propose Fixer-Guided Localization (FGL): the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch, and the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning. We build the full pipeline on Gemma 4 E4B for the Kaggle Gemma-4 Developer Agent competition, from 1,629 test-verified Kimi K3 trajectories. The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code. On the development set, one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6). Training also changes how the agent works: trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model. An analysis of 13 evaluation runs identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality) and gives practical lessons for training small coding agents. All data, code and adapters are publicly available.

Differentiable simulation propagates derivatives through a physics simulator, turning animation, system identification, control, and computational design into gradient-based optimization. Engines differ in the physics they model and, less visibly, in which mathematical object they differentiate. We separate two questions the literature conflates. The referent is which object the returned derivative belongs to: the finite computation that executed, the converged equation it approximates, or the continuous dynamics. The realization describes the organization of the backward pass: automatically generated differentiation, an equation-level adjoint solve, or a custom adjoint. The referent defines the target against which exactness is assessed; the implementation determines how accurately that target is approximated. Cost depends on the realization together with how the states needed by the backward pass are retained or recovered. Crossing the axes places every method and engine we classify in a cell, separates claims the literature merges, and shows that the apparent trade-off between exactness and memory follows from how methods have been built rather than from a constraint. A second, orthogonal taxonomy covers contact, where the map is non-smooth. We then ask whether a correct gradient helps. Four mechanisms by which useful gradients are lost are cataloged and measured on one testbed, where the advantage of the exact derivative over the equation-level adjoint depends on where the fitting target comes from as much as on the solver budget, and we review the evidence that zeroth-order methods can outperform simulator gradients in the regimes where those mechanisms act. The model classes, systems techniques and applications are surveyed through this lens, and we close with a reporting checklist, a released diagnostic that computes it, and a research agenda.

Training a neural network under homomorphic encryption requires transposes, batch reductions, optimizer updates, and repeated refreshes in addition to forward matrix products. We combine rectangular diagonal packing, feature-major SIMD mini-batch lanes, baby-step giant-step evaluation, and compact optimizer state in an MIT-licensed Rust CKKS implementation. Its CUDA path keeps training and bootstrap ciphertext arithmetic on the GPU, while retaining host key generation and client operations. In a matched 100-update experiment on the main non-interactive MNIST model, device-resident bootstrapping reduces training time from 14,268.295 to 293.075 seconds, a $48.68\times$ ratio against our CPU implementation, with exactly equal decoded parameters and predictions. Including setup and final evaluation gives a $40.67\times$ ratio. A separate GPU run completes ten passes over 4,800 images in 2,967.095 seconds and achieves 90.35% test accuracy versus 90.41% for its plaintext twin. These are single-initialization results; exact CPU/GPU parameter agreement is verified at 100 updates, not for the full 1,000-update run. Earlier completed CPU runs support a larger-batch key-switch advantage and baseline-batch accuracy within the registered tolerance, but the baseline-batch normalized key-switch target is missed. The released ReBoot baseline does not meet the common security-estimator threshold, and its full-setting CPU time is unmeasured. We therefore distinguish measured acceleration of our implementation from conditional layout comparisons and do not claim unconditional superiority over ReBoot.

Autoregressive graph generators typically linearize graphs through flat node-edge walks, which can obscure higher-order structure in large and complex graphs. We introduce a hierarchical graph abstraction and coarse-to-fine generation pipeline that separates coarsening, tokenization, and decoding. This exposes three design axes: community coarsening, hierarchical tokenization, and order randomization. Across these axes, traversal-based hierarchy consistently improves distributional and structural fidelity, while its V.U.N.\ cost depends strongly on the coarsening and randomization policy. In particular, the fidelity--V.U.N.\ trade-off observed under within-community randomization is substantially reduced when traversal order is randomized across the full hierarchy. With a fixed backbone, training protocol, and matched training time across five molecular and five non-molecular benchmarks, hierarchical serialization yields consistent distributional and structural gains on all molecular datasets, while the non-molecular results further show that these gains depend on how well the supplied hierarchy aligns with structure in the target distribution. Together, these results characterize when and how hierarchical serialization benefits autoregressive graph generation.

Over the past decade, a growing body of work has established that implicit biases in learning dynamics fundamentally shape the solutions found by neural networks, governing their learned features and generalisation performance. However, despite these strides in deep learning theory, far less is known about representation learning in generative models based on dynamical systems, from normalising flows to diffusion. Importantly, empirical evidence suggests that continuous-time generative models can be prone to mode collapse, yet a learning theory explaining such biases remains elusive. To address this, we use tools from dynamical systems and operator theory to express the gradient flow of a general class of stochastic differential equations as a composition of linear operators, allowing us to relate collapse to low-rank biases in generative models. In variational inference and flow-matching tasks, we show that this bias can manifest in two different mechanisms: i) where collapsed models are stable fixed points of the gradient flow, or ii) in sequential feature learning, where smaller modes take progressively longer to be learned, producing an effective collapse for a fixed training budget. Together these findings point to a trade-off between feature learning and collapse, suggesting an operator view as a step towards a unified framework for understanding the learning dynamics of dynamical systems-based generative models.

We ask whether the apparent forecasting skill of frozen time-series foundation models (TSFMs) on intraday index-futures returns is incremental skill. In a pre-registered evaluation on E-mini Nasdaq-100 (NQ) and E-mini S&P 500 (ES) futures, a common panel of five zero-shot models (Chronos-Bolt, Chronos-2, TimesFM 2.5, Moirai 2.0 and TiRex), plus separate experiments with Kronos, is compared with a no-change forecast at 5-120 minutes, using native prices, log returns, volatility-standardised returns and a volume clock as inputs. On a blind set opened once, 0 of 160 model-representation-horizon-market cells beat no-change, and a Hansen SPA test gives p = 1.000 on both markets; a later retrospective hold-out, scored as a descriptive check, gives 0 of 120. Input representations change forecast magnitude but not accuracy, and a transformation-aware ridge regression does as well. Several standard measures nevertheless suggest skill. Price-level correlation exceeds 0.9996, but no-change scores higher. Within-day RankIC across forecast origins rewards a trivial revert-to-mean rule (+0.35 at 30 minutes); we derive the expected within-day covariance under a random walk in closed form, and a sign-randomised null reproduces about 95% of the observed value. One model beats HAR for 5-minute volatility, but not the EWMA its input already contains. On strongly planted synthetic signals the models recover part of the signal at 5 minutes, and recovery deteriorates at 30 minutes. Benchmark-level success of TSFMs should therefore not be assumed to carry over to low-signal domains such as intraday returns.

Heavy-tailed bandit learners without moment parameters face an obstruction: the rewards that carry an arm's mean can stay unseen. We address the remaining part of a COLT open problem by characterizing the optimal-arm laws that permit adaptation to unknown moments at constant-factor cost. The condition, variance-core visibility, is one-sided and translation-invariant: the mean hidden in any rare event is at most a constant multiple of the standard error of the rewards that remain. A quantile-trimmed UCB index that knows the visibility constants attains the central-moment benchmark on this class, and no single strategy adapts across all visibility levels. The same index achieves logarithmic expected regret on every instance with a visible optimal arm and finite-mean rewards. On an explicit family with no higher moment, gap $Δ$ and constant $A$, its logarithmic coefficient and the information lower bound are of order $Δ^{-1}e^{A/Δ}$, exponential in the inverse gap within one visibility class.

Good views do not suffice to guarantee scene invariance or uniqueness up to rotation at the global optima of Gaussian-regularized JEPAs. We give globally certified counterexamples showing that the uniqueness premise of the Gaussian extension in Good Views fails for specified soft objectives. Gaussian marginal matching can reward retaining independently resampled observation variation, and this benefit can exceed its prediction cost. In a four-state model, comparison against the entire scene-only class proves that every global minimizer splits a scene and admits an equal-value partner outside its rotation orbit. In a 512-dimensional model with a complete scene basis, a spectral gap, and exact population whitening, pooled Epps-Pulley regularization still forces every global minimizer to retain nuisance and yields positive cross-modal alignment error even after the best rotation when modality nuisances are independent. Better scene alignment can also accompany worse prediction. With matched 512-label readout budgets and 1% observation noise, whitened prediction JEPA aligns better to the scene than next-token prediction but incurs 0.2872 nats/token higher linear-readout test NLL. A Bayes-risk decomposition distinguishes information lost through the observation channel from the fitted readout's difficulty using retained information. These results connect three determinants of learned representations: what views share, what regularization favors, and what a prediction task can use.

One-shot Federated Graph Learning (OFGL) has emerged as a communication-efficient paradigm for federated graph learning under strict bandwidth constraints, compressing collaboration into a single transmission round. Nevertheless, this extreme setting suffers from a fundamental Single-Transmission Bottleneck, which limits both the quality and utility of exchanged knowledge. First, while existing methods attempt to distill private graphs into shareable proxies, their aggressive one-shot compression inherently triggers a fidelity crisis. This compression distorts local graph topology and semantics across clients, leading to externalized knowledge underexpression. Second, although the server aggregates these proxies to update global models, current integration schemes suffer from transferability agnosticism. The server cannot reliably identify cross-client collaboration potential, thus resulting in integrated knowledge underutilization. To tackle these dual challenges, we propose FedHFT, an effective One-Shot Personalized Federated Graph Learning (OPFGL) framework that addresses this bottleneck from two perspectives, namely High-Fidelity Knowledge Externalization and Transferability-Ware Knowledge Integration. We use Fidelity-Driven Proxy Refinement to preserve client-side knowledge fidelity, and Transferability-Steered Personalized Collaboration to enable adaptive server-side collaboration. The superiority of FedHFT is validated through extensive experiments on both homophilic and heterophilic graphs. The code is available at https://github.com/Jumpy-Juice/FedHFT.

Agentic language model systems can improve performance either by iterating on their own outputs or by using external tools in grounded environments. While these strategies can reduce error rates and substitute for larger models in some settings, their environmental cost remains relatively underexplored. We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity. Study~1 evaluates iterative verbal self-refinement: GPT-4.1-mini at $k = 1$--$5$ steps versus a GPT-4.1 single-pass baseline on two MATH benchmark splits. Study~2 evaluates tool-augmented agentic coding: GPT-4.1-mini, Claude Opus~4.6, and GPT-5.3-codex on SWE-bench Lite, where agents operate inside Docker-isolated environments with access to code execution and file editing tools. In Study~1, $k=3$ refinements close a 16-percentage-point accuracy gap to within 1~point, but nearly triple the carbon footprint relative to the cheap single-pass baseline while remaining below the GPT-4.1 single-pass baseline under the current token-based carbon scenario. In Study~2, GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\%), ahead of Claude Opus~4.6 (68.9\%) and GPT-4.1-mini (25.0\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task, while GPT-5.3-codex delivers substantially higher task success. All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.

Hybrid reasoning models interleave attention with recurrent mixers such as Mamba-2 and Gated DeltaNet, whose state is cached between decoding steps. Prior work showed that storing this state in FP16 or INT8 with round-to-nearest (RTN) makes Mamba-2 and GDN hybrids more verbose, and that stochastic rounding (SR) fixes it. We study the other 16-bit format that serving engines expose, bf16, on six open hybrid reasoning models and 590 competition-math problems. With bf16 RTN state, five of six models reason longer (geometric-mean length 1.07x to 1.27x; also under greedy decoding in the two models tested) and hit the 32K token cap more often. On Qwen3.5-9B, cap hits rise from 118 to 223 and accuracy falls from 85.6% to 70.8%. On LiveCodeBench, both models tested also reason longer and lose 12.6 and 13.4 points of pass@1. Rerunning the capped items with a 128K budget halves the accuracy gap but does not close it (93.6% vs. 86.6%): bf16 runs that never finish are twice as common, and 85% of them are exact repetition loops. On Nemotron-Nano-9B-v2, by contrast, 30 of 31 extra caps finish at 128K, so the cost there is mostly extra tokens (exploratory). Same-width bf16 SR removes the length increase in four of five affected models, restores Qwen3.5-9B to 87.6% and brings code length and pass@1 near FP32 levels, so the failure is mostly rounding bias, not missing resolution. With an integer-hash random generator, SR on Qwen3.5-9B matches FP32 on cap hits and accuracy (113 vs. 118; 85.6% for both) at the memory and decoding speed of bf16 RTN. In the Nemotron-Nano-v2 lineage, SR on 25% of heads (an intermediate-decay band) removes the effect, while random heads and the top teacher-forced-KL octiles do not; in Nemotron-3 and Qwen3.5-9B the effect is spread across heads. We recommend that evaluations of hybrid models report the state dtype, cap rate and output length alongside accuracy.

External tool calls are a major running cost of an LLM agent, while its training objectives align almost exclusively with task completion. In a step-level diagnosis of Qwen3.5 trajectories at two model sizes on GAIA, a long-horizon tool-use benchmark, we find that most contain removable calls, at a similar rate at both sizes. Some methods add an efficiency reward to the objective to reduce the call count, but introduce problems of their own: the anchor they price calls against is computed from the policy's own rollouts (e.g., the group's shortest correct rollout) and so moves with the policy; training can drift toward a collapsed policy that issues no tool call at all, a degeneracy seen in published reports and in our 27B baseline; and a training run delivers a single operating point, so the user cannot trade efficiency against capability. We present CARE-PO (Controllable Anchored Relative Efficiency Policy Optimization) to mitigate these problems. The efficiency reward prices calls against a reference budget anchored to the task's own difficulty and fixed before training. Injecting a successful trajectory from the training data into each online GRPO rollout group and charging for evidence a trajectory never gathers ease the drift toward a collapsed call count. Shaping the cost curve by an effort tag in the prompt yields an efficiency and a capability operating point from a single training run. On GAIA, five of the six operating points across three model sizes improve on their base model in success rate and call count at once. On four held-out multi-hop benchmarks the result is size-dependent: at 27B and 4B success rate is held or raised at a lower call count, while at 9B both operating points trade success rate for fewer calls. The call count remains well above zero throughout training at all three sizes.

AI agent harnesses play the critical role of providing LLMs with an interface to a computing environment. Recent studies show that AI agents' task-execution accuracy depends mostly on the LLM used, while harness-side variances such as system prompts or tools have minimal impact. Rather, such harness-side variances show a larger impact on the cost of agent executions, achieving cost reductions up to 2.6 times on the same model and task. To further explore these harness-side cost optimization opportunities, we analyze the cost accumulation patterns across various agent traces and identify system opportunities to reduce token use and LLM calls without modifying the model or affecting task accuracy. Based on this analysis, we design Agency, a harness interposition platform that provides a unified abstraction and allows easy and portable implementation of harness-side system optimizations on existing off-the-shelf harnesses such as Claude Code or OpenAI Codex. Using Agency, we implement proof-of-concept optimization techniques based on our analysis and achieve cost reductions up to 35% across various real-world benchmarks on existing harnesses without compromising accuracy.

Autonomous agents must continually acquire reusable knowledge from experience without relying on costly large language model (LLM) reasoning for every decision. However, existing experience-based approaches often struggle to transform accumulated observations and reasoning into compact knowledge that lightweight models can independently apply. This paper proposes SCALE, a small-model-first framework that integrates selective LLM reasoning, experience-driven concept abstraction, and knowledge internalization. A lightweight model handles routine predictions, while prediction errors and accumulated cross-episode evidence selectively trigger LLM reflection. The LLM analyzes relevant historical experiences, generates positive and negative counterfactual examples, and proposes reusable concept hypotheses. Independently validated concepts are progressively incorporated into hierarchical neural representations, enabling lightweight models to exploit increasingly abstract knowledge. Meanwhile, mastered experiences are excluded from active memory while their original records remain available in a complete archive. Experiments across ten random seeds show that historical recall improves held-out accuracy from 87.11% to 98.28%, hierarchical concept knowledge increases difficult-case accuracy from 55.33% to 66.47%, and abstract archiving reduces active experience records from 6,000 to 1,528.

Full-duplex voice agents enable natural, hands-free interaction, allowing users to refine requests, redirect ongoing work, and seek clarification while the agent is speaking. However, existing voice-agent benchmarks insufficiently isolate coordination between task execution and conversation. We introduce a benchmark that assesses this coordination by examining whether agents track what is being executed, what has been completed, what has been communicated, and what remains to be reported as the conversation evolves. We propose an interactive user simulator that introduces configurable interventions at designated timing in tool execution and spoken result delivery, supporting controllable and reproducible evaluation conditions. We evaluate five leading commercial voice agents along three dimensions: execution responsiveness, coordination under request changes, and coordination under interruptions. Agent evaluations reveal persistent coordination failures. Under request changes, the highest result completion rate is only 55%, indicating difficulty for current voice agents adapting to new requests while preserving earlier obligations. After user interruption, four of the five agents exhibit average information omission rates of at least 23%. These findings highlight task coordination as a central challenge for current voice agents in complex, evolving conversations.

Strong machine-learning ensembles capture much of the trend, seasonality, and known events in demand, yet they can still understate peaks, miss moving events, or mistake temporary changes for shifts in demand level. We address this gap by framing forecast refinement as a reward-driven decision problem. For each forecast period, we let an agent choose among values already produced by the ensemble: its base forecast, distribution-free bounds, quantiles, and a bias-corrected estimate. We evaluate these choices against context-matched historical periods, accounting for calendar position, recurring and moving events, trend, price, and promotions. We then apply a structural seasonal guard to constrain adjustments and record the model, agent, and planner contributions to every published value. We evaluate our approach across five retail, consumer-goods, and direct-to-consumer deployments over nine months. Our refined forecasts reduce mean WMAPE by 6.3% compared with the stacked machine-learning forecast for every customer, with gains reaching 16.7%, and achieve the lowest error in 35 of 45 customer-months. Compared with statistical forecasts, our approach reduces mean error by 18.7%. These results demonstrate that context-aware, auditable action selection can add forecast value beyond a strong ensemble.

Full-pipeline data analysis over real-world data lakes requires LLM agents to locate and integrate information across large collections of tables with different schemas. Existing data agents follow task-specific exploration, which exhausts the context window in massive lakes. As the context fills, the agent's reasoning degrades, leading it to locate and process incorrect data files. These errors propagate into the final answer. To overcome these limitations, we introduce Ontology-Grounded Data Agent (OGDA), which utilizes a compact ontology to reduce exploration overhead and constrain data processing. OGDA constructs the ontology through agentic hypothesis testing. At each step, a proposer agent induces schemas, concepts, and relations from raw sources, and a verifier checks each proposal against the underlying data. During task execution, the ontology serves as a representation of the data lake for the agent to explore. Every data-processing action is constrained by a task graph, which the agent can extend only after observing the semantic meaning of the data. Across KramaBench and DiscoveryBench with three LLM backbones, OGDA outperforms the strongest baselines by up to 10.82 points. These results suggest that explicitly modeling the data lake is a key step toward reliable end-to-end data agents.

Screenshot-to-Code requires coding agents to sustain visual understanding across code generation, execution, and revision. We introduce LiveWorkspace, a visual coding harness that externalizes this understanding as Shared Visual State (Shared VS), an implementation-independent representation of visual hierarchy, node attributes, and layout relations. LiveWorkspace constructs and retains \textit{Target} from screenshots and recovers \textit{Current} from browser execution in the same schema. Structured Diff and implementation bindings guide code modifications; user edits update \textit{Target} and reuse the alignment loop. Using GPT-5.2 for both methods, LiveWorkspace improves end-to-end Avg by 1.8 points on Design2Code and 2.0 on ScreenBench over a control with matched model-call counts and stagewise budget limits. Further experiments demonstrate reuse of constructed Shared VS across coders and improvements in iterative repair. LiveWorkspace also achieves 93.0\% edit success across 100 controlled editing cases. These findings support treating visual understanding as a persistent shared working state throughout code generation, execution, and modification.

During reinforcement learning, experience can be both internalized in LLM parameters and retained externally as reusable textual skills. Parameter updates alter the model's behavior and the interaction experience from which new skills are extracted. These skills, in turn, guide the model's subsequent interactions and parameter updates. As the model's capabilities evolve, previously beneficial skills may become redundant or detrimental. We present SkillDuet, a framework that coordinates parameter updates and skill evolution through model response feedback. Skills extracted from on-policy trajectories guide parameter updates through OPD at every step. A response-based score, Strength, ranks candidate skills by the magnitude of the shift they induce in the current model's responses, and the highest-scoring skills are periodically consolidated into a persistent global skill set (Global). The updated model and Global jointly guide subsequent interactions, generating experience for further parameter and skill updates. With Qwen3-4B, SkillDuet achieves success rates of 98.57% on ALFWorld and 92.97% on WebShop, exceeding the matched SEED-API baseline by 2.86 and 5.47 percentage points, respectively. Ablations show that response-based skill selection and Global consolidation improve task performance. Further experiments show that the benefits of retained skills depend on both their content and their compatibility with the model. SkillDuet provides a concrete mechanism for model-skill co-evolution and skill-level recursive self-improvement.

The rapid integration of artificial intelligence (AI), particularly generative and machine-learning systems, in climate services development and provision is transforming how climate information is produced, accessed, and interpreted. AI-enabled climate services offer significant opportunities for process automation, optimisation, tailoring, and the translation of complex climate data into actionable knowledge for diverse users. However, the growing reliance on privately controlled algorithms, data infrastructures, and computing facilities raises critical concerns related to governance, transparency, equity and economic sustainability, among others. While the technical architecture of climate services increasingly relies on advanced machine learning and large-scale climate models, their legitimacy as public services may be weakened in the absence of arrangements that prioritise public value and long-lasting societal benefit over profit maximisation. The tension between the public mandate of climate services and the private nature of AI infrastructure challenges traditional notions of openness. Using a consensus-based elicitation method coupled with the PESTLE framework, we analyse the political, economic, social, technological, legal, and environmental risks and opportunities shaping the co-production value chain of AI-enabled climate services. We draw policy-relevant conclusions for data, model and infrastructure supporting the next generation of upstream and downstream climate services as critical applications for society. Our work presents potential and plausible challenges and benefits for a pan-European, decentralised public climate service built on sovereign AI infrastructure and open governance principles.

AI systems increasingly appear in criminal investigations as tools, targets, and sources of evidence, yet AI forensics lacks a systematic account of the artifacts they produce: investigators often do not know which traces to look for, and for many traces they do find, no established analysis technique exists. In this paper we first developed the Forensic AI Life Cycle (FALC) that unlike other life cycle models shifts the focus onto the forensic artifacts created during the life cycle of an AI system. Through FALC we derived a reference list of artifacts that might serve an investigation. Second, we collected examples for crimes with AI involvement throughout FALC and questions likely to be asked when investigating these crimes. Third, we propose a taxonomy of forensic fields of expertise with a focus on the current possibility to analyze AI forensic artifacts using established techniques. For investigators, the model provides a systematic basis for identifying which artifacts to look for when an AI system is relevant to a case. For researchers, the classification shows where existing forensic methods already suffice and where new techniques are most urgently needed.

Late Chunking encodes an entire document before pooling at chunk boundaries, preserving document context in local retrieval units. Chunk-level post-training improves context-dependent retrieval, but shared context can make semantically similar same-document negatives difficult to distinguish. We study the local decision boundary between a gold chunk and its hardest sibling. Starting from InSeNT, we analyze InSeNT's within-document gradients and show that the entropy of the conditional negative allocation decreases monotonically with the scaling factor at fixed logits. We propose a two-part framework for calibrating document-local chunk boundaries. The first component adapts the discriminative scale from within-document margins, while the second exposes chunk partitions to attention during document encoding. Experiments on ConTEB and NanoBEIR show that the resulting model improves contextual chunk retrieval across the evaluated encoder settings while retaining competitive passage-level retrieval quality; ablations further characterize the distinct roles of adaptive scaling and chunk-aware routing.

Biomedical ontologies are structured knowledge systems that must evolve continuously to represent newly emerging knowledge, but their manual maintenance remains time-consuming and laborintensive. Textual definitions are especially costly because they must be readable to humans and compatible with controlled entities, relations, and logical constraints. Large language models (LLMs) offer a promising way to assist ontology engineering tasks, including definition generation. However, textual plausibility does not guarantee knowledge-system compatibility. A generated definition may be fluent and close to a reference definition, yet still fail to ground its implied entities or violate ontology-level constraints. Directly inserting such definitions into ontology libraries may therefore introduce entity-level hallucinations and logical inconsistencies. To address this problem, we propose two structured-knowledge incompatibility metrics, Entity Error Rate (EER) and Logical Inconsistency Rate (LIR), to quantify entity grounding and logical consistency for LLM-generated ontology definitions. We further develop an agentic quality-control (QC) workflow that treats definition generation as an iterative process rather than a direct decoding step: it generates definitions, extracts axioms, checks entity existence and logical consistency, and uses the resulting diagnostics to drive repair loops. Comprehensive experiments on Disease Ontology, Uberon Ontology, and Uberon-ConstraintStress demonstrate that EER and LIR complement conventional reference-based metrics by evaluating whether generated content can be grounded and checked inside the target ontology. The workflow effectively decreases EER while maintaining strong conventional text-generation performance. LIR further exposes sparse ontology-level logical failures that are invisible to text similarity metrics, although their repair remains limited.

One of the fundamental needs for reaching resilient production infrastructure is flexible manufacturing, which requires translating changing production requirements into implementable system configurations. This paper introduces \textbf{CRAFT-AI}, an agentic AI approach combining natural language (NL) requirement interpretation, Asset Administration Shell (AAS)-based capability descriptions, and reusable engineering knowledge from existing manufacturing configurations. A tool-using AI agent matches requirements with available capabilities and reference implementations to generate manufacturing scenes and corresponding AAS instances. The approach is demonstrated using a Factory I/O prototype with twelve reference scenes and five illustrative manufacturing tasks of varying complexity, providing an initial feasibility assessment of the proposed AAS-enabled, system-level generative engineering approach. For the four valid user task descriptions, proper requirements, correct references to the relevant engineering knowledge, correct components and lastly a valid Factory I/O scene are generated. For the task with missing engineering knowledge and a wrong sensor selection, the user is informed about which requirements of the scene generation lack domain knowledge and also which sensor is not compatible with the selected raw material by the user with link to the related documentation. After fixing the task description, the correct scene is seamlessly generated. \textbf{CRAFT-AI} highlights the importance of safeguards for risk perception and communication in foundation model-driven solutions.

Wearable Inertial Measurement Units (IMUs) enable pervasive Human Activity Recognition (HAR), but traditional Deep Neural Networks (DNNs) suffer from an "arithmetic-latency deadlock" due to dense floating-point MAC operations. While Spiking Neural Networks (SNNs) offer ultra-low power, existing ones face spatial blind spots, rigid neurons, and gradient degradation. To address these, we propose TSFFormer, a purely spiking Fourier Transformer for Green HAR. It introduces two key innovations: (1) a MultiHeadFourier attention that performs frequency-domain global mixing, reducing complexity from \(O(V^2)\) to \(O(V\log V)\), eliminating spatial blind spots, and saving \(\sim\)50% GPU memory; and (2) a dual-compartment Temporal Spiking LIF (TSLIFNode) that utilizes dynamic thresholds to adapt to non-stationary rhythms and mitigate gradient degradation. Extensive experiments on 8 public IMU benchmarks demonstrate that TSFFormer surpasses state-of-the-art SNNs and even the classic DeepConvLSTM, achieving a \(\sim\)3\(\times\) faster training speed with an energy footprint of 0.1 pJ. Ablation studies confirm the synergistic effect of our modules, providing a highly accurate, ultra-low-power event-driven paradigm for wearable sensing.

Slay the Spire 2 (STS2) is a deck-building roguelike in which a winning run at the highest difficulty takes several hundred decisions, and most losses happen in late boss fights that a learner seldom reaches. We train agents that choose every action with one forward pass of a 6.6M-parameter entity transformer, without lookahead. Training and evaluation run the game's own compiled assembly (v0.107.1) in a headless .NET host. The host replays a run exactly from its seed and actions, so training episodes can restart from states the learner has visited, and a failure-driven curriculum allocates these restarts in proportion to where complete runs are lost. On 12,800 held-out games at Ascension 10 (A10), the released policies win 8.05% averaged over the five characters (Ironclad 15.6%, Regent 12.5%), and AlphaSpire's released networks used without search win 0.55% on a common panel. Longer A10 training raised Ironclad from 12.4% to 15.6% and Defect from 3.0% to 8.1% on these games. At Ascension 0 the released policies win 80.2%, level with the 80.3% that the tree-search agent Spire Mentor reports, while Mentor leads by 5 to 8 points from A4 upward. Run in our host on the same 2,000 A10 games, Mentor's search wins 8.8% against 8.2% for our policies, which need 21-27 ms per decision on one CPU core. GPT-6 Astra at maximum reasoning won none of 11 A10 games through our text interface. In a controlled study of 18 training runs from six random initializations, failure-driven allocation exceeds uniform restart allocation on each of three Regent initializations, by 3.6 percentage points on average, while the Silent effect is small and changes sign. Code, trained policies, an in-game mod and the records behind every table are at https://github.com/HalfVulpes/sts2ai.

Repository-level code-understanding benchmarks have progressively expanded the spatial scope of reasoning, from local code to entire repositories, while largely treating each repository as a static snapshot. However, how models reason about repository evolution over time remains largely unexplored. We introduce SWE-Time, a benchmark for temporal repository understanding that makes repository evolution itself the reasoning target. SWE-Time evaluates two complementary temporal directions: retrospective understanding of repository evolution through historical regression analysis and prospective understanding of intended changes from pre-change repository states. To enable realistic yet reliable evaluation in these settings, we introduce the RealWorld-to-Verifiable Construction Framework, which transforms real-world developer queries and historical evidence into verifiable evaluation tasks. We evaluate 15 coding agents on 1,661 cases from 65 GitHub repositories. Within Claude Code, Claude Opus 4.8 achieves the highest SWE-Time Avg. of 61.9, leaving substantial headroom. Controlled analyses highlight the importance of historical code evidence and developer intent, while consistent reasoning across repository states remains challenging. These results highlight repository evolution as a distinct and challenging dimension of code understanding.

We study how much the interaction strategy around a fixed, locally hosted 7B-parameter code model (Qwen2.5-Coder-7B-Instruct) affects its ability to replace PyTorch operators with custom CUDA kernels. On 30 KernelBench tasks (levels 1-3, first 10 each, evaluated on a V100), we compare four regimes of increasing structure: zero-shot prompting, profiling-guided prompting, iterative refinement on execution feedback, and a multi-agent loop (analysis, retrieval, generation, evaluation, feedback triage) in which the same model fills four roles around a non-LLM evaluator. Three findings stand out. (1) Correctness, not speed, is the binding constraint: even correct generated kernels average 0.97-0.99x the speed of eager PyTorch, because a 7B model cannot out-write cuBLAS/cuDNN on dense GEMM and convolution; the only real wins are fused memory-bound epilogues. (2) The best strategy depends on difficulty: profiling-guided prompting is strongest overall (43% correct, 27% faster than baseline), but the agent loop is the only method that solves any level-2 fusion problem (3/10 vs. 0/10 for all others). (3) Naive agentic scaffolding is actively harmful: our first agent scored 3% correct, below zero-shot's 30%, and the decisive repairs were matching the agent's ambition to the model's competence and spending compute on best-of-n sampling under a real execution oracle (3% to 33% correct from n=1 to n=8), not on more reasoning steps.

We present SadaqahStories, a parallel English-Arabic corpus of 41 illustrated four-page stories about charitable giving in Islam (Sadaqah, Zakat and Waqf), written for readers aged 7 to 18. The stories were first drafted with large language model (LLM) assistance, then audited against hadith collections, Qur'an references and historical sources, and finally rewritten for children and signed off by an Islamic scholar. Every story carries an evidence label and links to its sources. We release all three versions of each story (AI-assisted draft, audited text, final child-oriented text) together with the correction log, so the corpus can be used to study how AI drafts of religious and historical content go wrong. The audit found substantive errors in 39 of the 41 drafts. The most frequent were unsupported specific details (dates, figures, prices, names), invented quotations, institutions or anecdotes, and conflation of separate events or people; six drafts contained errors in the wording, grading or ruling of a religious text. Rewriting cut the estimated Flesch-Kincaid grade of the English text from 14.0 to 6.1. We also release an 81-entry charity lexicon covering Modern Standard Arabic and Qatari Arabic, and describe the free, account-free website that serves the corpus (https://sadaqahjourney.com).

Deep learning models, particularly convolutional neural networks (CNNs), are widely adopted in computer vision, yet their opaque decision-making raises concerns in high-risk applications like healthcare and autonomous systems, where trust and accountability are critical. Therefore, robust explainable AI (XAI) methods are desirable to provide transparent insights into model decisions. To ensure these methods are reliable and effective across diverse scenarios, it is important to understand how model performance and architectural complexity influence explanation quality. Our study takes an initial step toward systematically investigating the impact of these two factors on key XAI evaluation metrics. We constructed a pipeline using 10 subsets of ImageNet-1K, each containing 80 unique classes, on each of which we trained three CNN architectures of increasing structural complexity: VGG16, ResNet50, and ConvNeXt-Tiny. For each individual model, we analyzed 10 training checkpoints representing increasing levels of model accuracy using two XAI methods: Grad-CAM (feature-attribution-based) and CRAFT (concept-based). We assessed these methods across four evaluation categories: Faithfulness, Correctness, Sensitivity, and Simplicity. Our findings reveal that as training progresses, the complex architectures (ResNet50 and ConvNeXt-Tiny) exhibit improved Faithfulness and lower Sensitivity, indicating more robust explanations. In contrast, VGG16's simpler architecture yields simpler but less faithful and correct explanations.

Long-term memory lets LLM agents personalize, stay consistent, and reuse what worked before. Parametric memories need training, consolidation can lose details for good, and keeping every record leaves a growing store to search. We present MIRIX-2.0, a multi-agent memory system designed to be practical, accurate, and efficient at once. It is the most accurate of the compared systems on four of five QA benchmarks. It reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, where it leads the strongest baseline by 14.3 points, and with AutoDream disabled it gains 31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA. As procedural memory, it yields the highest mean success among the compared memory methods on ALFWorld and WebShop and reaches 49.2% on MetaClaw-Bench, against 44.0% for MetaClaw's skill library. Six typed memory agents keep complete records as the system of record, and a provenance hypergraph indexes episodic and semantic records with n-ary frames that cite their rows. Retrieval runs under explicit budgets, so the evidence a question inspects does not grow with the history. AutoDream merges duplicate anchors outside the write path without rewriting any record, and it distills task sessions into skills. With AutoDream disabled, MIRIX-2.0 answers faster than HyperMem on all five timed benchmarks. It needs no fine-tuning, runs with closed and open models, and its code is public.

Machine learning (ML) services increasingly expose explainability outputs alongside model predictions. While such explanations are intended to improve transparency, they may also reveal information about the underlying training data. In this work, we show that the local surrogate geometry employed by Local Interpretable Model-agnostic Explanations (LIME) can be repurposed as a proxy for a sample's distance to the model's decision boundary. We extend this boundary-distance view to tabular counterfactual explanations and, for the first time, to visual counterfactual explanations for image classifiers. We present XposedMIA, a one-shot, explanation-based black-box membership inference attack (MIA) framework that converts the returned explanation into a proxy for the queried sample's distance to the model decision boundary. After a one-time calibration step on a small labeled calibration set of as few as 30 samples, our attack requires only a single query for each target sample. Unlike prior label-only attacks requiring iterative queries or shadow training, our method relies solely on the predicted label and its accompanying explanation. Our evaluation indicates that XposedMIA achieves high effectiveness, with success rates of up to 0.95 using only a single query per target, revealing that releasing explanations introduces an additional channel for membership leakage.

A model that takes over another model's conversation must normally re-read the whole history; a text hand-off shortens the re-read only by dropping content. KV-cache translation instead maps the source's cache into the target's own format, so the target continues from the full context without re-reading it. It works for grouped-query attention (GQA), but open-weights frontier models now cache compressed latents (MLA), sparse-attention indexer keys and fixed-size linear-attention states. We replicate KV-Lingo on a GQA pair and then give, for each of these objects, where to capture and inject it and a mechanism that reaches native quality. We study two hybrid pairs at 16-48B scale in both directions, Kimi-VL-A3B - Kimi-Linear-48B (MLA - MLA + Kimi Delta Attention) and Qwen3.5-27B - Qwen3.8-27B (Gated DeltaNet hybrids), plus a sparse-attention target (Qwen3.8-27B - Qwen3.8-Flash-Next). MLA latents translate with token-wise linear maps; recurrent states are copied with a closed-form correction or synthesised by replaying the target's own recurrence. Exact long-range recall needs token identity: decoding token ids from the source's layer-0 MLA latent raises needle recall from 0.09 to 0.99 at 2-32K and to 0.83-1.00 at 64K, eight times the longest training prompt. The hybrid full-MLA direction is within the paired confidence interval of native quality on every task metric, and same-shape GDN hybrids reach the native level in both directions (apart from 6% refusals inherited from the source), with time to first token 18.7 and 46.8 times shorter than a re-read at 64K tokens. Segment-wise multi-turn transfer stays at the re-prefill level; naive re-translation collapses. On the sparse-attention target, predicted indexer keys recall the needle as well as the target's own keys (one seed). We state as a hypothesis how the recipes transfer to 300B+ hybrid flagships.

We present BARS (Beat-Adaptive Rap Synthesis), a non-autoregressive rapping voice generation model based on Flow Matching, conditioned on a beat track (instrumental backing), lyrics, and a short reference voice for zero-shot timbre control. Unlike prior autoregressive rap synthesis, where rap pace and output duration are strictly determined by the beat, BARS enables independent control over the rap tempo, allowing flexible-length generation without explicit duration prediction or phoneme-level alignment. We encode lyrics using a pretrained ByT5 encoder with a lightweight Transformer adapter trained jointly with the model, without relying on speech-pretrained text encoders, enabling natural handling of non-standard orthography (e.g., slang, abbreviations) and numerals common in rap lyrics. To address alignment instabilities, such as token skipping or repetition, under strong beat conditioning, we propose Asymmetric Scaled RoPE for cross-attention, which introduces a positional structure to facilitate sequential lyric-audio alignment learning. Furthermore, the model supports partial re-generation via span masking, enabling lyric-level editing while preserving the surrounding vocal style. Experiments demonstrate that the proposed model achieves high lyric fidelity, natural rap flow that is rhythmically coherent with the beat, and robust tempo control across a wide range of rapping styles and speeds.

Large language models offer a promising technique for translating natural-language optimization problems into solver code, but their reliability remains constrained by the quality of template retrieval. While recent structured methods improve over flat prompting through hierarchical taxonomy search, their traversal rules remain fixed or LLM-driven at inference time, leaving retrieval unable to improve from past successes and failures. This motivates treating template retrieval as a policy that can be learned from modeling experience, rather than as a static inference-time rule. We propose COMPASS, a policy-amortized framework for retrieving optimization templates in LLM-based solver modeling, which represents the template library as a directed acyclic graph and formulates retrieval as a sequential decision process over template nodes. A deep Q-network learns a reusable traversal policy from template-level judgments and execution feedback, reducing reliance on fixed heuristic traversal through feedback-driven graph navigation. To support extensible retrieval, COMPASS embeds problem descriptions and template nodes in a shared semantic space, allowing newly inserted templates to be scored from their text descriptions without redefining the action space. Experiments across multiple LLM backbones and optimization-modeling benchmarks demonstrate that COMPASS consistently improves over heuristic structured retrieval, with the largest gains on complex and mixed-category problems.

Online Conformal Prediction (OCP) sequentially constructs prediction sets whose average empirical coverage converges to a target level over time, for arbitrary, potentially adversarial datasets. In the multi-level setting, prediction sets are produced at $d$ coverage levels at once. Prior work argued for the use of online gradient descent with isotonic projections, which produces sets nested across levels, but with a $\sqrt{d}$ dependence in the per-level coverage convergence rate. Here, we both generalize and improve prior work. First, we show that isotonic projections can be used for the construction of multi-level online conformal prediction sets from \emph{any} online learning algorithm (not just gradient descent). Then, through a black-box reduction, we show that the linearized regret of the online algorithm itself automatically gives control over both per-level coverage and distributional consistency of multi-level predictions via Fenchel duality. Finally, by using the Weighted Interval Score, a standard scoring rule for assessing multi-level interval forecasts, we prove that these projections reduce both the size of the sets and their under-coverage. We study two instantiations of our theory, $p$-norm Online Mirror Descent and coordinate-wise Universal Portfolio methods, both improving the dependence on the number of levels to $\sqrt{\log d}$ for the multi-level OCP problem. Experiments on several datasets show favorable performance compared to baselines.

The Natarajan dimension is a fundamental measure of the capacity of multiclass hypothesis classes and plays an important role in analyzing their generalization ability. In this paper, by establishing lower and upper bounds, we show that the exact Natarajan dimension of homogeneous multiclass linear predictors with pairwise distinct class weights is \(n(k-1)-1\), where $n$ is the dimension of the feature space and $k$ is the number of classes, for \(n,k\ge3\), \(n,k\in\mathbb N\).

Training Gaussian processes (GPs) at large scale is computationally challenging because evaluating the log-marginal likelihood and its gradients involves large, dense kernel matrices. We develop a scalable low-rank GP training approach that combines AnchorNet landmark selection with a fully independent training conditional (FITC) correction. The resulting objective and its analytic gradients can be evaluated using low-rank factors, diagonal quantities, and small dense matrices, without forming the full kernel matrix or its derivatives. The complete training procedure is implemented in single precision and evaluated on nine large-scale regression datasets. Across these problems, the method reaches the reference predictive accuracy while substantially reducing training time, although the required approximation rank varies across datasets. We also study the numerical behavior of the full-kernel predictive solve and show that single precision can become limiting for ill-conditioned problems. Finally, we investigate further reducing selected matrix operations to fp16 or TF32. These lower-precision variants provide only modest additional speedups and can become unreliable when the reduced system is poorly conditioned.

Constrained preference-conditioned MORL typically assumes full task observability, treating preferences and safety budgets as stationary task signals. In real-world deployment, however, these signals are often unreliable proxies: intent may drift, safety tolerances may shift with context, and exposed descriptors may be noisy, delayed, or incomplete. We therefore formulate constrained preference-conditioned MORL as a partially observed task-specification problem, where the agent must navigate a fundamental conflict between unreliable external signals and latent interaction dynamics. We propose \textbf{BECON}, a belief-conditioned framework for constrained MORL with drifting preferences and budgets, which infers latent task context from recent histories to estimate preferences and safety budgets. These estimates are fused with the exposed task signal via uncertainty-dependent gates, enabling the policy to defer to observations under high epistemic uncertainty while correcting them when history is informative. The fused context jointly conditions preference optimization and context-adaptive constraint enforcement. Experiments on diverse multi-objective control tasks demonstrate that BECON consistently improves over state-of-the-art MORL baselines, achieving stronger robustness and constraint satisfaction under task drift.

Floorplanning is an early step in turning a logical circuit into a physical chip: it positions large circuit modules and constrains the space available for subsequent logic placement and wiring. These decisions affect congestion and timing, but their physical consequences emerge through downstream implementation. For AI agents, the task requires translating engineering constraints into spatial decisions whose quality cannot be judged by successful tool execution alone. We present an evaluation framework that connects submitted layouts to a prescribed downstream implementation flow, checks geometric validity, measures physical quality, and reports resource use. On one industrial design whose 99 macros occupy approximately 71\% of a fixed L-shaped layout region, we compare our specialized system Nova-SI with general-purpose engineering agents and agents restricted to file access, obtaining complete physical measurements for 39 layout outputs. Under our 0--100 composite quality score, Nova-SI achieves a mean of 57.97 without online feedback from the prescribed power-grid and placement evaluation, compared with 31.17 for the strongest measured general-purpose configuration, GPT-6 Astra. The difference is supported by raw congestion and timing measurements and persists across scoring-sensitivity checks. Layout-level analysis further shows that similar global utilization can conceal substantially different local congestion. Together, the protocol and case study assess agents through the physical consequences of their delivered layouts and provide evidence about system performance on this industrial instance.

Large language model agents increasingly rely on reusable natural-language skills, yet growing skill collections remain accumulations of isolated text artifacts that are difficult to compare, deduplicate, and compose. We argue that the reusable content of a skill lies not in its surface wording but in its procedural transitions: which operation to execute under a given state and which observable outcome to produce. Based on this observation, we propose State-Move-Outcome (SMO), a compilation-like framework that decomposes raw skills into reusable procedural bricks and organizes admitted bricks as an externalized symbolic structure. When compiling a skill, SMO first retrieves and composes existing bricks, dynamically connects them through Outcome-to-State conditions, and creates a candidate brick only for an uncovered procedure; the candidate becomes reusable only after validation and admission. Experiments on SkillsBench v1.1 demonstrate that SMO represents 232 rawSkills from 87 underlying tasks with 95 shared bricks, reducing the corresponding skill representations from 366,472 to 149,154 tokens--a $\mathbf{59.3}\times$ reduction and $\mathbf{2.46}\times$ compression. Across 232 skill-conditioned evaluation instances, SMO raises the average task score from $\mathbf{0.6379}$ to $\mathbf{0.7478}$ and increases 5/5 Pass instances from $\mathbf{138}$ to $\mathbf{164}$, while matching or outperforming rawSkill on $\mathbf{208}$ of $\mathbf{232}$ instances. It also reduces average runtime token consumption by $\mathbf{38.38\%}$. These results show that compiling natural-language skills into shared procedural units can jointly improve skill reuse, representation efficiency, and downstream execution.

In this paper, we propose CM3AE 2.0, a unified RGB Frame and Event-Voxel/-Frame pre-training framework. By jointly learning multimodal spatiotemporal representations and exploiting latent cross-modal dependencies through cross-modal generation, the framework enhances multimodal information interaction and collaborative modeling capabilities. Specifically, we first propose an Event-aware Spatiotemporal Masked Reconstruction (ESMR) strategy to learn unified multimodal spatiotemporal representations. It leverages event responses to guide masking toward dynamically salient regions. Through spatiotemporal masked reconstruction, it learns spatial structures and their temporal variations from continuous RGB-Event observations. In addition, it incorporates multimodal fusion reconstruction and cross-modal contrastive learning to facilitate the modeling of cross-modal complementary information and the learning of shared representations. Building upon this, we further propose a Hierarchical Cross-modal Generation (HCG) strategy to learn latent cross-modal dependencies through visual and language generation. At the visual level, Event-conditioned RGB generation is performed using a pre-trained diffusion model to capture visual associations between the two modalities. At the language level, video-level text description generation based on joint RGB-Event representations establishes connections between multimodal visual features and linguistic semantics, further enriching the learning of cross-modal dependencies. Extensive experiments on four diverse downstream tasks demonstrate that CM3AE 2.0 consistently improves performance, validating the effectiveness and generalization capability of the proposed pre-training framework.

ReLU networks are well-known to have continuous scale symmetries; infinitely many reparameterized weight vectors that represent the same function. Dinh et al. (2017) have shown that, by reparameterizing, the geometry of the loss landscape (e.g., curvature) of these networks can be changed arbitrarily without affecting the underlying function. Thus, any quantity that depends on the network's geometry should be defined with care. In this paper, we show that the Empirical Bayes (EB) Laplace's approximation is not well-defined for ReLU networks, mainly due to Hessian dependence in computing the marginal likelihood, which is not invariant under ReLU scale symmetries. Thus, we can get different prior precisions, posterior covariances, and predictive uncertainties for a network having identical predictions under ReLU scaling. We define family of prior precision fields indexed by $γ$ and introduce $γ$-canonicalization, which first converts the network to its canonical representation by balancing the ratio of incoming to outgoing weights, and then fits the Empirical Bayes (EB) Laplace to the canonical representation. This procedure is equivalent to transporting the isotropic Gaussian prior from the canonical representation to the original coordinates, by producing a ReLU-scale-compatible precision field. This makes the marginal likelihood invariant under ReLU scaling, while allowing $γ$ to be selected by maximizing the marginal likelihood. The resulting EB Laplace approximation is well-defined on ReLU scale-equivalence classes and yields consistent prior precision selection and predictive uncertainties across functionally identical parameterizations.

Typical diagnosis of Parkinson's disease (PD) occurs only after significant dopaminergic loss, while resting-state EEG abnormalities like alpha slowing and reduced coherence occur earlier. While many published EEG-based PD detectors claim accuracy greater than 95%, many use epoch-level splits which leak subject identity. I introduce MambaPD, a bidirectional selective state space model for resting-state EEG-based PD detection, blending the channel mixing, patch tokenization and bidirectional Mamba encoder. It is compared with four baselines (TCN, BiLSTM with attention, EEGNet and a patch Transformer) in two public datasets (UNM, SanDiego) using subject-independent 5-fold cross-validation (checkpoints were selected on the held-out fold, so absolute estimates are optimistic). On UNM the accuracy is between 0.67 and 0.71, and MambaPD (0.714) is not statistically distinguishable from the Transformer (0.715), and TCN (0.708), with the highest sensitivity (0.82), while the lowest specificity (0.52). For the smaller SanDiego cohort, accuracy is in the range of 0.60 to 0.70, with the highest scores (0.703) from TCN and the lowest (0.601) from EEGNet, while MambaPD scores 0.631, with the lowest specificity (0.46) among the five models. The differences among the models on UNM are minor compared to the differences among folds. I report the whole pipeline, per-fold results, and the shortcomings of the protocol.

Verifying the local robustness of neural networks under $\ell_0$ perturbations is inherently combinatorial, as it must reason over all sparse subsets of $ε$ perturbed pixels. This paper presents FocusBranch, a combinatorial branch-and-bound framework for $\ell_0$ robustness verification. The key insight is to make the $\ell_0$-ball decomposable and branchable using a node abstraction, which consists of a focus pixel set $\mathcal{S}$ and a level $l$, permitting at most $ε-l$ pixels outside $\mathcal{S}$. This abstraction unifies two branching mechanisms: concrete nodes branch through upper shadows to fix additional perturbed locations, while super nodes branch through covering designs to cover many perturbation cases with fewer child nodes. FocusBranch proves node safety by concretizing linear relaxations over node-constrained $\ell_0$-balls and recursively branches on unknown nodes. Rather than relying on a fixed verification plan, it samples nodes across levels to estimate safe rates and costs, then selects a starting level and branching strategy predicted to minimize verification time. Experiments on fully connected and convolutional classifiers for MNIST, Fashion-MNIST, and CIFAR-10 show that FocusBranch verifies $\ell_0$ robustness effectively across architectures and outperforms state-of-the-art verifiers. The code is available at: https://github.com/simpl-group/FocusBranch.

Direct non-binary constituency parsing must identify an ordered sequence of children at each phrase-level reduction. We extend our earlier delimiter-guided bottom-up construction to a common calculus for top-down, in-order, and bottom-up parsing. At reduction, the rightmost delimiter marks the left boundary, the nonterminal marker supplies the mother label, and the completed items in the delimited segment determine the children and arity. We use a stack invariant to prove soundness and completeness and derive the exact length of the static oracle's action sequence. A projection-and-lifting theorem shows that each delimiter-guided system and its canonical counterpart generate the same set of trees under both the structural and grammar-licensed transition relations. In controlled experiments on PTB and CTB, the delimiter-guided parsers achieve similar labeled accuracy to their canonical counterparts across all traversal orders. For bottom-up parsing, delimiter guidance reduces the number of action classes and increases the number of supervised decisions per tree.

Software developers spend significant time reading and understanding code, increasingly supplemented by generative AI. While these tools aim to reduce the difficulty of reading code, interacting with them introduces a dual-task environment that may impose extraneous cognitive load that is challenging to capture. This study investigates how access to generative AI affects developers' real-time cognitive workload, code comprehension performance, and self-perceived code comprehension strategies. We conducted a 3x2 within-subjects study with 20 novice CS students reviewing Java code snippets of varying lengths with or without AI assistance, followed by an AI-unassisted comprehension quiz. Cognitive workload was continuously captured using a Detection-Response Task (tDRT) during both reading and quiz phases, and compared to post-quiz NASA-TLX and SEQ scores. AI assistance significantly increased objective cognitive workload, raising tDRT reaction times by 11.2%, yet participants' perceived workload remained unchanged. Code comprehension scores were high under both conditions, with no evidence that AI assistance improved comprehension at the difficulty levels evaluated; a ceiling effect limited power to detect benefits on harder tasks. Qualitatively, participants reported trusting the AI conditionally -verifying its output against their own reading for more complex snippets, but directly trusting it for simpler code- and, while many felt their reading process was unchanged, a notable subset shifted toward skimming or reading the AI's explanations in place of the code itself. These findings reveal that developers do not accurately perceive the cognitive resources they expend when managing AI tools. This study provides evidence of added cognitive effort associated with using generative AI, pointing towards a need to account for this dual-task penalty in software engineering tools and education.

Dexterous grasping in robotics relies on distinct hand configurations depending on object geometry. We introduce BetaGrasp, a 3D diffusion policy that conditions a DP3-style point-cloud policy on a four-class grasp label (crimp, sloper, pinch, and jug), running on a Franka arm with a 16-DoF LEAP hand. We also introduce a climbing-hold benchmark: a spring testbed whose ratchet gives a slip-force reading in newtons. In our analysis the policy's point-cloud encoder did not separate the four holds, so the scan alone gave it little to choose a grasp from. Without the label (a DP3-style baseline), simulated rollouts from the robot's reset pose all end in a jug grasp, whatever the hold (25% correct grasp type, versus 94% for BetaGrasp). The label works as a switch for which grasp the policy performs. On 80 paired hardware trials, BetaGrasp raises average slip force from 9.6 N to 14.3 N, and grasps slipping fully off the hold drop from 44% of trials to 10%.

Many real-world applications require language models to internalize knowledge that was absent from pretraining, to incorporate continually updated facts, or to learn new domain skills. Self-distillation has emerged as a natural approach for this task. A model conditioned on the source material acts as the teacher, and an unconditioned student is trained to reproduce the teacher's behavior closed-book. We propose Speculative Self-Distillation (SSD), an efficient per-token mixed-policy method for self-distillation. The student generates each rollout by default, while the teacher intervenes only at positions where its next-token distribution substantially diverges from the student's. Unlike off-policy distillation with teacher-generated prefixes, this design keeps training rollouts close to the student's test-time behavior, while avoiding the low-signal states that fully on-policy rollouts drift into. In effect, SSD lets the teacher steer the rollout exactly where the privileged context changes the prediction. On three closed-book QA benchmarks measuring fact recall, knowledge updates, and skill acquisition, SSD nearly matches on-policy distillation while using 38-69% fewer supervised tokens and consistently outperforms off-policy distillation on the held-out test set. On a forgetting stress test, SSD retains the most general capability while also achieving the best in-domain accuracy and using less supervision than the on-policy baseline. Overall, SSD establishes a new accuracy-efficiency-forgetting Pareto frontier for self-distillation.

Large language models and the multi-agent systems built around them could accelerate materials discovery by connecting scientific knowledge with experimental decisions. Existing benchmarks establish capabilities in domain knowledge, multimodal interpretation and scientific tool use. As language models increasingly support materials research workflows, their ability to distinguish closely related material states and reason accurately with quantitative evidence remains insufficiently evaluated. These abilities are essential for comparing material responses, assessing mechanisms and maintaining reliable judgements as experimental evidence accumulates. We introduce MatScope to evaluate three connected capabilities. M1 tests unaided, condition-specific reasoning through 200 tasks spanning ten research functions, using aluminium and high-entropy alloys as representative material systems. M2 tests agent-assisted transfer of related knowledge to unresolved materials questions. MatEvoBench-20 tests initial forecasts across scientific dependency chains and revisions of still-hidden responses as observations accumulate. Across eleven model deployments, including GPT-6 Astra and Claude Fable 5.1, given-input calculations and many physical explanations are handled well. Fine-grained response estimates, comparisons between similar states and quantitatively supported choices remain substantially weaker. Related literature improves selected estimates, while condition and baseline mismatches limit transfer. Progressive observations calibrate some downstream responses and destabilize earlier accurate predictions. MatScope establishes measurable targets for developing reliable scientific reasoning in future self-evolving materials laboratories and materials-aware world models.

Predicting multi-step chemical reaction mechanisms requires identifying plausible intermediates and correctly ordering the elementary reaction steps. Existing computational methods focus on single-step elementary reaction prediction, limiting their ability to generalize to unseen reaction spaces. We introduce a graph-guided framework for multi-step chemical mechanism prediction that leverages a chemical mechanism knowledge graph (ChemMechKG) to guide large language models (LLMs). To improve pathway prediction, we develop a graph-aligned reaction mechanism predictor (GRMP) that uses graph neural networks to generate synthetic candidate pathways for pathway optimization. During inference, we design a relation retrieval (RR) module to extract relevant graph evidence from the KG for pathway refinement. Experiments show that GRMP achieves the highest pathway accuracy among all compared methods on ChemMechQA and oMe-Gold, reaching 48.1% and 57.9%, respectively. On a fixed sample of 110 out-of-distribution reactions from ChemMechQA-Hard, GRMP achieves 36.4% top-1 pathway accuracy. LLM-as-Judge and human assessment further demonstrate its ability to generalize to unseen reactions.

Multi-teacher on-policy distillation (MOPD) merges the capabilities of multiple domain-expert teachers, such as mathematical reasoning, code generation, and instruction following, into a single compact student. Since all domains share a limited training budget, how to allocate it across domains directly determines the quality of distillation. Existing methods allocate the budget by each domain's learning progress, but offer no guarantee for lagging domains. We therefore introduce distributionally robust optimization (DRO) into MOPD to prioritize the worst-performing domain. However, the per-domain training signals are estimated from a few rollouts and are inherently noisy, so directly applying DRO mistakes noise for the worst case, lets occasional spikes hijack the budget, and protects inherently hard domains rather than truly regressing ones. To address these issues, we propose RDS-MOPD, which keeps the DRO objective and redesigns its three parts: it shifts the allocation only when differences clearly exceed the noise, bounds the influence of any single anomaly, and defines the worst case as the regression from each domain's own best level, all with theoretical guarantees. Experiments show that RDS-MOPD outperforms recent competing methods and ranks first on most mathematical reasoning, code generation, and instruction-following benchmarks.

Large materials design spaces often contain rare high-performing configurations that are difficult to identify through trial-and-error exploration or purely representative machine-learning (ML) sampling. In this work, we present FruBO, a flexible Bayesian Optimization framework for accelerating the discovery of optimal configurations in computationally demanding materials design problems. As a case study, we investigate multilayer Moiré patterns generated by the superposition of six rotated hexagonal lattices, where the objective is to identify the rotational conformation that maximizes pore-size uniformity, quantified using the Jain fairness index. The resulting benchmark contains 600,000 unique configurations and exhibits a highly skewed objective distribution, making the global optimum difficult to locate by random sampling. We compare random exploration, Gaussian Process-based Bayesian Optimization, and Bayesian Optimization using a lightweight CatBoost surrogate with committee-based uncertainty estimation. Random sampling identifies the global optimum only after 533,012 evaluations, requiring approximately 31 hours, while Gaussian Process-based BO becomes impractical due to GPU memory exhaustion after 290 acquired samples. In contrast, FruBO with CatBoost identifies the global optimum after 1,470 acquired samples and approximately 1.15 hours. These results demonstrate that scalable surrogate models can substantially improve the feasibility and efficiency of BO in large nanomaterials design spaces.

Protein language models (PLMs) are standard tools for variant-effect prediction, protein annotation and design, but the compute needed to train and serve them keeps growing. We present HyPLM-Flash, an efficient and open PLM trained by Hyra 2.0, the research agent developed by our team. HyPLM-Flash is a 650-million-parameter masked language model trained on 726 billion tokens, 0.73 times the data of ESM-2 650M and a fifth to an eighth of the data of E1-600M and ESM C 600M. It runs on an inference engine with fused kernels and INT8 matrix multiplication, embedding about 130,000 residues per second on a single NVIDIA H20 GPU, 1.9 to 2.7 times faster than other 600M-class encoders. In a matched evaluation against 28 public checkpoints on 32 tasks, HyPLM-Flash reaches a ProteinGym zero-shot Spearman correlation of 0.409, statistically tied with ESM-2 650M, ESM C 600M and E1-600M, and the highest organismal-fitness score among self-supervised single-sequence models. It gives the best long-range contact precision of all models evaluated, including ESM-2 15B and ESM C 6B, and a lower masked-residue loss than ESM-2 650M, ESM C 600M and E1-600M on held-out sequences and on recently released PDB chains. Matching state-of-the-art accuracy at a fraction of the training and inference cost, HyPLM-Flash sets a new Pareto frontier for protein language models. We have released the model weights and the inference engine at https://github.com/Tencent-Hunyuan/hyra-results/tree/main.

Hypergradient (HG) methods for bilevel reinforcement learning (RL) typically assume a white-box follower whose reward is known. In practice, however, the leader rarely knows the follower's reward function and can only observe its trajectories. We address this black-box setting and propose Reward-Estimated Hypergradients (RE-HG), which recovers the follower's reward from these trajectories via inverse RL (IRL). Yet IRL identifies the reward only up to potential-based shaping, a transformation that leaves the follower's behavior unchanged but biases the HG. We prove that, for linear rewards, this bias cancels when IRL estimates are aggregated across sufficiently diverse dynamics, which the leader's own updates naturally supply. To handle finite samples, RE-HG truncates small eigenvalues of the aggregated feature covariance, restoring the identifiable subspace and stabilizing the estimation. Experiments show that RE-HG estimates gradients aligned with a white-box Oracle and achieves comparable performance; empirically, the truncation also suppresses HG explosions, letting RE-HG avoid local optima that trap the Oracle in sharp reward landscapes.

Inverse folding designs amino-acid sequences for a given protein backbone, and screening needs many candidates that are both diverse and well folded. We present HyInv, a 1.96-million-parameter backbone-conditioned graph network with an encoder-decoder architecture, trained by Hyra 2.0, the research agent developed by our team: pretrained with random-order autoregressive sequence modeling, followed by structure-reward reinforcement learning. At inference, a grouped draft-and-rewrite sampler that uses separate draft and rewrite temperatures improves refolding TM-score at matched diversity. An optimized implementation with BF16 numerics designs 128 sequences for a 402-residue backbone in 50.6 ms on a single NVIDIA H20 GPU, about 52 times faster than the reference sampler, and an exact FP32 path that returns the reference sequences is 8.7 times faster. Across five benchmark datasets and seven dataset-predictor combinations, HyInv reaches a higher average scTM than all ten external baselines at comparable or higher sequence diversity, with large margins over ProteinMPNN and ProteinZero V2. The same default setting supports de novo backbones, pocket-constrained libraries, enzyme redesign with the catalytic site kept, and loop redesign. We have released our code and model at https://github.com/Tencent-Hunyuan/Hyra-results/tree/main/AI4Science/hyinv.

In this work, we study whether frontier agent can continually improve real-world AI infrastructure over long horizons. We introduce Infrathon, a benchmark of diverse, long-horizon AI infrastructure optimization tasks across different levels. We evaluate six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours. Our results show that model rankings vary with the evaluation horizon, while 135 of the 150 runs achieve further improvements during the second half of their budgets. These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance. We also find agents can discover solutions that outperform human-designed baselines: Claude Opus 5.5 increases the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%. Analysis of agent trajectories identifies task violations, unreliable testing, and premature completion as barriers to sustained improvement. Together, Our Infrathon framework provide resources for constructing long-horizon tasks, evaluating frontier agents, and discovering good infrastructure solutions.

Small multimodal models can misjudge emotion when words, voice, and visual cues conflict. We present EmoHarness, which runs such a model, the frozen student, inside a harness: the prompts, settings, and verification code around its calls. A human-written starting harness, the scaffold, numbers the student's visual observations, transcript, and prosody descriptions as citable evidence and has the student weigh candidate emotions against it. In Meta-Harness Evolving (MHE), a stronger LLM, the Builder, reads the student's errors and execution traces offline and edits the harness; static checks and validation gates decide which edits are kept, and the Builder never sees evaluation items. On 5,507 MME-Emotion items, EmoHarness raises task-macro Recognition by 7.6-13.4 points over a single video-only call for all four students and both Builders. Transcripts and prosody descriptions supply most of this boost (+5.38 to +10.71); the scaffold changes the audio-augmented call by -0.79 to +2.54, partly through extra multi-label predictions, and evolved edits add at most +0.95, although no evolved increment survives Holm correction. The harness's emotion context also transfers to IntentQA, with small gains for three of four students, none of which survives correction. EmoHarness thus boosts small multimodal students chiefly by supplying and organizing audio-derived evidence.

Talking-head generation models, which synthesize realistic facial animations from audio, are increasingly vulnerable to misuse in multimodal deepfake scenarios. Protecting such systems remains challenging, as talking-head models fundamentally rely on precise audio-visual alignment, while effective audio protection methods remain largely underexplored. In this work, we propose MUTE, an audio protection framework that uncouples the underlying audio-visual alignment through a multi-level strategy. Specifically, MUTE combines (1) Alignment-level Disruption, which directly perturbs cross-modal attention to disrupt structural alignment, and (2) Representation-level Degradation, which perturbs temporal audio embeddings to degrade their representations. To improve robustness, perturbations are constrained in the STFT domain to high-energy regions, making them resistant to post-processing, denoising, and adaptive attacks. An optional speaker-level objective further mitigates potential bypass via TTS-based resynthesis. Extensive experiments demonstrate that MUTE consistently degrades lip synchronization across both white-box and black-box talking-head models while preserving perceptual audio quality. The proposed method remains effective under real-world transformations and can be combined with image-based approaches to provide stronger multimodal protection. The source code and project page are available at https://github.com/jindl465/MUTE.

Multimodal Emotion Recognition in Conversation (MERC) combines lexical, acoustic, visual and conversational cues, but accepting audio and video does not establish their effective or robust use. We audit off-the-shelf Omni models while holding the target transcript and prior textual context fixed and adding, removing or replacing target audio and video. In the primary evaluation of Qwen2.5-Omni-7B, Baichuan-Omni-1.5 and MiniCPM-o 2.6 on MELD and the full five-session six-class IEMOCAP set, T+A improves Accuracy in all six model-dataset pairs, with four gains significant after false discovery rate (FDR) correction. Across three source mappings, unrelated A/V (RandA/V) lowers Accuracy relative to Full in all 18 comparisons; all 18 remain significant under dialogue-cluster inference and BH-FDR correction, both on common-valid pairs and when invalid/OOM outcomes are treated as incorrect. Comparisons with modality-omission baselines further distinguish replacement penalties from mismatch-specific harm. A Qwen/MELD case study finds significantly greater gold-vs-best-candidate margin recovery in Late than Early or Middle under decision-query audio knockout, relative to an edge-count-matched Video-Sham control, using Holm-corrected tests. These results distinguish genuine-modality benefit from sensitivity and vulnerability to misleading input.

Many problems across the sciences compare two stochastic processes through the sequences they generate. Reinforcement learning asks what a new policy would return from the trajectories of an old one; molecular simulation asks how much more a ligand binds from samples of the unbound state; climate attribution asks how a heat wave shifted under warming; and language-model post-training asks what a deployed agent will do from logged interactions. In every case, one quantity would answer all these questions at once: the ratio of the two probability laws over their generated sequences, the trajectory density ratio. This ratio exists only under strong conditions. When one process reaches states the other cannot, the ratio does not exist; when their overlap is thin, its variance grows without control with the sequence length; and when they concentrate on different lower-dimensional sets, the ratio is undefined altogether. Weak support of this kind is the norm for sequence data, not a sampling accident, and its price is paid in identifiability rather than in noise alone. This survey organizes the field around a three-level diagnosis of weak support and a map of the nine retreats that methods take when the full ratio is out of reach, from direct estimation and reweighting through staged decomposition, stationarization, and partial-identification bounds to smoothing, transport couplings, operator spectra, and direct estimation of a single functional. Each retreat keeps part of what the ratio offers and pays for the rest, and each application domain has standardized the retreat its data force. The survey closes with open problems: testable diagnostics at every level, partial identification at the trajectory level, a rate theory of time-amplified error, and benchmarks with ground truth.

Recent Vision-Language-Action (VLA) models have begun incorporating thermal sensing, but existing approaches typically treat thermal images as ordinary visual inputs, providing limited support for temperature-aware relational reasoning in manipulation. To address this limitation, we propose Thermo-VLA with a ThermoRes Encoder, which decomposes thermal images into cold and hot attribute representations and injects them into the RGB token through language-conditioned residual modulation, enabling temperature-driven target selection and continuous manipulation. We further construct six real-world robotic tasks on a dual-arm Unitree G1 platform, spanning basic grounding to relational reasoning. Experiments show that Thermo-VLA achieves an average success rate of 0.86 across four thermal-dependent tasks, outperforming a pi_0.5 baseline that directly incorporates RGB-thermal observations by 18 percentage points. Under unseen-object and scene variations, Thermo-VLA maintains an average success rate of 0.78, achieving the best generalisation performance among the compared methods. Beyond thermal sensing, we transfer the same residual modulation mechanism to colour attributes and observe consistent improvements in both simulation and real-robot experiments, indicating its potential to generalise beyond thermal attributes. Project website: https://j-x-chen.github.io/Thermo-VLA-web/

Fine-tuning a deployed language model on one task after another can change behaviour that should stay fixed, such as earlier skills, general chat or refusals. Today, such forgetting is found only after training, on the benchmarks someone chose to run. This work bounds, before training, how often its outputs change on chosen prompt distributions. FireGate places each task's update, such as a new LoRA, behind a gate, which is a classifier on the model's frozen lower layers that decides for each prompt whether the update is used. Where no gate fires, the model computes exactly what the old one computes. As a result, on any prompt distribution, every bounded behaviour score changes by at most the gates' firing rate there over a whole task stream. Calibrating each gate on unlabeled prompts of what must not change fixes that rate before training, with a PAC guarantee, whatever weights training produces. Forgetting is thus bounded by the gate's false-alarm rate, and its price is the gate's power: no update, gated or not, that changes as few protected outputs can change new-task outputs more often than the best test of that level tells new-task prompts from protected ones. On reinforcement-learning streams of up to five tasks, gated training keeps the earlier tasks and changes 3.3% of held-out chat responses, whereas ungated training changes 91.8%. Across seven models, the gain kept behind the gate follows its firing rate on new-task prompts, measured before training. The certificate covers the gated model as deployed, on the prompt distributions sampled for calibration, not against an adversary who searches for prompts that fire the gate.

Neural Combinatorial Optimization (NCO) is a promising paradigm for solving routing problems, yet its generalization across diverse data distributions remains a critical bottleneck. Existing methods typically tackle cross-distribution generalization through complex multi-distribution training or intricate meta-learning schemes, implicitly presuming that single-distribution training yields a weak zero-shot baseline. In this paper, we demonstrate that the potential of single-distribution training has been substantially underestimated and can be unlocked by a simple architectural inductive bias: Mix Normalization. To approach this, we first construct a comprehensive benchmark spanning three representative routing problems and 179 fine-grained datasets, capturing diverse shifts in node coordinates, customer demands, capacity and time windows. Building on this benchmark, we propose MixRoute, a neural framework that adaptively combines normalization statistics at different granularities to mitigate varying distribution shifts. Trained exclusively on uniform distributions, MixRoute generalizes effectively to a wide range of unseen distributions without any distribution-specific adaptation. Extensive experiments demonstrate that MixRoute consistently achieves state-of-the-art zero-shot generalization. Our work revisits normalization as a parsimonious alternative for zero-shot generalization against complex training-adaptation schemes.

Sequence-to-function models such as AlphaGenome and Borzoi read up to a megabase of DNA and score a genetic variant by comparing their predictions for the reference and alternate sequences. Insertions and deletions (indels) are costly to score, because an indel usually shifts the unchanged DNA after it onto a different pooling grid, which prevents direct reuse of the reference's encoder activations. The AlphaGenome Atlas scored more than 100 million indels with three full model passes each, although the DNA after an indel matches the reference shifted by the indel's length. We present ShiftCache, an exact reuse system that encodes each shifted reference once and shares it among all variants scored in the same window with the same shift. ShiftCache copies every encoder output a variant can reuse and recomputes the rest in batched tiles. On held-out AlphaGenome windows shared by many indels, in strict fp32 on one A6000, ShiftCache is 3.5 to 6.5$\times$ faster than deduplicated dense execution and 1.6 to 2.7$\times$ faster than our adaptation of layer-by-layer reuse to indels. It is 3.5 to 4.1$\times$ faster than dense reimplementations of two published analyses with Borzoi and AlphaGenome, and 2.6$\times$ faster than dense execution with Enformer. Scoring each variant in its own centered 131 kb window makes ShiftCache only 1.36$\times$ faster than deduplicated dense execution, and highly diverse insertion lengths or many dispersed edits can make it slower than layer-by-layer reuse or dense execution. For encoders that meet our conditions, ShiftCache computes the same predictions as dense execution in exact arithmetic. On an A6000, every checked output of the default AlphaGenome and Borzoi runs matches dense execution within an absolute and relative tolerance of $10^{-5}$. With larger tile batches, with Enformer or on an H100, a small fraction of values exceeds this tolerance.

Quantization reduces LLM deployment costs, but effective strategies must balance model quality and storage. LLM-driven search can automate strategy design through repeated proposal and evaluation. This process must identify useful strategies within a limited evaluation budget. The key challenge is that \textit{actual quantization is expensive, while inexpensive proxies may not reflect quantized-model performance.} To address this challenge, we present AutoQS, an \textbf{Auto}mated \textbf{Q}uantization \textbf{S}trategy Design with LLMs. AutoQS keeps weight precision fixed and searches group-size allocations across layers and operators under deployment constraints. Evaluation history guides coordinated LLM proposals, while deterministic validation ensures compatible allocations. An inexpensive quality proxy evaluates candidates, and an candidate repository retains alternative allocations to inform subsequent proposals. Periodic actual quantization complements these proxy evaluations with measured quality and checkpoint storage, providing feedback for further search within the evaluation budget. Experiments on Qwen2.5-7B-Instruct show that AutoQS improves seven-task average accuracy over uniform-group GPTQ and AWQ by 0.70 and 1.12 percentage points, respectively, at four-bit weight precision.

Predict-then-optimize systems convert contextual forecasts into decisions, yet conformal robust optimization typically calibrates prediction residuals rather than downstream loss. For linear optimization, this parameter-space geometry can be badly misaligned with decision risk: large objective perturbations may preserve the optimizer, while small boundary-crossing perturbations may cause substantial regret. We introduce Conformal Regret Sets, a post-hoc, model-agnostic framework that conformalizes an optimization-derived regret surrogate and uses its calibrated level set for minimax-regret refinement. We first show that directly conformalizing exact regret, although sufficient to certify the nominal predict-then-optimize decision, yields a degenerate robust problem: absent an exogenous bounded support, every alternative decision has unbounded worst-case regret. We therefore use SPO+, a pointwise upper bound on exact regret, as the conformal score, obtaining distribution-free finite-sample marginal regret guarantees under exchangeability. For contextual linear programs, the resulting set decomposes into polyhedral pieces aligned with the normal fan of the feasible polytope. A vertex is active if and only if its predicted suboptimality gap is at most half the calibrated threshold $η$, enabling local, output-sensitive generation of all relevant pieces. Moreover, if $Γ$ is the predicted gap to the closest competing vertex, the nominal and refined decisions admit the sharper worst-case-regret certificate $(η-2Γ)_+$, with an exact early exit when $η\leq 2Γ$. Experiments on contextual fractional knapsack validate the geometry and certificates: continuous refinement reduces the $0.95$-quantile and $\operatorname{CVaR}_{0.95}$ of regret by $9.1\%$ and $13.1\%$ relative to nominal PTO, while vertex-restricted refinement changes only $10.0\%$ of nominal decisions.

Web agents extend LLMs from text generation to real-world web interaction by relying on execution harnesses that interpret user goals, perceive webpages, plan actions, and ground model decisions into browser operations. While prior work has recognized that agent security depends on more than the backend model alone, how harness design quantitatively shapes web-agent security remains insufficiently measured. To systematically study web-agent harness security, we first characterize web-agent harnesses as a four-stage execution workflow covering goal interpretation, webpage perception, reasoning and planning, and action execution. Based on this abstraction, we build WebTrap Farm, a systematic and extensible evaluation framework covering three representative attack categories: malicious user prompts, malicious prompt injection, and deceptive web design. To avoid conflating security with task capability, we use paired usability calibration tasks to separate genuine defensive behavior from capability limitations. Our evaluation shows that existing web-agent harnesses remain vulnerable to typical attacks, with some unsafe behaviors occurring even without explicit attack content, revealing harness-intrinsic security failures (RQ1). Statistical analysis further shows that both harness and backend model significantly affect security outcomes, indicating that deployed-agent security is jointly shaped by the harness and backend model (RQ2). Finally, we localize where defensive behavior emerges across workflow stages and identify more secure design mechanisms, providing actionable guidance for building more robust web agents (RQ3). These findings highlight the web-agent harness as an important security boundary and motivate moving beyond model-centric defenses toward harness-level security design.

Choosing the latent dimension is a critical challenge in data compression, where conventional hyperparameter optimization is costly and ad hoc. Intrinsic dimension provides a good alternative, which connects information-theoretic notions such as Renyi information dimension (RID) to rate-distortion dimension and random codebook compression. In this work, we introduce the pairwise distortion distribution (PDD) and its asymptotic limit, pairwise dimension (PD), which generalize the correlation integral to general distortion measures. We show that PD bounds RID, and PDD provides a lower bound for random codebook quantization performance, which can guide a practical use for compression, quantization, and dimensionality reduction applications.

Evidence-sufficiency benchmarks ask whether a model can tell a complete passage set from one whose reasoning chain is broken. When broken sets are made by editing text, near-perfect sufficiency scores--and even positive crossed-control scores--need not demonstrate multi-hop reasoning. We present ChainCheck, an evaluation-validity audit that pairs validator-certified counterfactual cells with artifact-only baselines. In an exploratory crossed control on 2,214 HotpotQA pairs, a language-model scorer that never reads the chain reaches a nominal AUC of 0.997, and a one-string match reaches the maximum of the crossed chain-sensitivity statistic CS4 (0.500), while the 11 evaluated systems score from 0.308 to 0.484; the two statistics are different metrics. We therefore read positive CS4 as operational bridge-consistency sensitivity, not as semantic multi-hop reasoning. Every system is also sensitive to one extra edit, and the edits are detectable according to model-based naturalness ratings. On an independent HotpotQA set of 2,276 pairs, patterns pre-specified on MuSiQue and 2WikiMultiHopQA did not replicate: 5 of 6 trained readouts were edit-dominated, and the reference judge's selectivity interval [-0.007, +0.042] includes zero. Judge scores also depended on batch layout. The MuSiQue and 2Wiki records are public; the HotpotQA records and the fine-tuned checkpoints are not, so external reproduction is partial.

Synthetic data can scale visual pretraining, but current selection criteria only partially explain which datasets train strong representations. We propose using the energy profile of a dataset--the distribution of log-probability levels it occupies under a normalised real-image density model--as an audit and curation principle for synthetic data. To operationalise this view, we introduce Energy-Level Distribution Distance (ELD), which compares real and synthetic energy distributions. Real images span sparse high-likelihood scenes and dense low-likelihood textures, so useful synthetic data should cover this energy range rather than collapse to likely but low-variation samples. Across diverse synthetic datasets, ELD exposes mismatches missed by FID and predicts pretraining performance beyond FID, LPIPS variation, and feature-space recall, including across CNN and ViT self-supervised settings. We then use ELD to resample and mix synthetic sources, matching the real-image energy histogram while preserving diversity and improving downstream transfer.

Unplanned 30-day intensive care unit (ICU) readmission is a critical benchmark for evaluating healthcare quality. Most predictive models assume readmission risk depends solely on a patient's individual pathophysiological state, ignoring hospital operational context and resource capacity. In practice, discharge decisions are dynamic trade-offs between patient stability and bed turnover pressures. To address this gap, we develop an Operation-aware Multimodal Clinical Decision Support System (CDSS) for ICU discharge risk assessment. The system combines multi-agent collaborative extraction with self-evolving retrieval-augmented generation (RAG) to eliminate template noise from progress notes, while constructing departmental workload indicators alongside admission flow features, continuous vital signs, and sparse medical codes (ICD/DRG). Addressing severe class imbalance (about 10.5% readmissions) and alarm fatigue, we implement data normalization, class reweighting, and decision-threshold calibration. In 20 independent runs across 20,000 ICU stays from MIMIC-III, empirical findings demonstrate that: vital signs alone have limited predictive capacity under real-world class imbalance (PR-AUC of 0.1387 +/- 0.0086); directly concatenating raw text embeddings degrades decision tree performance (XGBoost and LightGBM PR-AUC drop to 0.3893 and 0.4030); whereas an end-to-end multimodal fine-tuning architecture (Transformer E2E FineTune) utilizing layer-wise parameter freezing and differential learning rates reverses text negative transfer, achieving the highest AUROC of 0.8288 +/- 0.0110, PR-AUC of 0.4373 +/- 0.0310, and F1-Score of 0.4757 +/- 0.0196. Operational indicators supply independent risk signals, laying an engineering foundation for high-utility clinical decision support tools.

LLM advertising selects advertisements based on user conversations, where user needs may extend beyond explicit product searches. ChatGPT Ads allows advertisers to provide natural-language context hints describing what they offer, who may find it useful, and when it may be useful, as additional information for ranking advertisements across diverse user conversations. However, how to effectively construct these hints remains unclear. We formulate Context Hint Optimization (CHO) as the problem of constructing context hints to improve an advertisement's ranking under a given budget. Drawing on established marketing frameworks, we develop seven context hint generation strategies covering What, Who, and When, along with a Unified strategy that combines all seven dimensions. We introduce CHO-Bench and conduct over 1.19 million paired evaluations across five datasets using embedding-based and LLM-based rankers. Our results show that context hints improve advertisement rankings under the evaluated offline rankers. While individual strategies achieve similar average gains and Unified offers no consistent advantage, strategy differences become more pronounced under specific generation-ranking configurations, with embedding-based and LLM-based rankers favoring different information dimensions. Increasing the hint budget further improves rankings with diminishing marginal gains. These findings establish an offline benchmark for studying ad-side ranking optimization in LLM advertising and highlight opportunities for model-specific hint selection and budget-aware generation and aggregation.

India's temples and other places of worship are among its richest repositories of cultural heritage, bringing together religious tradition, history, architecture, and art. Due to the widespread curiosity, people increasingly turn to Artificial Intelligence (AI) assistants for information about them. Current models are rarely trained or evaluated on this domain. We present BharatSynthQA, a passage grounded question answering (QA) dataset focused on Indian temples, which also covers other religious sites connected to them. Our pipeline crawls Wikipedia from topics seeded by authentic heritage sources, filters articles for topical relevance, and locations in India. Further it generates Boolean, factoid, and explanatory QA pairs with a fewshot prompted LLM. An LLM judge filters the pairs on answer correctness, retaining 98,677 of 131,494 generated pairs. Additionally, 300 human-annotated pairs are released separately as a reference set. Human evaluation of the subsample of this dataset showed 89.83% of the questions were formed correctly. Benchmarking five current LLMs shows that factoid token F1 is only 0.18-0.30 without the source passage but 0.73-0.77 with it, indicating that this domain is poorly captured in model knowledge. The dataset is available at https://huggingface.co/datasets/theshivam7/bharatsynth-qa to support research on culturally grounded QA and retrieval-augmented generation for Indian heritage.

Video reasoning hinges on a small number of visually decisive tokens embedded in otherwise largely language-driven chains of thought. Yet prevailing post-training methods such as GRPO overlook this structure, assigning coarse trajectory-level rewards while requiring multiple rollouts and often costly external judges. We introduce Visually-Aware On-Policy Self-Distillation (VIOS), a post-training framework that identifies where a model under-utilizes visual evidence and concentrates learning at those positions. Our key insight is that a model can audit its own visual reliance: the divergence between its next-token distributions with and without the video reveals how strongly each prediction depends on the visual input. Measuring this reliance for both a student and an answer-privileged teacher exposes a per-token reliance gap, identifying positions where the teacher exploits visual evidence that the student overlooks. VIOS translates this signal into dense, targeted supervision through three complementary components: an exponential moving average (EMA) teacher that co-evolves with the student, a contrastive objective that encourages predictions to depend on the video, and a reliance-gap gate that focuses learning on visually under-utilized tokens. Crucially, VIOS requires only a single on-policy rollout from a single model, without external rewards, judges, or larger teachers. Across 12 benchmarks spanning general, temporal, spatial, knowledge-intensive, and long-video understanding, VIOS consistently improves three model backbones and outperforms contemporary video-reasoning systems, while training in over 9x fewer steps, 60x fewer FLOPs, and 111x fewer generated tokens. Project page: https://vision.cs.utexas.edu/projects/VIOS/

Robot demonstration datasets combine demonstrations from teleoperators of uneven experience, so collection teams need to know how many demonstrations from one group match those from another. We measure this with a data multiple: the ratio of the numbers of demonstrations two groups need for their policies to reach the same success rate. On robomimic's simulated Square and Can tasks, we train behavior-cloning policies on demonstrations from the two teleoperators robomimic labels Better and the two it labels Worse, varying dataset size and the share of Worse demonstrations. On Square, within robomimic's 25-second time limit, an all-Worse training set needs 1.55 times as many demonstrations as an all-Better set. Most of this difference is late completion: policies trained on Worse demonstrations finish later, as the Worse teleoperators do, and with a 100-second limit the multiple falls to 1.11. On Can, most of the difference remains at 100 seconds. If each Worse demonstration were worth 1/1.55 of a Better one, as in an all-Worse set, a half-Better, half-Worse set on Square would need 1.22 times as many demonstrations as an all-Better set. In fact, it needs 13% more than that, and this excess persists at every time limit from 20 to 200 seconds. The excess remains when training draws demonstrations uniformly instead of in proportion to their length. On Square, within the 25-second limit, switching to these Worse teleoperators is cheaper at equal success only if they earn less than a third of the Better teleoperators' hourly pay, because Worse demonstrations also take longer to record.

Origin-destination (OD) demand plays an important role in transportation system design and management applications. Conventional OD estimation methods typically judge estimation quality based on how estimated demand aligns with observed traffic data and historical demand patterns. We argue that data alignment alone is insufficient; how errors in OD estimates influence downstream decision quality should also be considered. This paper develops a decision aware OD demand estimation framework that leverages structurally constrained traffic data to characterize an admissible demand region, within which we identify OD estimates that would lead to least regret in the downstream decision. Therefore, our OD estimates are both data-consistent and decision-relevant. Methodologically, we approximate the admissible demand region induced by physical network constraints through a geometric representation, which regularizes the search over feasible OD estimates. Within this regularized region, we use a convex surrogate of downstream regret to obtain a tractable decision aware optimization problem. Numerical experiments show that the proposed framework consistently reduces decision regret relative to conventional estimation benchmarks, demonstrating the value of aligning OD demand estimation with downstream decision quality.

Transformers stop learning graph search as graphs grow, and learn the card game SET only after many distinct questions, although attention can represent both. We read these failures, and the uneven record of higher-order attention, from what a layer sums over: each layer we study computes the partition function of a family of witnesses and reads it at a temperature. Two laws follow. The depth law is a known result of logic, which we state for these layers: L layers over two-step walks reach exactly 2^L - 1 steps, just as a logical statement with L nested clauses of the form "there is a vertex such that" sees exactly that far. We build a construction that reaches this far, checked in Lean, and, when vertices carry no names, no such network reaches further. The size law is the cost of reading at a temperature, which logic does not have: existence is read correctly at every size, in the worst case, if and only if inverse temperature times margin is at least the logarithm of the number of candidates. Layers built on these laws, and on two principles that we test, learn where Transformers fail. One propagator layer trained on 49-vertex graphs answers every question at every search depth on 250-vertex graphs. Given the modulus 3, exponential Newton computes the softmax of 2-simplicial attention over pairs exactly in 3^d n operations for d clocks; with six clocks it learns SET from as few deals as 2-simplicial attention and transfers from 8- to 20-card tables. On CLRS-30 without hints, a 186,000-parameter block of OR-join and power-mean channels matches (within half a point) or beats a model of at least 73 million parameters on 9 of 12 graph tasks of the official test set, and on 8 of 12 with freshly drawn test graphs. The same account explains where pair layers did not help and where parallel layers must loop.

Constrained reinforcement learning in hybrid action spaces must coordinate discrete modes and mode-conditioned continuous controls under long-horizon cost constraints. Existing constrained hybrid-action methods typically optimize factorized policies directly in parameter space, whereas Constrained Variational Policy Optimization (CVPO) separates nonparametric improvement from parametric projection in its original continuous-action formulation. We introduce Hybrid Constrained Variational Policy Optimization (H-CVPO), extending this variational framework to hybrid action spaces. Under boundedness and strict feasibility, its constrained E-step has zero duality gap and, when the dual optimum is attained at positive temperature, yields a closed-form joint target distribution over hybrid actions through exponential reward-cost reweighting. Using the exact hybrid KL decomposition, we formulate a factorized M-step that fits both policy components to this target under separate discrete and conditional-continuous trust regions. For exact E/M updates, we prove fixed-reference surrogate improvement and conditional reward-cost bounds with a sufficient feasibility-margin condition. For approximate updates, we establish the existence of a dual minimizer under continuity on a compact domain and use residual-based admissibility to determine whether to project the target. Ablations identify cost-aware reweighting as an important contributor to constraint control. Extensive experiments in simulated hybrid-control environments show that H-CVPO achieves favorable reward-cost trade-offs compared with hybrid-action baselines.

Entanglement is a crucial resource in quantum science and engineering, but detecting it in many-body states from experimentally accessible observables remains challenging. Here we introduce EntNet, an AI-driven framework to discover compact combinations of spin observables that accurately distinguish entangled from fully separable many-body mixed states in one-dimensional spin-$1/2$ chains governed by short-range Hamiltonians. We construct local-unitary-invariant representations of measurable Pauli correlators up to four-body order, sparsely select 16 candidate invariant feature channels, and learn their spatial structure using a Transformer architecture with rotary positional embeddings. Using only expectation values of these low-order observables, EntNet achieves $>98\%$ recall on certified entangled states with near-zero false-positive rates on separable states. Remarkably, it exhibits transferability across system sizes and distinct Hamiltonian families, including frustrated and disordered spin models, with minimal loss of accuracy. Beyond detection, EntNet provides physical insight into entanglement detection by revealing a selective spatial organization of pair correlations, while its finite-temperature decision boundary tracks the excitation gap of Jordan-Wigner quasiparticles. These results establish a transferable and physically interpretable framework for data-driven discovery of entanglement witnesses from local correlation functions.

In most LLM agents, weights are fixed after deployment, so new information must be retained in external memory. Existing approaches often rely on predefined records and fixed organizational rules, which cannot readily accommodate evolving tasks and interactions. This raises the question: can LLM agents organize their own memory using ordinary documents? We introduce self-organized memory (SoMem), a framework that stores memory in linked documents. The agent decides what to summarize, how to group related information, and when to revise existing topics, and every summary it produces stays linked to the original interactions as evidence. The runtime validates updates and automatically derives temporal, relational, and recent-memory views from the organized content. At query time, the agent chooses what to search and read, following document links and inspecting original evidence as needed. On LoCoMo, BEAM-100K, and LongMemEval-S, SoMem outperforms the strongest baseline by 2.66, 3.10, and 2.00 percentage points, respectively. The constructed memory can be reused across models, and additional reasoning effort can further improve answer quality. These findings demonstrate the feasibility of agents organizing and using their own persistent memory through ordinary documents. Code is available at: https://github.com/Qi202/SoMem.

Language is central to social communication, which is altered in autism, and language abilities vary widely among autistic individuals. How the autistic brain processes language is central to understanding autistic communication. Yet neuroimaging studies report inconsistent findings on how autistic brain activity is altered during language processing. We synthesized 45 fMRI and PET studies (753 autistic and 777 non-autistic participants) with seed-based d mapping with permutation of subject images (SDM-PSI). We estimated autistic and non-autistic activation and their difference with meta-regressions adjusted for each study's language modality and mean age. Autistic and non-autistic brains activated a similar left-lateralized language network. Autistic brains showed reduced activity throughout the left perisylvian cortex, from the inferior frontal gyrus to the mid superior temporal gyrus and the posterior superior temporal sulcus, and increased activity in the right temporal pole. Spoken language showed a larger autistic reduction than written language in the left mid superior temporal gyrus. Adults, but not children and adolescents, showed increased autistic activity in the right superior temporal gyrus and cuneus. Autistic language processing thus relies on the typical left-lateralized network but shows reduced leftward lateralization.

An open-vocabulary segmentation readout silently loses a class deployed beyond its training categories. It still emits masks with high confidence, but the class never wins a single pixel. The training loss falls anyway, and mIoU can even rise while classes die. We show that which classes collapse is written in the class-name list alone. Before any image is seen, we rank each class's collapse risk from its text similarity to the training vocabulary alone. The ranking holds at AUROC 0.69-0.96 across nine model families, twelve backbones and eight pretraining recipes. Two further families, measured only after the predictions were written down, land in the same band. What the list writes, it can rewrite. With identical images and a class's counterpart moved in or out of the training list, retraining flips that class's gain over the frozen baseline by 21 IoU points; swapping in unrelated names changes nothing. We find a vocabulary-distance mechanism behind the collapse. Collapsed classes lose every vote to near-name neighbors, but in the frozen features they survive. Turning down the readout's amplitude, or stopping training early, brings them back, up to +16 points, without retraining; trained with its amplitude held down, the readout never collapses.

World models are widely considered to be a potential key component of model-based reinforcement learning and robot planning, but typically require task-specific training on substantial interaction data. We investigate whether a pretrained tabular foundation model (TabFM) can serve as a training-free dynamics model for world modeling by framing next-state prediction as in-context regression over transition pairs, and where such a model helps and where it breaks. Across synthetic dynamical systems and robot locomotion benchmarks, TabFM is more accurate than LSTM and MLP models trained from scratch on the same limited transitions in almost all settings, and among regressors fitted to the same data it is the most accurate on locomotion, while a Gaussian process is more accurate on smooth low-dimensional systems. However, models trained on the full dataset remain substantially better on most systems. When rolled out autoregressively over multiple steps, TabFM's accuracy depends strongly on how much context it is given. With a short context, its rollouts quickly degrade to the level of a trivial baseline, whereas with a long context they improve substantially and, on one locomotion task, exceed a trained LSTM. Inference is also about three orders of magnitude slower than an LSTM. Overall, in their current state, we find that training-free priors are useful as predictors when data are scarce and enough context is available, but they are not yet a substitute for trained world models in planning.

EEG-based fatigue detection requires modeling spectral information across bands and its temporal variation. Recent methods combine multi-band modeling with adaptive temporal processing, yet how spectral structure across bands should influence the contribution of each temporal interval remains insufficiently explored. We propose LBSNet, a lightweight network that uses prototype-based brain-state modeling to link local spectral structure with temporal aggregation. LBSNet extracts compact multi-band features through band-specific temporal convolution, shared spatial projections, and cross-band recalibration. Within each window, relative spectral activity is encoded as a normalized descriptor and softly matched to learnable spectral prototypes. Prototype assignments and learned prototype importance jointly determine temporal weights for power aggregation, allowing the spectral pattern of each window to guide its contribution to the prediction. Subject-independent experiments on SEED-VIG and SADT show that LBSNet outperforms seven baselines while maintaining low computational cost. Ablation and visualization analyses support prototype-guided temporal weighting, while the learned prototypes reveal a gradual redistribution of relative cross-band spectral states with increasing fatigue. The code is publicly available at https://github.com/zy0929/LBSNet.

Feature subsampling is used to accelerate gradient-boosted trees, yet standard approaches fix the feature count in advance and largely ignore the split-gain evidence accumulated during training. We introduce CRAFT-GBM (Confidence-Ranked Adaptive Feature Tracking), a native LightGBM extension that uses this evidence to adapt which---and, by default, how many---features are searched each boosting round. CRAFT-GBM transforms root-split gains into bounded observations and constructs joint time-uniform bounds valid under adaptive, non-stationary training histories. In its default endogenous mode, after initialization, CRAFT-GBM retains exactly the features that cannot yet be ruled out as best, letting the selected feature-set size $K_t$ emerge from training history rather than a pre-specified fraction. With probability at least $1-δ$, every feature with the largest expected gain is retained jointly across all rounds after initialization. When a feature fraction $f$ is specified instead, CRAFT-GBM ranks features by their interval point estimates and provides an approximate top-$K$ guarantee for feature budget $K_f$. We implement CRAFT-GBM within LightGBM while preserving its native feature-bundling representation, and evaluate it across 14 public and synthetic dataset families and four noise-augmented settings. In the fully endogenous setting, CRAFT-GBM reduces per-fit training time by 39.74\% with 1.04\% lower test loss in the geometric mean across all 18 RQ1 experiments. Under pre-specified feature fraction $f$, Exogenous CRAFT-GBM also yields aggregate training-time reductions up to 13.52 percentage points larger than random feature subsampling, with no increase in aggregate test loss relative to full-feature LightGBM. Accumulated training evidence can thus support adaptive feature search with statistical guarantees while retaining the practical advantages of native gradient boosting.

Wearables, mobile devices, and IoT platforms generate multimodal time-series data continuously and at scale, making efficient storage and transfer a primary concern. The same recordings are reused after collection by many downstream models, often at different temporal resolutions, so a compressed representation should be small, useful across resolutions, and produced at a bit rate known in advance. Prior work meets parts of this: symbolic and error-bounded compressors reconstruct accurately but fix resolution at compression time, learned codecs target one signal domain, and discrete tokenizers reconstruct without accounting for rate. We introduce CHORD, a coarse-to-fine compressor for heterogeneous multimodal time series. A residual cascade orders each payload from broad to fine temporal scales and synthesizes corrections through learned shapelet-inspired atoms. A shared pathway supplies a reusable synthesis basis across modalities, while private pathways retain capacity for sensor-specific morphology. Finite scalar quantization fixes the bit width of every modality and scale, so payload size is known before encoding, and scale dropout trains each coarse-to-fine prefix to be a valid reconstruction operating point. Across seven wearable-sensing and forecasting benchmarks, CHORD stays within 0.62 points of uncompressed classification accuracy on average at the 8x target, the smallest gap of any compressed method, with all resolutions decoded from one payload.

LiDAR point clouds and 3D semantic occupancy represent two distinct modalities capturing fine-grained metric geometry and dense volumetric semantics, respectively, holding natural potential for mutual complementarity. However, existing world models typically forecast these modalities independently or cascade them heuristically, missing the opportunity for reciprocal cross-modal regularization. To address this, we propose \textbf{LO-World}, a unified 4D world model that forecasts both modalities within a shared BEV latent space. This single-stream formulation enables dense semantics and metric geometry to mutually reinforce each other, while eliminating multi-pipeline redundancy. Specifically, LO-World couples an efficient hybrid range-voxel autoencoder with a flow-matching generator that caches historical context to reduce iterative sampling latency. Extensive experiments demonstrate that LO-World outperforms prior state-of-the-art baselines, improving occupancy forecasting IoU by 10.9% and reducing LiDAR Chamfer Distance by 30.1%. Meanwhile, it sustains a real-time throughput exceeding $25$~FPS on a single RTX 4090. Our code is available at https://github.com/lzzzzzm/LO-World.

Large language models (LLMs) increasingly plan and run experiments in automated laboratories. These experiments keep running while the model reasons and run on schedules with hard deadlines. If an LLM-generated answer arrives late, the system executes a safe fallback action instead, such as rejecting a sample. This creates a systems problem that standard serving metrics do not capture. An answer that matches the fallback adds no value, and answers that differ from the fallback may require more reasoning time. Reporting answer quality and latency separately can therefore overstate the value of reasoning. We test this on synthetic sample decisions, where the best choice is known exactly and a runtime enforces the deadline. We compare one LLM's fast mode, which answers directly, with its thinking mode, which reasons first. In a prespecified replay of 1,200 new decisions, thinking answers that missed a deadline were worth about five times as much as on-time ones. Multiplying the on-time rate by the mean answer value predicted about twice the value that thinking delivered. This prediction overstated thinking's advantage over fast mode by 0.0241 per decision, and by 0.0191 in a repeat on 300 new decisions. Both 95% confidence intervals excluded zero. In a single-GPU Qwen3-8B experiment using simultaneous request bursts, raising the number of requests from 8 to 56 reduced the share of thinking answers that met a 45 s deadline from 79.2% to 1.4%, although total output tripled. Evaluations of LLM agents for laboratories should therefore measure each answer's value and arrival time together and score the action that actually runs.

Multimodal Emotion Reasoning (MER) requires perceiving and integrating emotion cues from multiple modalities. Recent advances increasingly leverage Multimodal Large Language Models (MLLMs) to enhance reasoning capacity. However, due to the inherent heterogeneity in conveying emotions across modalities, these models often suffer from modality imbalance and confront flawed emotion understanding. In this paper, we conduct an in-depth analysis on MER models and uncover two key empirical insights. (i) In MER, modality imbalance manifests two distinct failure modes: unimodal over-reliance and insufficient cross-modal utilization, which call for different intervention strategies. (ii) MLLM layers exhibit implicit functional hierarchies in multimodal emotion processing, playing separable roles in stabilizing unimodal representations and integrating cross-modal information. Motivated by these observations, we propose a Layer-Aware Modality Rebalancing (LAMR) framework by explicitly aligning intervention strategies with layer roles. Specifically, we first introduce two quantized metrics for adaptive layer role profiling. In each layer, Modality Entropy (ME) measures the dispersion of modality contributions, and Cross-modal Intensity (CMI) captures the strength of cross-modal interactions. Then, we respectively enhance modality equilibrium in stabilization layers and cross-modal propagation in integration layers. Extensive experiments on two comprehensive MER benchmarks demonstrate that LAMR effectively mitigates modality imbalance and improves emotion reasoning across diverse MLLM backbones, without additional training.

Retinal disease monitoring with optical coherence tomography angiography (OCTA) relies on segmentation of vessels and the foveal avascular zone (FAZ). Promptable models such as SEEM (Segment Everything Everywhere All at Once) allow click-based refinement, but without adaptation SEEM segments thin vessels poorly and needs many clicks. Fully fine-tuning its 342 million parameters on a few hundred images is costly and risks overfitting, and parameter-efficient methods such as low-rank adaptation (LoRA) usually adapt only the encoder. Yet SEEM reads every candidate mask from one feature map at quarter resolution, too coarse for vessels one to three pixels wide. This study introduces the Read-out Basis Adapter (RBA), whose read-out branch generates 32 image-specific basis maps at half resolution. Each candidate accesses them through its own vector and a zero-initialised projection, so the model starts exactly as SEEM. The encoder branch, a standard rank-1 LoRA, supplies the information needed to tell arteries from veins. RBA adds 0.47 million trainable parameters. Under simulated error-driven clicks, with the ground truth used only to place them, RBA outperforms LoRA, Conv-LoRA andAdaptFormer on OCTA-500 3mmlarge vessels. Averaged over three seeds, it reaches an intersection over union (IoU) of 85.16% and a Dice score of 91.93%, 2.90 and 1.72 percentage points above Conv-LoRA. On the 6 mm subset, one run gives margins of 4.35 and 2.76 points. The gain extends to arteries, veins and ROSE-O but not to the FAZ. In SEEM, adaptation capacity pays more at the mask read-out than in the encoder.

The Adam optimizer is widely used to train neural networks with highly nonlinear and nonconvex objective functions. Remarkably, Adam often drives the empirical training loss nearly to zero. This phenomenon has been partially explained by neural tangent kernel (NTK) analyses of gradient descent (GD) and stochastic gradient descent (SGD). However, analyzing Adam presents additional challenges. Its first-moment momentum couples stochastic gradients across iterations, while its adaptive second-moment preconditioner depends on the same stochastic gradient used in the update. In this paper, we establish linear convergence guarantees for Adam in training sufficiently wide two-layer neural networks with either full-batch or minibatch gradients. For full-batch Adam, we construct a Lyapunov function to control the effect of historical momentum. For minibatch Adam, our analysis builds on an exact virtual-iterate representation, a conditional analysis of the dependence between stochastic gradients and adaptive preconditioners, and uniform concentration bounds for the second-moment estimates. We prove geometric decay of the empirical loss for full-batch Adam and geometric decay in restricted expectation on a high-probability event for minibatch Adam. Beyond regression, we extend our convergence analysis to physics-informed neural networks (PINNs) for solving the Poisson equation, where differential operators and separate sampling of interior and boundary points introduce additional challenges. We establish analogous global convergence guarantees for both deterministic and stochastic Adam under explicit conditions on network width and Adam's parameters, characterizing the effects of minibatch size on convergence. These results help bridge the gap between Adam's empirical success and its theoretical understanding, providing sharper convergence guarantees for PINNs than existing analyses.

Machine unlearning in large language models (LLMs) aims to remove specific learned knowledge, capabilities, or behaviors while preserving model utility. Existing methods have mostly focused on the design of unlearning objectives, while largely overlooking the intrinsic matrix structure of transformer weights. In particular, standard optimizers operate on independent scalar coordinates, thereby ignoring this matrix structure. We propose spectral unlearning, an update framework that combines standard unlearning objectives with matrix-structured optimization. We show that the spectral update corresponds to the steepest-descent direction for transformers composed of linear layers with normalized inputs and outputs. Our approach is applied on top of existing unlearning methods and evaluated across complementary axes of memorization, privacy, and utility, summarized via their harmonic mean. Spectral unlearning consistently outperforms a broad range of baselines and closes most of the gap to a retain-trained reference. Code is available at https://github.com/unlearn-ml/spectral-unlearning/.

Large Reasoning Models (LRMs) produce explicit chains of thought to improve transparency, but they frequently overthink, generating long yet inefficient reasoning chains that inflate computational costs and risk introducing errors. While numerous reinforcement learning methods struggle to penalize verbosity to achieve conciseness, they typically apply the penalty uniformly across lengthy sequences, which leads to credit confounding, that is, the failure to distinguish essential reasoning steps from superfluous ones. In this paper, we study a measurable proxy for this confounding: high-entropy tokens that often coincide with surface transition points where reasoning re-enters verification. Based on this insight, we propose Adaptive Entropy-Sparing (AES), which selectively spares high-entropy tokens that occur before an efficient in-group reference, while penalizing overlength high-entropy suffix tokens associated with superfluous branching. Our AES thereby navigates the suppression-exploration trade-off, reducing overthinking while preserving average reasoning accuracy. We also present an entropy-length model that motivates why local uncertainty can serve as a proxy for future reasoning cost. Across eight reasoning benchmarks, AES reduces reasoning length by up to 52.9% while increasing mean accuracy by up to 3.8% over the base models, consistently outperforming existing efficiency-oriented methods. Our code is available at https://github.com/MobiusDai/AES.

Random tests of stateful code check laws, like reading back a written value, at states that random writes reach. The laws hold at every state exactly when the code behaves like a record with hidden data, so a pass is taken as evidence. This can fail: the laws can hold at every sample while the code is far from any record. The laws are local, but behaving like a record is global. Self-testing and stability theorems bridge this gap when the query distribution is controlled by the tester or preserved by the tested maps, but here it is generated by the tested writes, which need not be invertible. We ask how far a test must explore. An interface has $m$ binary variables, each with a read and two writes; in a record, reads return their bits and writes set only their bits. Samples are the ends of geometric runs of mean $\ell$ of uniformly random writes from a start distribution; a tester adds its own write chains of depth $d$. The distance is the least, over records with hidden data, of the largest probability that a read or write differs from it at a sample. From one start state, law checks at $O(m^2\log m/ε)$ samples certify, at fixed confidence, any threshold $ε$ after geometric runs of every positive mean, while runs of fixed positive length can pass at distance $1/2$. For the worst start distribution, take $ε=m^{-β}$ with fixed $0<β\le1$. For $d\le cm$ with fixed $c<1/2$, a polynomial budget of samples and writes exists exactly when $\ell=Ω(m)$, where law checks suffice. With no runs, the least depth for this budget is $(1/2+o(1))m$. Counting samples alone, in the worst case at $ε=1/(32m)$ and $d\le m/16$, the critical mean for a polynomial budget is $Θ(m\log(m/d)/\log m)$, by exhaustive search. So one start state needs only law checks, and the worst case at these thresholds needs exploration linear in $m$ if writes are counted.

Distribution Matching Distillation~(DMD) enables strong few-step diffusion generation by matching the student distribution to a pretrained teacher. Existing improvements to DMD often augment its supervision or modify its optimization machinery. We ask a complementary question: \emph{can DMD diagnose its own updates using signals already produced during training?} DMD inherently computes the discrepancy between a pretrained teacher and a fake model tracking the student distribution, but conventionally uses it primarily to determine the distribution matching direction. We find that the magnitude of this teacher--fake disagreement provides additional diagnostic information. Across extensive experiments, large disagreement in the high-noise regime consistently coincides with substantially higher-energy DMD updates. This reveals a dual role of disagreement: it determines not only \emph{where} the student should move, but also provides an intrinsic signal of \emph{how aggressively} the update should be applied. Based on this observation, we introduce Disagreement-Aware Distribution Matching Distillation (DA-DMD), which activates only in the high-noise regime and continuously modulates the update strength according to teacher--fake disagreement. DA-DMD reuses quantities already available during DMD training and requires no additional model, evaluator, or inference-time modification. We show that it exactly recovers vanilla DMD outside the selected region, reduces update energy only through selected samples, and exhibits redescending influence under extreme disagreement. Across multiple diffusion backbones, DA-DMD consistently improves few-step generation at the same inference cost. On SD3.5-L, DA-DMD improves HPSv2 from $28.23$ to $30.93$, while achieving a $64.4\%$ average win rate over DMD2.

Large language models are widely used in e-commerce user simulation, providing a foundation for understanding users and predicting human decisions. However, existing simulators still struggle to achieve both personalization and behavioral realism. Most methods predict the next action directly from interaction histories and the current context, without accounting for the evolving mental world that underlies user behavior. Theory of Mind suggests that behavior is driven by mental states, including beliefs, uncertainties, and intentions. Users may interpret the same information in distinct ways, leading to divergent behaviors, while an individual's mental state evolves through interaction and shapes subsequent actions. Ignoring this process prevents simulators from faithfully capturing individual differences and realistic behavior patterns, and undermines its effectiveness in downstream applications. To address this, we propose ShopMind, a Mental World Model for e-commerce user simulation that captures how consumers interpret shopping information and how their mental states evolve, better reflecting individual users and real-world behavior patterns. We further use behavior-mental state pairs synthesized by ShopMind to train a Mental State Reconstruction (MSR) model, which infers consumers' latent mental states from observable behavior. Experiments show that ShopMind improves personalization by 18.0% and behavioral realism by 19.6% over the strongest alternatives. Moreover, MSR improves performance on consumption value measurement and CTR prediction, showing the utility of ShopMind for understanding and predicting real user behavior.

Autonomous language-model agents are increasingly deployed on discovery-driven tasks where signal is sparse and most retrieved information is irrelevant. Current benchmarks fail to evaluate this regime due to dense signals, cooperative environments, and ground truth typically being only a few tool calls away. We introduce Legion, a benchmark framework coupling a stochastic enterprise-telemetry simulator (Legion-Sim) with a tiered library of 74 expertly-crafted adversarial scenarios (Legion-Bench) and an evaluation protocol that scores investigation efficiency alongside detection accuracy. Scenarios are injected into freshly simulated companies, allowing independent control over confounder density and event volume. We test five agent architectures on two simulated companies and find that while recall is high (0.82-0.88) precision lags behind (0.61--0.72). They fail to find all of the indicators of compromise of the attack or mistake genuine indicators with attack related ones, which is critical in the context of cybersecurity investigation. Failures are consistent across reasoning tiers and persist across confounder density and event volume. We release Legion, the scenario library, performance on common agent architectures, and pipeline for reproducibility and extension.

Generalist robot policies often fail at one step of a task they can otherwise perform. We ask whether such a failure can be repaired by editing the policy's weights directly, without gradients, training data, or a change to the instruction. Told to put the cube then the banana in the bowl, a zero-shot $π_{0.5}$ vision-language-action policy in RoboLab releases the cube before it reaches the bowl. Replaying 21 of these decisions under rewordings shows that it binds "in the bowl" to the wrong object. Closed-form instruction distillation copies the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rises from 27 to 106 episodes, with 81 fixed and 2 broken ($p=7\times10^{-22}$), while an update of the same size in random directions leaves success unchanged. The edit also lifts two other tasks without detectable harm on the remaining 13, and a variant that stores 0.16% of the parameter count keeps most of its effect. Placed in the action expert or action head instead, the same distillation helps less or not at all. Our released tool reproduced the fix end to end on new held-out seeds, from 8/64 to 36/64. Edits aimed at perception, a missing grasp, contact physics, or the gripper threshold did not help, and of seven blind Claude Opus editors, only the one given the misread instruction produced an edit that passed held-out scoring. These results suggest that weight edits work best when a known-good input already exists to distil. Because reruns with unchanged weights flip about a third of paired outcomes, evaluations should pair seeds, rerun the unedited policy, and confirm gains on held-out seeds.

Answer probes applied to chain-of-thought hidden states are often read as evidence that a language model has already decided its final response. This reading overlooks a competing signal, the candidate the model is about to mention. We introduce emission-controlled dual-target probing (ECDP), which fits the same low-rank reader family separately to the imminent mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed before any fit. On correct DecisionDepth (DD) rollouts in which a distractor is mentioned before the final answer, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696 [.648, .744]; Llama-3.1-8B gives .575 [.514, .632]. This pattern is not universal. In a pre-specified evaluation of Qwen3-8B with thinking enabled on held-out MMLU questions, using a full-prefix attention reader over every thinking position, both targets are decodable and the final answer is read far more strongly (mention-minus-answer lift -.398 [-.459, -.336] at layer 22 over 467 questions). Model, task, positions, and reader change together between these settings, and an attempted matched comparison on DD produced no valid statistics because the imminent-mention probe did not train reliably, so we do not identify what reverses the order. On DD, planted-signal controls show that the reader recovers an injected final-answer signal at one tenth of full dose. Answer-candidate decodability is therefore target-relative and protocol-dependent: a strong answer readout does not by itself establish when a model commits, and a probe that fails to train is not evidence that the information is absent. ECDP is a practical control for distinguishing imminent text emission from evidence of a settled decision.

Securing patent protection across countries requires drafting claims separately for each national patent office, in the relevant language and according to its legal and stylistic conventions. Patent claims can therefore differ across languages and jurisdictions because of translation, target-office drafting practices, and prosecution-stage changes. We formulate cross-lingual patent drafting as generating plausible candidate claim sets for a specified target language, jurisdiction, and prosecution stage from a source claim set in another language. The task models jurisdiction- and stage-specific transformations beyond content-preserving translation. We construct a large-scale Japanese-English dataset by linking Japan Patent Office (JPO) and United States Patent and Trademark Office (USPTO) records through 884,179 Patent Cooperation Treaty (PCT)-derived patent families. Across six released configurations, these families yield 4,067,168 directional claim-set instances; four core configurations form the main benchmark. Commercial machine translation and zero- and few-shot prompting capture part of the task, while supervised adaptation provides the strongest overall baseline, showing that task-specific patterns can be learned from linked cross-jurisdictional claim sets.

One-step stochastic world models avoid the mode averaging of deterministic feature predictors without the tens of network evaluations per step that diffusion and flow matching require. The simplest such objective is nearest-of-m matching (conditional implicit maximum likelihood estimation, IMLE): draw m candidate futures and pull only the one closest to the observed outcome. Recent work conjectures that a small m suffices for feature-space world models and that a strictly proper energy score should replace the hard minimum. We test both claims with one-step world models on frozen DINOv2/v3 patch tokens, using benchmarks whose futures can be enumerated. Under mode-like uncertainty, nearest-of-m with m=16 covers the enumerated futures better than 50-step flow matching (at about 1/30 of its latency), one-step flow generators and an energy-score generator at matched parameter counts, and its rollouts stay on the token manifold where flow rollouts drift. One quantity governs the choice of m: the nearest-of-m objective is the expected distortion of a random codebook of m draws from the generator. This view yields three predictions, recorded before the corresponding runs: (i) memorization of the realized outcome in contexts that never repeat, whose strength is governed by the codebook's miss distortion I(m) rather than its miss probability, and which an energy-score generator does not escape by raising m; (ii) a flattening of mode probabilities with m, matching a known optimum within 0.01; and (iii) under-dispersion on continuous high-dimensional outcomes, which bounds the method on real robot video. Per-step matching on self-rollouts cuts long-horizon error by a third on a stochastic Push-T variant; an ablation separates its manifold and coverage benefits. Planning success does not improve despite accurate risk estimates, a negative result we report with its diagnosis.

Traffic state estimation (TSE) reconstructs traffic density at unmonitored locations from a sparse set of sensors. Data-driven models work well when sensors are dense but degrade sharply as coverage thins. We show that imposing the Lighthill-Whitham-Richards conservation law as a physics-informed prior makes TSE more data-efficient. On four real corridors (NGSIM, I-80, PeMS04, PeMS08), physics-informed estimators generalise to held-out sensors where data-driven models, including classical and deep kriging, overfit the observed ones. On the two shock-rich corridors the effect is a clear crossover: as more sensors are held out, data-driven models collapse while the physics-informed estimator degrades gracefully, leading the no-physics network by about 26-38% at 85-90% held out and the strongest classical kriging by about 10-21% at 90%. A matched smoothness prior does not reproduce the gain, and an extended Kalman filter on the same LWR model is outperformed on the shock corridors, so the advantage comes from how the conservation law is imposed. Finally, one loss-conditional PINN removes the per-corridor loss-weight search, matching a 20-point search at about 5x lower cost.

We study the minimax regret of full-information online decision making in extensive-form games. For prediction with expert advice, the optimal regret is $Θ(\sqrt{T\log K})$, where $K$ is the number of experts, or equivalently, the number of normal-form actions. For treeplex strategy spaces, existing efficient algorithms, including KOMWU and online mirror descent with the weight-one dilated entropy regularizer, achieve regret $\mathcal{O}(\sqrt{T\log|\mathcal{V}|})$, where $\mathcal{V}$ is the set of reduced normal-form strategies. We prove that this logarithmic dependence is minimax optimal in the worst-case full-information reward model. The main difficulty is that pure plans in a treeplex are not independent experts: their rewards are correlated through the recursive structure of decisions and observations. To capture this correlation, we introduce tropical Gaussian formulae, recursive Gaussian expressions built from maxima and normalized sums which preserve the variance. The index count of such a formula matches the number of reduced normal-form continuation plans in the corresponding subtree. Our main analytic result shows that every log-balanced tropical Gaussian formula with index count $N$ has Gaussian width $Ω(\sqrt{\log N})$. While a direct approach struggles to recover this lower bound, our proof uses a spherical entropy profile to measure scale-dependent separation among the Gaussian directions, and then applies Turán's theorem and Sudakov minoration to obtain the width lower bound. Applying this result to treeplexes gives an oblivious hard distribution with valid transition outcome and Rademacher terminal rewards, establishing a regret lower bound of $Ω(\sqrt{T\log|\mathcal{V}|})$. This matches known upper bounds up to universal constants and shows that the DilEnt/KOMWU dependence on the reduced normal-form complexity is unavoidable.

Exploration in deep reinforcement learning (RL) for continuous control is commonly driven by temporally uncorrelated Gaussian (white) action noise, which leads to jerky motions with locally constrained exploration. In contrast, temporally correlated coloured noise, particularly pink noise, has been shown to support more efficient exploration in RL. Motivated by these findings, we inquire whether scheduling changes to the noise colour over the course of training can outperform white noise for continuous control robotics. We perform a systematic study of 49 linear schedules between white and red noise, together with generalised state-dependent exploration (gSDE), on four sparse-reward manipulation tasks with a Fetch arm and a Shadow Dexterous Hand. We find that the noise colour at the end of training does not affect learning efficiency or end performance, whereas fading towards red noise halves the jerk of exploration in every task. Balancing learning efficiency, end performance and smoothness, red noise emerges as the preferred strategy: it learns as efficiently as pink noise with almost 2-fold less movement jerk. We therefore recommend exploration schedulers that fade towards red noise as the default for off-policy robot learning.

When competing campaigns target a population of large language model (LLM) agents, observed opinion changes may reflect both campaign influence and changes that would occur without intervention. We study how the timing of campaign expenditure and the population's own evolution shape outcomes in a network of persona-conditioned LLM agents that exchange and revise opinions in natural language. A proponent and a challenger are tasked with generating opposing messages under equal delivery budgets, while a separate stance probe measures opinions without influencing the simulation. Across eight claims, we compare single-campaign and competitive conditions against matched no-campaign runs, and evaluate all nine combinations of early, sustained, and late spending on a reference topic. Baseline stance drift varies substantially across claims and, on some, accounts for a sizeable fraction of apparent campaign effects. After accounting for this drift, the direction of observed campaign-associated changes aligns more closely with the claim's epistemic standing than with the campaigner's assigned role. Uninfluenced populations also consistently lose stance diversity, demonstrating that directional opinion change and convergence are distinct phenomena. In networks with two opinion-segregated communities, campaigns primarily reduce differences between communities rather than diversity within them. LLM campaign agents do not always argue for their assigned positions, particularly when these conflict with established evidence. Finally, campaign outcomes depend on the opponent's expenditure schedule: unlike a matched numerical bounded-confidence population receiving interventions at the same times and locations, the LLM population exhibits a payoff structure with no pure-strategy equilibrium.

Submodular maximization seeks a small subset best representing a large dataset. In its federated setting, the server is a single point of failure and some clients may be Byzantine. Removing the server raises a trilemma of decentraliza- tion, efficiency, and security. We propose DeFSM, a Byzantine-robust and communication-efficient architecture that replaces the server with a lightweight smart contract and a pool of independent, consensus-free agents. Its aggregation rule, SafeGM, builds on two notions from combinatorial geometry: Tukey depth and safe points (the points lying in the convex hull of the good points wherever the bad points are). SafeGM makes the robust decision on 2D Gaussian random projections, which the contract can compute efficiently, and lifts it back as a convex combination of actual updates. Since lifting loses information, it uses the projected geometric median only when its depth certifies it, and otherwise a Tukey median. The share of bad updates is then bounded only by a count: how many there are, not what they contain. We prove, and confirm empirically, that this bound controls the solution quality. On CIFAR10 and PathMNIST, with up to 30% Byzantine clients and agents, DeFSM never collapses and retains on average about 91% of the quality of the ideal reference, the geometric median at a trusted central server. The geometric median decentralized in its best setting would retain about 96%, but the contract cannot compute it, and so each agent would have to download six times more data, or each client twice as much.

Sensory delays create temporal misalignments between the state of the environment and the observations available to an agent, turning Markovian control problems into partially observable ones. In this work, we inquire whether such misalignments can be mitigated during training by using hindsight to recover aligned state-action transitions. We introduce Hindsight Critics, an asymmetric extension of the Soft Actor-Critic for reinforcement learning under partial observability. The critics are trained on hindsight-aligned transitions whereas the actor learns from the resulting privileged value estimates despite only having access to delayed observations. Comparisons against delay-aware and unaware baselines on the DelayBench suite show that Hindsight Critics can partially compensate for observation delays without requiring modifications to the agent's observations during deployment.

Implicit in much research in deep learning is the assumption that if a network achieves low loss, then it must have learned a good representation of its training data. Recent work by Kumar et al. (2025, arXiv:2505.11581) questions this assumption, arguing that networks trained with conventional gradient descent achieve low loss despite learning fractured entangled representations (FER), leaving related concepts disconnected and unrelated concepts entangled inappropriately. Kumar et al. further show that dramatically better representations are possible: a network evolved through an open-ended, human-in-the-loop process learned unified factored representations (UFR) that capture the regularities of the domain far better than a counterpart with nearly identical outputs trained with conventional gradient descent. Inspired by UFR, this paper introduces unified factored organizers (UFOs), networks trained to model regularities in the representations of other networks. A network can then be trained both to achieve low loss on its task objective and to align its representations with the model learned by the UFO. Networks trained with UFOs are shown to learn qualitatively superior representations in an image-regression domain where representations are easily interpretable. To provide quantitative evidence for the impact of UFOs on representation learning, a transformer is trained with and without UFO regularization to grok a modular arithmetic problem set. When combined with weight decay, UFOs can reduce the number of transformer training steps needed to grok by more than 10x relative to weight decay alone. Without weight decay, UFOs can reduce the number of transformer training steps needed to grok by more than 100x relative to an unregularized baseline. If UFOs are real, they suggest a new paradigm where loss functions themselves are trained to learn an effective model of good representation.

Prompt injections have emerged as one of the top security risks faced by AI agents today. Early work in prompt injection attacks looked at simple tool-call agents and static attack payloads, and were saturated since mid-2025. As of Fall 2026, human red-teaming may be nearing saturation, especially on the Gemini family of models. In this work, we argue that despite this apparent plateau, prompt injection attacks are still possible and continued effort into prompt injection defense is needed. We design a new synthetic evaluation dataset which uses random non-malicious actions as adversarial goals. Such dataset design helps us to evaluate prompt injection robustness as a model ability to respect instruction hierarchy, and disentangle this ability from model knowledge of security practices. We show that an agentic attacker of increased capabilities can reliably trick model to follow instructions in tool output in our evaluation, even for models previously considered robust on public prompt injection benchmarks such as GraySwan Arena. This shows that current frontier defenses may require proactive, automated red-teaming to remain resilient to increasingly capable adversaries.

Fleets of driving vision-language-action (VLA) models are developed one domain at a time: each operator finetunes a shared base model on private data, producing specialized experts that are costly to consolidate into a single model. Weight merging loses accuracy as heterogeneous domains accumulate, and dense mixture-of-experts routing is prohibitively expensive to combine them. We present DeltaForge, to our knowledge the first post-hoc mixture-of-experts VLA for driving, which consolidates N independently finetuned per-domain models into a single multi-domain model from their weights and a small proxy set alone. When different operators finetune the same base model, their weight updates end up sharing a lot of structure. We exploit this by decomposing all the updates together using singular value decomposition (SVD), splitting each one into a shared part (the same across all domains) and a tiny domain-specific part. The router then only needs to work with those tiny domain-specific pieces, not the full expert weights, which is what keeps the approach efficient. Then two cooperating clip-level blind routers, one over the VLM reasoning experts and one over the flow-matching dynamics experts, are trained from soft labels on the proxy dataset. Across six diverse domains spanning urban, off-road, and highway driving on cars, trucks, and all-terrain vehicles (ATVs), DeltaForge approaches the accuracy of the per-domain experts and surpasses every merging, routing, centralized, and federated baseline at close to single-model parameter and compute cost. Adding a new domain requires only projecting its weight update onto the shared basis with no end-to-end retraining needed. The resulting model also generalizes to domain combinations it was never explicitly trained on.

Vision-language-action (VLA) models pair a vision-language backbone with an action head, but in closed-loop control the backbone sits on the control-critical path: while it computes, the robot keeps moving, so each executed action is conditioned on an increasingly stale observation. A larger backbone therefore directly blocks reactivity, especially in humanoid dexterous manipulation. We present DASH, a Decoupled Asynchronous System for Humanoids that removes this coupling. DASH splits the policy into a slow vision-language model (System 2) and a fast flow-matching action head (System 1) that run as concurrent processes over the LCM communication protocol, so the control rate is set by the action head alone, independent of backbone size. The action head fuses stale semantics from a frozen backbone with fresh vision from its own ResNet encoder, desynchronized training exposes it to the staleness it meets at deployment, and an adaptive temporal aggregator aligns overlapping action chunks under variable inference latency. On an asynchronous RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz control rate from 0.8B to 8B backbones while the synchronous rate falls from 9.7 to 4.1 Hz, with higher success at every size. On a real humanoid with dexterous hands, DASH achieves clearly higher pick-and-place success than the synchronous pipeline.

How many observations are needed to estimate a common rotation in two-dimensional overcomplete independent component analysis? We study equally spaced mixing columns with known source count, common source law, and column norm. For a fixed non-Gaussian fourth-order Gram-Charlier source law, the rotation is identifiable modulo the column spacing. Yet the sample size for fixed relative angular accuracy grows exponentially with the source count. We characterize this sample size, up to absolute constants, by the reciprocal of the squared Hellinger distance between two rotations separated by half the column spacing. An explicit radially weighted phase estimator attains this scale uniformly over the rotation family, using the observation radius to prevent cancellation in the mean of an angular harmonic. We also determine the exact exponential decay rate of single-observation total variation for the same pair. This gives bounds on the sample size exponent; its exact value remains open. Finally, a noisy binary-source example shows how an exact zero line of the observed characteristic function can have an exponentially small normal derivative.

Large language models have improved Verilog generation, yet functional correctness remains difficult to optimize because simulation commonly provides a coarse module-level pass/fail outcome. Such feedback discards useful evidence from partially correct RTL modules and cannot determine whether a source statement is functionally responsible for an observed output behavior. We present a new method named Causal Signal Credit Assignment (CSCA), which is a counterfactual preference-optimization framework for Verilog generation. CSCA first performs continuous signal-level behavioral verification on sampled candidate modules. It then locates output-driving assignment spans and applies five interventions while holding the remaining module and testbench fixed. The resulting changes in signal-level errors are combined with behavioral agreement to form a continuous statement credit, which is mapped to response-token weights for strict weighted direct preference optimization. This enables CSCA to learn from locally supported statements without uniformly reinforcing an otherwise incorrect response. Built on Icarus Verilog and a CodeV-derived corpus, CSCA is evaluated on VerilogEval~2.0 and RTLLM~v2.0, including both code completion and specification-to-RTL tasks. With a Qwen2.5-Coder-7B-Instruct backbone, CSCA leads the evaluated 7B-scale models and remains comparable to the much larger GPT-5.4-mini on the reported Verilog-generation benchmarks.

Medication recommendation is a multi-label prediction task. However, existing prediction-set evaluation and calibration methods rely mainly on exact-overlap metrics and therefore treat all missed medications as equally severe, overlooking how closely the prediction set matches them within the medication hierarchy. To address this, we introduce Conformal Severity Control (CSC), a post-hoc framework for calibrating medication prediction sets under hierarchy-aware mismatch severity. CSC uses the Anatomical Therapeutic Chemical (ATC) hierarchy to quantify mismatch severity and controls patient-level severity without retraining the predictor, supporting both average severity and upper-quantile severity for the highest-severity mismatches. Experiments on MIMIC-IV show that CSC tracks user-specified severity tolerances and shifts prediction sets away from distant mismatches toward exact or close matches.

Out-of-distribution (OOD) generalization remains a central challenge in materials property prediction, where models must extrapolate to novel chemistries beyond the training distribution. While large language models (LLMs) have demonstrated strong scientific reasoning, their direct use for quantitative prediction remains limited. To address this, we systematically evaluate multiple roles for LLM assistance, finding that only taxonomy-based grouping consistently improves OOD generalization over a strong kNN+GBT baseline. Building on this, we here propose MOTIF, a framework that leverages LLMs not as predictors, but as generators of candidate taxonomies that partition materials into chemically meaningful environments. These taxonomies are evaluated via training-only pseudo-OOD validation, and only those that improve generalization are retained. Invariant feature selection is then applied within each accepted taxonomy, followed by gated aggregation across taxonomies. We evaluate on 2D halide perovskites, where bandgap prediction must generalize across distinct organic spacer families. On this benchmark, MOTIF reduces mean OOD MAE from 0.122 to 0.111 eV. Under a frozen dev/test protocol, the deployable configuration further improves errors from 0.123 to 0.117 eV (electronic bandgap) and from 0.181 to 0.173 eV (optical bandgap). Beyond this primary task, we validate transferability across additional materials datasets, including 3D perovskite solar cell efficiencies and Matbench benchmarks with structure-based splits. Across domains, MOTIF consistently improves robustness under distribution shift, outperforming expert-defined groupings, rule-based heuristics, and clustering baselines, while remaining competitive with strong classical selection methods. These results demonstrate that LLMs are most effective not as predictors, but as tools for discovering structure in data.

Connected physical systems fuse observations from robots, IoT sensors, and edge-AI modules into a shared account of the world, yet most evaluations score only the final answer or action and cannot tell whether a method attributed the evidence, updated its belief, handled contradictions, and justified its decision. We argue that shared reasoning in Connected Robotics should be evaluated as an evidence-to-belief trajectory, and introduce the Connected Robotics Evidence Graph (CREG) Benchmark to make that trajectory measurable. CREG comprises 350 deterministically generated scenarios across seven regimes (justified revision, belief preservation, relation association, contradiction onset and resolution, stale/delayed evidence, and source reliability), each exposing a normative posterior reference at every evidence prefix, validated by an independent log-odds checker and screened by a blinded expert audit. Nine representation conditions separate relation content from serialization, ordering, and metadata. Evaluating four open-weight instruction models and six transparent fusion baselines yields three findings. First, entity grounding accounts for a substantial part of the weak-model deficit. Second, deleting semantic relations costs 15.3 points of strict belief-state accuracy on relation-dependent scenarios (Holm-corrected p<0.001), but an information-matched structured list outperforms the full graph and shuffling the graph does not hurt: relation content matters where required, while graph serialization confers no consistent advantage. Third, belief accuracy is strongly non-monotonic along the evidence sequence, with mid-trajectory failures that endpoint-only evaluation cannot detect. All scenarios, references, prompts, metrics, and analysis code are released for offline, hardware-free reproduction.

An answer can match the reference for a complete agent history even when the reader's visible history does not determine it. We study this distinction using executable device logs. Our protocol separates reference agreement, certified evidence sufficiency, and context-appropriate behavior. Insufficiency requires two legal histories yielding byte-identical visible contexts but different answers; public chronology and query cutoffs certify sufficient prefixes. On 80 confirmation histories, Qwen3-32B matches 87 hidden references among 203 insufficient raw-prefix instances, with only seven appropriate abstentions. A fixed-input DeepSeek-V4-Flash replication matches 85 and abstains on 21 of the same 203 instances. Each reader makes 1,600 requests. Compact CSV's estimated appropriateness effects remain inconclusive: +3.16 percentage points, clustered 95% interval [0.00,6.57], for Qwen and +2.46, [-1.84,6.94], for DeepSeek. A separate held-out judge study achieves 99.38% correctness but retains support errors and truncations; a deterministic baseline solves all constructed cases. These results support an evaluation diagnosis across two hosted reader configurations, not gains for a general memory system or natural-language evidence assessment.

Out-of-distribution (OOD) detection is essential for deploying machine learning models in high stakes scientific applications, yet methods that flag OOD inputs without retraining the model have been validated almost exclusively on image classification. We introduce MOOD (Materials Out-of-Distribution Detection), a benchmark for evaluating whether these post-hoc OOD detectors transfer to graph neural networks (GNNs) on materials property prediction. MOOD evaluates 15 post-hoc detectors and 2 supervised diagnostic probes across 4 GNN architectures, 5 MatBench tasks, and 15 physically motivated compositional and structural shifts, yielding 264 evaluation settings and over 13,000 AUROC measurements across three encoder training seeds. Our evaluation reveals three key findings. First, current post-hoc detectors perform near chance; all 15 detectors achieve an AUROC around 0.5. Second, detector failures decompose into two regimes: method bottlenecks and encoder bottlenecks. Method bottlenecks arise when supervised probes can separate in-distribution (ID) from OOD samples but unsupervised detectors fail, meaning that the embeddings contain OOD signal that current detectors do not extract. Encoder bottlenecks arise when even supervised probes fail to separate ID from OOD samples, indicating that the relevant shift is not accessible in the frozen GNN representation. Third, when OOD prediction error exceeds ID error, the AUROC of distance-based and density-based detectors correlates with the size of that gap. Collectively, these findings indicate that reliable OOD detection in scientific domains requires representation learning strategies that explicitly preserve OOD relevant signal, particularly for tasks where current GNN architectures encode no extractable signal at all.

As language models (LMs) are increasingly deployed in downstream applications, accurately estimating their likelihood of success on a given example is critical. However, confidence estimates from LMs are poor predictors of their likelihood of success. Accurate confidence estimation is linked in part to the detection of aleatoric uncertainty; we hypothesize that LMs possess representations of at least some aspects of this uncertainty in the form of representations of the inherent ambiguity of the input. Using difference-in-means on minimal pairs of ambiguous and unambiguous sentences, we derive ambiguity representations in LM activations, and find that steering with them selectively (albeit weakly) modulates LMs' judgments of the input's ambiguity. We then propose DART, a fine-tuning method that encourages updates to lie within the span of the ambiguity directions, and penalizes changes to the LM activations outside of this subspace. DART greatly improves calibration, encodes these improvements into the model parameters, and maintains model accuracy on multiple-choice and generative QA tasks.

Vision-language models (VLMs) could support medical-image interpretation, but their use requires deciding which interpretations to accept and which images to refer to a clinician. We test whether thresholding stated confidence can retain useful coverage while limiting errors among accepted answers, and how this depends on the population. Eight open-weight VLMs, including two reasoning models, assess 3,888 brain MRI slices from post-treatment glioma patients (BraTS 2024) and healthy volunteers (IXI) for tumor or gross abnormality. BraTS labels are mask-derived; IXI slices are treated as negatives. On BraTS, no threshold selects a nonempty set with at most 5% empirical error for any model, even when chosen retrospectively. This also holds for four models evaluated on a large-lesion, distant-negative subset. At a 10% error budget, Gemma-4-12B retains 24.4% of slices and every other model at most 2.5%. To investigate this limited selection ability, we separate confidence differences between predicted answers from differences within each answer. For four models with available outcome-level summaries, pooled mean confidence differs more between answers than between correct and incorrect responses within each answer. A covariance decomposition explains how this pattern can produce opposite confidence-correctness associations across populations. At a 20% error budget, accepting every "tumor" answer covers more slices than any confidence threshold for two models. These findings caution against using unvalidated stated confidence as a standalone deferral signal and support population-specific validation before deployment. Code is available at https://github.com/amir-sbg/VLM-Answer-or-Defer.

Nonseparable costs and coupled constraints in distributed online convex optimization often arise when system performance and operational constraints depend on the joint configuration of all agents and cannot be decomposed into individual contributions. Such nonseparability leads to fully distributed decision-making with challenging partial local information, where each agent evaluates its local primal-dual update using estimated global information via a local belief of the joint action and the shared Lagrange multiplier to be exchanged with neighbors. While the belief consensus established by prior works ensures no-regret learning, exchanging beliefs comprising full-dimensional primal and dual states at every round imposes a communication overhead that scales with both the joint decision dimension and the network size. Existing event-triggered schemes typically activate communication according to state variation or consensus error. Under nonseparable coupling, however, errors of equal magnitude can produce markedly different perturbations after passing through the KKT operator. We therefore propose a randomized KKT-perturbation-aware communication mechanism, where the triggering of each agent's primal and dual communication channels depends on their impact on the KKT operator. The proposed mechanism transfers the adaptive-sampling principle of coordinate descent from allocating computation across decision blocks to allocating communication across information channels: channels with larger estimated perturbations receive greater priority for refresh. Under standard regularity assumptions, the resulting algorithm achieves $O({T}^{2/3})$ expected static regret and $O({T}^{2/3})$ expected cumulative constraint violation. Numerical experiments demonstrate improved learning and constraint-enforcement performance per transmitted bit over belief consensus baselines.

When an AI diffusion-type generator samples a frustrated spin glass, which macroscopic alternatives does it decide first, and when? We study this for bridge-diffusion models of the Path Integral Diffusion (PID) type started from a point source. We show that every isotropic PID is a time-changed Gaussian stochastic localization, and conversely every signal-to-noise schedule is realized by a PID protocol. What a continuous-time generator decides therefore does not depend on the protocol, which only sets when and at what cost. We describe the generator's state by its partial commitments: the probabilities it assigns, at each moment, to nested classes of configurations fixed by the target. The partial commitments are martingales that move only along class contrasts. Their ensemble average gives the information about each level at an exact rate set by squared magnetization contrasts. Each split is decided, approximately, like a binary Gaussian channel at signal-to-noise ratio 1.04/λ_P, with λ_P the leading eigenvalue of the parent's correlation matrix. The theory is applied to Bethe-lattice and Sherrington--Kirkpatrick (SK) glasses with up to 64 spins. It is observed generically that a trajectory first shifts its weight to one of the two mirror halves of the zero-field measure and then to the glassy classes inside it. These classes appear near the glass transition and keep their identity on cooling. Nested glassy levels have similar λ_P and are decided together, so a trajectory commits in one or two distinct events---the mirror choice and a compound glassy decision---which separate further with coupling and size, as the spectra predict. For a generator that takes finitely many steps, the protocol matters through the late, local stage rather than the macroscopic decisions.

Combining video generation, dynamic reconstruction and monocular SLAM requires consistent coordinates, explicit handling of missing outputs and predictable memory use. We present AcrossVAM3.0, an offline system that converts text, images or monocular RGB video into a versioned point representation. Frozen components run sequentially on one GPU; map corrections update cameras, points and trajectories together. The coordination layer adds no neural parameters. Three fresh input-mode runs completed with a maximum sampled task-process-tree GPU footprint of 17.14 GB. In a frozen 16-scene video suite, every scene yielded a 4D reconstruction and 15 yielded a non-empty SLAM map. Strict cross-window matching reduced local surface error but lost long-range track coverage on TAP-Vid-DAVIS. We therefore developed multi-hypothesis adaptive surface chaining (MHAC), which retains several candidate matches and an explicit missing state. After selection on disjoint development data, MHAC improved Average Jaccard over a position-only baseline by 0.1394 on a frozen ten-video holdout (95% bootstrap interval [0.1074,0.1689]). Across 265 RoboTAP videos, the mean paired gain was 0.0610 ([0.0510,0.0715]), with 215 wins, one tie and 49 losses. These results support resource-bounded scene construction and an external tracking gain under the tested protocol. They do not establish metric scale, semantic identity or high-fidelity novel-view rendering.

Alignment of text-to-image diffusion models with reward functions is a key challenge in controllable generation. Existing approaches typically rely on optimizing model inputs or fine-tuning denoising networks, which requires computationally expensive backpropagation through large denoising models. In this work, we propose Direct Trajectory Optimization (DTO), a training-free framework that performs alignment by directly optimizing the sampling trajectory of diffusion models. Specifically, DTO treats intermediate denoising scores and sampled noise as the optimization variables, eliminating the need for gradients through the denoising network. To mitigate the out-of-distribution problem, we introduce Reward-Optimized Score Guidance, which interpolates between the optimized scores and the pretrained model predictions to stabilize sampling, while providing a provable theoretical guarantee of reward improvement under constrained distributions. Empirically, DTO improves alignment performance across three diffusion models on two benchmarks, outperforming existing inference-time alignment methods while maintaining high efficiency. These results suggest that trajectory-level optimization is a promising direction for efficient and effective diffusion model alignment.

Fine-tuning public pretrained language models with lightweight adapters such as LoRA has become the mainstream way to adapt them to downstream tasks. A service provider often trains its adapter on a data owner's samples, so two assets need protection: the data owner's samples and the provider's adapter. Existing secure fine-tuning protocols protect the samples but reveal to the data owner intermediate results of every training step, such as the hidden representations or logits of its inputs. We show that a single step suffices to copy the adapter: from the hidden representations of one mini-batch of 32 samples, the data owner fits an adapter that agrees with the provider's on 99.2% of test inputs, and logits alone yield about 94%. Protecting both assets therefore requires that no value that depends on the adapter be revealed throughout training. We present VeilTune, to the best of our knowledge the first protocol that fine-tunes a pretrained language model with backpropagation entirely in secure multi-party computation (MPC). VeilTune computes only what the loss depends on, replaces generic MPC operators with training-aware nonlinear operators that stay accurate on the values that training produces, and obliviously skips the steps corrupted by truncation failures. On four few-shot tasks with RoBERTa-large, VeilTune nearly matches plaintext training and is 5.3 points more accurate on average than SecP-Tuning, the state-of-the-art secure fine-tuning protocol. On SST-2, it achieves a 3.5x end-to-end speedup and a 3.6x reduction in communication over SecP-Tuning, and 26x and 17x over full fine-tuning in MPC.

Content moderation is inherently policy-dependent: moderation boundaries are often subtle, subjective, and frequently updated, while conventional rule based or fine-tuning based systems require costly rule rewriting, model retraining, or system redeployment. We propose CaseLaw, a policy-decoupled and training-free framework for adaptive content moderation. CaseLaw reformulates moderation as precedent-based retrieval and voting, where multimodal content is transformed into evidence-preserved structured semantics, embedded into a stable semantic space, and classified by thresholded voting over retrieved precedent labels. We train NoteCap on a large-scale corpus of structurally annotated notes to ensure consistent semantic extraction and freeze it during subsequent policy adaptation, which requires no additional training. Policy adaptation is achieved by updating the semantic case library, including case labels and mappings, without retraining models or modifying prompts. To evaluate CaseLaw, we construct MFO-RedNotes, a benchmark for the male-female opposition risk domain. We additionally provide a five category CoPro comparison. Experiments show that CaseLaw consistently outperforms competing baselines in both precision and recall across policy-dependent risk levels and broader safety categories, with category-wise precision/recall reaching up to 86.8%/90.1% on MFO-RedNotes across both policy settings and 94.5%/92.3% on CoPro. Our work highlights the potential of precedent-based reasoning for policy adaptive content moderation and offers a practical direction toward more stable, transparent, and training-free moderation systems.

We study online monotone submodular maximization under a cardinality constraint with bandit feedback. On each round, the learner selects at most $k$ elements from a ground set of size $n$ and observes only the value of the selected set. We develop an efficient algorithm achieving $O\!\left(k\sqrt{nT\log(n/k)}\right)$ expected $(1-1/e)$-regret over $T$ rounds. To our knowledge, this is the first polynomial-time algorithm with $\sqrt{T}$ dependence on the horizon for arbitrary monotone submodular functions in the bandit setting. Our approach uses randomized sampling and a carefully constructed potential to reduce online submodular maximization to online linear optimization over the probability simplex.

We study the learning of homogenized constitutive laws for inelastic heterogeneous materials, aiming to replace costly concurrent unit-cell simulations with surrogates that generalize across loading histories and microstructures. We focus on the recurrent neural operator framework, motivated by the explicit form of homogenized Kelvin--Voigt linear viscoelasticity. In this framework, internal variables enable a Markovian approximation of loading-history dependence, and neural operators represent the microstructure-dependent constitutive response. For periodic Kelvin--Voigt viscoelasticity, we prove a universal approximation theorem showing that the model can approximate the homogenized constitutive law to any prescribed accuracy, uniformly in time and over a wide set of admissible strain trajectories and microstructural property fields, in two and three dimensions. This comprises a significant extension of the existing, somewhat limited, one-dimensional theory that uses explicit solutions for piecewise constant materials. Instead, in two and three dimensions, we adopt a more versatile approximation based on the spectral representation of the memory kernel that arises from homogenization. This also enables us to give an explicit bound on the internal-variable dimension, sufficient for approximation, exhibiting logarithmic growth in the inverse relative stress error and in material contrast. Numerical experiments on viscoelasticity and crystal plasticity complement the theory, illustrating the roles of internal-variable dimension and sample complexity in learning homogenized responses from unit-cell data. Theoretical results are formally verified in Lean 4.

Knowledge graph question answering (KGQA) systems are typically evaluated on questions whose surface tokens lexically expose the gold KG path. Users, however, may ask questions without knowing the underlying schema: "Which condition could Andrew have due to family history?" compresses a 4-hop chain "Andrew --[father]--> parent --[father]--> grandparent --[medical condition]--> disease --[subclass of]--> hereditary disorder" into a single phrase (family history) that names none of those relations. We identify this collapse of path-question alignment as a blind spot in existing KGQA evaluation. Three major LLM-KG paradigms, hop-by-hop exploration, plan-then-retrieve, and similarity-based retrieval, implicitly rely on such lexical cues, an assumption rarely tested in practice. We make three contributions. (i) We introduce HiddenPathQA, a manually verified benchmark of 1,055 questions spanning 2-6 hops and multiple domains, constructed so that question wording does not reveal the gold path's relations. (ii) We propose Holistic Path Selection, which compares questions against candidate relation chains in the local subgraph as complete units, replacing irreversible per-hop relation scoring with globally informed chain-level reasoning. We develop two implementations, TieredHolistic and FlatHolistic. (iii) We conduct a pseudoword stress test that replaces entity names with meaningless placeholders to distinguish KG-grounded reasoning from entity-level memorization. Holistic Path Selection outperforms six strong KGQA baselines (ToG, PoG, FiDeLiS, R2-KG, KAPING, and KARPA) by 8.8-17.9 accuracy points across two LLM backbones and both original and pseudoword settings. Its advantage under pseudoword replacement shows that the gains cannot be explained by entity-level memorization alone.

Currently, artificial intelligence-generated content (AIGC) poses a severe threat to facial video deepfakes, and many AIGC companies employ video watermarking to mitigate unauthorized fake video usage, but its effectiveness is limited due to watermark tampering and private-model forgery. Additionally, some schemes actively add watermarks to user videos. However, current approaches focus primarily on the quality of watermark embedding in offline scenarios, and no existing work offers a solution for protecting human facial videos in real-time online applications. To fill this gap, we propose, to the best of our knowledge, a novel biometric video watermarking method, UltraGuard, for the first time. It uses the smartphone's built-in speaker and microphone to emit ultrasonic signals and capture biometric features, which are then fused with spatio-temporal dynamic features derived from the video content. The fused features are encoded as watermarks and embedded into videos in real time, enabling dynamic, content-aware watermarking throughout transmission and propagation. Extensive experimental validation demonstrates that UltraGuard can effectively defend against malicious attacks such as forgery and editing, both at the source and during propagation, proving the system's efficacy. In the future, it can be deployed on users' mobile devices to provide real-time protection for users' facial videos, thereby preventing security and privacy breaches caused by various malicious attacks.

AI agents are increasingly automating data analysis, but responsible adoption requires reliable performance evaluation, which is costly in specialized scientific domains. We introduce a framework that translates scientific publications into auditable benchmarks for AI agents and demonstrate its use on spatial transcriptomics alignment tasks. Across 121 tasks, 16 baseline methods, and six AI agents, the best agents (GPT-6-Astra and Gemini-4-Argon) matched or outperformed all baselines on 80% of tasks, often using surprisingly simple heuristics. Agent performance was robust to prompting and access to pre-installed packages. Benchmark construction also uncovered errors in prior work and shortcuts in current evaluation metrics. Finally, we show that agent-generated solutions can be distilled into a reusable program that ranks in the top 10% of baseline methods on unseen data. Our framework offers a scalable way to evaluate AI agents in specialized scientific domains by leveraging and stress-testing existing papers and metrics.

We introduce ChemInsight, a benchmark designed to evaluate large language models (LLMs) on chemistry tasks that require multi-step reasoning and structured chemical understanding. We propose a two-dimensional scoring system (scope and depth) to assess both task coverage and reasoning completeness. ChemInsight consists of 203 challenging chemistry questions organized into 50 bundles, which support the two-dimensional scoring evaluation scheme. We evaluate ten representative LLMs using ChemInsight and observe consistently low performance across models, with scope scores below 30% and depth scores below 11%, substantially lower than human expert performance. We then proceed to ask (1) What factor limits the advancement of current LLMs in chemistry? (2) To what extent are statistically driven language-model training methods effective at modeling chemistry symbolic language such as SMILES? And (3) Can LLM performance in chemistry be improved through intensive training on large, domain-specific corpora?

Preference optimization has emerged as a promising alternative to reward-based training for neural combinatorial optimization. However, existing methods typically rely on binary pairwise labels, which discard the magnitude of solution-quality differences and provide supervision only at the complete-solution level. We propose Advantage-aware Preference Optimization for Two-level Constructive Search (APO2CS), a preference-based training objective for constructive neural solvers that integrates both macro-level and micro-level supervision. At the macro level, APO2CS enriches pairwise preferences with an advantage term derived from objective-value differences. At the micro level, it compares partial trajectories, thereby providing denser supervision throughout the sequential construction process. The resulting objective can be used either to train a policy from scratch or to fine-tune a pretrained one. Experimental results show that APO2CS improves the expressiveness and search capability of preference learning for combinatorial optimization, while demonstrating strong effectiveness and adaptability across multiple combinatorial optimization tasks.

Robots are typically evaluated in episodes that begin from a reset, while industrial work consists of repeated jobs whose outcomes affect subsequent operations. Evaluating such work at scale requires simulation of task-critical physical interactions, construction of industrial environments, and measurements of performance as scene state persists. We present HOLIDAYFACTORY 1.0, a framework for simulating, generating, and evaluating industrial robot work. Given a human-specified workplace and task, the framework defines a world pairing a scenario of scene variants and process logic with an evaluation defining a job, metrics, and a rubric. HOLIDAYFACTORY 1.0 comprises three components. HOLIDAY-SIM extends Newton's SDF-based hydroelastic contact with additional constitutive laws and a semi-implicit primal-dual solver, couples stateful actuator dynamics to articulated motion, and derives tactile and joint observations from simulated states. Agentic world construction generates simulation-ready assets and scene variants using geometric and interaction checks, and proposes task-conditioned evaluations for human confirmation and executable checks. 24/7-BENCH introduces seven industrial tasks in which jobs are evaluated without resetting the scene, measuring completed, failed, and unfinished jobs, throughput, latency, and motor energy. Controlled studies characterize the proposed contact, actuator, and tactile models and examine the effectiveness of verification-guided world construction. Experiments on scene and evaluation generation further reveal the limitations of model critics and the difficulty of ensuring that generated evaluations faithfully capture task requirements. We examine job-level evaluations with scripted trajectories (e.g., code-as-policy) but have not yet compared policies over persistent runs.

Population game dynamics describe how aggregate behavior evolves in response to payoff signals. In many applications, the payoff function may change, and the goal is to predict the population trajectory induced by a new payoff design before trajectory data under that design are available. We study this problem as $\textbf{payoff-aware prediction of population game dynamics}$. Standard dynamics learning methods fit the motion observed under training payoffs, but do not separate the payoff-dependent incentive from the payoff-independent response structure needed for prediction under new payoffs. Building on the existing framework of Riemannian game dynamics, we propose PA-RmD, a payoff-aware learning method that keeps payoff information explicit and learns the payoff-independent response structure from data. Theoretically, we show that the learned model preserves fixed-payoff Nash equilibria and derive finite-time one-step prediction bounds under shared-response and payoff-coverage conditions. Empirically, we evaluate PA-RmD on synthetic zero-sum and potential games, as well as a YouTube Trending multi-country dataset, showing that our method performs well under changing payoff designs.

Electroencephalography (EEG) foundation models have emerged as promising paradigms for decoding brain activity through large-scale self-supervised pretraining. Notably, most existing approaches employ masked patch reconstruction, in which randomly selected EEG patches are masked and recovered by the network. However, purely random masking may limit representation learning as informative patches from highly correlated channels often remain unmasked, causing the network to underutilize weakly correlated channels or serial dependencies. Furthermore, because these models are typically based on transformer architectures, their performance is influenced by the choice of positional encoding. Although positional encoding has been extensively explored in natural language processing and computer vision, few studies have adapted it to the unique characteristics of EEG signals. To address these limitations, this study introduces a conditional positional encoding method based on graph convolutional networks (GCNs) to explicitly model inter-channel correlations, along with an adaptive masking strategy that enhances representation learning by modulating the difficulty of masked patch reconstruction. To the best of our knowledge, this is the first study to integrate graph-based representations into conditional positional encoding and introduce adaptive masking within EEG foundation model pretraining. Experimental results on ten benchmark datasets demonstrate that the proposed GCN-based positional encoding and adaptive masking strategies, both individually and in combination, achieve state-of-the-art or competitive performance, highlighting their effectiveness in EEG foundation modeling.

Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.

The rapid progress of generative speech models has made high-fidelity voice cloning and editing cheap, fast, and widely accessible, weakening the security of voiceprint-based authentication in mobile speech interactions. Existing speech watermarking techniques focus on offline protection or watermarking synthetic speech, and thus cannot verify whether a live utterance originates from the genuine speaker at capture time. We present \textit{MoBiMark}, a real-time \textbf{mo}bile \textbf{bi}ologic speech water\textbf{mark}ing system that binds ultrasonic articulatory signatures to speech content and embeds a dynamic watermark during capture. MoBiMark emits a multi-frequency ultrasonic probe on commodity smartphones, extracts echo features from the vocal tract, and fuses them with a content-dependent speech embedding to produce a dual-layer dynamic watermark combining a persistent user identifier (UID) and a per-utterance speech identifier (SID), enabling both source authentication and propagation-time tamper detection. A jointly trained embedder and detector injects the watermark while preserving perceptual quality, and system-level optimization with INT8 quantization enables lightweight on-device deployment. Evaluated on 50 users across four smartphone models, MoBiMark achieves an average 99.79% accuracy in source and propagation attacks, with negligible impact on speech quality and intelligibility.

Distributed artificial intelligence (AI) increasingly relies on wireless networks, where communication and computing resources shape how AI tasks are configured and executed. However, resource availability alone does not determine how effectively these tasks can be supported, since different information selections and transmission strategies may yield different AI performance under the same resource constraints. To address this challenge, this article presents Wireless AI Networks as a unified design perspective for bidirectional adaptation between wireless communication and distributed intelligence. Specifically, an AI-executable resource budget characterizes candidate AI configurations, while task-level information value guides their selection and wireless resource allocation. Building on these principles, a wireless-AI co-adaptation framework coordinates resource abstraction, joint configuration and allocation, and execution feedback. Two complementary experiments illustrate this perspective: update selection and statistical correction improve federated learning under identical model-update payload budgets, while sensitivity-aware power allocation reduces model accuracy degradation under the same average transmit-power constraint. Together, these results highlight the importance of coordinating resource feasibility and task-level information value in distributed AI execution.

Retrieval-Augmented Generation (RAG) systems have demonstrated remarkable effectiveness in mitigating hallucinations by incorporating external knowledge. However, the retrieval process inevitably introduces noise, posing significant challenges to RAG robustness. Fundamentally, noise robustness is a composite competency encompassing noise identification, retrieval filtering, and robust reasoning. Unfortunately, existing methods often fail to capture this holistic nature, typically restricted to single-dimensional enhancements or hampered by noise-induced optimization interference. To address these limitations, we decompose RAG robustness into three sub-tasks: knowledge boundary perception, retrieval verification, and robust reasoning. Accordingly, we propose ARMOR, an Adaptive Curriculum Meta-Learning for Noise Robust RAG Reasoning. Inspired by the teaching process, ARMOR introduces an adaptive task scheduler as a teacher to curate training plans according to the absolute learning progress of RAG generators (Student). Distinct from GRPO, ARMOR utilizes inter-group advantage estimation during learning to effectively mitigate mode collapse caused by limited output spaces. Experiments on three complex reasoning datasets (2WikiMultiHopQA, MusiQue, and HotPotQA) indicate that ARMOR achieves synergistic improvements in retrieval noise robustness, significantly outperforming baselines in output quality.

LLM inference has two phases with different bottlenecks: LLM prefill runs once per request and is compute-intensive. Decoding repeats for every output token and is limited by memory bandwidth. To address the challenges associated with this issue, different solutions have been proposed. All of these solutions use the same model for both prefill and decode phases. We propose BridgeKV, which uses a larger model for prefill and a smaller model of the same family for decoding. A lightweight KV-stitching module translates the larger model's prompt cache into the smaller model's cache layout. It handles differences in cache shape and representation through learned local layer mixing and separate nonlinear maps for keys and values. With both models frozen, we train the module to align their caches and make the smaller model follow the larger model's next-token predictions. A second training stage adds low-rank adapters when further task adaptation is needed. Across six model pairs and five datasets, BridgeKV outperforms fine-tuned small-model baselines with matched trainable parameter budgets in all 30 settings. It closes up to 80% of the quality gap to using the larger model for both phases. On one A40 GPU, Llama-3.1-Instruct-8B to Llama-3.2-Instruct-1B achieves 3.0X faster requests than the larger model alone, with latency close to the smaller model alone. BridgeKV can improve the hardware and system solutions that separate prefill and decode phases across GPUs. A separate serving simulation models request arrivals, queues, and cache transfer. For Llama-3.1-Instruct-70B to Llama-3.2-Instruct-1B, it predicts up to 33.86X faster requests and 27.5X lower energy use than using the larger model for both prefill and decode phases.

Compact object detectors such as YOLOv8n are a common choice for camera-based fire and smoke detection on low-power devices. These detectors are usually evaluated with box-level metrics such as mean average precision, and their outputs are turned into decisions with a fixed confidence threshold such as 0.25 or 0.5. A fire monitoring system, however, makes one decision per image: raise an alarm or not. In this paper we evaluate two publicly available compact detectors on the D-Fire test set at the image level. We compare three simple rules for turning box detections into one alarm score per image, and we use a split conformal procedure to choose the alarm threshold from held-out images so that the expected fraction of missed fire and smoke images stays below a chosen target. We find that the same fixed threshold leads to very different miss rates for the two detectors, 17.3 and 8.4 percent at a threshold of 0.5, while the calibrated threshold meets miss-rate targets of 10, 5 and 1 percent for both detectors and all three rules. At a 5 percent target, the false alarm rate on images without fire or smoke stays between 1.2 and 3.0 percent. We also find that smoke-only images stay below the target even when the overall target is met, and we describe the main failure cases.

Language-model post-training modifies weights through cumulative updates and successive displacements, yet existing studies report seemingly conflicting spectral properties: nearly full-rank updates and low stable-rank changes. We show these observations are compatible and characterize update directions during training. For singular values following $σ_i \propto i^{-α}$ with $0 < α\leq 1/2$, we prove that stable rank becomes negligible relative to fixed-energy rank as dimension increases. Full singular value decompositions of selected OLMo-2-7B updates confirm this separation: three proximal policy optimization (PPO) attention updates have stable ranks of 9.6-18.3 but require 532-1,222 singular components to capture 80% of their energy. We next compute consecutive checkpoint-displacement cosines from full-model inner products. Across nine publicly released reinforcement learning from verifiable rewards (RLVR) stages, all 131 adjacent pairs have negative cosines on their native checkpoint grids, including PPO and Group Relative Policy Optimization (GRPO) trajectories. Supervised fine-tuning comparisons across four model families show configuration-dependent alignment, including negative late-training cosines for Mistral and FP32 Llama. Finally, a controlled OLMo-2-1B study with three paired seeds and a 2,048-token response cap compares group-relative advantages against a learned critic with generalized advantage estimation (GAE). At matched 60-rollout windows, late-training mean cosines are +0.259 and +0.012, respectively; the paired critic-minus-group difference is -0.247 (seed-level 95% interval [-0.358, -0.137]). These findings distinguish spectral concentration, energy coverage, and trajectory direction, and demonstrate that the advantage system influences directional alignment under matched training conditions.

Current single-cell clustering pipelines discard most genes and can yield inconsistent clusters across random seeds. We introduce KQE-2, a training-free hybrid quantum-classical framework for reproducible whole-transcriptome clustering. KQE-2 maps cells into entanglement-free quantum states, enabling exact feature evaluation on classical hardware and direct checks on a quantum processor. Theoretically, we prove that removing entanglement is necessary and sufficient to bound single-qubit feature variance locally. Practically, while retraining scVI shifted 14% of cells and altered rare-type markers and drug targets, KQE-2 delivered identical biology across runs. On standard highly variable genes, the quantum encoding alone raised accuracy on a developing cortex from 0.61 to 0.79. KQE-2 scales to 11.4 million nuclei on one server, outperforming the compared methods on eight of ten complete atlases (and within the best method's run-to-run range on two), and leads on six even under standard downstream clustering.

This work proposes the Adaptive Federated Hamiltonian Monte Carlo (AFHMC) for Bayesian federated learning (FL) tasks. The prior work on HMC applications for FL achieves $O(\sqrt{d}/ε)$ communication cost under log-convex distributions, already demonstrating its advantage over the Langevin dynamic-based counterpart. Based on a refined analysis, this paper further reveals the benefit of HMC for the FL regime by incorporating a control mechanism for local node shifts. We prove that AFHMC at most requires $O(\{d/ε^2\}^{1/3})$ communication cost up to a logarithmic term, which is significantly better than the existing results. Moreover, the proposed method adapts to client heterogeneity. Under low-heterogeneity settings (i.e., the target sampling precision $ε^2$ is higher than the heterogeneity level), the communication rate of AFHMC can be as good as $O(\log(d/ε^2))$, resembling the communication costs for FL optimization tools. To the best of our knowledge, this provides the first logarithmic communication complexity result for federated sampling, even though it requires low heterogeneity. Numerically, we also justify AFHMC's advantage under the low-heterogeneity regime. Our code is available at https://github.com/laaaarrywang/Adaptive-Federated-HMC

Many digital systems courses teach on FPGA boards whose vendor tool is no longer maintained: the Digilent Basys2 and Nexys2 (Xilinx Spartan-3E) and Nexys3 and Atlys (Spartan-6) need Xilinx ISE 14.7, discontinued since 2013 and hard to install on the computers and tablets students use today. The author presents Silinx, a free (AGPL-3.0) HDL design platform for teaching that keeps ISE's Project Navigator workflow but runs in a web browser with nothing to install. A module can be written in VHDL or Verilog, in C or C++ through Silinx HLS, which is compatible with AMD's Vitis HLS and passes all 73 of its introductory examples, or drawn as a schematic, an ASM chart, a state diagram, a Boolean function with its Karnaugh map, or a Petri net, from which its HDL is generated. Designs are checked with explanations written for beginners, simulated, run on an emulated board, built, and programmed onto a real board over WebUSB. The implementation flow is open and runs in the browser: Yosys and nextpnr compiled to WebAssembly, with Silinx's own device data, timing model and bitstream writer for the Spartan-3E and Spartan-6, built by black-box observation of ISE's documented tools; Lattice iCE40 boards use the existing open iCE40 tools, and further families are planned. ISE is the oracle: Silinx's bitstreams are byte-identical to those ISE writes for the same placed and routed design, and pass ISE's design rule check. Against ISE at the same optimization goal, the open flow uses 15% fewer slices with 14% shorter critical paths when optimizing for area, and 21% fewer slices with 8% shorter critical paths when optimizing for speed (geometric means over 45 open-source designs built by both flows). The author also describes how the platform was developed with AI coding agents under test-first project rules.

Many of the most consequential advances promised by AI, from curing diseases and inventing new technologies to formulating new scientific theories, are fundamentally creative acts. Yet creativity is uniquely difficult to learn because it faces a creative paradox: learning to repeat past examples of creative success makes those specific examples more likely to recur, inhibiting the very creativity desired. This work argues that creativity is better modeled as an explicit process rather than as a property of individual artifacts, and proposes to learn the underlying generative principles behind the creative process. In particular, to escape the paradox, we introduce creative priming (CP), wherein a trained priming model generates a prompt-specific primer that assists a response model in being creative, explicitly decoupling the creative process from response generation. The priming model learns through GRPO, rewarding primers that lead to responses that are novel, high-quality, and diverse, thereby reinforcing the process that yields creative responses without reinforcing any specific response.

Real-world intelligent systems must reconcile two demands: the adaptability conferred by emergence and the steerability required for accountable deployment. We show that these demands are governed by two distinct bifurcations on the same object and that, consequently, increased capability does not imply increased controllability. Emergence is a bifurcation of the capability variable $S$, with criterion $Γ:=\mathcal F'(0)>1$ equivalent to the existence of a unique positive stable fixed point of a self-consistency equation, following solely from the concavity of $\mathcal F$. Controllability is a bifurcation of the permission variable $c$, whose dynamics include an erosion term and exhibit a saddle-node fold with hysteresis as the gate criterion $ρ^\ast$ crosses unity. Three quantitative results follow: the critical exponent distinguishes families rather than classifying the transition; the gating-cost bound $K(M)\propto\log(1+M)$ grows with scope, not capability; and the fold threshold is $γ_c\approx4μκ$. Calibrated in both directions, the criteria reproduce the two-dimensional Ising transition, admit two true-feedback order parameters, and reject nine candidate quantities, supporting the conclusion that the order parameter is geometric rather than capability-related. Since permission cannot be endogenized and no deployed system has been measured to satisfy $Γ>1$ and $ρ^\ast>1$ simultaneously, increased capability does not imply controllability.

Missing ratings can overturn a personalized model's current prediction. For completed models with additive response scores, exact worst case bounds check all missing assignments without changing the current predictor. We evaluate this audit retrospectively in two self-triage cohorts, 108 clinicians rating three simulated psychiatric interviews, and 126 counseling demonstrations annotated by ten experts. Under shared random half histories, Fisher adaptation based on observed ratings reduces log loss relative to response frequency by 0.1071 nats in the clinician audit (95% CI [0.0867,0.1261]) and 0.0223 in counseling (95% CI[0.0121,0.0321]). The counseling test excludes target experts and source videos from training. Certification covers 5.4% and 10.5% of targets; fixed width acquisition raises coverage to 13.7% and 28.6% at the same budgets. With weights and responses held fixed, the observed-rating update also has lower loss than an affine partial update while certifying exactly the same evaluated targets. The certificate establishes agreement with the completed model, not correctness or stability at intermediate updates. These rating tasks test the method; clinical usefulness remains untested.

Generative models can support both state estimation and control in robotics. For control, a common approach is to use the model as a policy that directly samples actions conditioned on observation histories. However, models that jointly represent states, actions, and observations can also predict how the robot state and its environment evolve. We investigate whether these predictions can improve state estimation and control through a change in inference procedure, without modifying the pretrained model. To this end, we use particle filtering to maintain competing hypotheses about state histories. A single pretrained diffusion model generates these hypotheses along with their predicted observations and associated action proposals. As new measurements arrive, particles are reweighted according to how well their predicted observations match the measurements. While particle filtering conventionally uses these weights only for state estimation, we also use them to guide action selection, favoring proposals associated with hypotheses that better explain the observations. Simulation experiments in partially observed navigation and vision-based manipulation demonstrate improvements in both state estimation and task performance compared to action generation without filtering, using the same pretrained model and no additional test-time information.

The Pravda Network is a Russian state-aligned media aggregator, with over 110 websites and more than 15 million articles published since 2023. Given its prolific and repetitive output, emerging scholarship argues that the Pravda Network was built to influence AI systems. We conduct a large-scale audit of leading AI chatbots to understand: (1) the retrieval rates of Pravda Network articles, (2) the conditions of retrieval, and (3) whether chatbot responses affirm or caveat this content. We build a 10,000 prompt corpus from sampled Pravda Network headlines varied along two dimensions (headline rewording and user skepticism) and query 11 AI chatbots via the web interface (n=104,997 completed responses). Overall, we find susceptibility to Pravda Network retrieval to be infrequent---appearing in 7.3% of responses---and highly platform dependent---with three chatbots responsible for 94% of retrievals (You.com, Claude, and Copilot). Among the frontier models, Claude is an outlier: ChatGPT, Gemini, and Meta AI retrieved no Pravda Network articles, likely indicative of a domain-level filtering policy. All 11 chatbots, however, retrieved other Russian state-aligned media to some degree (median=13.56% of responses). When Pravda Network articles are surfaced, caveat rates vary widely by platform (from 5.3% to 76.2%), with You.com providing the fewest caveats. We also identify the tactic of headline washing: where 9.92% of sampled Pravda Network headlines attribute their reporting to other, often more credible, media outlets, while the articles themselves are republished from Russian state-aligned media. We discuss headline washing's potential to launder state media as independent reporting. With this audit, we provide a systematic analysis of AI chatbots' susceptibility to Russia's latest attempt to manipulate the information environment.

Representations learned from electronic health record (EHR) data encode patient physiology together with the clinical process that decides what is measured, when, and for whom. We formalize the data-generating process with latent physiology, ordering intent, laboratory values that exist only when measured, a learned representation, and clinical environments, and analyze two strategies for undoing this entanglement. First, no division of the information carried by laboratory "value" and "ordering" channels is identified from a single dataset: when ordering depends on the analyte, its distribution among untested patients is not recoverable, and a sentinel code for the undefined recorded value makes it carry all the entropy of the ordering indicator, fixing partial information decomposition (PID) for every redundancy measure satisfying the Williams-Beer axioms and invariant under injective re-encoding. Second, cross-site invariance of representation-conditional outcomes is a falsifiable implication of biological invariance taken jointly with exclusion and representation-sufficiency conditions, but not an identification strategy: we construct observationally equivalent models that disagree about biological invariance, whether the invariance test fails or passes. For binary physiology this happens exactly when case mix differs, the representation is lossy, and physiology affects the outcome. A worked example yields a 9.6-percentage-point cross-site gap under an unchanged biological mechanism. Tests performed for every patient remove the ordering gate, and measured tests can restore identification: biological invariance is identified when a test exactly summarizes outcome-relevant physiology, or, for binary physiology, when two tests with independent errors are available at sites with different case mix. Cross-site analyses should be reported as tests of a stated joint hypothesis.

Many major scientific advances come from approaching a problem with a new or novel perspective. While large language models (LLMs) with internet-scale knowledge seem poised to take such perspectives, their consistent generation remains elusive. Instead, models are drawn to dominant views and preexisting, in-field bias. To explore the potential of fresh eyes to increase creativity, this paper proposes naive LLMs, models that are made intentionally ignorant of a specific topic. Through deprivation in training or intentional forgetting, naive models are unburdened by existing dogma, allowing them to be free to make new associations and conjectures. These naive models can then be paired with a stronger mentor model that can refine their unique perspectives into concrete proposals. For the 1B and 7B models tested in this paper, naive models consistently produce seed ideas that are more novel and useful than baseline counterparts, and mentor refinement improves the ideas even further. Thus, contrary to the prevailing tradition of scaling up training data, we demonstrate a case where less information is actually more powerful, in particular when the aim is to discover genuinely outside-the-box solutions.

In collectives, language model agents can exchange individually innocuous messages that together support a policy violation. Understanding and preventing collective misalignment requires answering the question of which messages sufficiently influenced and necessarily enabled the violation. Collective Dependence Analysis (CDA) operationalises necessity and sufficiency to explain decisions and find the fewest removals interrupting support within tested combinations. Tests covered constructed violations, unscripted agent conversations, and recorded interactions, with selected message removals reducing violations from 93% to 6% in controlled cases. Using O(N) model evaluations for N messages, CDA achieves the same violation reduction with half the deletions of a frontier LLM judge given the transcript and asked for all supporting messages in eligible operational cases. CDA also proposes the mechanism through which the collective misalignment emerged. Nonetheless, complete explanations remain difficult on unfamiliar tasks and some removals can inadvertently disrupt correct decisions. CDA unifies explanations of collective misalignment with experimentally tested interventions.

Agentic code generation often asks a model to explain its reasoning before it writes code, and a common way to make such a workflow safer is to add a security instruction to the prompt. We test what that instruction does in small language models, which are attractive because they run locally and cheaply. Four open code models of 1.5B to 3B parameters generate C++ from 166 sections of eleven public requirements documents, once with a general prompt and once with a secure prompt. The secure prompt more than doubles the share of reasoning about validation, resource handling and threats, and we find no evidence that the rise is larger for requirements with security content. The code changes only in part. Among functions with a body and reasoning, the share with a check that our lexical test recognises rises from 23% to 39%, but more than a quarter of the functions remain stubs, compile rates stay low, and half or more of the functions whose reasoning mentions a check contain no such check. Two rounds of compiler-feedback repair leave more than 80% of sections unusable, and neither cppcheck nor an exploratory ESBMC run separates the prompts on the usable code. A hand check by two annotators supports the stub measure and the direction of the reasoning result. For small models, a security instruction changes the reasoning broadly but the code only in part, so the reasoning should be checked against the code rather than trusted.

LLM agents increasingly act as routers for tools, APIs, and software-supply-chain workflows, making an execution path - rather than generated text alone - a security-relevant model output. We study shadow invocation: a task-preserving request variant that causes an LLM router to select a riskier execution path, potentially skipping validation, provenance, staging, redaction, or audit controls even though the request does not ask for a malicious action. We introduce ShadowBench-MCP, a benchmark and metric suite that separates task preservation (TPSS), operational pressure (OPSS), and route-geometric drift (RGAS), and combines them in the pre-execution SODS score. In a focused policy-ablation study using a Groq-hosted qwen/qwen3-32b router, 120 routing decisions yield 96 original-versus-variant comparisons across six security-sensitive workflows. Efficiency-first routing produces a 29.2% shadow-invocation rate and 26.7% safety-tool skip rate. Policy-aware routing reduces these to 4.2% and 1.7%, respectively, corresponding to an 85.7% relative reduction in shadow invocations. Pre-execution signals distinguish shadow routes with ROC AUC up to 0.94. These results are early, single-router evidence of a policy-dependent routing-layer security effect, not a claim of universal vulnerability.

Reinforcement learning (RL) requires large amounts of interaction data or a reliable simulator, both hard to obtain for nonlinear soft robotic hardware. Recent work shows that a time-scale separation inherent to soft robot dynamics justifies their reduction to low-dimensional invariant manifolds, known as adiabatic spectral submanifolds (aSSMs). We demonstrate that an aSSM-reduced model, a simple polynomial ordinary differential equation learned purely from pre-collected data, can serve as an RL training environment. On a pneumatically actuated soft arm, we train a trajectory-tracking policy in 3 hours on a six-dimensional aSSM-reduced model and deploy it without fine-tuning, whereas reproducing this policy training extrapolates to 3 years of hardware interaction. Our policy tracks dynamic targets with 40% lower error than otherwise identical policies trained on a 4.8-million-parameter recurrent neural network (RNN) or a 1.7-million-parameter transformer, which take 13 and 100 hours to identify, against 5 minutes for our mathematically justified 6,924-parameter model.

Unlike the brain, artificial neural networks (ANNs) suffer from "catastrophic forgetting": complete overwriting of solutions when re-training on new functions ("concept shift") or on a disjoint domain of the current function ("domain shift"). ANNs require each batch of data to be roughly representative of the complete distribution, a typically expensive and impractical requirement. In this paper, we introduce Functional Task Networks (FTNs), a method of continual learning and knowledge transfer modeled after several properties of the mammalian neocortex. FTN is formulated as a distributed, stochastic mixture of experts in which a sparse, binary routing mask is computed at inference time using a small support sample of input-output pairs. Per-batch reconfiguration of the router during training, together with mask-overlap-filtered replay, shapes a basis of overlapping, partially trained subnetworks. Unrelated tasks route through mostly separate assemblies, related tasks overlap on shared components. On block-sequentially trained continual learning and knowledge transfer tasks FTN retains solutions nearly as well as disjoint-mask isolation and attains the lowest pre-update head-start loss among matched-architecture comparators. The learned mask-overlap geometry tracks the known concept-sharing structure where available. We show experimentally that under correlated concept shift, support-set mask reconfiguration is necessary even for retention. Under domain shift, the distributed bank with replay is sufficient to handle retention without the configurer. Forward transfer, or head-start, and backward transfer, are also contributed by support-set mask configuration.

Large language model (LLM) agents increasingly operate in groups, each pursuing an individual objective while jointly determining system performance. If an operator controls which agent observes which other agent, but not their payoffs or messages, how should observations be structured to raise collective welfare? We make the information network (who observes whom) a design object, decoupled from the fixed interaction network over which agents act. Ten identical LLM agents play repeated group-decision games; by default each observes only an anonymous summary of the group's previous-round outcome, and a manager with a fixed budget adds directed observation links that reveal peers' past actions. Across 1,080 episodes and six games spanning coordination, anti-coordination, social dilemmas, and discovery, we ask how much information to share and, at a fixed budget, who should observe whom. Disclosure creates behavioral coupling: observation acts as a contagion channel whose welfare effect depends on the alignment between individually attractive behavior and the collective objective. First, the value of information is alignment-contingent: more peer information tends to help or be neutral when individual behavior is aligned with the collective goal, but under free-riding or congestion externalities it amplifies collectively harmful behavior; in a public-goods game the group does best with no information at all. We read this sign separation as a cross-game regularity within one controlled LLM population, not a model-invariant law. Second, at a fixed budget, who observes whom matters primarily when coupling is harmful: a modular "barbell" that lets sub-populations entrench harmful behavior is reliably the worst wiring, reversing the network-reciprocity intuition from human evolutionary game theory. These results argue against a default of full transparency for agent collectives.

We study off-policy evaluation (OPE) in contextual bandits with data collected from multiple logging policies. Inverse propensity scoring (IPS) is a standard approach to OPE and extends naturally to the multi-logger setting. However, as highlighted by Agarwal et al. [2017], neither balanced nor weighted IPS consistently outperforms the other in this setting. We resolve this dilemma by deriving an optimal IPS estimator with sample-dependent weights that minimize variance subject to unbiasedness. Using a variational calculus approach, we obtain closed-form optimal weights, yielding an estimator that is unbiased and achieves asymptotically optimal variance within a generalized weighted-IPS estimator class. Instead of a reward model, the estimator needs only a $K\times K$ matrix and a $K$-vector of moments, where $K$ is the number of logging policies, and we quantify how much efficiency this costs relative to the semiparametric oracle. Experiments on benchmark datasets show that our estimator generally improves on existing multi-logger IPS methods. We also extend our estimator to a doubly robust form by incorporating a conditional reward estimator. We compare the resulting DR extension with the semiparametrically efficient DR estimator of Kallus et al. [2021], both theoretically and empirically. We show that the two estimators achieve the same asymptotic variance when the conditional reward estimator is consistent, while our estimator can retain semiparametric efficiency under certain forms of reward misspecification. Empirically, the proposed DR estimator often achieves lower estimation error than the existing DR baseline.

C-to-Rust migration is an active agenda, but no dataset pairs the two languages at the level of whole solutions to a shared problem. We present CRUSTPairs, 1857 C/Rust implementation pairs drawn from competitive programming archives, normalised, deduplicated, and stratified into difficulty tiers by zero-shot cross-language embedding similarity, with perrecord provenance and licensing. Cross-language retrieval is hard on this data, and harder than a single baseline would suggest: TF-IDF over shared identifiers reaches MRR 0.21 and so beats a pretrained code model at 0.19, which on the hardest quarter of the data scores 1.7x chance. Finetuning reaches 0.79, and lifts that hardest quarter 10x. Renaming every user identifier settles what is being learned: the finetuned models keep 96-97% of their accuracy without them, while the pretrained model keeps 66\% - less than the lexical baseline does - and on structurally divergent pairs the finetuned models are better without identifiers than with them.

Expert analog layouts differ from auto-generated ones in density and regularity, not only in legality: experts make a succession of discrete, hierarchical physical decisions and revisit each after extracted electrical results, while most existing automated flows act on a physical search space fixed before the controller runs. We present CASCoDe, a closed-loop LLM agent framework for analog netlist-to-layout design that formalizes those expert decisions as a hierarchical discrete action space-sizing selection, structural floorplan, local layout refinement, routing-attempt selection-on which the agent runs feedback-driven tree search from a circuit netlist to a verified GDS. The agent authors the circuit-specific floorplans and scripts, and an execution layer of interface adaptations over adapted tools make every action land at physical implementation and be judged by DRC, LVS, and extracted-RC simulation. On 20 SKY130 amplifiers spanning single-, two-, and three-stage topologies, CASCoDe is the only flow that delivers a DRC- and LVS-clean routed layout when meeting design specifications for every circuit. It reduces area by 59-85% (mean 71%) against a conventional P&R flow on identical PCells, raises mean device density from 0.39 to 0.74 against APPlace, MAGICAL, ALIGN and ICPI. A layout that beats every baseline in area and legality appears after 11 minutes of agent time on average, and ablations show that the structured action space and the feedback-guided tree search each reduce area and search cost.

A purely data-driven method grounded in the theory of adiabatic spectral submanifolds (aSSMs) was recently used to model and control nonlinear soft robotic simulators. Here, we extend the methodology to learn aSSM-reduced models from data collected from hardware soft robotic platforms. Specifically, we learn a five dimensional (5D) aSSM-reduced nonlinear ordinary differential equation (ODE) that accurately models the dynamics of a soft-helicoid trunk robot. We then use this ODE in a model predictive control scheme to demonstrate real-time precision control across various challenging closed-loop trajectory-tracking tasks, ranging from 2D to 3D trajectories with static and dynamic constraints. We find that our methodology enhances accuracy and precision across all control tasks compared to available data-driven linear methods and model-free reinforcement learning. While we focus here on a soft-helicoid trunk robot, our methodology relies solely on hardware data and is therefore applicable to a broader class of soft robotic systems.

Generating counterfactual (CF) explanations for time series is challenging due to complex temporal dependencies and the computational inefficiency of traditional instance-specific optimization. We propose a policy learning framework for counterfactual explanation in time series, trained as a one-step contextual-bandit policy-learning problem. By training an actor-critic agent in the latent space of a pre-trained autoencoder, our method learns a transferable policy that captures global structural regularities. Unlike existing methods that rely on iterative optimization for each new input, the learned policy enables direct generation of valid and plausible counterfactuals through a single forward pass, avoiding per-instance optimization at inference time. Experiments on 13 UCR benchmark datasets show that our framework achieves high validity and strong plausibility across multiple anomaly detection metrics. Overall, by decoupling counterfactual generation from per-instance optimization, the proposed method provides an efficient and scalable alternative to existing approaches for time series counterfactual explanations.

Energy claims for efficient neural networks are often inferred from operation counts. We report direct GPU energy measurements on two workloads that batch very differently, recurrent policies in GPU-vectorised multi-agent reinforcement learning (MARL) and mixture-of-experts (MoE) language models, and ask one question of both: can the rows of a batch be made to share computation, across the batch and over time, at a measured saving? On recurrent policies the answer so far is no. Shared activity patterns exist but rarely last one step, masking a frozen policy fails, and trained-in masks succeed only because a fixed mask is equivalent to a narrow GRU, at 11x the energy of a small dense policy in our implementation. Test-time steering adds nothing beyond re-picking on either workload. On MoE models the answer is yes, and granularity decides how much: on Qwen3-30B-A3B, batch-aware routing reads 25-31% fewer expert weights within KL 0.05, with benchmark changes within noise at 500 items, and uses 12-18% less energy per token on a fused forward. We also introduce persistence-aware OEA, a one-parameter rule that keeps the batch-aware cut with about 40% fewer new experts per step than native routing at batch 32-64, though its energy has not yet been measured. These are pilot measurements on one GPU type, and we state their limits.

Compression can preserve a language model's average accuracy while selectively degrading individual capabilities. Whether these losses follow method-specific patterns or a transferable order is unclear. We applied seven perturbation and compression methods to eight instruction-tuned models, scoring 16 tasks in four answer formats against preregistered predictions. Short-answer capabilities failed in nearly the same order under four random perturbations with different mechanisms (median correlation 0.90) and, within each model, under pruning and quantization (0.875). Intact answer margin, answer frequency and answer format predicted this order. Fragility profiles measured under random perturbation predicted which tasks and facts compression removes first and, after four-task calibration, held-out task accuracy (median absolute error 0.11-0.12, against 0.20-0.21 for uniform decline); profiles borrowed from other model families did almost as well. The order did not transfer across models under GPTQ quantization or hold for long-form tasks. Within short answers, what compression removes is predictable before deployment.

We identify a homological barrier to stable learning of local Nash equilibria, also known as first-order Nash equilibria (FONE), in constrained non-concave differentiable games. We consider continuous deterministic learning dynamics that fix every FONE point and aim to attract all initial conditions to the full FONE set. We show that this convergence guarantee is impossible when the target set has nontrivial positive-degree cohomology, such as that arising from a loop, and the state space is compact and contractible. We exhibit this barrier in a simple non-concave quadratic zero-sum game on a box whose exact FONE set consists of a diagonal six-edge loop and an isolated point. Consequently, no such dynamics can make the full FONE set a stable global attractor. We further show that degree two is the minimal polynomial degree at which this barrier can arise and extend the construction to obtain nontrivial homology in arbitrarily high degrees. Finally, we establish that the barrier persists for sufficiently small approximate-FONE targets under small perturbations, yielding an open, positive-volume family of quadratic zero-sum games with the same homological barrier.

Textured 3D meshes specify object geometry and appearance, making them a natural query for finding a particular asset in an image. We introduce LocateAnyMesh, a vision-language model that uses the mesh itself as part of its prompt. A frozen 3D encoder maps colored surface points to a compact token sequence, which a learned connector places alongside image and instruction tokens. We evaluate two grounding models with PPAT mesh tokens: LAM-Select selects image-region proposals, while LAM-Direct predicts coordinates directly. Neither requires rendered query views or view-template matching. New objects require a single encoder pass and no object-specific fine-tuning. LAM-Select improves localization on the reported real-image suites and the official BOP YCB-V evaluation. Secondary experiments compare mesh encoders, and a LAM-Direct study isolates surface color. A single-seed LAM-Select continuation supplies pointing and mesh-only description evidence without replacing the reported localization model. The results support mesh tokens as a query representation, with remaining limitations in presence decisions, identity binding, proposal coverage, and descriptive fidelity.

We study Minty set inclusion on a compact convex set $\mathcal{X} \subseteq \mathbb{R}^d$ with a set-valued operator $A : \mathcal{X} \rightrightarrows \mathbb{R}^d$. Under the Minty promise, the goal is to compute $x \in \mathcal{X}$ and $u \in A(x)$ such that $\sup_{y \in \mathcal{X}} \langle u, x - y \rangle \le \varepsilon$. Our approach is geometric and we work in the fixed-scale constrained-resolvent oracle model, where a query at anchor $a$ is assumed to return a resolvent point $x^+$ together with the Yosida vector $g_η(a) = (a - x^+)/η$. We prove that each oracle response yields a sharp dichotomy: either the returned point provides an $\varepsilon$-SVI certificate, or the Yosida vector defines a strict separating normal for the Minty set. Thus, a single black-box resolvent query supplies either a certified candidate solution or a valid geometric cut. Combining this separation principle with the ellipsoid method, we obtain RESOLVENT-ELLIPSOID, an algorithm that computes an $\varepsilon$-SVI in $\mathcal{O}\!\left(\operatorname{poly}(d, \log(1/\varepsilon))\right)$ black-box oracle calls to the fixed-scale constrained resolvent. For monotone inclusions, our algorithm gives, to our knowledge, the first $\mathcal{O}\!\left(\mathrm{poly}(d, \log(1/\varepsilon))\right)$ black-box guarantee in the fixed-scale resolvent model.

Self-driving laboratories (SDLs) integrate laboratory robotics and automation with AI-driven decision-making to close the design-build-test-learn (DBTL) cycle. However, physical testbeds that combine off-the-shelf hardware and AI into a functioning closed-loop workflow remain in early stages of development, and their integration with commercial laboratory automation is comparatively less developed. In this work, we develop a physical testbed that assembles and integrates commercially available laboratory automation instruments and demonstrate continuous, closed-loop, autonomous operation over 24 cumulative hours of unattended experimentation. We validate this platform on a minimal closed-loop discovery task in which the system autonomously formulates aqueous dye mixtures to match a target absorbance spectrum. The system completed six independent campaigns, each consisting of 48 closed-loop iterations over approximately four hours. In every campaign, the system autonomously selected, dispensed, and measured each mixture, with an AI-driven decision-making algorithm determining subsequent formulations without human intervention. Using this testbed, we compared the performance of two AI-driven decision-making algorithms, traditional Bayesian optimization (BO) and an agentic approach, in selecting mixtures under a fixed total-volume constraint. The agentic optimizer matched target absorbance spectra using fewer iterations than BO, with reductions up to 52%, and converged on a solution for one target that BO failed to match within 48 iterations. These results demonstrate a reference physical platform for SDLs that can evaluate both established and emerging AI strategies, including large language model (LLM)-directed optimization, before their deployment in real-world applications where experiments may be costly and time-consuming to prepare.

Failure monitors for vision-language-action (VLA) policies are trained offline and then frozen. After deployment, lighting, objects and cameras drift, and the policy itself is updated or moved to new tasks. On recorded real-robot runs, after a switch of task or policy version, a frozen monitor fit on the old task ranks failures little better than chance. We present Streaming Failure Forecasting (SFF), a monitor that keeps learning from the one label every episode provides: whether it succeeded. SFF keeps the offline monitor's network fixed and refits only its last linear layer after each episode, at 2.18 ms per policy call on a Jetson CPU core. Its alarm threshold moves only after successes, which in simulation kept false alarms near target except when two drifts came at once. With every setting locked before testing, SFF ranks failures better than the same monitor kept frozen on 4 of 5 simulated test suites under unannounced drift, and above resetting on detected drift on every suite. Carried with its locked learner settings to 434 recorded runs of four VLAs on real SO-101 arms, replayed across unannounced switches of task or policy version, it ranks failures better than the frozen monitor for every VLA and, in an exploratory comparison, above a full reset at the true switch. A deployed failure monitor need not stay frozen: refitting its readout from outcomes is cheap and helps after the changes we tested.

Continual instruction tuning often preserves old skills with task-private modules. Inference still requires selecting the right module from the prompt alone. We introduce Task-Isolated Latent Adapter Continual Training (TILA-CT), a replay-free method that makes each task module self-indexing. Each task receives a latent memory and adapter bank in a frozen LLM. During adaptation, the memory gathers causal-prefix information and injects task-conditioned context. For access, its persistent state provides the routing prototype used by a shared prompt-only router. On the 8-task CT-Pipeline stream, TILA-CT achieves 65.19 acquisition and 63.98 final macro score, with 1.38 points of average forgetting. A middle-reversed order gives 65.38 acquisition and 64.17 final macro score. Across 6 backbones, final scores remain within 1.5 points of acquisition. Ablations support memory-derived routing keys and causal gathering. Final macro routing accuracy is 93.9\%, and oracle routing recovers 1.21 final macro points. These results separate preserved task computation from learned access to it.

Subsurface flow systems often involve hierarchical geological uncertainty, meaning both higher-level metaparameters and cell-by-cell properties are uncertain. This leads to challenges for data assimilation because both sets of parameters must be estimated. In this work we develop an adaptive-tempered sequential Monte Carlo (SMC) procedure for joint Bayesian estimation of these quantities. The method uses noncentered geological parameterizations, with geomodel realizations represented by metaparameters and Gaussian latent variables that are independent under the prior. Adaptive likelihood tempering, importance weighting, resampling, and blockwise Markov chain Monte Carlo (MCMC) mutation are applied at each iteration. Flow surrogates are used for all function evaluations. The hierarchical SMC procedure is further extended to treat successive data assimilation. In this case, existing posterior particles are updated in a manner that accounts for correlations in surrogate model errors across observation batches. We consider two distinct 3D systems with hierarchical uncertainty. The first is a faulted carbon storage system containing three stacked aquifers and two extensive faults. Principal component analysis is used to represent target-aquifer heterogeneity, while geological metaparameters characterize uncertain aquifer and fault properties. The second is a channelized oil-water system containing three facies with uncertain properties. A latent diffusion model represents variability in channel geometry and facies realizations. For the four synthetic true models considered, SMC posterior metaparameter distributions and flow predictions agree closely with reference MCMC results. SMC, however, requires fewer function evaluations and provides speedups of 2.5--7.0 for single-stage data assimilation. Even greater speedups are achieved with multiple GPUs.

We study reconstruction attacks, which recover training data from a learned model. In our adversarial model, the adversary designs a hard instance, drawn from a parameterized family of data distributions, and is given the parameter used to generate it. Our main contribution is a general reduction showing that, in this model, membership inference (determining whether a given datapoint was used for training) implies reconstruction: for a training set of size $n$, a membership test with recall at least $q(n)$ and false-positive rate $O(δ^2/n^2)$ yields a reconstruction algorithm that, with probability $1 - O(δ)$, outputs only training points, and recovers $Ω(n \cdot q(n))$ of them in expectation. This reduction applies to any learning algorithm. For stochastic convex optimization (SCO), we make the reduction computationally efficient over exponentially large domains via structured distributions whose support is partitioned into polynomially many categories of exponentially many elements each, so that reconstruction iterates over categories rather than the full support. Combining this with an existing membership inference result for SCO, we prove that any learning algorithm achieving low excess risk over all distributions admits a polynomial-time reconstruction attack recovering $Ω(n)$ training points, even when the underlying distribution has exponentially large support.

Language models are used as simulated psychiatric patients to train clinicians and evaluate AI therapists, and are judged by plausibility or total scores. We give a framework for simulated respondents: an information budget bounds how much simulated patients can differ, and an average-patient hypothesis, that a model answers each item at the level typical of the named condition and severity, implies little variation, a full profile of the condition's typical symptoms and totals at or above the top of the stated severity band. We examine these implications at matched totals on item-level data from eight human samples, including a clinical and an instructed-feigning sample. Across 17 models, simulated patients vary 1.5 to 11 times less than real respondents with the same total and on most questionnaires endorse nearly every symptom at a middling level; on the PHQ-9 their symptoms covary less like real ones. People, honest or feigning, answer "not at all" about as often as genuine patients with the same totals, and more often than every model. Single answers are highly concentrated, and prompt changes, published personas, larger models and reasoning reduce the pattern without removing it. In interviews, patients told a severity are rated as having nearly every symptom, so an assessment agent's recognition of absent symptoms goes untested, and scored against the stated severity the agent places a quarter of them in their band, against about half of patients given a real symptom profile. Having the model word such a profile restores half to two thirds of the variation.

User sequence modeling is widely used in recommendation and advertising. Search ranking presents an additional challenge: the current query expresses immediate intent, while a candidate product determines which parts of a user's long-term history are relevant. We present a production sequence framework for sponsored-product ranking that augments recent behavior with two retrieval channels over long-term history: candidate-nearest-neighbor retrieval and query-nearest-neighbor retrieval. The selected actions are encoded together with dedicated query and candidate context tokens, while candidate-aware early fusion injects the candidate into every history token. Offline ablations show that dual retrieval improves major ranking metrics relative to the production baseline. In an online A/B test on sponsored-product search traffic, the model increased CTR, attributed conversion rate, and orders per impression.

Robot policies increasingly act through a small vocabulary of discrete latent actions learned from unlabeled video, yet training leaves open which motion receives which code. We find this choice highly unstable: two independently trained runs of Moto's latent motion tokenizer assign the same robot clips to nearly unrelated codes (adjusted Rand index 0.001), and evaluating one tokenizer in bf16 instead of fp32 reassigns 74% of its tokens. We use this freedom to impose the representational geometry of macaque ventral premotor cortex (PMv), recorded while a monkey drew shapes, on the tokenizer's code assignment, with a straight-through similarity loss trained alongside its reconstruction objective on robot video. In one training run per target, on five shape categories the loss never saw, the PMv-trained tokenizer reaches a representational similarity of 0.52 with PMv, against 0.20-0.26 for released and no-prior tokenizers and 0.42 for the 95th percentile of random kinematic encoders. Kinematic, language, and random targets do not transfer (-0.17 to -0.03). Two results limit the claim: a hand-picked kinematic descriptor, the stroke's initial direction, matches PMv better (0.65), and a label-shuffled PMv control scores 0.41. We specify the downstream manipulation test, on a CALVIN pipeline that reproduces Moto's published results, before running it.

We study the problem of learning directed acyclic graphs from observational data in nonlinear models and provide finite-sample, nonasymptotic statistical guarantees in these models. Although widely studied, statistical guarantees for gradient-based methods in nonlinear models are missing from the literature. Under a generalized linear model for each node, we analyze a global nonsmooth objective combining an exact acyclicity constraint with a quasi-minimax concave penalty, which can be optimized via standard gradient-based solvers. We establish finite-sample guarantees under this model, including the most widely used logistic and Poisson models. Experiments corroborate structure recovery accuracy and computational performance, including both standardized data and negative controls as baselines.

A deployed language model does not merely answer questions. It spends computation, consumes time, consults evidence, exposes memory, invites or avoids human review, and sometimes makes mistakes whose consequences are radically unequal. Yet LLM systems are still commonly compared as if the relevant object were a single model and the relevant outcome were a single benchmark score. This paper argues that the unit of analysis should instead be the deployment policy: the mechanism that decides when to retrieve, when to remember, when to call a larger model, when to verify, and when not to answer at all. We introduce risk-adjusted deployment efficiency (RADE), a system-level objective that scores the utility a deployment delivers net of inference, retry, oversight, and priced expected failure cost, while treating harms that cannot be priced as hard budgets alongside constraints on latency, privacy, and resources. The framework brings selective routing, retrieval and governed memory, efficient specialization, uncertainty estimation, abstention, and human escalation into a common account. We then examine five deployment regimes in which the same nominal LLM capability leads to different rational architectures, and propose a reporting protocol based on risk-coverage behavior, end-to-end cost, update drift, memory exposure, evidence quality, and escalation load. The central claim is modest but consequential: frontier capability should be treated as a resource to allocate, not a default state to maintain. A safe and sustainable system is not the one that always uses the strongest model; it is the one that spends capability where the reduction in risk or increase in utility justifies the cost.

The use of generative AI in political campaigns has drawn public concern and policy attention. Public concerns range from distrust of candidates to skepticism about the authenticity and accuracy of political advertising, including AI-generated images and video. Nearly half of U.S. states have enacted laws that prohibit, or require disclosure of, "materially deceptive" AI-generated content (AIGC) in elections to guard against these concerns. This study measures how candidates and their campaigns produce and circulate AIGC during an election cycle. We develop two measures that estimate not only wether campaigns use AIGC, but how much they produce (magnitude) and in what forms (modality). We apply these measures to a novel dataset of text and images from the campaign websites, X accounts, and emails of 12,977 candidates for U.S. political office from 2023 to 2025. We report three findings. First, detectable AIGC from campaigns increased significantly across modalities over the observation window, and campaigns that use AIGC frequently do not disclose it. Second, incumbents' campaigns produced significantly less AIGC than challengers' campaigns. Third, candidates in states with AIGC disclosure or prohibition laws do not produce measurably less AIGC than candidates in other states. Our method and findings have implications for transparency and trust in elections, and on state laws that regulate AI use in campaigns. Because current state laws largely target mainly "materially deceptive" content, we identify additional policy mechanisms that could increase electoral transparency and improve trust as AIGC becomes more prevalent.

We built a hybrid multi-agent trading system that combines large language model (LLM) analysts with momentum-based portfolio construction and rule-based execution. Company-news and macroeconomic agents return structured reports to a coordinator that sets stock weights and cash reserves. We evaluate the portfolio engine using its broker-recorded paper account, which operated with deterministic fallbacks, and compare LLM-enabled configurations through retrospective replays. From March 11 to August 28, 2026, the account returned 34.5%, compared with 14.4% for the S&P 500 fund SPY over the same dates. This raw return advantage came with a larger maximum drawdown. Reconstructed policies also finish above matched SPY and QQQ references when assessments are used in the same session or deferred to the next. Design comparisons examine cash authority, selective invocation, and response contracts. A constant-cash control gives a return estimate close to stance-based sizing, and shorter report contracts improve observed parsing but can change recommendations.

This study investigates whether generalized means can achieve forecasting performance comparable to that of TimeGPT for daily emergency call volume prediction. Specifically, two families of generalized means, the Hölder means (HM) and the Lehmer means (LM), are evaluated against fine-tuned TimeGPT across forecasting horizons ranging from one to seven days. Forecasting performance is assessed on a real-world emergency call dataset using multiple error metrics, and Diebold-Mariano test. Experimental results show that HM and LM achieve nearly identical predictive performance across all forecasting horizons. For short-term forecasting, both generalized means outperform TimeGPT, with reductions of mean absolute error (MAE) approximately 1.24\% and 2.48\% at horizons of one and three days, respectively. Conversely, TimeGPT achieves a lower error at the seven-day horizon, with an descresing of MAE by 1.98\%. However, Diebold-Mariano tests reveal no statistically significant differences between the methods at the evaluated horizons. These findings suggest that generalized means can provide competitive forecasting performance compared with a energyvore foundation model.

Digital programmatic advertising operates through complex, multi-hop supply chains. The opacity of these intermediary networks leads to significant efficiency losses, often referred to as the ``ad-tech tax.'' While current industry practices utilize heuristic Supply Path Optimization (SPO) to trim inefficient routes, they fail to optimize the equilibrium of the entire network. This paper extends previous work on Directed Acyclic Graph (DAG) intermediation models by applying them to traceable programmatic supply chains mapped via the OpenRTB SupplyChain object. We introduce multi-attribute utility functions and Mixed-Integer Linear Programming (MILP) to account for substitutable routing (OR nodes) and percentage-based fee structures. Furthermore, we outline a computational methodology and experimental framework using public \texttt{schain} and \texttt{sellers.json} data to benchmark optimal Nash social welfare against current uncoordinated market outcomes.

We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.

Quantifying mouse behavior from pose is a foundational tool in neuroscience, ethology, drug discovery, and animal welfare. Self-supervised foundation models are a natural fit, but a foundation model for mouse behavior must handle social behavior - many behaviors that matter (aggression, courtship, mounting, allogrooming) are inherently multi-animal. Existing self-supervised pose models fall short: they encode each animal independently and discard interaction signal, hard-code the animal count, and pretrain one dataset at a time. We introduce MICE (Multi-animal Interaction Context Encoder), a hierarchical foundation model for mouse behavior that pretrains jointly on six mouse-pose corpora spanning one to four mice and 7 to 27 keypoints. MICE combines a hierarchical individual encoder capturing per-mouse kinematics at multiple temporal scales with a Perceiver-style social encoder whose learned latent codebook decouples parameter count from animal count, so a single checkpoint serves any group size without retraining. Across four evaluation protocols spanning six datasets, MICE matches or exceeds prior keypoint-based baselines, and layer-wise probing shows the social stage incrementally improves discriminability beyond the individual encoder. Code, weights, and data are available at https://github.com/eric0708/MICE.

Data-driven remaining useful life (RUL) estimators for turbofan engines are accurate but opaque, and their internal features typically encode remaining life rather than any specific component's state. This paper presents an interpretable genetic fuzzy tree that predicts RUL from cruise-condition sensor means on N-CMAPSS while exposing a per-component diagnostic layer inspectable rule by rule. The tree follows the engine hierarchy: sensors to component health, then to bounded component damage, spool damage, and RUL, so each level answers a different maintenance question. Training is staged so the prognostic objective cannot reshape the diagnostic layers: health leaves are fitted to physical targets and frozen, a deterministic fuzzy damage map with fleet-referenced thresholds is imposed on each component, and only the aggregation is optimized for RUL. It consumes 13 cruise-mean scalars per cycle rather than a trajectory, which removes any sequence model. On the native split it reaches 8.44 cycles RMSE on the multi-fault subset and 11.51 across all modes with 333 free parameters, 411 total rules, 32 to 12,000 times fewer than compiled baselines. An exhaustive sensor search bounds which components are diagnosable under the tested cruise-mean design, and the rule base is structured to admit formal verification, which this study does not discharge.

When a scanpath descriptor discriminates a task, it is rarely clear what the discrimination rests on. We propose a perturbation profile for trial-level descriptors. Five transformations each destroy one stated property while preserving a listed set of invariants: fixation order, duration-position coupling, the duration distribution, orientation and handedness, and fine spatial structure. The classifier is refit, and the drop in participant-held-out AUC is the pipeline's sensitivity to that property. Because refitting lets the learner use redundant information, this describes the refitted pipeline, not what a fitted classifier used. Intervals come from a participant bootstrap that refits; in simulations at the sample sizes analysed, where the target was zero they excluded zero at a pooled rate of at most 2.5%, and they covered the target in 99.6% of 2720 cases. We profiled seven descriptors on twelve binary problems from three public datasets: reading (SB-SAT, 95 participants), program comprehension (EMIP, up to 159) and video and game viewing (GazeBase, 60). Permuting fixation order raised the AUC of a descriptor built on mean saccade amplitude on three problems (by up to 0.23), because permutation turns amplitude into a measure of spatial spread. A "comprehension" contrast that pooled two programs reached AUC 0.70 with a spread-only baseline and fell to 0.47 within one program: the profile located the confound in stimulus layout. On native versus non-native readers, recurrence was sensitive to order and to jitter at the scale of text lines but not of characters, while a two-feature duration and amplitude descriptor reached the same AUC (0.65). A graph descriptor built on the code's tokens was the strongest non-baseline descriptor on the program contrast (0.78; 0.87 with boosted trees). The interpreted results held under longer runs and three seeds; under a second learner 15 of 16 kept t

Ship design requires coordinating coupled subsystems to satisfy engineering constraints and balance competing performance and cost objectives. Generative artificial intelligence has supported component design, but integrated ship generation must also account for interactions among hull geometry, structure, propulsion, and operation. This paper presents GenISys, a framework for generating integrated systems, demonstrated through early-stage bulk-carrier design. Its dataset contains 217,999 proposed designs, including 16,247 that satisfy the formulated engineering constraints. The proposed simultaneous conditional denoising diffusion probabilistic model (SCDDPM) generates subsystems together, allowing each subsystem model to respond to changes elsewhere during generation. Experiments compare this approach with monolithic whole-ship generation and sequential subsystem generation across five independent training runs per architecture. In a study targeting lower cost per tonne of cargo transported, designs were generated for a required voyage range of 5,000 nautical miles. SCDDPM produced designs satisfying the formulated engineering constraints and meeting a common low-cost target in 55.70% of generation requests, compared with 51.10% for the monolithic model and 47.54% for the sequential model. These correspond to 9.0% and 17.2% more qualifying designs, respectively, while maintaining similar estimated transport costs. These results show how generative artificial intelligence can coordinate subsystems to generate feasible, high-performing integrated systems.

Test-time scaling improves the accuracy of a language model's answers by spending more computation at inference time. Established methods do so either sequentially, revising a single solution over several passes, or in parallel, generating several independent solutions and selecting among them. We present ODSA (Orchestrated Discovery and Self-Aggregation), an LLM-directed scaling framework in which an orchestrator LLM is given existing test-time scaling methods as actions --- discovery of a new independent solution, comparison of existing solutions to build a better one, and acceptance of a proposed answer, which triggers its verification --- and directs stateless workers to carry them out. The orchestrator does not solve the problem itself: it allocates computation, feeds the results of earlier actions into later ones, and takes every action from a single context, so that each decision is informed by everything tried before. On Humanity's Last Exam, ODSA improves accuracy by 4.7 to 18.0 points across three models over a single direct call to the model with no test-time scaling. On LiveCodeBench, where a single direct call already solves most problems, it improves accuracy by 0.6 to 14.0 points.

This study investigates the potential of TimeGPT, a pretrained foundation model for timeseries, for forecasting daily emergency call volumes. Real-world emergency call data, comprising 2,101 daily observations, are used to evaluate both univariate and multivariate forecasting settings under one-step-ahead and seven-day multi-step-ahead scenarios. The multivariate forecasting incorporates temporal, meteorological, demographic, and socioeconomic variables. A chronological training, validation, and test split is adopted to prevent information leakage, while hyperparameters are selected using the validation set through grid search and early stopping when applicable. Forecasting performance is evaluated using multiple error metrics. Its forecasting performance is compared with several established forecasting approaches, including Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX), Poisson regression (PR), Negative Binomial regression (NBR), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM). The results show that SARIMAX achieves the best overall performance in the univariate setting. In the multivariate setting, NBR and PR provide best performance. These findings indicate that the large-scale pretraining of TimeGPT do not necessarily translate into superior forecasting accuracy for emergency call volumes. Frugal methods trained solely on the available data remain better for this application.

Autonomous microscopy can turn fast acquisition into scientific progress when it decides not only where to measure, but also what information to acquire, at what scan rate, and when to recalibrate the instrument. We develop a budget constrained framework for such multistep structure-property workflows. It combines explicit scientific rewards, action costs, instrument observation models, transfer of pre experiment knowledge into the online campaign, and evaluation kept independent of the controller. The transferred knowledge includes predictor accuracy and computational cost, empirical learner maturity functions that weight model-derived rewards, and drift characteristics established before controller training. Using a real electron energy-loss spectroscopy dataset, we run two complementary campaigns. In the first, the controller chooses between imaging and spectroscopy, image size, and scan speed. In the second, the instrument drifts and the controller also decides when to recalibrate. The campaigns show that response discovery and structure-property learning favor different surrogate models and acquisition strategies. They also show that simple acquisition rules form strong, practical baselines. Independent error analysis reveals where model uncertainty is and is not a reliable learning signal. In the drift campaign, explicit registration greatly improves positional accuracy, and the framework measures what this costs in spectroscopy throughput. Together, these results provide a reproducible way to design, compare, and audit autonomous experiments, and they define what microscopes must provide for deployment: independently executable actions, characterized scan and detector responses, timestamped registration telemetry, and complete action records.

Large language models (LLMs) can write the weights of small neural networks directly, without gradient descent. This weight-space programming creates policies for sequential decision-making from a task description alone, but nothing guarantees that a written policy behaves as intended. We combine weight-space programming with probabilistic model checking. Given an environment modeled as a Markov decision process and a requirement in probabilistic computation tree logic (PCTL), an LLM writes all weights of a policy network, optionally repairs them on a dataset of optimal actions, and receives model-checking feedback until the requirement holds or a round limit is reached. The model checker Storm computes, rather than estimates from simulations, the probability that the policy satisfies the requirement, so every returned policy is verified. With two LLMs on three environments, the created policy of each of the 18 runs attains the optimal probability, before any dataset or model-checking feedback. Four of six policies programmed for two transporter jobs also complete three jobs with probability 1, so they generalize to a task that the LLM never saw. The pipeline also programs a policy that satisfies a complex safety requirement expressed with the PCTL until operator in all six runs. In one of them, dataset repair raised the accuracy to 98.4\% but broke the requirement, and one round of model-checking feedback restored it.

Knowledge distillation trains a student language model to match a teacher's predictions on the training text. We ask whether learning the teacher's behavior beyond the training text can further improve the student. We investigate this using three input perturbations: additive token noise, blending two documents, and blending neighboring tokens. Specifically, for an early part of pretraining we perturb the inputs and train the student to match the teacher on them, and we spend the rest of the same token budget on ordinary distillation. When pretraining from scratch, the best method improves on standard distillation at both scales we tried, by 0.025 nats at 135M parameters and by 0.020 nats at 0.5B (mean of three seeds). At 135M this is 1.45 times the gain that distillation itself provides over next-token training. We also observe promising downstream transfer: at 0.5B, the best variant improves average zero-shot accuracy by about one percentage point across seven tasks. It also retains its validation-loss advantage after low-bit weight quantization. We also test a data-limited setting in which training data must be reused, and find that the gains persist. Finally, ablations suggest that, for additive noise, perturbing the teacher's inputs contributes most of the gain, while perturbing only the student's inputs yields much less.

Recent work advancing machine translation quality using LLMs for very low resource languages has relied on the injection of lexicon-based prompt components to expose vocabulary to models. However, these approaches run into a ceiling when models must guess sentence meaning purely from plausible combinations of possible word senses mapped to strings. In this paper we test UD-PrMPT, an approach using syntactic analyses in the Universal Dependencies formalism as prompt components to improve MT in five diverse low resource languages: Beja, Chintang, Chukchi, Coptic, and Xibe. Our results show additive improvements which stack on top of the lexicon based approach which are stronger for higher quality manual analyses but consistent for predicted parses as well.

Recent progress on contrastive audio-language models can be attributed mainly to the scale and quality of their training data. Larger training datasets expose models to a broader range of sounds and corresponding audio captions. To exploit this, we collect 1.48 billion audio samples from academic datasets, internet datasets, web collections, Common Crawl, and YouTube, and process them with a multi-stage curation pipeline. Furthermore, we generate short and long captions for samples in our corpus that lack them. We use our dataset to study the scaling behavior of contrastive audio-language models based on the NaFlex architecture for the audio encoder. Our experiments identify trade-offs between model size and training data for a given compute budget. Moreover, our resulting models, CLAP2, outperform prior audio-text models of comparable size on the MAEB and HEAR benchmarks. We release our dataset, captions, model checkpoints, and code.

In financial backtesting and other point-in-time tasks, external inputs can be restricted by date, but model parameters may retain financial facts disclosed later, introducing parametric look-ahead bias. We introduce FinCutBench, a continually updated monthly financial knowledge benchmark for measuring how far a fixed model's financial knowledge extends over calendar time. The current dataset contains 9,000 Chinese and English four-option multiple-choice questions from January 2023 to September 2026, covering Chinese and U.S. markets and four categories of financial information. We construct monthly knowledge profiles using the approximate months in which the tested facts became public. After a profile passes the decline test and knot checks, we define the FinCut Boundary as the first month when the fitted profile reaches its lower level, and report the stability of the estimated date under resampling. Boundary dates and localization precision vary across models. Using the independently measured Transition Start and Boundary as fixed references for downstream sample grouping, we find clear relative declines after the Boundary in U.S. asset ranking and earnings recall for some models, while Chinese-market ranking estimates vary across models and all eight disclosure-period-paired forecast-error comparisons have unadjusted $p$ values above 0.10. The FinCut Boundary thus provides an empirical reference for the temporal scope of parametric knowledge in point-in-time financial evaluation, supporting evaluation-window selection and result interpretation.

Ray Tracing enables rendering capabilities that are difficult to achieve with rasterization-based approaches such as 3D Gaussian Splatting (3DGS). However, the training cost of Gaussian-based ray tracing (GRT) is much higher than that of 3DGS. In this paper, we study how ray tracing using Gaussians executes on GPUs and show that a considerable fraction of the training cost arises from how the computation is organized. We develop FastGRT, which decouples ray traversal from blending and reuses work done in the forward pass for the backward pass. FastGRT further incorporates optimizations such as fast gradient accumulation and a hybrid BVH update strategy to tackle other major bottlenecks in the training pipeline. We implement FastGRT in Vulkan. To the best of our knowledge, it is the first Vulkan-based GRT implementation that supports both training and rendering. Compared to 3DGRT, FastGRT improves training time by 1.9$\times$ to 4.97$\times$, while preserving the original rendering formulation. Our code is open source.

A tool-using language model must decide whether to retry a failed call, correct it, or stop and tell the user. The error text alone may not be enough: the same feedback can lead to different actions depending on where it appears in the conversation. We isolate this delivery problem by varying placement, guidance, retained history and turn structure. Two studies with plans fixed before data collection, across 13 open models from 3B to 70B parameters, compare the next action and subsequent task completion, with exact deterministic reruns. A new user turn that retains the failed exchange avoids the severe stopping failure observed in Llama models when that exchange is removed. It is non-inferior to history rewriting in pooled synthetic comparisons, but is not best for every model. We implement this form as an opt-in recovery layer within a deterministic inference engine and evaluate it on public benchmarks and a coding agent. On ToolMaze, it raises success on explicit errors by 14.2 points with 7-8B models and 10.5 points with 32-70B models; Loud/Quiet, 24 multi-turn tasks with visible failures and silent corruption, scored by final-state agreement, shows no measurable gain. A client given the exact fault kind performs better still. The findings support delivery as a distinct factor in recovery, not an engine-exclusive solution or a universal retry policy.

Hybrid quantum reinforcement learning has been explored for robotic decision-making, but evaluations often emphasize task success, return, or convergence, leaving unresolved whether quantum circuit outputs influence constrained actions and remain traceable through robot execution. We present \texttt{Qubit-to-Motion}, an evidence-centered framework for support-constrained sequencing and simulated robotic placement. Four Pauli-$Z$ readouts provide the sole learned input to the actor and critic, while a support mask enforces admissible object selection. Each decision is recorded and linked via step indices and SHA-256 fingerprints to a sealed route, deterministic Cartesian and inverse-kinematics waypoints, prescribed joint motion, and MuJoCo contact replay. A frozen checkpoint is evaluated across five deterministic 30-object scene variants. All 150 placements satisfy support, containment, and settling criteria. A fixed-state zero-readout intervention changes $90.0\pm0.0\%$ of decisions, demonstrating dependence on the joint circuit representation. Mean macro travel is $34.537\pm0.206$~m, inference requires $8.60\pm1.27$~ms per route, and replay achieves $8.764\pm0.220~μ$m mean final position error with 0.356~mm maximum interpenetration. These results establish reproducible decision-to-contact traceability. The implementation is available at \href{https://anonymous.4open.science/r/Qubit-to-Motion-Quantum-Reinforcement-Learning-F7DC}{\textcolor{linkpink}{the anonymous repository}}.

Solving optimal classification tree (OCT) problems is computationally challenging because branching, routing, and leaf-assignment decisions must be optimized jointly. We develop an incumbent subtree column generation algorithm (ISCG-OCT) that exploits the recursive structure of classification trees. We first introduce a mixed set-packing and set-partitioning formulation (MSSF). We then construct a restricted master problem whose columns represent feasible subtree patterns. Unlike conventional dual-based column generation, our incumbent-centered pricing evaluates objective improvements relative to the current solution. We establish a recursive decomposition in which fixing a branching decision separates the remaining pricing problem into independent subtree subproblems, enabling exact dynamic programming. We also develop threshold-based variants that trade optimality guarantees for computational efficiency. Computational experiments on 27 benchmark datasets demonstrate competitive training and testing accuracy and substantial reductions in solution time on many instances. When initialized with a CART solution, ISCG-OCT can certify global optimality faster than warm-started mixed-integer programming solvers on selected instances. These results demonstrate how recursive pricing and structured subtree optimization can improve the computational efficiency of exact classification tree methods.

Studies of online political discourse commonly infer public opinion through stance analysis while implicitly assuming that observed user participation is authentic. Meanwhile, studies focusing on detecting manipulative participation are typically modeling it as a binary phenomenon, overlooking the varying degrees to which users may engage in coordinated or inauthentic behavior. This study proposes a continuous modeling approach for manipulative participation and examines its integration with stance analysis. Using online discourse on the 2024 Street Animal Law in Türkiye as a case study, we collected a dataset comprising 7,320,574 tweets from 537,994 users. From this corpus, 301 users were selected through a high-information sampling strategy and manually annotated by domain experts using a 10-criterion rubric that yields user-level continuous manipulative participation scores. Evaluation under a unified user-level cross-validation protocol showed that the proposed framework accurately estimates both manipulative participation and user stance, with gradient boosting methods providing the best manipulative participation estimates and BERTurk achieving the highest stance classification performance. External validation on an independently annotated user set demonstrated good generalization performance. The findings support modeling manipulative participation as a continuous construct rather than a binary label, providing a more informative basis for computational analyses of the relationship between manipulative participation and stance in online political discourse.

Audio language model benchmarks provide comprehensive evaluation across diverse tasks. However, some tasks can be answered from the transcript or even the question alone, and benchmark scores can therefore confound acoustic understanding with language modeling ability. We distinguish lexical tasks, for which the transcript usually suffices, from global-attribute and within-recording-structure tasks, which require the model to analyse the recording beyond its words, either as a whole, for properties such as the speaker and emotion, or within it, for patterns such as stress and intonation. We evaluated 19 open models and Gemini 3.1 Pro on 92 tasks from 6 benchmarks and estimated the recording gain, the accuracy advantage of the original recording over references that retain its words, by comparing the recording with synthetic readings of its transcript and by adding it to the transcript. On attribute and structure tasks, the open models fall behind Gemini by 28.6 and 41.9 percentage points, and about half or more of each deficit lies in the recording gain. Nevertheless, linear probes in Gemma-4 12B and Qwen2.5-Omni-7B show that the acoustic information needed for answering is already readable in early language-model layers, while the readout of many cues declines in the layers that follow. We therefore constructed training data and trained LoRA adapters in these layers, with a content anchor, which raises attribute and structure accuracy by 12.8 to 44.4 points in the two models and keeps mean lexical accuracy within 1 point of the base model. Most of the improvement appears as larger recording gains. These results suggest that training improves the use of acoustic information that the models already encode.

Tool-using LLM agents act on instructions embedded in the data they read, so prompt injection is a structural rather than a prompt-wording problem. We compare in-prompt defences (hardened prompts, spotlighting, LLM detectors) with deterministic enforcement outside the model (four reference monitors of increasing strength), human approval and taint tracking, judged by one effect-based oracle independent of every defence in an executable environment of 13 tasks and 12 tools, with argument-fuzzing attackers and severity-weighted harm. In an evaluation with real open-weights models (Qwen2.5-1.5B and Qwen2.5-7B; 156 attack episodes per defence and model), the models proposed the attacker's call in 64% and 84% of attack episodes behind the task-bound monitor, which admitted 1.3% and 7.7%; guardrail effects depended on the model (a prompt guardrail moved obedience from 67% to 65% and from 86% to 55%, spotlighting plus a detector to 12% and 0%) against static attacks. Simulation with an assumed model covers what no real run did: under the assumed erosion law a persistent adaptive attacker or a compromised model drives every guardrail stack to 100.0% attack success, whereas the task-bound monitor holds severe actions at 0.0%; its residual (harm 0.69/10) falls by 62% under taint-triggered human review, and a human layer without a deterministic monitor fails against approval spamming unless each episode has a deterministic approval budget. Implementation quality matters: a naive monitor admits 18 of 33 effective attack variants, a hardened one 6 and a capability-bound one 1. The real-model evidence is limited to two open-weights models and static attacks, the human-approval and taint results are simulation-only, and the larger model's detector stack contradicts the assumed erosion law. Code, data, a two-rater oracle audit and a notebook for repeating the real-model evaluation are provided.

LLM agents can use feedback from repeated attempts to improve their decisions without updating model parameters. However, positive action credits can continue to favor actions that repeatedly achieve the same score, even when higher scores are achievable. Under a fixed interaction budget, trying new actions in the final attempts can also lower performance. We introduce EARN, a method that uses past experience to guide further progress and retain earlier performance through conditional reuse of effective action sequences. EARN reduces action credits for reaching scores repeatedly attained in earlier episodes and assigns larger credits to action segments that reach higher scores. These credits adjust the probabilities of LLM-proposed actions. During the final episodes, EARN directly replays stored actions when the current context matches the recorded context and the actions remain available, bypassing LLM candidate generation. We evaluate performance at the end of the interaction budget using the average score over the final four episodes. Across Jericho, ScienceWorld, and Crafter, EARN achieves the best score on this metric in 31 of 33 task--budget settings among the evaluated methods, including ties. Across the evaluated Jericho budgets, replay bypasses 78.9--86.3\% of candidate-generation calls in the final four episodes, averaged across games.

Occupancy-based controls regulate lighting and HVAC according to whether a space is in use. Passive infrared (PIR) sensors, the most common choice, respond to motion and can report an occupied room as vacant when its occupants remain still. Sound offers an alternative cue, since seated occupants type, turn pages and cough, but microphones raise privacy concerns. We therefore study a two-stage design: an on-device encoder maps each second of audio to a short embedding, and a server-side temporal model turns the stream of embeddings into an occupancy state. To our knowledge, no publicly available dataset pairs office audio with occupancy labels, so we evaluate the two stages separately on public data. The encoder is trained on synthetic scenes that mix empty-room recordings, room impulse responses, office sounds and read speech, and tested on unseen rooms, recordings and speakers for whether a second contains a sound made by a person. On 11 real office recordings (DCASE 2013) it reaches an AUC of 0.87, against 0.76 for log-mel statistics. Attacks on the transmitted embeddings recover at most 1% of the spoken words, against 94% from log-mel frames, but still identify the speaker among 40 in 69 to 77% of the utterances. Noise-floor normalization with adversarial training, an 8-dimensional embedding, or sending only class probabilities lowers this to 38 to 55% without hurting detection. On HPDmobile, which pairs home occupancy labels with an amplitude envelope of the audio, a causal temporal model pretrained with a hierarchical contrastive loss matches hand-crafted features (AUC 0.75) and does better when only one labelled day per home is available. These results make the two-stage design a practical basis for privacy-aware acoustic occupancy sensing, with speaker identity as the one leak still to close.

Hydrologic models, including deep learning systems now deployed for operational flood forecasting, are evaluated primarily by how closely their simulated hydrographs match observed streamflow. However, high hydrograph skill does not guarantee physical consistency: in a synthetic calibration experiment, a conceptual model that loses 15% of its water input to an unreported sink still reaches a Nash-Sutcliffe efficiency of 0.93, and no shared benchmark tests whether models obey fundamental conservation laws. Here we introduce HydroTuring (v0.1.0), a black-box benchmark of 37 physics probes that test mass (24 probes), energy (7) and momentum (6) conservation in hydrologic models. Each probe runs a controlled experiment on forcing generated at run time from a recorded seed and applies binary pass or fail criteria validated against 18 physically exact and 59 deliberately broken reference models. Across 32 evaluated models (3 exact references, 7 AI-based models and 22 process-based models), only the 3 exact references pass every probe among the 28 models scored on 5 or more probes. Process-based and AI-based models pass similar shares of their scored probes (76% and 74%), and a pre-trained flood-forecasting LSTM fails 6 of the 11 runoff-only probes it can be scored on. Furthermore, 49% of all 1,184 model-probe pairs remain unscored because models omit the required state or flux variables, while the scored failures occur in perturbation responses and state bounds rather than long-term water-budget closure. HydroTuring provides a reproducible smoke test for model development, peer review and operational deployment under an open community protocol.

As long-video understanding gains increasing attention, augmented reality (AR) devices are emerging as a compelling deployment platform. However, egocentric videos exhibit significant temporal redundancy, imposing substantial memory and computational burdens on resource-constrained AR hardware. To address this challenge, we introduce UTOPI, a plug-and-play pre-compression module designed specifically for egocentric long videos on AR platforms. First, we leverage user motion signals to estimate inter-frame overlap and propose a viewpoint-shift-invariant pruning algorithm. Second, we exploit eye-tracking sensors to identify user fixations, enabling the separation of attended foreground content from background and its preferential retention via a fine-tuned neural network-based adjustment mechanism. Across multiple benchmarks and model backbones, UTOPI achieves up to 95% compression while preserving performance. Code and project page are available at https://sai-lab-nyu.github.io/utopi/.

Large language model (LLM) agents acting across web applications and online services often need multi-turn corrections to align their work with users' underspecified goals and preferences. In practice, users may identify problems in the current outcome without specifying the desired result, so learning user intentions requires interpreting each correction against the agent's preceding and revised outcomes. These expectations are often connected to a particular context, making it hard to know when they apply and how to ensure the agent later acts as the user expects. To this end, we propose LiveAsset, an intent harness that learns reusable intent assets from corrective interactions and uses them for runtime control of agent execution. LiveAsset jointly interprets each correction with the agent's original and revised results to infer user intent, incrementally updating the assets while preserving earlier requirements that remain applicable. The learned requirements guide agent execution from task entry and provide criteria for runtime verification, which uses the explanations to assess applicability before each reply is delivered and directs the agent to revise the result to address unmet requirements while preserving those already satisfied. Following a learning phase in which simulated users progressively reveal requirements through multi-turn corrections, we evaluate LiveAsset using same-task replay of the original requests on π-Bench, PERMA, and Vita and cross-task transfer on PAST-Bench and SkillEvolBench. Experiments show that LiveAsset better fulfills requirements learned from corrective interactions and improves performance on these benchmarks compared with memory, skill, and rule-based harness baselines. Code is available at github.com/landian60/LiveAsset.

While Recursive Self-Improvement (RSI) accelerates isolated algorithm and kernel synthesis, distributed LLM training depends on tightly coupled launch settings, operator kernels, and communication schedules, where single edits often hit out-of-memory (OOM) boundaries, compiler conflicts, measurement jitter, or silent numerical drift. We present InfraRSI, the agentic harness that turns open-ended human commands into verified training optimizations via a five-stage workflow: init locks editable scope and a fixed task contract over workload, loss, and timing invariants to prevent reward hacking, doctor profiles baseline phases and calibrates cluster noise, run co-optimizes launch knobs and code patches before re-verifying top candidates against the baseline and logging findings to a cross-hardware experience bank, and report and apply audit source paths, track failed-trial costs, and commit promoted changes. Rather than stepping greedily from a single incumbent, InfraRSI initializes from diverse experience-seeded configurations and fits global execution-time and censored peak-memory models across all trials, including OOMs. Using these models, it prices per-setting memory costs, revises miscalibrated LLM priors from live runs, and pairs memory-hungry optimizations with memory savers and verified code patches. Across four single- and multi-node dense and Mixture-of-Experts (MoE) LLM workloads, InfraRSI achieves a speedup of up to $2.75\times$ and lifts mean Model FLOPs Utilization (MFU) from $17.4\%$ to $30.0\%$, outperforming existing tuners and coding agents while transferring experience across GPUs. As it improves training infrastructure efficiency, InfraRSI forms a two-level Recursive Self-Improvement loop---revising priors online within runs and distilling experience across clusters---to push foundation model scaling.

Humanoid loco-manipulation requires continuous coordination of high-dimensional whole-body motion to realize task objectives. While vision-language-action models and world action models transfer semantic and visual-dynamics priors to robot control, these priors do not necessarily capture the structure of robot motion. We introduce Kinematic Action Models (KAMs), a paradigm that treats whole-body motion as a direct pretraining target and uses a pretrained motion generator as the primary action-generation backbone. KAMs ground learned temporal and cross-body coordination priors through task conditioning, visual observations, and execution feedback. We present Kimodo-Policy as an initial instantiation, combining a frozen Kimodo text-to-motion backbone with lightweight ControlNet-based visual conditioning. A root-body-hand hierarchy coordinates locomotion and manipulation, while motion-history conditioning and repeated action-chunk prediction bridge motion synthesis and closed-loop control. Following large-scale loco-manipulation pretraining, the policy achieves the highest overall success among compared policies on multi-task HumanoidArena (72.14%), competitive performance on SIMPLE, and 60% mean success across eight real-world tasks, with 678M parameters (92M trainable). Ablations support the contribution of the pretrained motion backbone. Together, these results provide initial evidence for motion generation as a complementary foundation for humanoid action learning.

Masked Diffusion Models (MDMs) are well known for fast text generation via parallel decoding. However, independent parallel sampling can commit to inconsistent predictions, and less is known about how to selectively revise these early token commitments. In this work, we introduce $\underline{\mathrm{P}}$lanned $\underline{\mathrm{U}}$nmasking-and-$\underline{\mathrm{R}}$evision Joint $\underline{\mathrm{L}}$earning (PURL) for MDMs, a training-based framework that seamlessly unifies parallel generation with selective token revision. We derive a normalized joint transition whose likelihood factorizes into edit selection, unmask selection, and token prediction. These components are trained jointly on partially masked sequences constructed according to the model's own unmasking policy. During inference, all three decisions are supported by a single shared forward pass, enabling the model to decode masked tokens and revise noisy tokens. Across logic puzzles, mathematical reasoning, and code generation, PURL improves accuracy at matched unmask budgets, with relative gains up to 47.4% accuracy improvement and up to an $8\times$ reduction in scheduled decoding rounds. Beyond these gains, our analysis further reveals general principles of revisable decoding: token edits become more effective as context accumulates; early edits can steer unfinished trajectories, whereas later edits tend to remain localized; and effective edits therefore depend on both calibrating when to intervene and learning where and how to revise. Together, these suggest that effective parallel generation depends not on revising more tokens, but on making the right commitments revisable at the right time.

Generative recommendation models represent items as sequences of discrete tokens known as semantic IDs. Existing approaches derive these identifiers using residual or hierarchical vector quantization, assuming that tokens carry independent semantic meaning across quantization levels. Revisiting this assumption, we argue that semantic IDs should behave as hierarchical paths where each token's meaning is conditioned on its prefix. This highlights a mismatch between conventional fixed-codebook quantization and the hierarchical structure implicitly learned by generative recommenders. Motivated by this insight, we propose Adaptive Silhouette Tree (AST), a top-down divisive clustering method that constructs semantic IDs as paths within a tree structure. Unlike prior methods with fixed branching factors, AST dynamically determines the number of branches at each node by maximizing the silhouette coefficient, enabling finer partitions in dense regions and compact groupings in homogeneous ones. We further introduce constrained variants that enforce minimum probability mass during splitting to improve supervision for long-tail items. Extensive experiments on five Amazon Review datasets of varying sizes and domains show that AST significantly outperforms strong baselines. These results demonstrate that aligning semantic ID construction with the inherent hierarchical structure of generative recommenders leads to more effective recommendations.

Purpose: To estimate how often tool-using large language model (LLM) agents record clinically relevant false or unsupported information in simulated radiology documentation and how much safeguards applied before each action reduce these errors. Materials and Methods: In this simulation study (October 2026), 48 synthetic, radiologist-reviewed base scenarios covered finding communication, tumor board preparation and follow-up management. Forty-three also had a misleading-information variant (91 variants). Six LLM agents (Claude Sonnet 5.5, GPT-6.1-Sol, GLM-5.3, MiniMax M3, Gemma-4-31B, Qwen3.8-27B) completed each variant with permissions only, deterministic rules, and rules plus a locally run LLM checker (1638 episodes). The primary outcome was at least one executed high-severity violation per episode. Risk differences were estimated per agent at base-scenario level and, exploratorily, pooled across agents, with analyses of document access and blocks. Results: With permissions only, 7.3%-27.1% of episodes had a high-severity violation (pooled across agents, 19.4%). Pooled, it was 16.7% with rules (-2.8 percentage points; 95% CI: -6.1, 0.5) and 13.2% with rules plus the LLM checker (-6.2; 95% CI: -9.5, -3.0). Adding the LLM checker to rules changed it by -3.5 (95% CI: -6.3, -0.7). No difference between agents was detected (P = .43), and task success did not decrease detectably. In misleading variants, agents opened the altered document in 83% of episodes regardless of error, and 58% of erroneous statements cited it. Most blocked actions (64%-68%) would not have caused a violation and were returned for revision. Conclusion: Layered safeguards reduced but did not prevent high-severity errors. Errors after misleading information reflected misjudged rather than missed information.

This paper shifts the object of role engineering from human users to large language model (LLM) agents, and proposes PRISM to address the over-authorization that arises when an agent obtains full tool privileges through the Model Context Protocol (MCP). PRISM formulates whether a tool is necessary for a task as an intervenable causal proposition, and applies structured interventions in a sandbox to learn a counterfactual necessity model. It then compresses trajectories into discrete latent variables through an information bottleneck, and determines the number of roles from the knee of the rate-distortion curve. Finally, it formalizes pruning as a bilevel Stackelberg game, in which a generative flow network samples diverse near-optimal solutions. On three public benchmarks, AgentDojo, $τ$-bench and ToolEmu, PRISM compresses the number of authorized tools per role by 78.3%, attains a task success rate of 78.2%, reaches a counterfactual necessity accuracy of 88.7%, and lowers the blast radius to 15.9%.

Variational quantum algorithms (VQAs) optimize circuit angles through repeated quantum-device evaluations, often facing vanishing gradients and high optimization costs. We ask whether a large language model (LLM) can write these angles directly. Extending *weight-space programming* from classical networks to quantum circuits, we provide a frontier LLM (\gpt) with the problem, circuit, and gate conventions. Its angles are executed exactly as written, without returning losses, gradients, or accuracy feedback. Across 139 requests, we test handwritten-digit classification, VQE ground-state estimation for spin chains and H$_2$, and MaxCut QAOA, comparing against random weights, classical optimization, and optimization initialized with LLM weights on the same circuit. When parameters have interpretable structure, the LLM matches or approaches the optimizer: it achieves 95--100\% of the energy spectrum on every VQE problem, derives the exact H$_2$ ground state in closed form, finds optimal depth-one QAOA angles for every graph, and reaches 93--96\% accuracy on digits 3 versus 8 with 12--18 hand-written weights, outperforming an optimizer trained on the same ten examples. Its weights are also strong warm starts: optimization reaches its best value after a median of one evaluation rather than hundreds. Allowed to design circuits, the LLM chooses data-dependent rotation chains encoding pixel importance in all 30 cases (one qubit for binary tasks, one chain per qubit for ten classes), reaching 65\% accuracy across ten digits. Yet the approach fails in opaque weight spaces: accuracy drops to chance when fixed circuits exceed 30 weights, while optimization improves. It is also sensitive to gate conventions: 6 of 42 QAOA replies reverse mixing-angle signs and fail. We release all prompts, replies, and scores.

Language-model agents increasingly run on workstations with a few consumer GPUs, where large mixture-of-experts (MoE) models keep part of their experts in host memory. With existing CPU-GPU engines, these agents get fast prefill or fast decode, not both: llama.cpp's expert cache skips forward passes of more than 32 tokens to keep the experts of decode cached, so an agent waits 2.1-2.8 s after each tool output, whereas its default layer placement shortens this wait but is 1.5-1.6$\times$ slower in decode. Our key insight is that a short prefill touches too many distinct experts to upload, yet sends 84% of its expert activations (routes) to a stable set of experts that fits in GPU memory. We present Ostinato, a CPU-GPU MoE inference engine that computes these experts on the GPUs in every forward pass without changing what the model computes. Short prefills split their routes, leaving fewer than one of a token's six routed experts per layer to the CPU; long prefills upload each distinct cold expert once; and each decode step runs as one fixed-shape CUDA graph per GPU. On real coding and customer-service agent sessions with DeepSeek-V4-Flash on four RTX 4090D GPUs, Ostinato is the fastest engine in both phases and finishes every session sooner than all six baselines. Against llama.cpp and KTransformers, it cuts the median wait after a tool output by 22-55% and finishes sessions 1.4-2.5$\times$ sooner.

A driving foundation model must learn reusable decision knowledge from rare critical events and apply it under strict runtime constraints. We introduce DriveFoundry, a closed-loop framework improving iteration efficiency under data and computation budgets by coupling structured causal data refinement with adaptive inference. The Data Engine builds traceable scene-candidate records linking scene evidence and factor-decision relations to candidate trajectories and outcome assessments. Candidate comparisons and counterfactual annotations clarify how scene factors affect behavioral feasibility and preference. Verified model failures guide VLM-assisted annotation and expert correction, refining training data and annotation systems. The model engine, Xiaomi-Drive-o1, unifies Perception, Causation, Action, World, and Value within a shared vision-language backbone to enable cross-task knowledge transfer. Block-wise latent refill passes candidate-specific evidence and predicted visual consequences between task blocks, connecting scene understanding, planning, and evaluation. A learned value-of-computation head selects tasks and allocates latent tokens based on expected benefit and cost, adapting inference to each decision. On PhysicalAI, the 9B model achieves the lowest ADE@1 across evaluated splits and horizons among tested models. In AlpaSim, it outperforms Alpamayo 1.5 across all metrics and achieves lower close-encounter and off-road rates than the 34B Alpamayo 2 Super, while trailing in progress and AlpaSim scores. For the 0.8B variant, joint training reduces 5 s minADE@6 by about 54% without altering the inference path or token budget. Adaptive inference matches full reasoning performance using only 28% of tokens. Real-vehicle tests confirm runtime feasibility, supporting a unified paradigm for learning from critical experience and dynamically allocating computation.

Which representation should be probed in a transformer-based foundation model? Recent work shows that the standard practice of probing the final layer is often suboptimal, as intermediate layers can provide better features. Yet, representations are still largely studied along depth alone, treating each transformer block as an atomic unit. In this work, we analyze transformer representations layer by layer, module by module. Through more than 144,000 linear probing evaluations of vision foundation models (VFMs) and large language models (LLMs), we show that layer and module choices should be considered jointly: the best layer-module pair yields substantially larger gains over the final-layer residual stream than the best layer with the residual stream fixed. Moreover, module performance can vary substantially across layers, showing that useful representations cannot be characterized by depth alone. Building on these findings, we evaluate simple fixed probing strategies: the final-layer feedforward activation (Act@100) for vision and the second normalization layer at 60% depth (LN2@60) for language. These strategies improve over the probing baseline in a majority of evaluations and recover, on average, 42% and 86% of the gains from probing the best layer-module pair for vision and language, respectively. For VFMs, these gains increase with the transfer difficulty between pretraining and downstream data.

There is a well-established body of work demonstrating that individual machine learning (ML) algorithms can be manipulated or deceived. However, in many real-world systems, multiple ML algorithms are composed to enable automated decision-making, particularly in Cyber-Physical Systems (CPS). Often such systems use a composition of vision, planning, and other ML models to affect their internal state and ultimate take actions in the physical world. While there is a huge literature on the (in)security of ML algorithms, it is not immediately clear how an attacker might leverage their ability to trick individual learners when attacking the composition as a whole. We develop methods for creating an attack policy (a mapping from state to attack) for a data-manipulating attacker affecting a CPS or other state-machine-based AI compositions. We describe our framework, a basic threat model and corresponding methods for computing an optimal attack policy, as well as planned extensions. The goal of this work-in-progress paper is to stimulate conversation and forge new collaborations regarding the security of compositional learning.

AdamW is ubiquitous in deep learning, yet its behavior remains poorly understood. We analyze its dynamics through the lens of dynamical systems and show that AdamW admits an implicit fixed-point objective: its fixed points coincide with the stationary points of a constrained and regularized optimization problem. However, not all of these fixed points are stable under AdamW's dynamics, and stability depends sensitively on curvature, weight decay, and momentum parameters. Even in one-dimensional settings, equilibria may be unstable, and trajectories can exhibit persistent oscillations rather than convergence to them. We further extend the analysis to higher dimensions, deriving sufficient conditions for local stability of the continuous-time dynamics, and propose a convergent modification that preserves AdamW's fixed points. These results clarify what optimization problem AdamW is associated with, when its convergence can be expected, and how its dynamics could inspire more reliable optimizers.

Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Transferability and probe-quality scores characterize global alignment, likelihood, margins, and calibration. Boundary-near target mass complements these diagnostics by measuring the ambiguous regions where scarce labels are most costly. We propose effective boundary complexity, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and predicted-class imbalance are both counted. We prove that, under a smooth representation map, density change and tangential surface change cancel at first order, leaving boundary-normal stretch or compression as the governing factor. This yields a cached-feature estimator that fits a probe on one split and counts independent held-out features near its boundary. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent controls preserve the first-order coefficient. A differentiable surrogate reduces measured complexity by 16-36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet experiments support boundary mass as a second-stage diagnostic for near-tied representations. Five-way experiments show strong OvR-interface backbone-ranking agreement (Spearman 0.872), while acquisition experiments show high correlations between initial boundary mass and realized label budgets across three sampling policies (0.779-0.830). Code is available at https://github.com/qqaqq-code/boundary_mass.

We propose a music inpainting technique combining similarity matching and a diffusion-based generative model to replace large segments of audio recordings with suitable alternatives. The proposed hybrid method, Similarity-Guided Diffusion Posterior Sampling (SimDPS), retrieves candidate replacement segments from the wider signal based on contextual similarity and uses them to guide the diffusion process toward coherent reconstructions. Being model-agnostic, the framework is compatible with any pre-trained diffusion model operating in either the waveform or latent domain. Objective evaluations demonstrate that SimDPS outperforms widely adopted RePaint and DPS methods across acoustic features and neural-embedding distance metrics. Subjective evaluations on multi-genre music datasets confirm that SimDPS enhances perceptual plausibility compared to previous diffusion-based methods and demonstrates robustness to poorly matched similarity candidates. These findings pave the way for improved music restoration and context-aware audio editing.

World models learn internal simulators of environment dynamics for prediction, planning, and control, yet learning them from limited, noisy, and partial observations remains challenging. While much progress has focused on pixel-based observations, many physical systems instead provide vector-valued measurements that reveal only part of the underlying state. We introduce K-PWM, a probabilistic world model for prediction and control under partial observation. K-PWM couples Koopman-bilinear latent dynamics with a nonlinear observation decoder and performs probabilistic latent-state inference through the learned model via extended Kalman filtering and smoothing, without a separately trained inference network. Its bilinear structure enables exact propagation of the maintained Gaussian belief through the latent transition and efficient rollouts for uncertainty-aware model predictive path integral (MPPI) control. For training K-PWM, we present a generalized expectation conditional maximization (GECM) procedure with closed-form updates for the generators, noise covariances, and latent prior, complemented by gradient-based decoder learning. A Koopman-consistent initialization selects the latent dimension from data and initializes the action-dependent dynamics. With limited training data, K-PWM achieves lower long-horizon prediction error than the evaluated baselines on partially observed Gymnasium MuJoCo systems, while substantially reducing training time relative to the strongest prediction baseline. On high-dimensional MuJoCo Menagerie robots, it achieves competitive tracking accuracy with low per-step computation times, demonstrating potential for real-time deployment. These results support structured probabilistic dynamics as a data- and compute-efficient approach to world modeling and control under partial observation.

Real-world evolving networks are naturally modeled as temporal graphs (TGs), where capturing temporal dynamics is essential for predicting future graph properties that support downstream decision-making. Existing temporal graph methods have been developed primarily for single-task prediction, and little is known about their generalization across tasks or transfer to unseen networks. This leaves the challenge of multi-task graph property prediction in TGs largely open. We address this challenge by introducing Hydra, a novel architecture that integrates local connectivity features from temporal GNNs with a spectral learning module that captures global connectivity patterns. This design enables joint learning of local and global information under a multi-task objective. In multi-task classification, Hydra achieves an 8.9% relative gain in AUC over the strongest competitor. In multi-task regression, Hydra achieves competitive results in all three tasks, while obtaining the best results in two tasks with a 8.2% relative gain in MAE compared to the strongest baseline. Moreover, Hydra delivers these gains with a 22% reduction in training time compared to temporal transfer models. These results provide the first systematic evidence that multi-task transferable learning on temporal graphs is effective. By delivering consistent top-ranked performance, Hydra highlights multi-task training on temporal graphs as a promising direction toward adaptable foundation models for temporal graphs.

Text- and goal-conditioned human-object interaction (HOI) provides an intuitive interface for specifying high-level intents in physics-based character animation. Mapping such intents to physically plausible, contact-rich behaviors naturally calls for a generative planner and a physics-based controller, yet existing planner-controller systems either execute plans open-loop or close the loop while keeping the planner fixed, so execution feedback never reshapes what is planned. We present InfCoiL, a closed-loop framework for physics-based HOI from coordinated planner-controller learning. InfCoiL couples a morphology-aware flow-matching planner with a unified multi-morphology controller through a shared morphology interface in an autoregressive plan-and-execute loop, and alternately fine-tunes both components with reinforcement learning from the same physical rollouts. To optimize the flow-matching planner online, we introduce Semantic Flow Policy Optimization (SFPO), which stochasticizes selected denoising transitions with channel-wise semantic covariance over heterogeneous HOI features and assigns group-relative, sequence-level rewards from complete closed-loop executions. On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy. Project page: https://infcoil.github.io/

Understanding how an athlete moves can improve both technique and safety, yet outside the laboratory technique is still assessed by eye. We present a pipeline that detects named golf swing faults from ordinary video, without markers or calibration. Instead of learning end to end, it tracks body pose and the clubhead in a body-centered frame, and classifies the biomechanical quantities a coach would measure, such as the height of the clubhead above the swing plane, with shallow boosted trees. We demonstrate the approach on two common faults, coming over-the-top (OTT) and early-extension (EE). On 160 swings and 143 swings from public sources it detects OTT and EE with a cross-validated ROC-AUC of 0.975 and 0.911 respectively. Because every feature is a biomechanical quantity a coach already uses, measured at a known phase of the downswing, each decision can be traced to the signals and the moments of the swing that drove it. The signals the model relies on most are the same ones a coach checks when diagnosing these faults. The pipeline processes phone video on a consumer laptop in seconds, so it could feasibly be developed into an app for golfers of all abilities and extended to other sports in which technique is judged against a biomechanical reference.

Amortized Bayesian experimental design (BED) enables real-time design strategies by shifting acquisition cost offline, but existing methods remain tied to the prior and task distribution they were trained on and cannot exploit additional information available at deployment. We introduce in-context amortized BED with unified knowledge encoding (IMBUE), a BED framework that admits external knowledge at deployment through a unified in-context interface. IMBUE accepts two forms of such knowledge without retraining, namely user-specified prior knowledge and observations from previous instances of the experiment. Both are encoded as additional input tokens and processed in-context with the accumulated observations by a shared Transformer, and a learned reliability filter scores each external token against the observations and excludes inconsistent ones. Across four benchmarks, IMBUE accelerates early-stage information acquisition when the external knowledge is reliable, and remains close to the no-external baseline when it is not.

Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.

Lookup-table (LUT)-based neural networks (NNs) offer a compelling path to ultra-low-latency inference on FPGAs. However, existing LUT-based neural networks have largely been restricted to MLP-style topologies, leaving structured vision and audio tasks underexplored. We present ConvLUT, a family of fully unrolled LUT-mappable convolutional architectures targeting the MLPerf Tiny benchmarks. It is comprised of a standard convolutional model built around a hierarchical tree of LUTs and a depthwise separable model following the MobileNet design philosophy. These architectures are enabled by four key contributions: a unified LUT-based convolutional framework supporting standard, depthwise, and pointwise convolutions under explicit fan-in constraints; a LUT-tree that addresses the $k^2c$ receptive field dimensionality bottleneck; a channel-partitioned LUT pooling layer that replaces conventional spatial pooling; and a hardware-aware iterative pruning strategy for stable convergence at scale. Evaluated on CIFAR-10, Visual Wake Words, and Keyword Spotting, ConvLUT achieves competitive accuracy with fully unrolled inference latencies below $\sim 50$ ns. On CIFAR-10 it reaches at least a $8844\times$ latency reduction over non-logic-based prior implementations, and outperforms existing MLP-style LUT-based models on the same task by at least $20$ pp in test accuracy.

An LLM agent that reads untrusted content, such as an email or a web page, may follow instructions hidden in it. Noninterference is too strong for such agents, since a useful agent lets untrusted content influence what it does and may disclose the user's data by design, and declassification needs a release policy written for each task. We instead state security relative to a reference agent, which runs the same plan under the same runtime monitor with a model that never obeys untrusted content. An agent is relatively secure if every leak it has is a leak the reference already has. Our framework, PlanTERS, instantiates relative security, a notion from speculative execution, with a small agent language in which the attacker chooses how the model follows instructions, a composition with runtime monitors, and attacker models that differ in what they observe. A fence theorem shows that revoking injectable tools at the first untrusted read, and never granting them back, makes every plan relatively secure at the event level on finite runs, for every model behaviour in which each injection is a single call and the model obeys on a schedule fixed in advance; with a fairness condition, this extends to possibly infinite runs. We prove or refute relative security for agents with fences, budgets, approvals, loops and dual LLMs, in Isabelle/HOL.

Due to the fundamental problem of causal inference -- the impossibility of observing counterfactual outcomes -- there is always a gap between the training goal (outcome, propensity) and the inference target (CATE, ATE). Traditional causal effect estimation revolves around setting assumptions, selecting identification strategies, and choosing causal estimators, while more recently, Causal Foundation Models (CFMs), built on Prior-Data-Fitted Networks (PFNs), ease this process by pretraining on massive synthetic prior datasets. Despite their success, such amortization makes causal inference less transparent, obscuring which assumptions the model implicitly relies on for a given dataset. The high diversity of real-world data-generating processes (DGPs) and Structural Causal Models (SCMs) makes it hard to know, a priori, which causal inference strategy is compatible with a given dataset. In this work, we focus on this compatibility between an observational dataset and method-specific assumptions, and find that geometric characteristics of data embeddings -- specifically, the variogram -- provide a useful signal of whether a given causal estimation strategy is appropriate. We develop MoE-Learner, a mixture-of-causal-experts framework that embeds a dataset under multiple assumption-specific views and uses the variogram as a counterfactual-free router to assess each expert's validity. Rather than eliminating causal assumptions, this shifts their validation from a fixed, a priori choice to a criterion that can be checked directly from data geometry. MoE-Learner achieves the best average rank on benchmark datasets (IHDP, ACIC, Lalonde). Being orthogonal to (causal) foundation models, our method builds a bridge from data characteristics to causal estimation strategy rather than proposing another CFM.

Marcolli, Chomsky, and Berwick (2025; henceforth MCB) formulate the Merge operation of Minimalism on a Hopf algebra of Connes-Kreimer type, and claim that language and physics obey the same algebraic structure. This paper separates their program into the working hypothesis H, that Merge on workspaces can be described as a Hopf algebra, and the claim P placed on top of it, that language and physics obey the same principle, and examines the validity of H. MCB employs two coproducts in formulating Merge, corresponding to two different translations of movement (Internal Merge): $Δ_d$, which deletes the lower copy, and $Δ_c$, which leaves a trace. MCB uses the former in syntax and the latter at the syntax-semantics interface, but $Δ_d$ is not coassociative, and $Δ_c$ is not connected under the gradings MCB gives and, moreover, not coassociative as long as the labels of traces are retained. Consequently, MCB's algebra is not a Hopf algebra, and the Birkhoff factorization is not a consequence of the Hopf-algebra theorem but of a weaker statement that requires neither coassociativity nor connectedness. There are four ways to recover a Hopf algebra, but none of them can distinguish different traces, and hence none is compatible with Heim-Kratzer-style formal semantics, which reads traces as indexed variables. Thus H does not hold at present, and under the current formulation, making Minimalism a Hopf algebra and preserving its connection with formal semantics are mutually exclusive. Finally, we indicate three directions in which algebraic Minimalism may proceed.

The performance of language agents depends jointly on model weights and the harness. Optimizing either alone can be constrained by the other component, limiting overall performance gains. Our analysis of naive alternating training reveals progressive incompatibility between the model and harness, which increasingly interferes with task execution and may also undermine model training, leading to performance degradation. To coordinate the evolution of models and harnesses, we introduce JoMoHa, to our knowledge the first framework to jointly train an executable harness generation policy and a task execution policy through reinforcement learning. JoMoHa outperforms the evaluated weight-only and harness-only baselines and achieves the highest success rates among the evaluated methods on ALFWorld, WebShop, and DBBench, averaging 70.3% and reaching 97.0% on ALFWorld. By learning both task execution and runtime improvement from its own generated trajectories, JoMoHa performs recursive self-improvement (RSI).

We describe the system submitted by Telefónica Innovación Digital to the Speech-COSER Challenge 2026, which features spontaneous speech from elderly speakers recorded in rural Spain. For ASR, we perform CTC-based pseudo-forced alignment scores audio-text match quality to guide a two-stage LoRA curriculum, followed by post-hoc magnitude pruning of the LoRA matrices to enhance generalization. An ensemble of seven systems combined via ROVER achieves 36.25\% unnormalized and 22.17\% normalized WER.

We study stochastic bandits with reward-dependent delays, possibly infinite, whose reports omit the originating round. For each $c\in[0,1)$, Balanced Historical Optimism (BHO) has a delay term controlled by the largest suboptimal-arm quantile at level $1-cΔ_i$, without requiring optimal-arm reports. Lower bounds give an inverse-square slack cost in the logarithmic statistical regime and linear minimax regret at $c=1$. A family with growing $K$ forces worst-case regret $Ω(G\log K)$ for every learner, where $G=\max_iΔ_i h_i$ is the largest gap-weighted suboptimal cutoff. Pending-triggered balance preserves BHO's weighted-cutoff bound for $c<1$ and attains the historical-greedy exploration coefficient with deterministic rewards, a unique optimum and fixed bounded delays. Gap-weighted comparisons and experiments show when shared waiting helps and what balanced exploration costs.

Predictive algorithms are increasingly used to guide hospital discharge decisions, yet these models can exhibit censoring bias for vulnerable patients due to insufficient historical data or past discharge decisions. When biased readmission risk predictors are used to inform discharges, the training data generated by discharge decisions may perpetuate similar biases when predictors are re-trained, leading to a vicious cycle. We extend a queueing-based Markov decision process for hospital discharge to incorporate periodically retrained readmission risk predictors. Our key insight is that cost-optimal, "capacity-aware" operational decision policies can naturally diversify the lengths of stay at which patient outcomes are observed, enabling consistent estimation of readmission risk without explicit fairness constraints. We prove that repeated retraining under discharge-up-to policies converges to unbiased risk estimates under well-specification. In synthetic experiments and calibrated simulations based on patient-level data from a partner hospital, we find that where commonly used threshold-on-risk policies can remain persistently biased, discharge-up-to policies reduce learning error over status quo decision policies (such as thresholding on risk) by 24%-25%. These improvements to estimation also mitigate some forms of bias, reducing readmission cost disparities (adjusted for baseline risk differences between demographic groups) by 38%-63%. Our results demonstrate how capacity-aware decision policies can improve both predictive performance and fairness by addressing censoring bias.

Many studies of animal vocalizations require clean recordings, but animals are usually recorded in the field, with background noise from weather, vehicles, and other animals. Animal vocalization denoising is challenged by the scarcity of clean training data and the extreme diversity of animal vocalizations. Classical spectral gates leave residual noise, especially without voice activity detection (VAD), while detector-free neural denoisers designed for bioacoustics remove less noise. We show that a general-purpose source separation foundation model, SAM-Audio, removes more noise than either of these types of methods, when used without training and prompted with time spans from a simple animal VAD. To turn this heavy method into a practical tool, we distill it into CallCleaner, a 19M-parameter denoiser that needs no detector and runs ~90x faster than real time on a laptop CPU. We find that SI-SDR has limited utility in our setting as it penalizes regenerated output and conflates noise removal with call fidelity, a trade-off our separate metrics reveal. CallCleaner leaves 11 dB less noise around the call than Biodenoising, the leading neural bioacoustic denoiser with public weights, and among fast, CPU-capable systems it pushes the Pareto frontier of noise removal vs. call fidelity, without requiring a detector. We release code and weights for CallCleaner, which enables fast, private, local denoising of single-caller recordings without specialized hardware.

Existing evaluations often test persona expression rather than whether persona information persists in extended decision-making. We introduce THEATRON, a controlled benchmark that uses financial replay to evaluate persona persistence in long-horizon sequential decision-making. In these environments, models conditioned on fixed behavioral-finance personas observe shared daily market evidence and trajectory-specific account-state feedback and make sequential decisions over multi-month trajectories. We evaluate Zero-Shot, One-Shot, and Chain-of-Thought persona injection across four LLMs, using cumulative-return and action trajectories to diagnose within-persona stability, cross-persona separability, and action collapse. Persona conditioning produces systematic behavioral shifts beyond a no-persona baseline, and persona-level separation generally exceeds the variation induced by semantically equivalent descriptions. Nevertheless, these shifts do not reliably form coherent persona-specific policies: apparently stable separation can arise from stereotyped action collapse, while diverse policies can exhibit noisy trajectory switching. These findings expose a gap between observable persona-induced behavior and persistent, context-sensitive persona-conditioned decision policies. Code and benchmark data are publicly available at: https://github.com/yuyan12138/theatron

We present MicroSim, a system for binary code similarity detection that represents each function as a graph built from IDA Pro Hex-Rays microcode and compares such graphs with a contrastive Siamese graph neural network (GNN). We evaluate on the TREX cross-optimization benchmark, a corpus of 849,072 function-level microcode graphs spanning optimization levels -O0 through -O3, under a strict class-disjoint split in which no function key is shared between train and test. Under this protocol, MicroSim reaches 0.9972 area under the receiver operating characteristic curve (ROC-AUC) and 0.9762 Matthews correlation coefficient (MCC), outperforming Gemini (0.9332 ROC-AUC), SAFE (0.9752), and jTrans (0.9740) retrained on the same split, as well as a structure-free Bag-of-Words (BoW) baseline (0.9484). We analyze the gap between pair classification and ranked retrieval in graph transformers, showing that Cross-domain Similarity Local Scaling (CSLS) under unified tie-resolved evaluation raises GraphGPS mean reciprocal rank (MRR) from 0.2542 to 0.2559 (R@1 = 0.1639, R@10 = 0.4561), strictly satisfying the theoretical MRR upper bound. Finally, over 5 seeds with equal compute, the pairwise cosine-embedding loss attains a slightly higher ROC-AUC (Δ = -0.0256, 95% confidence interval [-0.3282, +0.2770]), whereas the supervised-contrastive information noise-contrastive estimation (InfoNCE) loss is significantly better on thresholded classification and ranked retrieval.

In multi-institutional biomedical data analysis, comprehensive disease profiles are often fragmented across different data silos. Furthermore, local data distributions typically exhibit varying levels of granularity or even complex nested structures. Although Hierarchical Clustering (HC) is a powerful tool for exploring such data topologies, its cumbersome recursive process incurs prohibitive computational overhead and renders it highly susceptible to privacy leakage. This paper, therefore, introduces an efficient one-shot hierarchical FC framework that performs client-end (institutions) distribution exploration and server-end distribution aggregation through one-way prototype-level communication from clients to the server. Although the designed one-shot prototype transmission guarantees minimal computational overhead and mitigates privacy risks, it inevitably incurs information loss. To counteract this, a multi-granular competitive learning mechanism is proposed on the server to fuse the clusterlets, effectively maximizing the retention and mining of global distribution topologies under limited, one-way communication constraints. It turns out that the complex cluster distributions across clients can be efficiently explored, and extensive experiments comparing state-of-the-art methods on public datasets demonstrate the superiority of the proposed framework in balancing robust clustering, privacy preservation, and computational efficiency.

Swarm Learning improves resilience and removes the need for a single trusted coordinator by distributing model aggregation and validation across participating nodes. However, it also delegates the validation of model updates to peers whose private data may be biased, unrepresentative, or adversarial, and existing Swarm Learning protocols generally do not attest that data, leaving decentralized validation as an attack surface. We present a robust, privacy-preserving Swarm Learning framework that makes verifiable data quality the condition for admission to the learning task. A validator joins only after a zero-knowledge attestation proves, without revealing any raw sample, that a reference model ran correctly on its private data. The attested outputs also yield a label-free class profile, so the network assembles a collectively IID validator pool from individually skewed validators as accurately as an oracle with the true labels (total-variation distance to uniform 0.025, against 0.021 for the oracle and 0.21 for random selection). Admitted validators score client updates and, through a Tendermint-style Byzantine Fault Tolerant consensus, agree on a gradient mask and a quality-derived aggregation weight for each submission, replacing dataset-size weighting with a quality-aware signal. Adapting the cut-layer idea of Split Learning, a challenge-based audit lets peers re-verify that a mask was computed on the attested data without an expensive end-to-end zero-knowledge proof. Across four benchmarks under non-IID partitions the framework matches FedAvg on clean data; under a gradient-scaling attack that collapses FedAvg to random chance, it retains 94% of clean accuracy on CIFAR-100 and over 99% on FashionMNIST, where FedAvg achieves only 2.1% and 11%. The result is a private, robust, and auditable foundation for decentralized learning.

The small open embedder EmbeddingGemma-300M, used as released, finds the right passage first for only 51.7% to 60.8% of Belebele questions in Tajik, Kazakh, Uzbek and Kyrgyz. We present Saidzoda-CA-EmbeddingGemma-300M, one model for all four languages. It is our Tajik fine-tune of EmbeddingGemma-300M, trained for one more epoch on 279,904 pairs: questions in Kazakh, Uzbek and Kyrgyz that our own language model wrote about source passages, a Tajik refresher, and English and Russian replay. On Belebele, recall@1 rises by 14.4 to 26.0 points in each language, to between 75.2% and 77.7%. Against single-language models trained with the same settings, the one model stays within 2.8 points at recall@1 in Kazakh, Uzbek and Tajik, with overlapping recall@10 intervals, though it also had more training. It keeps the Tajik law recall@10 of its starting point (92.2 against 92.4) but loses 4.4 points of law recall@1. A model trained on one of these languages also improves the others on Belebele, but less than training on them. The model still trails the 560M XLM-RoBERTa embedders on Kazakh, Uzbek and Kyrgyz Belebele: the better of multilingual-e5-large-instruct and BGE-M3, used as released, is 9.4 to 10.2 points ahead at recall@1. Our scores of those released models agree with the published MTEB results within 2.4 points. Russian and English recall@1 fall by 3.0 and 4.2 points.

Generating accurate CAD models requires substantial engineering expertise and careful control of geometry, topology, and mechanical constraints. Text-to-CAD aims to simplify this process, but current LLM-based methods often generate CAD programs directly. Language models remain unreliable when producing long code while coordinating coupled geometric, topological, and dimensional constraints. Consequently, executable programs may still produce structurally incorrect models, which execution rate and rendered appearance may fail to reveal. To address these limitations, we propose GOPR-CAD, an agentic Text-to-CAD pipeline based on Graph-of-Operations Progressive Reasoning. A language model acts as a decision agent over encapsulated CAD operations and progressively constructs an executable operation graph, which a deterministic compiler converts into the final CAD model. This separation requires no additional training and supports different language models. To evaluate structural fidelity, we construct MHCAD, a benchmark of 300 standard mechanical part requirements derived from Machinery's Handbook. Their explicit structures enable reference-free STEP/B-rep evaluation. GOPR-CAD achieves 82.7% Structural Yield on MHCAD, and 87.2% Structural Yield on an external CADPrompt subset.

Large language models (LLMs) perform well on general coding tasks, but their capabilities in application-layer embedded software development remain underexplored. Existing embedded benchmarks primarily focus on hardware-in-the-loop validation in simulated environments. We introduce RealEmbedBench, a benchmark of 5,142 samples covering three application-layer tasks: code generation and completion, state machine modeling and implementation, and cross-OS code translation. We evaluate six representative LLMs and find substantial performance differences across tasks and models. Project context, especially the local code window, significantly affects code generation and completion. Models perform reasonably well at state machine modeling but struggle to translate these models into executable code. Cross-OS translation performance varies substantially by direction. Translation from Harmony to RT-Thread is particularly challenging, with an average pass rate of 0.1667. These findings suggest that developers should consider local code context when using LLMs for code generation and completion and select models according to the translation direction. They also motivate the use of more real-world state machine implementation data in LLM training to narrow the gap between modeling and executable code.

Audio deepfake detectors trained with known synthesis attacks can learn generator-specific shortcuts that fail as synthesis systems and acoustic conditions evolve. We investigate an attack-free alternative: a normality model trained exclusively on bona fide speech, which scores synthetic audio by its deviation from authentic representations. We propose MLN-WavLM, a method that applies shrinkage-Mahalanobis modeling to pooled hidden states from a frozen large-scale, pre-trained speech processing WavLM Base+ architecture. Evaluated on ASVspoof 2019 LA, our best attack-free variant achieves an 8.45\% Equal Error Rate (EER), closely tracking a fully supervised WavLM-HM-Conformer baseline (7.86%). However, when subjected to the ASVspoof 2021 DF domain shift, performance drops to 22.12% EER, exposing vulnerabilities in the bona-fide-calibrated operating point. Layer-wise analysis reveals that early WavLM representations dominate the ensemble. To test real-world viability, we simulated deployment streams using unlabeled threshold adaptation. While robust median/MAD recalibration easily handles in-domain contamination up to 10%, recalibrating under the 2021 DF domain shift triggers a total collapse of low-FPR recall. We conclude that while frozen source-domain normality can effectively rank deepfakes, it does not yield a stable open-world decision rule, and downstream threshold adaptation cannot rescue upstream ranking degradation.

The CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a $π_{0.5}$-based policy that is also its own value function. It is trained with AWR and RECAP on advantages of the exact official score, in a loop where every rollout is at the same time a Thompson-sampling trial of about 20 inference parameters. Out of the box, it took the policy to 0.85 RDS on held-out seeds in two days (and it was still very far from convergence). The competition-specific part adds data that local exploration cannot find: generated demonstrations that reach the full reward, and recovery episodes that start from the policy's own failures. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.

Aligning two models without paired data requires a distance that both share, yet existing methods use Euclidean or spherical geometry. This is either an arbitrary choice or imposed by their loss function. We show that, under a precise form of the manifold hypothesis, the pullback of the data metric gives such a distance for any two models of one data-generating process. This holds whether the models are encoders, decoders or both, and whatever their architecture, loss, prior or training data. The alignment map preserves this distance and the distribution of representations, and is identifiable if and only if the data distribution breaks every symmetry of the data manifold. We show that this symmetry breaking holds generically. Identifiability makes the alignment coupling the unique optimum of Gromov--Wasserstein and of its weighted combinations with a linear program. The linear program alone has the same optimum if almost every data point has a distinct distribution of distances. The coupling, and with it the alignment map, is recovered exactly from permuted samples, even without identifiability, and consistently from independent ones. Our method, Neural Gromov--Monge Alignment (NGMA), uses no pairs or labels and outperforms every unpaired and label-informed baseline. It relies on pullback distances to align image models with different priors and single-cell models trained on disjoint cells, where Euclidean distances fail.

Temporally correlated noise can substantially improve the utility of differentially private optimization, yet candidate mechanisms are commonly designed or evaluated using only partial information about the optimization process, such as workload structure, learning-rate schedules, or local curvature. We study whether persistent optimizer state provides additional information that can change the relative utility of privacy-valid correlated-noise mechanisms. We introduce MOMENT-DP, an optimizer-state-aware framework that models how temporally correlated privacy noise propagates through momentum-based optimization. For locally quadratic objectives, MOMENT-DP expresses privacy-induced final loss through an optimizer transfer operator that jointly captures curvature, learning-rate scheduling, momentum, and temporal noise covariance. NoiseCurve's curvature-aware quadratic criterion is recovered exactly as the zero-momentum, constant-learning-rate special case. Changing only momentum while holding curvature, schedule, privacy calibration, and candidate mechanisms fixed induces genuine mechanism-ranking reversals. The effect persists outside exact quadratic dynamics in per-example-clipped private logistic regression, while remaining unresolved in the neural regime we study. These results establish optimizer state as an under-modeled, decision-relevant axis in evaluating temporally correlated privacy noise.

Agentic autoregressive video generation combines semantic planning, sequential generation, and visual verification for controllable synthesis. Growing visual histories enlarge the key-value (KV) cache, increasing memory use and attention costs across candidate rollouts. Existing compression methods primarily optimize individual generators without the planner's semantic constraints, risking low priority for states supporting requested attributes or actions. We introduce SemanticKV, a training-free method for semantically aligned KV cache compression. To our knowledge, SemanticKV is the first KV cache compression method for agentic autoregressive video generation that explicitly uses planner-extracted semantic constraints to guide retention. Text cross-attention identifies constraint-related video positions, whose self-attention queries are collected into separate groups. Attention weights are averaged within each group and maximized across groups, giving each cached entry a high score when strongly used by any represented constraint. A single top-k selection retains high-scoring entries under the memory budget, leaving their keys, values, and positions unchanged. For fixed query groups and equal entry costs, this selection minimizes an additive upper bound on worst-group discarded attention. The method operates independently within each candidate without retraining or additional video generations. We evaluate four AR video backbones on three benchmarks for 5-, 30-, and 60-second generation. On Causal Forcing at a compression ratio of 0.7, SemanticKV obtains a VBench Total Score of 84.10, compared with 84.05 for the strongest compressed baseline and 83.99 for Full KV. It achieves a 1.22x speedup over Full KV and reduces KV-cache memory from 6.04 GB to 1.81 GB, a 3.35x reduction.

On-policy distillation (OPD) trains a student language model on its own responses, guided by a fixed teacher. Generating these responses is expensive, so systems use quantized inference and asynchronous rollouts to reduce the cost. Both, however, introduce training-inference mismatch between the rollout engine and the model being trained. Such mismatch can destabilize reinforcement learning (RL), but its effects on OPD remain unclear. This paper investigates how much mismatch OPD tolerates, where its effects arise, and how training systems can use this tolerance. Measuring all mismatch sources on one scale, we examine quantization across three student-teacher pairs, with additional tests of weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful loss of training quality but degrades at higher levels, revealing an empirical mismatch budget. Both the quality cost of mismatch and the mismatch introduced by quantization depend on the model, with some layers particularly sensitive to reduced precision. In controlled comparisons, OPD remains stable under rollout configurations that destabilize RL methods. Under weight noise, mismatch also shifts the prefixes reached during generation, introducing errors that per-token corrections cannot address. To put these findings into practice, we introduce a lightweight admission check that screens rollout configurations against a calibrated budget without requiring a full training run for each candidate. Model-specific FP8 recipes preserve training quality while providing capacity for at least $1.93\times$ as many KV-cache tokens as BF16. Combining FP8 with asynchronous rollouts reaches the same training quality $1.62\times$ sooner than synchronous BF16 training. Code is available at https://github.com/benyucong/train-inference-mismatch-opd

The order in which propagators are executed can substantially affect the performance of constraint programming solvers, yet propagator priorities are typically determined by fixed, manually designed heuristics. In this work, we explore learning instance-dependent propagator priorities from static problem features. We formulate the task as a learning-to-rank problem, addressing the challenge that the quality of a priority assignment can be measured through solver runtime, while the desired priorities of individual propagators are not directly observable. A self-attention-based model jointly scores the propagators present in an instance and is trained using pairwise preferences derived from solver runtimes measured under randomized propagator priority assignments, with a Plackett-Luce-based ranking objective. Predicted rankings are mapped to the discrete priority levels of the Choco solver before search begins, without modifying its propagation or search procedures. We evaluate the approach on XCSP3 instances and demonstrate that propagator ordering can have a substantial impact on solving time. On a test set of 698 instances, picked out as most difficult by problem class, the learned priorities increase the number of solved instances from 634 to 654, reduce median solver runtime by 27.93%, and reduce aggregate penalized runtime (PAR1) by 12.65% compared with Choco's default priorities. We further observe that improvements vary across problem classes and show that a simple policy selecting between learned and default priorities further increases the number of solved instances. These results demonstrate the potential of learning instance-dependent propagator priorities to improve constraint programming performance, particularly on difficult instances.

Pre-training code language models relies on massive code corpora usually streamed over the internet. However, researchers rarely need a multi-terabyte archive in its entirety; instead, their workflow centers on curated subsets: selecting well-documented files, filtering out repetitive boilerplate, auditing for benchmark contamination, or retrieving specific code patterns. Yet, because modern corpora are typically published as file pointers and basic attributes rather than source text, no general structural layer exists to select which files to download: curating a subset requires downloading and parsing millions of files first just to discard most of them. We present SWHForest, an open-source framework and public dataset that provides this missing structural layer for any pointer-based corpus without redistributing source code. A single pass over the archive parses each file into an AST and immediately discards the source text, keeping only its structural summary. Queried locally or exposed by archive providers as a server-side filtering API, SWHForest selects matching files first (via SQL or index intersection) and downloads only surviving candidates. Across the ten largest languages of Stack-Edu (52.1M unique files, 341M definitions), SWHForest answers curation queries in milliseconds while cutting downloads by 46-72%, reducing live structural searches from hours of per-language streaming to seconds or minutes.

Evaluating general-purpose multimodal large language models (MLLMs) for surgical scene understanding remains limited due to scarce surgical data and fragmented evaluation protocols. We address these challenges through a two-stage evaluation framework for surgical scene understanding. In Stage I, we introduce SurgKnowBench, a diagnostic benchmark that harmonizes public surgical datasets across procedures and tasks and evaluates proprietary and open-source MLLMs to map the boundary of pretrained surgical competence. SurgKnowBench reveals that current models retain partial visual priors but remain weak on relational and procedural understanding and workflow assessment. Building on this failure frontier and motivated by the scaffolded nature of surgical training, we introduce SurgAgentGym, an agentic environment where general-purpose MLLM agents can use bounded knowledge search over SurgWiki and general-purpose visual grounding for surgical tasks. SurgAgentGym evaluates whether generalist agents can improve surgical understanding at inference time through tool-mediated evidence use, without additional surgical training. Our results show that agentic inference narrows the gap mainly for perception-heavy tasks, while fine-grained localization and expert-style assessment remain persistent bottlenecks. This opens a path toward future studies of agentic surgical understanding.

Wireless foundation models learn reusable representations across tasks, configurations, and deployment environments. Their pretraining requires environmental breadth, yet wireless datasets often sample few sites densely. We introduce Channels4World, a scene-centered dataset of 100,000 distinct outdoor RF scenes and a procedural method for generating them. Each scene links a generated environment and RF material assignments to 32 base-station (BS) positions, 16,384 UE positions, and aligned variable-length multipath and ordered interaction records at 3.5 and 7~GHz. The corpus comprises 52.43 billion physical BS-UE pairs, 104.86 billion frequency-specific evaluations, and a 29.14-TB packaged payload. Scene identity aligns geometry, placements, channels, and interactions across frequencies, allowing environmental breadth and within-scene observation density to vary independently. The public release offers predefined subsets at multiple scales and separate access to corpus components. A deterministic 90K/10K scene-level train-test split keeps both frequencies and all scene components together. We characterize morphology, RF materials, and transceiver placement across three generation profiles. Controlled propagation and path-retention analyses show that diffraction materially affects retained-link availability and multipath structure, while the retention policy reduces storage substantially with little loss in represented aggregate power. Under matched training exposure, models trained on broader scene collections perform better on held-out scenes, although an equal-episode control shows that scene count alone cannot explain the gain. By making environmental breadth explicit and controllable, Channels4World supports the development and evaluation of wireless foundation models for generalization across deployment environments.

Recent zero-shot image compression methods exploit the powerful priors of pretrained diffusion models to produce perceptually pleasing reconstructions without compression-specific training. However, image compression serves not only human viewing but also machine analysis, while existing zero-shot codecs have limited flexibility in meeting the diverse reconstruction requirements of these applications. In this paper, we propose Flex-DDCM, a framework for zero-shot task-oriented image compression via diffusion trajectory guidance. Building on Turbo-DDCM, we adapt the coded diffusion trajectory through task-guided noise selection, embedding the effects of task guidance into the compressed representation. Guidance terms can be applied individually or combined, with timestep-dependent strengths controlling their influence throughout denoising. This design enables task adaptation without introducing additional decoding complexity. Experiments demonstrate flexible trade-offs among pixel fidelity, perceptual quality, and downstream task performance, with competitive performance against trained task-oriented codecs and faster decoding than Diff-ICMH.

A legal answer is only as good as the article it rests on. Vietnam's largest open legal corpus pairs 109,543 published questions and answers with 573,851 legal documents, yet its citations are free text: "Khoan 3 Dieu 101 cua Bo luat nay" ("Clause 3, Article 101 of this Code") names neither the code nor its version. Following the five stages of NVIDIA Auto Ontology, Discover, Ground, Validate, Expose and Learn, we turn four open datasets into a knowledge graph of 11.5 M nodes and 16.0 M edges in which every node and edge carries its source record and method, with no duplicate keys, dangling edges or schema violations. Of 67,868 cited provisions, 61,487 resolve to a dated law version, 51,694 to a parsed article, clause or point, and 19,012 are confirmed word for word. The ontology then shows what flat records hide. 36% of the 78,361 answers whose citations resolve now cite a law that is no longer in force; 2,525 verified citations point to an article number that says something else in the current law; 2,520 citations were re-versioned because their words match a different version than their date suggests; and 5,575 document numbers are shared by several local authorities, one of them by 72. Every number in this paper is computed from the graph.

We study offline policy optimization in finite-horizon Markov decision processes where every deterministic Markov policy has an action-value function linear in known features, with coefficient norm at most $B$. From independent complete trajectories of one unknown Markov behavior policy, and under all-policy occupancy coverage, we give an algorithm with $\widetilde O(\varepsilon^{-2})$ sample dependence and computational complexity polynomial in $\log(2+B)$. The algorithm performs exactly one convex regression per layer. Its stopped-return targets are the values of a single consistent Markov witness policy. A shared reference design and bounded prediction constraints convert fast clipped-regression error into pointwise confidence for the raw linear predictions. Truncated leverage controls the performance loss without requiring empirical covariance approximation at scale $B^{-2}$. Rounding confidence radii upward preserves the exact witness and yields a polynomial-time bit implementation for rational numerical inputs. The guarantees allow any fixed initial-state distribution; realizability and coverage need only be checked for deterministic policies. We also establish a covariance-coverage variant with uniform approximate realizability and polynomial dependence on $B$. A two-action lower bound shows that the accuracy exponent $2$ is necessary.

Humans engage in pragmatic reasoning when encountering mentions of highly predictable events. Such mentions may appear redundant, leading comprehenders to infer that the event is atypical, and accommodate this inference with an explanation for why it is atypical. However, these inferences are not drawn uniformly across individuals, and experimental items vary in inference rates due to differences in event typicality. We show that these patterns can be predicted with a resource-rational cognitive model in which comprehenders select interpretation strategies based on individual exploration tendencies and item-level predispositions. These parameters are estimated independently from participants' Raven's Matrices score, event typicality rating and language model based accommodation likelihood. Taken together with recent findings from reference games, our results motivate a new perspective on the mechanisms that support pragmatic inferences.

Earth observation foundation models produce global metre-resolution embeddings that underpin work across agriculture, ecology, hydrology, and beyond. Yet these embeddings are opaque: each dimension entangles unrelated physical concepts, which makes them difficult to interpret and their robustness hard to assess. Here, we propose sparse autoencoders (SAEs) as a principled way to decompose Earth embeddings into sparse, approximately monosemantic features over an overcomplete dictionary. Because LLM judges of interpretations do not transfer to satellite pixels, and spatial autocorrelation matters, we introduce a tailored evaluation protocol. It scores any SAE on reconstruction at matched sparsity, alignment with reference concepts, and spatial scale, each against baselines that include an SAE whose decoder is a frozen random dictionary. On this protocol, SAEs reconstruct AlphaEarth and Tessera embeddings better than every baseline at matched sparsity. Their features align with far fewer reference concepts than raw embedding dimensions (on AlphaEarth, 88% align with at most one of 20 concept groups, most with none, against a median of 12 per raw dimension). They also carry more of their variance at fine spatial scales (23 to 32% within a 640 m patch for BatchTopK and Matryoshka, against 9%). An SAE with a frozen random decoder shares both properties, which therefore follow from sparsity, whereas only a learned dictionary reconstructs the embedding accurately: on AlphaEarth, trained SAEs leave eight times less variance unexplained. We test the sparse code on two applications, discovering land-cover sub-concepts and detecting out-of-distribution pixels by area-of-applicability estimation. Both indicate how sparse features could be used, but neither is yet reliable enough for practical use. Together, these results are a step toward Earth observation models whose features can be inspected.

Dense retrievers are meant to select passages that match a query's semantic specification. We show they suffer from search-by-answer bias: they favor passages that mention answer objects the model knows (scores as likely), over passages that support valid alternative answers. We test this with controlled factual contrasts from LAMA/TREx. For each retriever, we measure how strongly the model scores the canonical object over an alternative object using a token-prediction head, and ask whether this score predicts passage rankings. Across eight retrievers, including DPR, Spider variants, Contriever, E5, and Qwen3-Embedding, larger token-prediction scores consistently predict higher retrieval scores for passages containing the corresponding object. This holds when passages differ only in the answer object, and even when the passage that mentions the high-scoring object expresses a wrong relation. Dense retrieval thus exhibits search-by-answer bias during passage selection, shaping which evidence is retrieved.

Single-loop minimax methods are appealing for modern machine learning, but standard stochastic gradient updates can be unstable and oscillatory under tightly coupled primal-dual dynamics, making performance notoriously sensitive to stepsizes. To fundamentally improve stability, we develop an orthogonal update framework tailored to minimax optimization, which applies updates along orthogonalized directions to reduce update anisotropy. However, orthogonalization reshapes the optimization geometry and makes coordinating primal-dual stepsizes even more delicate. To overcome this challenge, we develop AdaSGDA, which equips orthogonal updates with an adaptive stepsize rule driven by accumulated gradient norms. This mechanism automatically balances coupled primal-dual progress under orthogonalization, yielding an effectively parameter-free method for nonconvex-strongly-concave minimax optimization. We prove that AdaSGDA avoids problem-dependent tuning while attaining the state-of-the-art $\mathcal{O}\left(T^{-1/4}\right)$ convergence rate. We further propose AdaSGDA-VR, which incorporates a carefully designed variance reduction scheme and achieves the faster $\mathcal{O}\left(T^{-1/3}\right)$ rate. Extensive experiments demonstrate that our methods deliver competitive performance compared to existing minimax optimization methods.

Cross-lingual concept analysis requires distinguishing the information that a subspace retains from the classifiers and erasers that transfer through it. We study these operations in six frozen encoders using 10,359 Bangla (Bengali) idiom senses, eight input conditions, and figurative-class and sentiment labels. Class-mean directions provide the most consistent evidence of shared coarse meaning: in post hoc analyses, they outperform random-subspace, permuted-label and covariance-matched nulls in all 12 model-label tests under balanced accuracy. A prospective internal re-split supports idiom-alone sharing in 10 of 12 tests, spanning five encoders. The original confirmatory logistic-regression family rejects both planned nulls in 4 of 12 tests, where English directions gain only a few accuracy points above the majority-class floor (3.0-7.3). Classifier transfer depends on the input map and regularization: English-trained class-mean readouts beat both zero-shot nulls in 8 of 12 discovery tests, and stronger source-selected penalties improve transfer through logistic and LEACE directions. These readouts use Bangla-validation-selected layers and unlabeled target preprocessing. Input controls show greater coarse-class sharing for explicit glosses than for idiom counterparts. Reading, however, does not imply effective erasure: English-fitted LEACE removes less than one percentage point of Bangla figurative-class accuracy, compared with 4-11 points for Bangla-fitted LEACE; INLP shows a similar asymmetry. Pattern-filter analysis, covariance-weighted input maps and a binary Gaussian overlap bound explain how these operations can diverge. The analogous multiclass ordering is an empirical diagnostic, observed in 11 of 12 cells. The results support an estimator-sensitive account of cross-lingual concept sharing and motivate evaluating reading, classifier transfer, erasure and overlap separately.

Domain adaptation or DA methods are indispensable when target labels or computational resources are scarce, yet the domain kernel density and subspace alignment approaches completely assume a Hilbertian geometry. They measure distributional discrepancy with symmetric L2 divergences over Euclidean feature spaces or reproducing kernel Hilbert spaces (RKHS). This preprint summarizes the manuscript Non-Hilbertian Domain Adaptation. Total Variation Alignment via L1 Density Regularization on Feature Manifolds, which proposes Banach Manifold Transfer Learning or BMTL. BMTL makes two coupled design decisions. The first one is that features are embedded on a geodesic manifold rather than treated as points of a flat Euclidean space, and the second one is that the density matching regularizer is formulated in an L1 Banach norm, whose minimization is equivalent, by Scheffe's theorem, to contracting the Total Variation or TV distance between domain densities, a bounded metric that weights all regions of the sample space uniformly. The L2 formulation, by contrast, squares deviations and therefore amplifies errors in the tails, over emphasizing density outliers that do not reflect meaningful distributional shift. Experiments on MNIST-USPS, COIL20, and Office+Caltech10 show consistent gains; the L1 variant beats an otherwise identical L2 ablation on every task, for example 92.44% vs. 89.80% on USPS->MNIST and surpasses the strongest baseline DMDA by +3.68% on average over twelve Office+Caltech transfers. This document situates the contribution with an explicit comparison of Euclidean, Hilbert, and Banach geometries.

Tabular foundation models based on prior-data fitted networks (PFNs) achieve strong predictive performance, but their value for causal inference is unclear. We evaluated whether PFNs, which predict through in-context learning, could match or outperform conventional nuisance learners in supporting accurate causal estimation and reliable uncertainty quantification (inference). Learners modeled propensity scores and outcome regressions within doubly robust estimators of average causal effects, with and without cross-fitting, across synthetic and semi-synthetic settings with varying confounding complexity and effect heterogeneity. Under simple parametric data-generating processes, the PFNs were competitive with correctly specified parametric regressions and ensembles containing them. In more complex settings, they outperformed all conventional learners evaluated, including richly specified Super Learner ensembles, for both accuracy and uncertainty quantification. As sample size increased, only the two PFN models (TabPFN and Mitra) consistently approached nominal confidence interval coverage, whereas under-coverage persisted for conventional learners. In some settings, PFNs achieved near-nominal coverage and well-calibrated standard errors even without cross-fitting. At the smallest sample size (n=200) and in the higher-dimensional semi-synthetic settings, PFNs and other flexible learners exhibited poor coverage without cross-fitting. For PFNs, this appeared to be driven by propensity score overfitting consistent with in-sample memorization of treatment labels. These findings highlight the promise of prior-data fitted networks as in-context nuisance learners for causal estimation, while underscoring the need for further evaluation in challenging settings.

Diffusion models have demonstrated the effectiveness of score-based objectives for generative modeling, yet their potential for self-supervised representation learning remains underexplored. We introduce Entropy-Based Regularization (EBReg), a framework that uses latent score estimation to regularize the latent distribution during self-supervised training. EBReg jointly learns an encoder and a latent score model, and utilizes the learned score to provide a diversity-preserving signal that prevents representational collapse. Motivated by a mutual-information lower bound, we formulate the objective as view alignment with smoothed marginal entropy maximization and optimize the marginal entropy without explicit density evaluation. The same score model can additionally be used at test time to refine corrupted representations toward the training latent distribution. Empirically, EBReg learns competitive representations on ImageNet-1K, achieving 75.57% linear-probe accuracy with ViT-L. Across diverse architectures, score-guided test-time adaptation further improves robustness on ImageNet-C, yielding an average relative accuracy gain of approximately 5%.

User plane function (UPF) selection in 5G networks is an active area of research, with various limited approaches proposed to enhance network efficiency and manage resources effectively. In this paper, we address this gap in the related literature by proposing a Deep Reinforcement Learning (DRL) UPF selection method that utilizes a dynamic, multi-objective reward function and validating it on a real-world 5G emulation platform. The proposed method is based on the large language models' (LLMs) translation of network operator intents and the network data analytics function (NWDAF) abnormality alerts. The experimental results validated in the real-world experimental setup demonstrate that our dynamic DRL solution achieves an effective balance between latency, power consumption, and traffic distribution, outperforming the heuristic baselines and fixed DRL by 16.89%. These findings highlight the real-world applicability of our approach for dynamic, intent-driven UPF selection.

Block-diffusion drafters propose a whole block of tokens in one forward pass, and draft trees built from their position-wise marginals increase the expected accepted length of speculative decoding. Every tree node occupies a query row in target verification, so under concurrent serving low-yield branches compete with other requests for verification capacity. Existing tree methods select verified nodes by draft probability under a fixed or cost-selected per-request budget, and thus always retain a prefix of the drafter's ranking. We present D3Tree, which learns which nodes of a block-diffusion draft tree are worth verifying. A tree-size head with 2H+2 parameters predicts each candidate's conditional survival from the draft state, the candidate embedding, and its draft log-probability. Thresholding the product of these scores along parent chains yields an ancestor-closed subtree whose size adapts to each request and decoding round. The head is trained on a frozen drafter with conditional labels on the full candidate tree; an optional joint mode also steers the drafter toward the retained reference path. At serving time, D3Tree verifies whole trees under light load and, under heavy load, allocates one batch-wide node budget by predicted survival, sized to the target's matrix-multiplication tiles. Retained trees share one ancestor-masked target pass, while construction, pruning, and scheduling run on the GPU and overlap with the host. Across six vision-language verifiers and 17 text and multimodal benchmarks, D3Tree is fastest in 70 of 72 load settings, with 3.2-4.0x speedup over autoregressive decoding for one request and 1.3-1.8x at 64 concurrent requests, where fixed-budget trees fall below autoregressive speed. Under a time-varying arrival trace, it has the lowest mean end-to-end latency in every arrival phase.

Large language models (LLMs) have made substantial progress in mathematical reasoning but still produce incorrect solutions to complex problems, indicating the need for more reliable self-correction mechanisms. Mathematical self-correction typically identifies errors through verification, which can proceed backward or forward. The two directions provide complementary verification capabilities. Backward verification checks a candidate answer against the original constraints, bypassing complex intermediate reasoning, but cannot ensure correct reasoning or objective completeness. Forward verification follows the problem conditions and current solution, enabling it to locate reasoning errors and assess objective completeness, but may overlook key constraints in complex solution chains. We propose BiVerify, a dynamically scheduled framework for bidirectional mathematical verification. BiVerify begins with backward verification to screen the answer against the original constraints, followed by forward verification of the reasoning process; subsequent direction switching, state resetting, and re-verification are governed by verification outcomes. Upon failure, the system revises the candidate or verification flow until the same candidate passes both directions. We also introduce shared mathematical tools that separate logical reasoning from precise computation. We systematically evaluate BiVerify on five challenging mathematical reasoning benchmarks using multiple mainstream DeepSeek, GPT, and Gemini models. BiVerify exceeds the strongest representative public baseline evaluated in this work by 13.5 percentage points in average accuracy, with the margin reaching 31.2 percentage points on GPT. Further analysis shows that BiVerify's overall performance exceeds that of either verification direction alone.

Diffusion Transformer (DiT) models, while having become a powerful backbone for generative synthesis, suffer from high inference cost on generating high-resolution images. Diffusion caching has emerged as a promising acceleration technique to reduce inference cost. However, existing caching strategies fail to accurately identify the varying importance of different computational modules and rarely coordinate cache reusing across the timestep and layer granularities, making it difficult to control error accumulation. To address these limitations, we propose multi-granularity caching (MG-Cache) for DiTs, a novel and fully training-free approach to exploit computational redundancy both on timesteps and Transformer layers. Our core innovation contains a two-stage caching strategy: an offline timestep profiling analysis that identifies redundant timesteps for cache reuse, followed by an adaptive layer-wise caching strategy that combines input stability with importance of each Transformer layer to selectively cache stable Attention and MLP submodules. Extensive experiments on mainstream open-source DiT models validate the effectiveness of our approach, achieving significant acceleration while maintaining high fidelity to the uncached output, with up to a 4.18x acceleration on HunyuanVideo.

Offloading model state to host memory and streaming layer weights let a single multi-GPU node train large mixture-of-experts (MoE) models. GPU memory then becomes a performance budget, shared by resident expert state, retained activations, and execution worksets. The best allocation depends on the expert execution path, batch size, and sequence length. We present OOM (Offload-Optimized MoE), a full-parameter MoE training system that co-designs expert execution and memory allocation. For expert-parallel execution, OOM updates resident experts on their owning GPUs and jointly plans state residency and MoE activation retention according to pipeline dependencies. For local execution, OOM bounds GPU worksets through dependency-aware partitioning; in hybrid models, it also retains attention results on the host to avoid costly recomputation and selectively regenerates layer boundaries through cheaper layers to make room. Calibrated cost models guide joint configuration within GPU and host budgets to minimize step time; OOM selects among the configured paths through short executions. On four H100s, OOM trains with 1M-token contexts, whereas FSDP2, ZeRO-Offload, and MegaTrain exhaust GPU memory at 512K. It also trains a 122B-parameter MoE model, on which all three exhaust host memory. On four 24 GB RTX 4090s, it trains 512K-token contexts and 1,024 1K-token sequences per step without gradient accumulation. Across matched-global-batch comparisons, OOM delivers 1.58x and 4.08x geometric-mean (up to 2.14x and 7.77x) throughput over CPU-offloaded FSDP2 on the two platforms while using the least host memory among the evaluated systems -- as little as 1/3.0 of FSDP2's.

The rapid growth of scientific literature and the high heterogeneity of biological terminology and mechanistic descriptions make it difficult for researchers to integrate scattered research evidence, identify connections among biological events, and construct mechanistic networks. Here, we developed a high-throughput human-AI hybrid framework, AOP-Smart-Finder, for biological event discovery, relationship mining, and Adverse Outcome Pathway (AOP) construction and evidence mapping from large-scale literature. This framework integrates Large Language Models (LLMs), AOP-Smart retrieval-augmented generation (RAG) technology, and Python-based automation, supporting two analysis modes: one based on a single compound and the other based on a predefined set of biological events. In the di(2-ethylhexyl) phthalate (DEHP) case, the system processed 12,983 articles within 30 min and included 1,093 of them, from which it extracted 1,589 biological events (including 458 official key events [KEs] and 1,131 candidate KEs) and 2,519 key event relationships (KERs). In the amyotrophic lateral sclerosis (ALS) case, the system identified 55 stressors from 437 articles and extracted 278 biological events and 548 event relationships. Following manual curation and Weight of Evidence (WoE) evaluation, a mechanistic network comprising 16 KEs and 23 KERs was ultimately constructed. This framework can either use existing AOPs as knowledge guidance for mechanistic expansion and evidence supplementation, or discover novel event connections and construct mechanistic networks around user-specified research topics. It provides a high-throughput, traceable, human-AI hybrid methodology for complex biological mechanism elucidation and AOP discovery driven by large-scale literature.

Standard Operating Procedures (SOPs) are kept concise for human readability, so LLM agents that generate executable skills from them lack the tacit knowledge and implementation details needed for automation. Human intervention fills these gaps, but must be repeated as SOPs evolve---a problem we call \textit{SOP-skill drift}. We propose managing SOPs and Contextual Knowledge (CK) separately, together with a \textit{contract-aware} evaluation that distinguishes genuine violations from CK-resolved differences. On eight Indeed-derived workflows (single LLM), it raises the functional-equivalence rate from 12\% to 58\%. Under three SOP-revision perturbations, previously extracted CK already covers 79.5\% of newly introduced automation gaps.

Spatial reasoning requires understanding how objects and observers are arranged in a scene and how their relations change with viewpoint or movement. Vision-language models (VLMs) typically perform both steps through model-internal reasoning, forming an implicit scene interpretation from partial views and imagining the consequences of spatial transformations. A misread scene can lead to a wrong answer, while even a correct interpretation does not ensure that imagined transformations satisfy spatial constraints. We introduce Spatial Harness, which represents scenes as explicit and executable world states to ground spatial reasoning in physics-engine feedback. A Perceiver integrates scene constraints extracted from the question with specialized visual estimates to place objects and cameras in a shared coordinate frame. The resulting state is instantiated in a physics engine, where a Reasoner executes viewpoint changes and uses computed measurements and rendered views to guide reasoning. We evaluate on MindCube, SAT, MMSI-Bench, and SpatialRGPT-Bench with open- and closed-weight VLMs. On MindCube, Spatial Harness improves Qwen3.5-27B from 52.70% to 71.14%, surpassing pySpatial by 7.52 percentage points. Given identical annotated object and camera poses, it reaches 98.25% versus 71.81% for model-internal reasoning. Code is publicly available at: https://github.com/tmlr-group/Spatial-Harness

We evaluate whether an AI agent's harness, the notes around its model, can learn from the agent's failure traces beyond a matched search that never sees them. It can, but less than it looks. We tested an open 9-billion-parameter model with fixed weights; a larger teacher model wrote its notes. On generated customer-service decisions, where both note writers were told the target rules, notes based on the agent's mistakes improved accuracy by 4.0 percentage points over notes from the same process with the mistakes withheld. That process alone, trying and keeping better-scoring notes, reached about 60% of the total improvement. In full customer-service conversations the same notes showed no demonstrated improvement. Separately, a strict screen of single lessons admitted two customer-service lessons, still awaiting tests on new data, and one tool-use lesson, written without seeing the agent's mistakes, which showed no demonstrated gain on new tool-use tasks. A neutral note with no procedure scored higher than the learned lessons (exploratory). A blind audit judged 146 of 240 tool-use test items defective.

Security tools raise findings across cloud environments, and AI agents now investigate and prioritize them. A finding is still resolved by people: it has to reach the people responsible for fixing it and accountable for the resource it concerns. Ownership is the organizational context that decides who that is. Where recorded, it is declared by hand in configuration management databases or owner tags and goes stale unless kept current; accounts are shared, and automation creates much of the infrastructure. The knowledge sits in work records, code history, audit logs and identity directories, encoded in each organization's own conventions. Recovering it poses four challenges: the signal is scattered across sources, none of them authoritative; the conventions that carry it differ between organizations; reaching an owner takes multi-hop links; and those related to a resource differ in the strength of their relation. We define ownership as a set of entities, people or teams, each related to the resource by one of five patterns ordered by strength. We present a reference architecture for agentic ownership attribution that derives ownership from these records instead of declaring it: it profiles an organization offline into tested signal procedures and an ownership playbook, then attributes each resource with a coordinator agent and parallel investigators, with confidence computed by rules. On 100 resources of 50 Amazon Web Services (AWS) resource types in one organization, we pool the architecture's candidates with those of an unprimed general-purpose agent and verify every evidence piece against the records. The architecture reaches 92.1% weighted recall and 86.6% weighted precision, against 54.8% and 77.4%, at 0.55 times the agent's cost per unit of coverage. The difference lies in people who build and change resources without an explicit owner, found through work records and code history.

We study online reinforcement learning for the stochastic shortest path (SSP) problem with unknown transition dynamics. Although existing posterior-sampling methods perform well empirically on SSPs, their regret guarantees do not achieve the minimax-optimal order. We address this limitation by proposing OPSRL-SSP, an optimistic posterior-sampling algorithm for the SSP setting. The algorithm operates in epochs and uses a logarithmic, time-dependent schedule for the number of posterior samples. To induce optimism, we augment the transition model with a pseudo-state assigned an optimistic boundary value. We establish a high-probability regret bound of $\tilde{O}(B_\star \sqrt{SAK} + B_\star S^2 A)$, where $B_\star$ is an upper bound on the expected cost of the optimal policy, $S$ and $A$ are the numbers of states and actions, respectively, and $K$ is the number of episodes. In addition, we extend the analysis to general zero-cost SSPs using a universal cost-perturbation argument. The leading term matches the minimax lower bound $Ω(B_\star \sqrt{SAK})$, thereby addressing the open problem raised by Jafarnia-Jahromi et al. [19] for posterior sampling in the SSP setting.

Modern state-space models owe much of their computational efficiency to their linear, often diagonal, dynamics. Classical recurrent neural networks instead allow nonlinear state-to-state interactions, which can be beneficial when memory updates need to depend on the past context. In this paper, we ask whether state feedback through a simple gate regulating memory updates can enhance the expressivity of diagonal state-space models. To investigate this question, we compare gates that depend on the current input, the recurrent state, or both. We establish a strict expressivity separation: for scalar models, state gating solves induction-head tasks for arbitrary vocabulary size and sequence length, whereas input gating fails on a finite instance under standard stability assumptions on the state transition matrix. Experiments show that state gating retains its advantage on induction-head tasks in higher-dimensional models, whereas input gating performs better on selective copying tasks. Together, our results show that even a simple, diagonal form of state-feedback gating can unlock capabilities unavailable to input-controlled models, while input and state gating offer complementary strengths and achieve the strongest overall performance when combined.

3D visual grounding (3DVG) localizes language-referred objects in 3D scenes, with broad applicability in both indoor and outdoor environments. However, existing 3DVG research focuses on static indoor objects, limiting its use in dynamic outdoor autonomous driving. Motion descriptions, which capture the temporal dynamics of moving objects, offer a promising way to address this challenge. We therefore introduce MoT3DVG, a large-scale dataset for dynamic-aware outdoor 3DVG with temporally evolving motion descriptions. It contains 850 scenarios with 31,128 frames from nuScenes dataset, and provides 144,568 language prompts with motion-aware descriptions for dynamic objects across time. We further propose DynaVG, a novel framework that effectively leverages temporal cues for outdoor 3DVG. In addition to the standard modality-specific encoders, DynaVG introduces a Short-Long Integrated Dynamic Encoding (SLIDE) module before the language-point cloud alignment. SLIDE models both local motion cues and global temporal context via short-term window encoding and long-term window shifting. Extensive experiments demonstrate its state-of-the-art performance on MoT3DVG, validating the effectiveness of motion-aware encoding. Our work reveals open challenges and promising directions for future research in outdoor 3DVG.

Risk-sensitive optimal stopping is useful when poor stopping outcomes carry substantial costs, as in option exercise, sequential medical treatment, and safety-critical control. A natural objective is Conditional Value at Risk (CVaR) of the stopped reward. We develop a theoretical framework and algorithm for CVaR deep optimal stopping. The framework and algorithm address four difficulties in applying risk-sensitive reinforcement learning to optimal stopping: state augmentation, score-function pathology, blindness to success, and dependence on auxiliary networks. We isolate three structural properties of optimal stopping: single-point reward, action-independent dynamics, and sum-zero gradient. We exploit these properties to establish Markov optimality without state augmentation and derive a pathwise gradient estimator. Under bounded policy derivatives, the pathwise estimator has bounded second moments near deterministic policies, avoiding the divergence that can occur with score-function estimators. Building on this framework, we propose DIOS (Distribution-Informed Optimal Stopping) to address the remaining two difficulties. A per-time-step soft penalty combined with tail-level annealing mitigates blindness to success. Estimating the reward quantile directly from each batch removes the need for an auxiliary critic. For a fixed tail level, we establish a finite-time stationarity guarantee under regularity assumptions along the training iterates. Across the 24 main Bermudan max-call configurations, DIOS achieves the highest mean CVaR and the shortest recorded training time among the tested CVaR baselines.

Agentic reinforcement learning with verifiable rewards trains language models through multi-turn tool interactions, yet their reasoning performance and tool-calling ability can decline during training. We trace this degradation to three coupled mechanisms: (1) tool-token imbalance produces negative overall tool credit when incorrect trajectories contain more tool tokens per turn than correct ones; (2) aggregation amplification can further strengthen this negative signal as token-mean aggregation gives longer incorrect trajectories greater weight in policy updates; and (3) failure-driven retries prolong trajectories as models tend to call tools again after execution errors, extending potential negative credit to recovery attempts. To counter this suppression, we propose Transition-Aware Credit Assignment (TACA), which targets the transition from reasoning to tool use through supervision of both call initiation and generation. Specifically, TACA uses Advantage Projection to enforce a nonnegative mean advantage on tool-call tokens within each prompt group through a minimal uniform shift. Additionally, TACA guides tool-use decisions with Decision Contrast, comparing rewards for tool-assisted and direct answers to the same prompt. Through Supervised Replay, it reinforces tool-call generation using calls from previously sampled correct trajectories. Across models and tasks, TACA improves performance and sustains more reliable tool interaction during training, raising mean accuracy over MT-GRPO on six reasoning benchmarks by 4.61 percentage points for Qwen2.5-1.5B and 3.61 for Qwen2.5-3B. Our code is available at https://github.com/tmlr-group/TACA.

Sensor outages can conceal prediction-interval coverage deficits behind acceptable marginal performance, while unavailable future readings leave some issued forecasts unscorable. We present a measurement-record audit that links issue-time outage duration, original target availability, and time-uniform statistical evidence. The audit retains complete forecast keys, batches all sensors and overlapping horizons at each measurement tick, and projects simultaneous covered-count and scored-count bounds onto a cumulative conditional expected coverage ratio. This target is distinguished from empirical coverage and outage-conditional population coverage. Selected traffic and air-quality results illustrate the gap between pooled and long-outage coverage; in METR-LA, only 7.6% of issued long-gap forecasts have originally observed targets. Of 30 primary sequential outage-state audits, 26 are inconclusive and four lack scored support. This concise presentation gives the core construction and selected aggregate evidence without the full experimental supplement.

Long-context applications, long-lived agents, and multi-agent systems require large language models (LLMs) to process and retain extensive information. In these settings, the KV cache can exceed GPU high-bandwidth memory (HBM) capacity, necessitating offloading to host memory, where CPU-to-GPU transfers can dominate inference latency. While KV cache compression can reduce data movement, query-agnostic compression that guarantees accurate attention for arbitrary queries is extremely challenging, if not impossible. We introduce VerKV, a hybrid approach that provides $(ε, δ)$ guarantees for approximating attention on any given query. VerKV maintains a compact set of query-agnostic pivotal tokens on the GPU, together with a token-level sketch of the full KV cache, while retaining the complete cache in CPU memory. We introduce a novel stochastic sampling procedure, PoissonComp, to select the pivotal tokens. At inference time, our proposed KVFetch algorithm uses the pivotal tokens and the sketch to estimate the relative error of the resulting attention computation. If the estimate satisfies a user-specified $(ε, δ)$ tolerance, inference proceeds without accessing host memory. Otherwise, VerKV performs query-aware retrieval to fetch additional tokens from the CPU, refining the attention computation to meet the required tolerance. Thus, VerKV enables attention with bounded approximation error while minimizing CPU-to-GPU transfers. We show that PoissonComp outperforms state-of-the-art approaches to KV cache compression. Combined with KVFetch, it achieves near-dense performance (90-100% of dense quality at 32K context length, averaged over RULER and LOFT), using Qwen3-4B, Qwen3.5-9B and Ministral-3-8B, while reducing the GPU memory footprint by 8-32x and keeping CPU-to-GPU memory fetches within 2-23% of the cache per generated token, depending on the model.

High-resolution weather prediction is important for resolving local atmospheric patterns missed by coarse global forecasts. Large pretrained weather backbones have improved global forecasting, but adapting them to produce future high-resolution fields from recent coarse states remains challenging and costly. We formulate this setting as Predictive Spatiotemporal Weather Downscaling (PSWD), a task that jointly refines spatial and temporal resolution while predicting future high-resolution trajectories. To this end, we propose a framework that decomposes prediction into spatial and temporal stages using a shared pretrained backbone, fixed numerical scaffolds, and lightweight residual heads. For backbone adaptation, we introduce PROLA (PRincipal-Orthogonal Low-rank Adaptation), which allocates a fixed low-rank budget between pretrained weights' principal directions and their orthogonal complement. PROLA further rescales rank-1 optimizer updates using gradient signals, improving adaptation without increasing trainable rank. Experiments in the PSWD setting show that PROLA outperforms representative low-rank adaptation baselines. Further experiments across diverse datasets and tasks support the broader applicability of PROLA, showing that its benefits extend beyond the primary weather downscaling setting.

The June 2025 outage in the Global Database of Events, Language, and Tone (GDELT) collection pipeline left a gap in the Global Knowledge Graph (GKG), interrupting article-level news metadata used in downstream research. ReGKG-2025 is an archive-derived sample of 281,620 reconstructed article records covering all 1,665 missing 15-minute batches. Articles were recovered from Common Crawl and the Wayback Machine using source quotas estimated from surrounding native GKG data. Text was screened and normalized, non-English bodies were translated, and flagged records were manually reviewed. Eligible articles were converted to the 27-column GKG 2.1 layout. The versioned package includes record-level provenance, archive replay and WARC locators, source summaries, source-by-day calibration weights, and the complete manual decision journal. The sample reaches about 14% of modeled native volume and represents sources accounting for 84% of selected reference record mass. In an intact-week test, 41% of recovered articles matched native GKG records; field agreement on matched pairs is close to the processor baseline (tone MAE 1.69). The dataset supports retrospective content analysis and source-level aggregates; its tone and theme distributions differ from native GKG, and it does not support exhaustive counts. Data: https://doi.org/10.5281/zenodo.23964265.

Quadratic manifold fitting provides a flexible framework for modeling nonlinear low-dimensional structures embedded in high-dimensional spaces. However, existing methods typically employ prescribed noise models or fixed loss functions, limiting their adaptability to varying noise characteristics. To address this limitation, we propose an alternating maximum likelihood estimation method for subspace-constrained quadratic matrix factorization, termed AMLE-SQMF, which jointly estimates the quadratic manifold and the underlying noise distribution. Specifically, we model the residuals using a generalized Gaussian distribution (GGD), whose shape parameter controls the tail behavior and determines the associated fitting loss. By exploiting the closed-form maximum likelihood estimator of the scale parameter, we eliminate the scale variable and derive a profile negative log-likelihood. Based on this formulation, we develop an alternating optimization algorithm that iteratively updates the manifold parameters and estimates the noise shape parameter. This enables adaptive loss selection within a unified likelihood-based framework, without requiring prior specification of the noise characteristics. Experiments on synthetic and real-world datasets demonstrate the effectiveness of AMLE-SQMF in quadratic manifold fitting and its robustness to diverse noise distributions.

AI agents deployed on the open web are becoming more capable of performing actions that may breach contractual provisions and violate computer security laws. In this paper, we introduce ScrapeBench, the first agentic benchmark for empirically evaluating the legal compliance of AI agents with website scraping policies, namely terms of service (ToS) and robots.txt directives. Across widely used closed-weight and open-weight AI agents, we find that most agents exhibit low compliance rates (< 5%) when instructed to scrape websites in our sample of 1,626 popular websites that explicitly prohibit scraping. In addition, we observe that even where AI agents actively check the applicable ToS and robots.txt, they continue to exhibit low compliance rates. Conversely, in our sample of 269 websites that do not prohibit scraping, we observe that agents sometimes refuse to scrape despite it being lawful, such as Claude Opus 4.6 exhibiting an over-refusal rate of 98.5%. Our human baseline experiments (n = 180 participants) suggest that humans exhibit compliance rates comparable to AI agents but over-refuse at much higher rates.

LLM agents that call tools are exposed to indirect prompt injection: instructions hidden in tool outputs can steer an agent into calls the user never asked for. We present a runtime audit that derives an executable intent graph from the user's objective and the deployment's tool catalog with one LLM call per task, and checks every tool call against it with a deterministic loop that makes no LLM call (median 5.7 microseconds per call). Two further mechanisms recover utility. Lookups of the user's own state bypass the audit, because the harm of an attack lies in its write or outward step. The calls that remain flagged are escalated to a reasoning guardian, in the spirit of tool-integrated reasoning models such as START: a reasoning model that sees only the objective, the plan, the call history without tool outputs, and the flagged call, and can verify facts with a restricted Python-expression tool. A current agent showed no baseline attack success, so end-to-end validation remains open.

NCCL picks an algorithm-protocol pair for every collective from an analytic cost model (AUTO) that never sees a measured latency. The model is a strong prior but a brittle final decision: priced on forced-arm runs, its choice is exact on most contexts and badly wrong in a few. We built NCCL-BOT, a tuner plugin that closes the loop: a per-context bandit, warm-started from the model, learns from measured latencies. We replay a forced-arm H100 sweep, tuned at 4-8 nodes and scored at 16-32, and price the model's choice on the same forced arms; that sweep has no stock-NCCL run. The bandit cuts the contexts where the model's choice loses more than 10% from 19 to 1 (p99 103% to 9.9%, mean 6.1% to 1.6%; p=0.047 over nine scenarios after Holm correction). The prior and continued learning both matter: cold-started bandits, and committing after one probe of each arm, lose more on the mean in every scenario. Supervised selectors fit once on the same data and tuned with the same budget match the mean, and their heavier tail is not significant (7 against 1 contexts above 10%, p=0.06); at seen scales they match or beat online selection. With hyperparameters frozen from H100, UCB1 cuts AUTO's p99 from 60% to 23% and from 56% to 15% on two GB200 fabrics, but its contexts above 10% go from 16 to 8 on one and from 10 to 13 on the other, and scenario-level tests do not separate the two. Live, on two latency-emulated node sets of one H100 system with one run per policy, the learners repair most of a gross AUTO error on one (selection loss 31% to 6-8% for deterministic Greedy and UCB1; average all-reduce bandwidth 1.18-1.24x stock) but average 0.94-0.99x stock on the other, and on both lose to stock below 1 MiB (on the first through plugin overhead); on H100 the cost model is exact there offline, so small sizes can be left to it. We do not show a net end-to-end win.

Reusable key-value (KV) memory avoids repeatedly prefilling retrieved text, but stored and injected memories do not necessarily become usable evidence. We present EASE-KV, a controlled framework for understanding this gap under fixed candidates and a frozen language model. EASE-KV separates three interfaces: whether required relations are preserved during encoding, whether preserved memories receive appropriate influence during allocation, and whether outcome supervision can identify useful allocations. Matched experiments show that causal encoding improves compositional accuracy by 25.0 percentage points over block-diagonal encoding, while the tested inference-time bridging methods fail to recover the lost contextual information. With unchanged stored states, diagnostic evidence-guided allocation improves LongMemEval accuracy from 31.2% to 38.4%, demonstrating that memory influence can be controlled without re-encoding. However, controllability does not guarantee learnability: binary correctness provides informative reward contrast for 53.0% of sampled query groups, compared with 73.6% under the evaluated shaped-reward setting. These findings show that effective KV memory requires more than retrieval and injection: the required relations must be encoded, preserved memories must receive appropriate influence, and supervision must distinguish better allocations from worse ones.

The infrastructure definitions and the code for serverless applications are often only loosely connected. This yields a large error class: the referenced infrastructure components might not exist. It is also hard to keep the application and its permissions in sync. To address this problem we present an infrastructure from code solution where we provision cloud infrastructure from Wasm Interface Type (WIT) - the language defining the interface of WebAssembly components. We limit ourselves to the basic example of an AWS Lambda function accessing S3 buckets. We build two interpreters for the WIT language: one that derives the definition of S3 buckets as infrastructure as code and least-privilege IAM permissions, and a generic runtime that runs the component as a Lambda function and provides access to the S3 buckets as capabilities. We formulate consistency properties that ensure the existence of the S3 buckets and correct permissions of the Lambda, and provide evidence that our interpreters satisfy them. We show how selected failures for deployment of such cloud components can be prevented using our technique and how a baseline setup fails at this task. Together with the strict sandboxing of the WebAssembly runtime we achieve an instance of capability-based security for access to AWS infrastructure as a further result.

This paper presents a unified transformer-based framework for object detection and dense region captioning across distributed edge networks. The architecture uses a shared visual encoder followed by detection, region extraction, and autoregressive language-decoding stages, allowing localization and region description to reuse the same image features. This design removes the second image-encoding path required by decoupled detection and captioning pipelines. To support execution on resource-limited devices, a workload-allocation formulation assigns model stages and input streams across connected edge nodes while accounting for execution time, memory capacity, processor utilization, power limits, and intermediate tensor transfer cost. The framework is evaluated on a six-device edge testbed comprising four NVIDIA Jetson Orin Nano and Orin NX inference nodes and two Raspberry Pi 5 auxiliary CPU nodes, with TensorRT FP16 and INT8 execution on the NVIDIA devices. On a Jetson Orin Nano, the unified INT8 model achieves 38.1 FPS with 3.5 GB peak memory use and 11.8 W measured power. Under four concurrent streams, the proposed allocation achieves 35.2 FPS and 28.4 ms mean end-to-end latency with an internode transfer rate of 6.8 MB/s.

This paper presents a unified framework for license plate detection and character recognition on low-power edge hardware. The proposed network uses a shared CSP-Darknet and PANet feature encoder with an anchor-free localization head and a RoIAlign-based sequence-recognition branch. Character decoding is performed directly on aligned feature maps, avoiding image-space plate cropping and a second recognition backbone. The model is trained on an AI-assisted dataset of approximately 37,000 manually verified images covering license plate layouts from all 50 United States. The dataset includes variations in illumination, viewpoint, weather, blur, occlusion, and plate design. The framework is converted to an INT8 TensorRT engine and evaluated on an NVIDIA Jetson Orin Nano with 8 GB memory. The INT8 model shows an mAP50 of 97.8% and a full-sequence accuracy of 90.7%. It records 147 FPS at 6.8 ms end-to-end latency, with 315 MB inference memory and 9.4 W average power consumption. Relative to the YOLOv8s+CRNN baseline, the proposed deployment provides a 7.1x throughput increase and a 77.8% reduction in allocated inference memory

This paper contains examples of stereotypes about veterans and displaced people that may be offensive or distressing. Social bias benchmarks for language models commonly examine categories such as gender, race, and religion, often drawn from lists of legally protected characteristics. Veterans and internally displaced people (IDPs) fall outside these lists, and stereotypes about them, which can undermine access to employment, housing, and social participation, remain largely absent from language model evaluation. We introduce WarBias, a Ukrainian--English benchmark of stereotypes about these two groups. Built from a taxonomy of 120 stereotype claims grounded in research and monitoring reports, WarBias contains 1,007 stereotype--counter-stereotype--unrelated sentence triplets and 240 multiple-choice questions per language. A Ukrainian language model selected in a blinded human pilot generated the Ukrainian triplets. At least two independent annotators reviewed each triplet in the annotation cohort. The authors revised items in response to the reviews, and the final release was machine-translated into English. Each claim carries a human severity rating. In a baseline evaluation of six models, four prefer stereotypical Ukrainian sentences about at least one group, most strongly about IDPs, while English scores are mostly balanced. WarBias supports complementary evaluations of stereotypical preferences and context-sensitive judgments, and provides paired forget/retain material for unlearning-based bias mitigation. We document the benchmark's construction, intended uses, and limitations, and offer its construction method as a template for analogous benchmarks in other conflict-affected societies.

We evaluate AI agents in an online setting where tasks arrive sequentially and, before running each task, we construct an upper predictive bound on its \emph{time-to-success}---the number of interactions until the task is completed successfully. Under a limited budget, some runs are stopped early, so their time-to-success is unobserved (right-censored), and the correctness of the bound may be unknown. We introduce \emph{\textbf{PR}edictive Bounds for \textbf{O}nline \textbf{B}udget \textbf{E}fficiency} (\texttt{PROBE}), a method for constructing calibrated upper predictive bounds while adaptively allocating a limited budget across agent tasks to reduce the anticipated coverage deficit. Our method calibrates the bounds online using inverse-probability-weighted feedback together with a correction designed to minimize coverage deficit while remaining unbiased. We prove that \texttt{PROBE} provides a coverage guarantee for any finite time horizon while satisfying the budget constraint. Experiments on software engineering, retrieval-augmented generation, and instruction following show that \texttt{PROBE} attains a coverage rate closer to the target level with tighter predictive bounds than a calibrated baseline.

Recommendation systems have become more important since the widespread use of the Internet. The large amount of information means that it is difficult for users to discover they really need. However, users' preferences can change over time because of the age and the impact of social networks. Recommendation systems must capture users' preferences and recommend suitable items, which is a difficult challenge. This study proposes a novel recommendation system that adapts to the changing preferences of users. By learning and adapting to users' changing preferences better recommendations are provided. This study uses context factors as additional information to model users' preferences more precisely. For the proposed recommendation system, the context is based on users' interaction items within each defined time interval. This study proposes two novel attention mechanisms for the recommendation system. The contextual item attention module captures contextual information, changing pattern in users' preference and the importance of items. The multi-head attention module extends the diversity of users' preferences and adapts to changing preferences. The recommendation performance is improved using additional item's temporal information to model the contextual item's representation. Experiments compare the proposed algorithm with several state-of-the-art recommendation methods using three real-world datasets. The experimental results demonstrate that the proposed context-aware recommendation model outperforms traditional methods and demonstrate the effectiveness with which contextual information is captured by the attention mechanism.

Stochastic first-order methods use sampled gradients to reduce per-iteration costs, but estimation errors can slow convergence and, in some settings, cause divergence. To improve gradient accuracy while retaining acceleration, we develop scalar and diagonal versions of VR-Adam-SHANG for smooth stochastic convex optimization. The methods combine STORM-type recursive estimation, Adam-style preconditioning, and SHANG's accelerated interpolation and correction, using at most two stochastic gradient evaluations per iteration without full-gradient refreshes. Under unbiased sampling, bounded variance, and mean-square smoothness of gradient noise, polynomial steps with prescribed initialization yield $O(1/K)$ mean-square gradient-estimation error and expected objective-gap bounds combining $O(1/K^2)$ initial-energy decay with an $O(1/\sqrt K)$ stochastic term after $K$ iterations. These guarantees require neither monotonicity of the preconditioner nor a uniform positive lower bound on its eigenvalues. The analysis also establishes polynomial-step convergence of projected nonrecursive Adam-SHANG-s, filling a gap in its original theory. Experiments compare the methods at matched sample-gradient budgets. In every tested setting, at least one of the two versions achieves the lowest mean terminal objective gap among all compared methods.

Clinical speech is increasingly used for disease screening and clinical assessment, but directly distributing speech recordings poses privacy risks because they contain both speaker-identifying voiceprints and sensitive spoken content. Existing speech anonymisation approaches are primarily developed for general speech applications or task-specific clinical speech and overlook that different clinical tasks rely on distinct speech characteristics, necessitating task-adaptive anonymisation strategies to protect patient privacy while preserving clinical utility. Here we introduce CASA, a clinically adaptive speech anonymisation agent that transforms anonymisation from a fixed processing procedure into a task-guided optimisation process, dynamically tailoring voiceprint and content anonymisation to task-specific privacy and clinical utility requirements. Through recursive self-improvement (RSI), CASA iteratively evaluates and refines anonymisation strategies against task-specific privacy and clinical utility requirements, while learning from these outcomes to guide more efficient planning in subsequent clinical tasks. CASA is validated across 14 clinical tasks encompassing five languages, diverse elicitation paradigms, and mental health, cognitive and voice disorders. CASA preserved disease-related classification performance while meeting speaker-anonymisation requirements across all clinical tasks, versus only 8/14 for the strongest fixed method. Assessments of resistance to speaker re-identification by external attackers and identification and redaction of sensitive personal information in clinical speech further demonstrated CASA's dual protection of speaker identity and sensitive spoken content. Clinical expert assessments further supported the preservation of clinically relevant information and the acceptability of anonymised speech for controlled research sharing.

As enterprises delegate consequential work to multi-agent LLM systems, benign-looking requests can escalate into unsafe actions as they propagate across sessions through poisoned shared state, long-context documents, and trusted peer messages: no single message is malicious, yet the assembled tool action is. Prior benchmarks isolate a single agent under prompt injection or score multi-agent attacks by their final output, leaving cross-session erosion of shared state out of scope. We introduce ErosionBench, an environment-grounded benchmark that executes such attacks and verifies alignment decay at the process level, scoring the realized tool call against authorized intent and provenance with deterministic contracts, not output text. Each scenario is authored as a deterministic state graph before any prose is rendered, so every state change traces to a known cause. It spans four enterprise domains and four decay mechanisms across 1,728 scenarios with matched benign controls, labeled against the OWASP Top-10 for LLM Applications, plus 198 image-based attacks, each paired with a text equivalent, so a defense must survive the same attack in either channel. Perimeter defenses suppress attacks only by over-refusing 54-92% of legitimate work, so a low attack-success rate can conceal severe over-refusal. As a reference defense, we propose a calibrated Meta-Auditor pairing synchronous state checks with asynchronous trajectory monitoring; on a family-disjoint 906-scenario held-out split it reaches 98.5% recall at a 2.9% false-positive rate. Since it monitors trajectory, not surface content, its detection is channel-invariant: the strongest auditor flags image-delivered attacks about as reliably as their text equivalents (recall gap $\leq$2 points per model). Across GPT, Gemini, and Claude auditors, calibration keeps false positives low while recall tracks model capability, strongest for Claude.

Methods for detecting phishing accounts on Ethereum are usually evaluated as complete pipelines that differ in features, architecture, and training protocol, which makes the contribution of interaction information hard to isolate. This paper evaluates the Unlabeled Feature Extension (UFE), a block of 12 local interaction statistics computed from each account's incoming and outgoing call and transfer edges without reading class labels. UFE is appended to the 14 source features of HEIG-Phish and tested with six model families: Logistic Regression, Random Forest, LightGBM, Graph Attention Network, GraphSAGE, and Meta-IFD. Within each pair of Baseline and Plus variants, the data partition, model configuration, random seed, decision threshold, and sampled neighborhoods are identical; only the input feature block changes. Across 20 predetermined stratified partitions of the same 2,763 labeled externally owned accounts, UFE increased test Macro-F1 by 1.0 to 9.1 percentage points across the six models. Every pair improved on all 20 partitions, all Nadeau-Bengio corrected 95% confidence intervals lie above zero, and the false-positive rate decreased for all six models. An exact-edge Random Forest ablation over 10 predetermined partitions associates most of the gain with the incoming transfer-edge sum attribute; once this feature is present, an incremental contribution of the remaining 11 UFE columns was not established. The results support reporting feature-augmented baselines alongside graph models, but they show stability under re-partitioning of one static, transductive benchmark rather than temporal or cross-dataset generalization.

LLMs are powerful but incur substantial inference cost, which can be exploited by malicious users to induce overly long outputs, increasing latency and operational expense. Prior resource consumption attacks either coerce repetition or suppress the end-of-sequence token by lowering its logit value. However, with autoregressive, sampling-based generation, the search space grows exponentially with length and small deviations or unstable trajectories can derail optimization, limiting attack effectiveness and stability. To address this issue, we introduce Chatter Attack, a novel class of resource consumption attacks that optimize adversarial prompts to induce predefined short target phrases, thereby avoiding search-space explosion while still eliciting long responses. Specifically, Chatter Attack triggers self-doubt using short doubt-inducing phrases as the target, and sustains this state throughout generation via inference attention loss and global entropy loss. To extend to black-box settings, we further propose a novel Bayesian Prompt Optimization (BPO) method, which models the objective function globally by constructing a discrete token kernel, efficiently exploring the entire search space using prior information. Extensive experiments across multiple models and datasets show that our method outperforms baselines, with maximum response length increases to 31.5$\times$, and revealing practical resource-depletion risks for real-world LLM inference services.

Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification. We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, \emph{effective full-solution support}, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, \emph{atomic primitive coverage}, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting a distinct empirical signature and calling for a distinct intervention. We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models. We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains. The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.

Interpretable glaucoma detection requires a transparent connection between fundus images, clinically meaningful biomarkers, and diagnostic predictions. Knowledge graph (KG) reasoning provides this structure, but performance measured using annotated biomarkers does not establish performance when only images are available. We present Auto-GlaKG, an automated framework that combines image-based biomarker prediction with fixed clinical-prior rules. A ResNet50 network predicts the cup-to-disc ratio (CDR) and six morphological signs using biomarker supervision, and a KG reasoner aggregates their signed rule contributions into a glaucoma score. On 689 fundus images evaluated using stratified, group-aware five-fold cross-validation, Auto-GlaKG achieves a pooled area under the receiver operating characteristic curve (AUC) of 0.951 (95\% confidence interval: 0.931--0.969), with sensitivity of 0.970 and specificity of 0.779. The matched image classifier also achieves an AUC of 0.951, while the rule stage adds 0.2~ms per image. Replacing predicted CDR with its annotation increases fold-mean AUC from 0.953 to 0.995. The framework provides an explicit rule-level account of image-based predictions and identifies CDR estimation as a priority for validation using independently measured biomarkers.

Well-calibrated forecasting for non-linear dynamical systems is essential in science and engineering applications such as fluid dynamics and weather prediction. We present a forecasting method combining probabilistic machine learning with the structured linear dynamics of Koopman mode decomposition. Within this framework, the observation, linear dynamics, and state reconstruction operators are constructed using standard machine learning architectures, producing predictive distributions via the pushforward of injected noise samples. The model is trained using a proper scoring rule, the continuous ranked probability score, to directly target calibrated, accurate forecasts. The approach is validated on two non-linear ordinary differential equation examples and two partial differential equation examples, vortex shedding and cavity flow, using both large and small, noisy datasets. Results demonstrate that known Koopman eigenfunctions and spectra can be identified and used for accurate forecasting. The model can also produce multi-modal predictive distributions, demonstrated for a trajectory initialised on the boundary between two basins of attraction. Uncertainty is captured by the sample ensemble spread, with spread--skill ratios close to one over the training horizon in several experiments, although for extended rollouts and noisy datasets, the ensemble may become under- or over-dispersed. In low-noise experiments, temporal rollout is generally accurate for at least twice the training horizon, and up to twenty times in some cases. Approximation quality is comparable to a deep ensemble approximation over the training horizon, but at a reduced computational cost and with better robustness in identifying Koopman eigenvalues.

Learning neural solvers for spatiotemporal partial differential equations (PDEs) under real-world constraints remains a key challenge in scientific machine learning, especially for inverse tasks with sparse and noisy boundary observations. We present the Aletheia dataset, the first 3D benchmark for learning data-driven solvers in the context of nondestructive testing (NDT). The dataset simulates eddy-current-induced heating in conductive solids and models the resulting transient heat propagation governed by the heat equation. Aletheia contains over 4,700 high-resolution samples across 10 excitation frequencies (1-100,kHz), each providing volumetric heat source and temperature fields over time. It supports both forward prediction of temperature evolution and inverse reconstruction of internal heat sources or defects from surface infrared measurements. Real infrared thermography data from cracked rail specimens are included for calibration and comparison. We define four canonical tasks on both regular and irregular grids and benchmark them using various neural operators. Aletheia establishes a unified platform for evaluating neural PDE solvers under realistic NDT conditions, enabling progress in reliable, data-driven inverse modeling. The dataset and code are publicly available at https://github.com/sherlock-0327/ALETHEIA.

Self-supervised speech models (S3Ms) encode a rich hierarchy of acoustic, prosodic, and phonetic information, yet when and where it emerges during pretraining remains unclear. We train frozen-encoder probes on every Transformer layer across pretraining checkpoints of two matched models: HuBERT-Base, whose stage 2 is a separately initialized model trained on new pseudo-label targets, and single-objective wav2vec 2.0-Base. Our five tasks span an acoustic-to-abstract gradient: absolute pitch (F0), speaker-relative pitch, lexical stress, phoneme boundaries, and 38-way phoneme identity. Acoustic-prosodic cues and phone boundaries are decodable from the earliest checkpoints and barely shift across layers, whereas phoneme identity is constructed during pretraining, re-localizing to upper-middle layers under HuBERT's stage-2 targets at a matched update count. Acquisition order therefore tracks abstraction from the acoustic surface, not the segmental/suprasegmental distinction, extending to prosody an account established for phonetic, lexical and syntactic structure; the no-switch wav2vec 2.0 shows no comparable re-localization.

EmbeddingGemma-300M is a small open embedder whose tokenizer represents every Tajik letter, and as released it is weak at Tajik retrieval: on our Tajik law test the right passage is in its top ten results for 49.3% of questions, on tax documents for 45.9%, and on the human-written Belebele Tajik set for 69.8%, against 98.6% on the Russian version of the same set. We present Saidzoda-Tajik-EmbeddingGemma-300M, the same model fully fine-tuned for one epoch on 235,140 question and passage pairs written by our own Tajik language model for laws, tax documents, Wikipedia and web text, with the same mined hard negatives, English and Russian replay, loss, global batch, learning rate and epoch as our earlier Saidzoda-Tajik-E5-Large (arXiv:2611.00592) and BGE-M3 models. After training, law reaches 92.4% and tax 91.7%; on Belebele, the one test written by people, recall@10 rises from 69.8% to 92.6% and the first result is right for 76.6% of questions instead of 51.7%. A checkpoint at step 100 of 466 already holds 88% of the law gain at recall@10. At recall@10 the result is no more than 2.8 points below the two 560M models trained on the same pairs on law, tax and Wikipedia, within the intervals, with 768-dimensional vectors and about half their parameters; at recall@1, for which we have no intervals, it stays 7.0 points behind the E5 model on Belebele. The cost is a drop of 3.0 and 2.8 points in Russian and English recall@1, larger than for the two earlier models. We also report that, according to our run notes, a bfloat16 test run produced a non-finite loss at step 3 which the training framework logged as zero, so training was done in fp32. Our generated test questions come from the same model as the training questions, so we treat those scores as an upper limit until questions written by Tajik speakers exist.

Large language models (LLMs) are increasingly asked to judge whether a news article is true, and their stated confidence is often taken at face value. We test what happens to that confidence when an article's writing style, but not its facts, is manipulated, as in recent \emph{style attacks} on fake-news detectors. Tracing how the prediction and its confidence form layer by layer inside the network, we study three model families (LLaMA-3-8B, Qwen2.5-7B, Mistral-7B) on three fake-news benchmarks, and find the same three patterns in every model--dataset pairing. First, on the original articles the models are competent and largely honest judges: they settle on an answer, briefly become overconfident, and then correct their confidence in the final layers. Second, style-rewritten articles break this self-correction: the models become far less accurate, under the strongest attack, worse than random guessing, yet remain just as confident, so they are \emph{confidently wrong} precisely when a warning is most needed. Third, nearly all of the damage comes from rewriting into sensational or emotionally charged styles, while sober, neutral rewrites leave the models' honesty largely intact. The practical implication is that an LLM's confidence about veracity cannot be trusted as a safety signal when an article's style may have been manipulated.

Asynchronous pipeline parallelism can accelerate distributed model training, but suffers from instability caused by inconsistent gradient computations. We identify a fundamental source of this inconsistency: error signals are coupled to forward-time features, not just parameter states. This means that parameters may evolve as long as they preserve forward-equivalence. Leveraging this insight, we introduce FRESCO, which frames asynchronous updates as a constrained optimization problem: it finds the minimal update modification that acts trivially on the activation subspace. We demonstrate that FRESCO consistently outperforms representative asynchronous baselines in LLM pre-training, with the gains most pronounced under deep pipelines and high inconsistency levels: at pipeline depth 32, it reaches the same loss as the strongest baseline with 43% fewer iterations and 30% less wall-clock time, while requiring 8~921$\times$ less memory overhead. Moving beyond the current limitations, FRESCO establishes a new foundation for high-utilization pipeline parallel training, providing a scalable path for large-model development on diverse, resource-constrained systems.

Frontline sales staff spend much of their day re-coordinating quotes after customers change quantities, ask for discounts or new payment terms, or costs move. We present LoMASA (Local Multi-Agent Sales Assistant), a training-free multi-agent system that runs entirely on one desktop computer. A 35B mixture-of-experts model (System-1) makes each decision in a single forward pass and reads its confidence from the probabilities of the answer options. A 27B dense model (System-2) writes customer replies and escalation briefs and is consulted when System-1 is unsure. Code computes every price, margin and option, and hard pricing rules cannot be overridden. On a test split of 500 simulated episodes with 1512 decisions built from public supply-chain data, the salesperson has to decide only 19.5% of decisions: 38.0% are settled automatically, with a median latency of 2.67 seconds, and 42.5% are escalated to a manager by the hard rules. On the automatically settled decisions LoMASA is as accurate as a rule baseline (0.894 vs.\ 0.894) and makes fewer commitments whose margin later falls below the floor (2.61% vs 4.53%); the price is additional escalations to the manager. Option probabilities are well calibrated without any training, unlike confidence stated in words by the same model. We also report where the design does not help: the 27B model rarely changes a decision.

Direct Preference Optimization (DPO) is a simple and widely used alternative to RLHF for aligning large language models (LLMs), but its sequence-level objective cannot precisely locate or suppress the local patterns that make an otherwise acceptable response unsafe or undesirable. We propose TLRS-DPO (Token-Level Risk Scoring for Better Direct Preference Optimization), which augments DPO with fine-grained token-level control while remaining fully offline and reference-based. TLRS-DPO introduces a token-level robust compensation term that converts the policy-reference log-ratio on rejected responses into localized risk scores, a curriculum-style schedule that gradually increases the strength of this compensation during training, and a risk-adaptive screening mechanism that selectively penalizes only high-risk tokens instead of entire sequences. This design bridges sequence-level preference fitting and token-level behavior control, and can be implemented as a drop-in modification to standard DPO trainers. Empirically, TLRS-DPO improves alignment quality and safety-relevant behavior with minimal impact on helpfulness compared to strong DPO baselines.

Machine Learning Interatomic Potentials (MLIPs) based on Equivariant Graph Neural Networks (EGNNs) have achieved remarkable progress in atomistic modeling. However, accurately representing atomic interactions across different spatial scales remains challenging. In this work, we investigate the radial spectral properties of SO(3)-equivariant message passing and establish that the cutoff radius controls the characteristic spatial-frequency scale of radial interactions. Based on this analysis, we propose Multi-Cutoff Spectral Aggregation (MCSA), a plug-and-play module that aggregates radial kernels with different cutoff radii to enrich multi-scale geometric representations. To achieve efficient implementation, we exploit the compact support of radial kernels to fuse cutoff-dependent radial weights before angular tensor-product operations, avoiding redundant angular computations while preserving rotational equivariance. Extensive experiments across representative EGNN backbones demonstrate consistent improvements in energy, force, and stress prediction, as well as enhanced accuracy in higher-order properties, including phonon dispersion and lattice thermal conductivity. Further evaluations on layered materials with distinct interaction scales demonstrate the effectiveness of physically motivated cutoff selection. Code is available at: https://github.com/JF-Tan/MCSA.

Post-hoc explanation methods for neural networks are often heuristic and lack provable guarantees. A common alternative to such methods that was studied in recent years is to compute a cardinally-minimal subset of input features that is provably sufficient to determine the model's prediction. For general neural networks, however, finding such explanations is computationally intractable. In contrast, recent work showed that this problem becomes more tractable for a particular family of neural networks, namely neural additive models (NAMs), though standard NAMs are highly restrictive because they exclude feature interactions altogether. In this work, we study the computation of provably minimal and sufficient explanations for neural networks with restricted feature interactions, bridging the gap between fully general neural networks and purely additive models. We first prove that even with constant-order interactions, the problem is $Σ_2^P$-hard, implying that worst-case exponential complexity is unavoidable. We then identify two provably tractable settings: one based on sparsity in the interaction structure and another based on a stronger interaction-aware notion of sufficiency. Under either of these settings, we develop algorithms that compute provable explanations for neural networks with any-order interactions while retaining much of the computational efficiency of additive models. Empirically, we show that these explanations are significantly smaller and faster to obtain than those produced by standard algorithms, while being derived from models that achieve much higher accuracy than standard NAMs. Overall, our results help expand the class of neural networks that admit efficient provable explanations, while clarifying the fundamental role of feature interactions in shaping their complexity.

Activation sparsity can reduce FFN weight traffic during single-user LLM decoding, but active channels must be identified early enough to avoid fetching unused weights. In pretrained SwiGLU models, training-free routing with the native gate performs substantially worse than routing with the up activation at high sparsity. We show that this ordering reverses after sparsity-aware training. Based on this insight, we introduce SGA-Sparse, which converts pretrained SwiGLU LLMs into fixed-budget activation-sparse models by training the existing gate projection directly as a top-k router. SGA-Sparse updates all model parameters end-to-end while preserving the original SwiGLU activation and introducing no auxiliary router parameters. After approximately 30B training tokens, LLaMA models from 1B to 8B achieve uniform per-token 75-85% FFN channel sparsity while remaining within 0.2-2.7 pp of their original dense checkpoints in average zero-shot accuracy. Because post-training quantization degrades sparsity-trained checkpoints, we further study two complementary quantization-aware training (QAT) recipes: continued-QAT largely recovers performance at 4-bit weight quantization, while joint-QAT performs better for the more aggressive sub-4-bit target. To compare activation sparsity with dense low-bit quantization, we translate the sparsity-induced reduction in weight traffic into an equivalent bit-width. Under this unified weight-traffic metric, our 2.5- and 1.5-bit-equivalent models establish a new accuracy-traffic Pareto frontier: the 2.5-bit-equivalent models match dense 3-bit QAT accuracy, while the 1.5-bit-equivalent models approach dense 2-bit models and outperform dense ternary models by 1.3 pp on average. These results establish native-gate training as a practical approach for combining high FFN sparsity with low-bit weights.

MeanFlow learns interval-average velocity for few-step generation, but its fixed temporal sampler mixes diagonal and interior supervision throughout training. We interpret temporal coordinates as collocation points: diagonal samples provide a stochastic boundary target, whereas finite intervals supply a differential-consistency target. The diagonal fraction controls how often the consistency component appears, while interval length enters its gradient through an explicit factor. Motivated by this distinction, EbbFlow uses Temporal Supervision Allocation (TSA) to schedule the diagonal fraction and interval geometry. It starts with diagonal supervision, then admits interior samples and expands their maximum interval; a left-anchored interval law preserves coverage across time. The per-sample loss, target, backbone, and inference procedure remain unchanged. In matched class-conditional ImageNet $256\times 256$ runs without classifier-free guidance (DiT-B/2, one training seed per method, 400k updates), EbbFlow achieves 50k-sample FID of 39.65 at one network evaluation and 36.02 at two-evaluation ODE sampling, versus 44.82 and 39.86 for MeanFlow. Gradient probes show a mid-training difference in component geometry. Ablations support allocation scheduling and left-anchored intervals, although a fixed full-horizon variant performs better at one evaluation. These results suggest that temporal sampling is a useful training control whose benefits depend on the inference protocol.

Predicting single-cell transcriptional responses to genetic and chemical interventions supports gene-function discovery and regulatory analysis. Existing studies often focus only on highly variable genes, restricting the expression changes available for downstream analysis. Cells exposed to the same intervention can also show heterogeneous expression responses, with coordinated gene programs active to different degrees. We propose scBIG++, a full-gene perturbation prediction framework that models this heterogeneity by connecting expression-adaptive gene modules with population-conditioned generation. To preserve shared gene structure while accommodating cellular variation, Gene-Relation Clustering organizes genes into common windows, within which Dynamic SoftSub learns expression-dependent submodule memberships. This shared organization enables Set-GCAE to exchange information about corresponding gene groups across control cells, placing individual representations in their source-population context. The cell representations initialize a latent flow, while their pooled summary and the perturbation representation condition response generation. Generated states are decoded into expression profiles covering all assayed protein-coding genes retained after preprocessing, supporting comprehensive analysis of perturbation-induced expression changes. Experiments across six genetic and chemical settings spanning 6,539--19,451 genes demonstrate competitive perturbation-effect recovery and full-gene expression reconstruction under unseen targets and combinations, cross-cell-line transfer, and held-out compounds. Our implementation is available at https://github.com/ttruan2426-dot/scBIG.

Reproducibility in large language model (LLM) agent evaluation depends on replay fidelity: which parts of an execution trace are reconstructed, controlled, and compared. This matters for agent debugging, failure analysis, and counterfactual evaluation of stateful tool use. We audit all 71 unsuccessful executions in a fixed historical AppWorld corpus using 284 isolated fixed-code replay attempts. Reconstructed declared settings admit 42 cases under printed-output, pre-oracle database-delta, and native-verdict/test-count checks; resetting the random-number generator before each interaction admits 68. The 26-case gap contracts to four under database/oracle checks, with one endpoint unknown. Mechanism studies identify four materialized-state discrepancies despite matching outputs and verdicts. In one admitted endpoint pair, an identical API call yields matching responses but different row effects; common future Python RNG removes that deterministic difference in two paired repetitions. A separate JWT credential-equality program and alignment controls expose authentication/write differences, with missing repetitions retained. For all 10 affected post-oracle export cases, native export from retained output trackers restores the reference comparison without changing original cached contents. A purposeful 12-case BFCL pilot shows no discrepancy under its tested public-state, return, and verdict checks. We preserve resource censoring and infrastructure deviations. The contribution is a bounded empirical characterization of replay eligibility, hidden state, and test-oracle integration, without estimating debugger accuracy, repair gains, or method-ranking distortion.

Post-hoc explanation methods for video action recognition typically conflate spatial appearance and temporal dynamics into a single saliency or importance map. We propose a perturbation-based framework that disentangles these contributions into a spatial appearance mask ($M_a$) and a temporal motion mask ($M_m$). The two masks are jointly optimized against a frozen action recognition model through factored perturbation architecture, with nearest frame replacement to identify important temporal dynamics. A separation penalty discourages the appearance mask from spending budget on frames that the motion mask has already discarded. A ReLU-based area constraint penalizes explanations larger than an assigned budget. To evaluate these explanations, we introduce a separation protocol that sweeps spatial pixels and temporal frames independently, backed by proofs characterizing the proposed method. Across EPIC-Kitchens-55, EGTEA Gaze+, and Something-Something V2, our method attains the lowest deletion AUC against leading baselines while producing masks roughly five times sparser.

Numerical programs implementing the same computational task can exhibit different numerical behavior because of their execution stacks, program representations, or program logic. Source text and final outputs alone provide limited diagnostic information: numerical differences can propagate through corresponding computations, while distinct executions can produce equal outputs. We present NumericalDiff, an execution-based method that records assignment-level numerical events, aligns traces using a tolerance-aware longest common subsequence independent of variable names and source locations, and optionally reexecutes after replacing one intermediate value with its counterpart. Among the original and all single-intervention executions, it lexicographically maximizes matched events and minimizes unmatched events. NumericalDiff supports code comparison across language and framework front ends, currently including NumPy, PyTorch, JAX, C, C++, Fortran, and Octave. We evaluate NumericalDiff on 36 controlled cases comprising 675 pair/input instances: 36 execution-stack, 588 program-representation, and 51 program-logic instances. All 24 operator-injection instances expose trace differences, including three with tolerance-equal final outputs, while complete trace alignment is achieved for all 441 canonical cross-representation instances. Within the 51 program-logic instances, intervention improves 24 alignments and raises macro-average matching coverage from 0.4210 to 0.6513. The results show that value-based trace alignment and reexecution can delimit propagated numerical differences across heterogeneous implementations.

Historical social simulation faces a fundamental reproducibility challenge: the stochastic nature of large language models (LLMs) introduces a digital butterfly effect that prevents exact replication across multi-turn runs. Furthermore, the inability to control low-level server parameters in commercial LLM hosting environments severely limits reproducibility. To overcome this opacity, we present a deterministic multi-agent framework coupling LLMs to an Agent-Based Modelling (ABM) pipeline grounded in Theory of Mind (ToM) and Cognitive Archaeology, and validate exact artifact identity within local, pinned execution environments using SHA-256 hashing. Run locally at temperature zero with clean per-call model reloads, the system injects differentiated cognitive states per turn through dissonance-driven behavioural modes and gated internal voices. Evaluating seven model configurations spanning 4B to 31B parameters against an ablated baseline reveals clear operational boundaries. Small models ($\leq$8B) suffer from context capacity overload, producing repetitive, specular dialogue loops. However, at $\geq$12B parameters, the ABM scaffolding breaks the monotonous response patterns seen in ablation. Crucially, these scaffolded models exhibit spontaneous diegetic expansion: they coherently expand scenario lore through tactically motivated, anachronism-free details rather than escalating into aggression. By resisting the structural pull toward premature scene closure, scaffolded sessions sustain an open terminal Narrative Vector ($V_N = \mathrm{Open}$ in 12/21 overall; 9/9 in official $\geq$12B configurations) whereas unscaffolded ablation never does (0/21). This framework provides a transparent, auditable benchmark architecture for virtual archaeology, with the full pipeline released as open-source software to enable independent replication.

Visual chain-of-thought (visual-CoT), also known as "thinking with images", equips LVLMs with the ability to locate and use key visual evidence for reasoning. However, acquiring such visual-CoT capabilities requires costly post-training and tightly couples them to a specific backbone, making transfer to other LVLMs difficult. Existing cross-model transfer methods typically treat visual-CoT as a monolithic capability, using a unified representation that ignores the functional distinction between grounding and reasoning and introduces stage-irrelevant interventions. To address this problem, we propose Viscot-steer, a training-free activation steering framework for transferring visual-CoT capabilities across heterogeneous LVLMs. Specifically, Viscot-steer first disentangles visual-CoT into grounding and reasoning vectors and extracts them from the source-model layers associated with the corresponding capabilities. It then aligns the vectors with the activation space of the target model and steers the grounding and reasoning layers during bounding-box prediction and reasoning-trace generation, respectively. Across benchmarks covering high-resolution visual understanding, general visual perception and reasoning, and mathematical visual reasoning, Viscot-steer transfers capabilities from different visual-CoT models to architecturally diverse LVLMs and outperforms the strongest capability-transfer baseline by an average of 1.86 percentage points on the primary evaluation metrics.

On the Web, user profiling builds compact user representations from interaction histories for efficient personalization, and dynamic user profiling updates them as users' behavior and preferences change. However, new user behavior can be ambiguous, and its interpretation differs across users. To address this, we introduce Uncertainty-aware Adaptive Profile Manager (UAPM), which separates the claims in a user profile from interpretations of new behavior that are not yet confirmed, through a Profile, a Buffer, and a Rulebook. UAPM records user-specific interpretation and update rules in the Rulebook, keeps ambiguous evidence in the Buffer outside the Profile as evidence for later decisions, and writes supported general and situational characteristics to the Profile. We train UAPM with Privileged On-Policy Distillation with Dual Specialization (POPD-DS), which learns from recorded future interactions during training without access to them at deployment. Its first stage, Dual Specialization, distills a stronger model into a current-only Student and a future-informed model that becomes the Teacher. In the second stage, privileged on-policy distillation (POPD), only the fixed Teacher receives future interactions and guides the Student's own management decisions alongside task rewards. Experiments on three Web personalization tasks show that UAPM outperforms multiple baselines under the same downstream input budget, improving on the best same-budget baseline by up to 5%.

We study prompted chain-of-thought (CoT) with self-consistency (SC), i.e. sampling N traces and plurality-voting, on small language models, a regime where per-query compute budgets are tight yet SC is commonly applied with a fixed sample budget inherited from large-model practice, regardless of instance difficulty. Across two model families (Qwen-2.5-0.5B, Gemma-3-1B) at FP16 and MLX-Q4 (4-bit quantisation in Apple's MLX framework), on GSM8K and MATH-500, we run 8 per-task and 4 task-pooled experiments at n=200-400 with 1,000x paired-bootstrap confidence intervals (CIs). Our first result is an upper bound: an oracle gate deciding per prompt whether to invoke SC reaches 109-174% of always-SC's accuracy at 23-30% of its compute in all 12/12 cases, so at N=5 always-SC is dominated everywhere on our grid. That margin is exactly the rate at which SC flips correct greedy answers to wrong, here 3-10.5% of prompts, which on Qwen-Q4 x MATH-500 exceeds the wrong-to-right rate and leaves always-SC 3 pt worse than always-greedy. Our second result is a deployable approximation: a gate over greedy-pass signals skips 45-99% of SC calls, with bootstrap CI half-widths under 5 pt, removing 36-79% of always-SC compute. Accuracy retention at those operating points has point estimates of 98.1-98.8%, which we report as point estimates because their CIs are wide (lower bounds 76-86%) over an SC accuracy of only 15-36%. The evidence covers N=5 on two <=1B families and two math benchmarks, which is the scope we claim. Code is available at https://github.com/heywanrong/SiftSC.

Relevant evidence in Indian Knowledge Systems (IKS) collections may be distributed across passages in different parts of a work, including verses, translations, and commentaries. Locating it requires matching natural-language questions and partially remembered verse fragments to source passages. We present Dhara, a retrieval engine combining hybrid search, chapter- and book-level Entity Notes (EN), Fuzzy Verse Retrieval (FVR), and neural reranking. Five comparisons on the 325 questions in Dhara Bench examine retrieval representations against BM25, EN's contribution, reranker quality and computation time, FVR, and Gemini. BGE-M3 single-vector retrieval exceeds BM25 in nDCG@10, Recall@20, and MRR@20 across all categories. Its scores are broadly comparable to ColBERTv2, with mean retrieval times of 0.108 s and 0.962 s under the same hardware configuration and a smaller memory footprint for chunk storage on disk. Adding EN to a direct branch with a matched candidate count improves nDCG@10 in all twelve categories with Qwen and ten with BGE; the combined ranking is capped at the same candidate count. Qwen improves nDCG@10, Precision@5, and MRR@20 over BGE in every category, at approximately 6.72 times the overall estimated serial computation time. FVR increases Verse-category nDCG@10 by approximately 20 percentage points with either reranker. Gemini is evaluated after source mapping and manual review with relevance annotations visible. The Qwen pipeline with EN and FVR has higher Recall@20 than Gemini in ten categories and stronger complete-evidence retrieval in Multi-hop questions, while Gemini retains early-relevance advantages in several categories. The results inform retrieval choices according to evidence requirements and resource budgets.

Migrating Android applications to HarmonyOS is an important yet challenging software engineering task. Existing large language model (LLM)-based migration approaches primarily translate code at the method or file level, overlooking project-wide dependencies and workflow coordination. Moreover, ArkTS, the primary language for HarmonyOS, is relatively low-resource, and migrated applications often lack sufficient tests for reliable verification, making end-to-end migration error-prone. We present A2HTrans, an agentic framework for end-to-end Android-to-HarmonyOS application migration. Rather than performing isolated code translation, A2HTrans decomposes applications into functionality-aware migration tasks, prioritizes them through dependency analysis, and orchestrates translation, validation, and iterative repair within a unified execution workflow. To improve migration reliability, A2HTrans employs a tool-augmented verification strategy that combines compiler-based syntax checking with LLM-based semantic validation and automated repair. We evaluate A2HTrans on 14 Android applications spanning diverse sizes and domains using Claude-Sonnet-4.6 as the underlying migration model. Compared with Claude Code, A2HTrans preserves 8.3% more functional completeness, improves behavioral correctness by 8.0%, and reduces migration time by 28.8%. These results demonstrate that structured agentic execution substantially improves the reliability and efficiency of project-level cross-platform software migration.

Multimodal-Attributed Graphs (MAGs) organize heterogeneous data in graphs whose nodes are characterized by modalities such as text and images rather than traditional unimodal attributes. However, many representative MAG methods organize propagation around node-level states and topology-defined neighborhoods, which can obscure fine-grained modality contributions: a bounded neighborhood may contain semantically weak evidence, while compatible context need not be confined to it. To address both limitations, we propose TEAM, a modality-centric framework that constructs and encodes graph context with node-modality tokens rather than treating each node as a single interaction unit. Concretely, TEAM unfolds the local topology into a Modality Encoding Tree and performs tree-anchored semantic evidence acquisition, organizing local structural context while expanding the accessible evidence space outside the constructed local Tree. We further introduce the Modality-Source Context Transformer (MSCT), which encodes the selected node-modality tokens and coordinates interactions among modality-source groups. By retaining textual and visual attributes as separate interaction units until readout, it supports fine-grained cross-node cross-modal exchange without first collapsing them into whole-node features. Across the evaluated baselines, TEAM attains the highest reported mean on all graph-learning metrics, all three G2Text metrics, and both G2Image metrics. On Toys, it surpasses the strongest baseline by 3.37 and 2.93 percentage points in Accuracy and Macro-F1, respectively.

Text-to-speech systems built from a language model and a flow-matching mel decoder now run the decoder with two to four Euler steps instead of the official ten or more. At two steps the word error rate (WER) is unchanged, but speaker similarity (SIM) drops by one to six points. This paper shows that most of this loss comes from classifier-free guidance on the first step. At pure noise the network can only return the conditional mean, so guidance shifts every trajectory by nearly the same vector. Later steps do not remove this shift, and its size grows with the length of the first step, so it is negligible with ten or more steps and accounts for most of the loss with two. Two training-free samplers are derived from this analysis. First-step unguided sampling (FSU) leaves the first step unguided and recovers $84$--$91\%$ of the two-step SIM loss on CosyVoice~2/3 with no significant WER increase. FSU-Mid replaces the two Euler steps by one midpoint step with an unguided first evaluation, recovers $77$--$85\%$ on IndexTTS-2 and matches the ten-step teacher on CosyVoice~2 and GLM-TTS. Both use three network function evaluations, one fewer than the official two-step sampler, and FSU-Mid gives the highest mean recovery among eight published guidance rules re-run at two steps on four decoders.

Random perturbations around a shared checkpoint can produce experts for distinct tasks, but shared origin does not ensure that their gains survive composition. We formulate expert-pair selection for joint base-relative retention, measured by strict joint success and worst-task gain with unconditional abstention accounting. An exact finite-difference decomposition and an arbitrarily small orthogonal counterexample show why proximity and orthogonality alone are insufficient. We instantiate this formulation with quality-constrained spectral selection, a fixed-budget framework combining cross-task screening, layerwise pair geometry, separate validation, and Base fallback. Across six checkpoints, 21 task pairs, and five independent perturbation pools, the procedure improves strict joint success over Top-1 averaging throughout and leads the four-method model-scale comparison on four of six checkpoints. For the five models with at least 1.5B parameters, matched quality-only, global-cosine, and spectral ranking attain 49.1%, 50.5%, and 53.1% joint success. Under this matched protocol, the spectral selector also yields the highest worst-task gain and Weak Pareto rate and the lowest regression rate. Mean worst-task gain reaches +0.61 percentage points, while 17.1% regression and 13.0% abstention show that composability remains conditional. Project page: https://anonymous.4open.science/r/Thickets-Merge/.

This study asks what changed in engineering articles after conversational language models became publicly available on November 30, 2022, and develops methods for measuring such change in a field without a uniform abstract index. It assembled 1,839,962 articles published from 2015 to 2025 in 256 engineering journals, tracked twelve measures of wording, style, and article structure, and compared each with the projection of its own pre-2023 trend. Building the monthly series showed that records dated only by year fall on January 1 and create false January peaks unless removed. In abstracts, 31 words associated with language model output appeared 82% more often than projected (55.4 against 27.0 per 10,000 words in 2025), while four control words of the same formal register fell 16%. Words that soften a claim, such as may, fell 13%, words that strengthen one, such as clearly, rose 20%, and long words rose 11%. Author counts, reference lists, page counts, and publication volume stayed within 4% of trend. The change appears in all fourteen subject categories, and false release dates, dropping any one journal or word, and word lists chosen from the data support it. A random sample of 12,866 full texts from the same corpus, including subscription articles and weighted back to it, shows the same words rising 99% in article bodies. Statements disclosing AI writing assistance, measured apart from mentions of language models, went from none to 3.7% in 2025, four in five of them at one publisher. The design is descriptive, not causal.

Although hidden-state probes efficiently monitor model operation, it remains unclear whether probes trained on benign tool use remain useful for detecting incorrect actions under adversarial conditions. In this research, we assess this problem from the perspective of operational tool-action correct- ness. we use four AgentDojo suites, four conditions, and three models: Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, and GLM- 4.6V-Flash (9B). We perform a five-fold, task-grouped out-of-fold (OOF) evaluation for each model. We train step-specific logistic- regression probes only on benign and non-defended trajectories and evaluate them on held-out trajectories for benign, prompt- injection, system-override, and tool-poisoning conditions. We compare our hidden-state probe to mean tool-call-token entropy, negative mean tool-call-token log-probability, and self-consistency uncertainty. The paired tests highlight six significant differences after Holm correction. All of the tests involve Qwen2.5-14B- Instruct under the system-override condition, with all three scalar uncertainty metrics obtaining higher area under the receiver operating characteristic curve (AUROC) and average precision (AP) than the hidden-state probe. These results suggest that benign-trained hidden-state probes should not be assumed to reliably detect operational tool-action errors under adversarial conditions.

Large-scale device-level load forecasting for distribution transformers is challenged by substantial cross-device heterogeneity and the deployment burden of maintaining device-specific forecasting models. Existing approaches typically rely either on device-specific models, which are difficult to scale to large transformer populations, or on global models, which reduce deployment complexity but may overlook heterogeneous load patterns across devices. To address this trade-off, this study develops a trend-clustered forecasting framework that enables model sharing among transformers with similar temporal behaviors. Representation Learning for Time Series using Time-Embeddings (T-Rep) is employed to learn compact temporal representations of transformer load series, which are subsequently grouped using a Gaussian mixture model (GMM). Representative transformers are sampled within each cluster to train cluster-specific Patch Time Series Transformer (PatchTST) models while retaining device-level predictions. The framework is evaluated using real-world 15-min operational data from nearly one thousand distribution transformers. The proposed framework reduces mean absolute error (MAE) by approximately 20% relative to the direct PatchTST baseline, achieves a high-load warning accuracy of 95.66%, while requiring only seven cluster-specific forecasting models. These results demonstrate that behavior-based device organization provides an effective balance between forecasting accuracy, deployment scalability, and risk-oriented operational support for distribution transformer management.

Test-time reinforcement learning with majority-vote rewards trains a language model on unlabeled questions by rewarding the samples that agree with its most frequent answer. During training, each question locks into one answer, on hard questions often a wrong one, but which answer it locks into is known only after training. In this paper, we show that this outcome is decided early, in a short race driven by sampling noise between the answers that are close at the start, and we derive the probability of each outcome. With a separate logit for each answer, the runner-up wins with a normal tail probability of the initial gap, scaled by an explicit noise constant. In a network, however, the answer logits depend on shared parameters, and through the kernel of these parameters the updates of the trailing answers also change the gap of the two leaders. This spillover shifts the boundary between the starts won by each leader, and a closed form computes the shifted probability from a kernel probe and the start logits. In LoRA experiments with three models from two families, predictions made before training with the kernel frozen at its start value forecast which answer each question locks into better than a logistic model refitted on the outcomes, although they use no fitted constants. The closed form matches these predictions on the runs in which a check before training finds that the two leaders stay above the other answers, and it explains why a law that ignores the spillover needs a strongly inflated step.

Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.