The Inference Report

Realistic and editable animal fur reconstruction from multi-view images is challenging due to fine-scale detail, self-occlusion and obfuscation, and, unlike human hair, the lack of animal-fur datasets. Fur usually covers most of an animal's body, with large inter-species and intra-species variability. We present FurE, an efficient strand-based animal fur reconstruction method that recovers a per-strand, editable groom by optimizing a root-conditioned latent field, decoded into strand geometry via a PCA-based decoder. We reconstruct a defurred animal body using local fur-thickness cues from a surface-constrained Gaussian Frosting representation together with part-based priors. We further show that a PCA-based decoder learned from human-hair strand data can alleviate animal-data scarcity while enabling substantially faster optimization. FurE achieves a 10x speedup in strand training over current SOTA dense per-strand optimization while retaining strand fidelity and generalizing across synthetic and real-world sequences, with quantitative and qualitative validation despite the reduction in training time.

One deployed language model must often serve many compute budgets, yet serving each budget still means a separate training or compression run per point. We train a Telescopic Language Model (TLM) to be that continuum: a nested-capacity Transformer supervised by stochastic prefix supervision with a full anchor. At every step, one randomly truncated prefix of the capacity axis is trained against the full next-token target, alongside one full-capacity pass, so the trained artifact is a valid language model at every depth. Two forward-backward passes per step, no architectural change, nothing extra at inference. Fixed-exit suites such as Matryoshka Language Model Suites (MLMS) occupy one point in this design space, and the point has a cost: supervising only a few fixed exits leaves the nested model at chance level everywhere else (perplexity 10^2-10^5 in our baselines). On a 200M proxy suite (20B FineWeb-Edu tokens, identical data stream for all methods), a single TLM run is a valid language model at every one of its twenty layer prefixes, in perplexity and on perplexity-sensitive downstream tasks, reducing the area under the quality-budget curve by 43-44% relative to the fixed-exit suites while matching them at full capacity, at ~12% lower GPU cost per run. The prefix sampling density is a dial: concentrating it on a few depths recovers fixed-exit quality there at the price of the continuum, so the operating points become a training-time choice rather than an architectural one. These results indicate that the training objective, not the nesting itself, is what makes a model elastic.

Modern video diffusion models require tens of denoising evaluations over long spatiotemporal token sequences. Distribution Matching Distillation (DMD) reduces the number of function evaluations (NFE) to just a few. However, DMD samples can degrade during training, exhibiting progressive oversaturation and artifacts. We trace this instability to critic errors, which enter successive student updates and accumulate over time. We introduce Projected Distribution Matching Distillation (PDMD) to filter critic errors. PDMD projects out the component of the DMD update parallel to the student-critic endpoint residual. At a fixed noisy query, we prove that this residual is an unbiased estimate of the critic's endpoint error. Under high-dimensional assumptions, this projection removes a constant fraction of critic error while discarding only a vanishing fraction of ideal DMD signal. Empirically, the projection stabilizes training and improves sample quality where DMD degrades and develops unnatural textures. PDMD requires only a one-line code change to DMD, with no extra loss, network, data, model pass, or multi-stage training. With Wan2.1, PDMD achieves a VBench total score of 83.73 at 4 NFE, surpassing matched DMD by 1.03 points. On MiniMax-H3 joint video-audio generation, PDMD achieves a VideoGen-Eval visual total score of 83.17, 0.41 points above the strongest distilled baseline. PDMD also achieves the best performance on all six audio metrics among the compared 4-NFE models. Qualitative comparisons and user studies favor PDMD over the distilled baselines in visual quality, motion, and audio quality. Code and models are available at https://pdmd2026.github.io/.

Unified multimodal models can both look at and render images, so in principle they can repair their own generations: diagnose what an image gets wrong, revise it, observe the result, and diagnose again. Whether a revision helps is known only after it is rendered, so the reflection text and the image generation must be learned jointly, over the whole loop. Supervised fine-tuning (SFT) on reflection trajectories gives a cold start but does not find the high-success repair paths, and naive RL that optimizes only the renderer or only one head leaves most of the gain untapped. We introduce UMM-Reflection, which applies reinforcement learning (RL) to complete reflection trajectories inside one unified model: sibling trajectories share one initial image, so the group-relative advantage compares reflection strategies, and one trajectory-level advantage updates both the reflection tokens and the flow-based revisions, avoiding the combinatorial blow-up of per-round credit assignment. Unlike single-round editing or pipelines with an external critic, credit flows across rounds and to both roles of the same model, and no verifier is needed at inference. On BAGEL, UMM-Reflection improves GenEval by 12.05 points over SFT, and the gains transfer to WISE (+10.97), OneIG-Bench (+3.48), and T2I-CompBench++ (+4.63), none of which is used in training.

Identifying intertextual references is central to literary scholarship, but computationally difficult when source material is transformed through paraphrase, allusion, historical language, and translation. We investigate this problem through biblical intertextuality in Karen Blixen's Seven Gothic Tales. Drawing on the commentary to a critical edition, we construct a benchmark of 189 annotated references and evaluate retrieval against all 31,170 verses of historically plausible Danish Old and New Testament translations. We compare TF-IDF and BM25 with multilingual and Danish sentence encoders, examine the effect of linguistic normalization, and fine-tune a Danish encoder using hard negatives and five-fold cross-validation. We analyze performance across automatically derived lexical-overlap strata representing quotations, paraphrases, and allusions. Linguistically normalized BM25 provides a strong zero-shot baseline, attaining an overall R@10 of 0.365 and retrieving every quotation within its ten highest-ranked verses. The best zero-shot dense model achieves a comparable overall score of 0.360 while performing better on allusions. Fine-tuning DFM-large raises its overall R@10 from 0.265 to 0.508 and more than doubles its performance on allusions, from 0.138 to 0.339. However, evaluation against editorial annotations alone understates the model's scholarly usefulness: a literary scholar judged seven of 30 selected rank-one predictions counted as false positives to be meaningful additional references. These findings show both the potential and the epistemic limits of computational intertextual retrieval. Rather than treating scholarly annotations as exhaustive or model outputs as discoveries, we propose retrieval models as heuristic co-readers that recover documented references and generate candidates for expert-led close reading.

\emph{Distributional training} provides collective supervision for one-step visual generation by matching real and generated features in frozen representation spaces. We introduce \emph{a unified theoretical framework} that separates distribution modeling from matching discrepancy and connects global objectives to pointwise feature updates through Wasserstein gradient flow. Under this framework, FD-Loss and Gaussian-kernel Drifting are recovered through Gaussian optimal transport and kernel-density-based KL matching, respectively. The framework motivates \textbf{MGFlow}, which models feature distributions with Gaussian mixtures at an adjustable granularity between global moments and sample-based representations. MGFlow supports both optimal transport and score-based matching, and couples mass-constrained sample assignment with paired component updates to address mode collapse that mixture expressivity alone does not resolve. On ImageNet $256\times256$, MGFlow substantially surpasses the FD-Loss baseline, achieving state-of-the-art results with \textbf{1.45} $\mathrm{FDr}^6$ on pMF-H and \textbf{1.64} on JiT-H. For text-to-image generation, MGFlow post-trains FLUX.2 [klein] 4B into a one-step generator that outperforms the original four-step model on both GenEval and PickScore. Project page: https://shihaoyang0423.github.io/MGFlow-website/

LLMs have demonstrated strong capabilities in creative writing. However, scaling them to full-length novels remains challenging, as maintaining narrative consistency becomes increasingly difficult. Existing story-generation methods typically focus on stories of up to about ten thousand words, leaving their ability to scale to full-length novels underexplored. In this work, we introduce Narrative State Tracking Agent (NstAgent), a training-free agentic framework that allows LLMs to track a structured narrative state including characters, past events and future requirements. We extend an existing benchmark to compare narrative consistency across lengths, and use it together with a writing-quality benchmark to systematically evaluate stories ranging from 10K to 100K words. We show that NstAgent achieves better narrative consistency and writing quality as stories grow longer, and neither of them degrades noticeably as length increases, suggesting that it provides an effective approach to scaling story generation toward full-length novels.

When a large language model (LLM) agent executes the same task, token consumption can vary by over an order of magnitude across runs. The agent chooses its next steps based on tool feedback and intermediate results, while the growing context steadily inflates the input size of every subsequent call. The total consumption of a task is therefore hard to predict before execution and the prediction must be revised as the run unfolds. In this paper, we propose TokenCast, which learns a composable cost representation for each execution segment, recording its own consumption and the context growth it introduces. Composing adjacent segments yields a cumulative estimate that captures the extra input cost incurred when context from earlier segments is re-read by every later call. As execution unfolds, newly observed evidence refreshes the forecast, requiring no additional LLM calls and incurring a mean cumulative prediction time of 32.8 ms per run on SWE-bench Verified. Across 4 task suites and 6 agent models, TokenCast's mean absolute error reduction against the strongest comparator averages 14.5% over 96 evaluated combinations. In offline budget-control replay, TokenCast uses 21.3% fewer tokens on average than a fixed-budget policy at matched trace completion. The code is available at https://github.com/DEFENSE-SEU/TokenCast.

We present a representation-learning framework for composite adaptive tracking control under dynamically coupled disturbances. The framework connects classical disturbance-accommodating control (DAC) to recent last-layer adaptive disturbance-rejection methods. Specifically, we introduce a statistically principled hard expectation-maximization (hard-EM) procedure, with a Kalman smoother in the hard E-step, to identify dynamical representations of disturbance whose latent evolution is uniformly contractive. The learned representation evolves a latent disturbance-excitation state from measured plant features and control inputs and decodes that state into the time-varying disturbance acting on the nominal plant, thereby extending prior "fixed-decay" last-layer adaptive methods to a learned, predictive DAC-style formulation. Combined with Bayesian filtering of the learned latent state, this representation yields a composite adaptive tracking controller with predictive capability and provable exponential convergence to a bounded neighborhood. We validate our approach experimentally on a slippery ground vehicle carrying a liquid-sloshing tank and a pendulum load, and we further assess its robustness on a system of coupled Duffing oscillators. Across both settings, the method achieves accurate disturbance prediction and improved overall tracking performance relative to fixed-decay representation-learning ablations, LTI disturbance-accommodating baselines, and model-based PD baselines.

We introduce Neural Harmonic Measure Operator (NHMO), a neural solver for elliptic PDE problems on variable-shape domains. The harmonic measure of a domain is the boundary probability distribution that, integrated against any boundary data, returns the Dirichlet Laplace solution. It depends only on the geometry, not on the boundary data. NHMO parameterizes the density of this measure as a transformer-based boundary kernel supervised by Walk-on-Spheres exit samples, so one trained kernel handles different boundary values on a shape with no retraining. We extend it to Poisson via a classical decomposition, with an auxiliary network amortizing the source-induced correction and avoiding the singular volume quadrature that breaks direct evaluation. At inference, new boundary values and new sources both yield PDE solutions by re-integration against the fitted kernel and lift, with no retraining. NHMO improves over four prior baselines on the MCB-B 3D variable-shape Poisson benchmark across all five categories, and is competitive with major neural-operator baselines on a controlled 2D testbed.

Looped Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of many experts for each token. Looped MoE bridges these two design philosophies and gives MoE models new potential for better expert usage, but it raises a question: how to loop a MoE? We answer it with Foil. With the expert parameters and the expert compute per token held fixed, Foil (1) flattens the experts, halving the expert layers, doubling the experts per layer and doubling the passes, so that every routing decision chooses from a larger pool, and (2) unties the attention, giving each pass its own attention parameters while the experts and routers stay shared. Experiments show that Foil clearly outperforms the unflattened looped baseline: at 20B tokens every Foil model has lower pretraining loss than the baseline; at 100B tokens the loss improves monotonically with the degree of flattening, the most flattened Foil ending 0.012 nat below the baseline at equal parameters and compute, with downstream accuracy on par or better; untying the attention also yields more balanced and more confident routing at equal shape. Our ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more passes. Code and configurations are available at https://github.com/SR-A-W/how-to-loop-moe.

Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU memory constant for a given trace. Unfortunately, most compaction strategies rely on prefilling the LLM context many times over, hindering training throughput. To alleviate this bottleneck and enable efficient trainable compaction, we propose KV-streams, a plug-and-play strategy compatible with any compaction strategy that substantially increases throughput while showing no evidence of hindering performance. KV-streams enable scalable compaction by streaming the KV cache forward rather than flushing it after each compaction. We show that KV-streams enable three different compaction strategies, achieving a 2.6 to 5x wall-clock speedup in training. Beyond efficiency, we find that the streamed KV cache can act as a recurrent state, carrying forward information that has long since disappeared from the context. Specifically, in a controlled setting we show that, contrary to prior work, RL alone is all that is needed for this behavior to emerge. Overall, we show KV-streams to be an efficient and lightweight plug-and-play addition to any post-training pipeline.

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, we propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding. We first characterize communication efficiency in terms of action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges. We design TACT to revise student-generated actions, test the revisions through partner responses, and distill useful feedback into the student. An expression specialist removes unnecessary detail while preserving the intended action, while a strategy specialist proposes alternatives that may better address the partner's constraints. To determine which revision helps, TACT samples a partner response for each candidate and selects a teacher reference by balancing local goal support against action-token cost. That reference guides on-policy distillation on the student's own generation prefixes, allowing the student to act independently at deployment. We evaluate TACT on SOTOPIA and AgentSense. On SOTOPIA, it achieves the highest Goal among the evaluated methods on All and Hard while using substantially fewer target tokens than SFT+SDPO. On AgentSense, it improves goal success over the initial student while reducing target tokens and interaction messages.

Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowledge, whether looping improves test-time scaling as outputs grow longer remains underexplored. Through post-training looped transformers, we study the accuracy-compute slope, measured as the accuracy gain per doubling of test-time decoding FLOPs. We find that existing looped transformers often yield steeper slopes than their non-looped baseline, yet underperform it at matched compute. While fixed-depth looping spends extra iterations on every token, our analysis shows that many tokens do not benefit from extra iterations. We therefore propose TaH2, which enables the model to focus extra iterations on the tokens that benefit from looping. It jointly post-trains the backbone and an iteration decider through lookahead depth supervision, which uses online labels indicating whether further iteration improves the prediction. TaH2 improves both the efficiency and attainable accuracy of test-time scaling. On challenging AIME benchmarks, TaH2 improves the accuracy-compute slope by 53% (2.74 vs. 1.79) over the non-looped baseline, exceeding the baseline's peak accuracy by about 3.4 points at matched test-time compute. As the maximum iteration depth increases, existing looped models largely plateau, while TaH2's gain over the non-looped baseline continues to grow from +2.8 points at depth 2 to +3.9 points at depth 8. Our code is available at https://github.com/thu-nics/TaH.

Linear Vision Transformers (ViTs) are designed to replace the attention in Softmax ViTs with the linear-complexity attention operator for more efficient token routing, but they require from-scratch pre-training and typically underperform the original Softmax version. How to initialize linear ViTs both efficiently and effectively still remains unclear. In this work, we explicitly ask: given that most foundation ViTs are built on the mainstream Softmax attention, can linear ViTs benefit from their pre-trained weights? Recent works on Attention Transfer show that attention is the effective transferable component between Softmax ViTs, suggesting attention alone suffices for such reuse. However, we find the opposite for Softmax-to-linear transfer. The attention weights are operator-specific: copying them barely helps, and is sometimes even worse than random initialization. Instead, the attention's token routing behavior can be recovered through distillation with a proper loss design, letting linear ViTs reduce the gap and even match Softmax ones. In contrast, the MLP weights, which carry the learned representation, are operator-agnostic: they can be transferred by simple direct copying, which already carries most of the benefit of the pre-trained weights. Thus, copying MLPs can serve as an effective foundation for Softmax-to-linear transfer: paired with the distilled attention, linear ViTs eventually close the remaining gap and even surpass Softmax ones. These findings hold consistently across various linear ViT variants, different model sizes, and diverse datasets. We hope this study deepens the understanding of reusing pre-trained weights across attention operators: copy what stays the same and distill what differs, to recover the benefit across the Softmax-to-linear boundary.

Evaluating finance research agents requires rubrics that reflect expert standards and fix the values correct as of an information cutoff. Expert-reviewed finance benchmarks rely on fixed, per-item rubrics, which are costly to extend and cannot encode each institution's own standard. In FinAutoRubric, experts specify reusable evaluation guidance, while agents and code carry out query-specific rubric generation, review, and validation. This expert guidance governs every agent, as prompts and as rules that code enforces, and a Task Bank of reusable criteria carries it across tasks. In long-horizon loops that follow the expert guidance, a writer agent researches every expected value and a reviewer agent verifies it, and failures escalate to a human. On three expert-authored finance benchmarks, its rubrics track expert scoring as closely as the strongest evaluated generator while stating the expert rubric's expected value for more criteria, their scores agree with human grading, and in-house analysts prefer them in a blind review. The released 100-query FinAutoRubric Benchmark, built from in-house analysts' key questions across 78 tasks and eight asset classes, shows that rubrics from an earlier model generation still leave headroom for a later one.

People learn not only by repeating successful actions, but also by recounting and explaining their experiences, revising their understanding to guide future behavior. Can a language-model agent improve its future actions by training only on explanations of its own experience? We investigate this question by studying Retrospection-Only Fine-Tuning (ROFT), a minimal online procedure designed to isolate the effect of explanation-only training on subsequent behavior. The agent attempts a task, observes available feedback, generates a retrospective explanation, and is fine-tuned with a next-token prediction loss on the explanation tokens alone. The procedure uses neither an external teacher nor a reward-based policy update. In software-engineering experiments with Qwen3.5-4B, ROFT is trained on problems with mixed successful and unsuccessful base-model attempts. On held-out SWE-bench Verified and Pro, it reaches 49.2% and 26.8% solve rates after 20 updates without using a verifier, compared with GRPO's 48.0% and 25.3% after 40 updates in the evaluated runs, and makes faster early progress in training time and sampled attempts. It also learns to solve individual tasks on which all 64 sampled base-model attempts failed, showing that learning can begin without any initially successful trajectories. Behavioral analyses find that ROFT indirectly assigns credit to actions, encouraging good actions and discouraging incorrect ones. Moreover, prompting retrospections to emphasize more direct solutions yields shorter subsequent attempts even without an explicit length penalty. Together, these findings show that learning to explain can also improve learning to do, establishing self-generated retrospections as useful training targets and motivating further study of explanation-to-action transfer.

A language-model agent is jointly defined by its model and its harness, the executable program that organizes model calls, tool use, and information flow. Because different tasks call for different ways of organizing these operations, the harness needs to be adapted using feedback from the task at hand. We introduce harness learning, which trains a proposer model to revise a solver's harness using execution feedback. We formulate this process as meta-learning over executable programs, with harness revisions playing the role of weight updates in gradient-based adaptation. We train the proposer with reinforcement learning, using the task performance of revised harnesses as the reward. At test time, the proposer uses feedback from successive executions on a new task to refine the harness, without performing any parameter-space update. Experiments on reasoning and multi-hop question answering show that harness learning improves revision quality and that the ability to adapt at test time transfers to unseen tasks. Policies trained on individual revisions can continue improving harnesses over multiple rounds, while the benefits of training on revision sequences vary across settings. These findings suggest a path towards continually learning agents that turn accumulated experience into generalizable improvements.

Tool-using agents can fail twice: a required tool can fail, and the agent can then report success without the evidence needed to justify it. Existing benchmarks often entangle this reporting failure with tool selection, recovery, and environment dynamics. We introduce Failure-Transparent Agents (FTA), a controlled benchmark that fixes the failed observation and required evidence state before generation, making post-failure claims directly auditable. FTA contains 100 tasks with deterministic failure traces spanning five failure families, a neutral control, and four user-pressure conditions, and evaluates unsupported claims alongside useful recovery. Across six models, three response policies, and 3,600 human-annotated responses, false-success rates are 22.8% under the baseline policy, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract. Fabricated-detail rates decrease from 28.3% to 14.3% and 0.8%, while useful responses increase from 74.9% to 89.2% and 98.8%, respectively. The tested evidence-contract policy is associated with substantially lower post-failure reporting errors while useful-response rates remain high within this blocked-task benchmark.

Reinforcement learning (RL) in simulation can train dexterous manipulation policies without robot demonstrations, but training a single generalist policy with task-agnostic rewards faces a severe exploration problem: approaching, grasping, and reorienting diverse objects with many degrees of freedom is difficult to discover from scratch. Prior works make exploration tractable with high-quality robot demonstrations, per-task reward shaping, or by restricting policies to narrow modes of behavior. We propose X-Reset, a framework that instead resolves exploration with human hand-object demonstrations. Rather than imitating or tracking retargeted human motion, X-Reset kinematically retargets hand-object states to noisy robot states, filters out states that are unstable in simulation, and samples the remainder as resets during RL training with general-purpose object-centric rewards. The resulting policy depends only on object state and goal, with demonstrations entering training through the reset distribution. We show that X-Reset trains generalist policies on 20 objects across three embodiments---a 22-DoF hand on two different arms and a parallel-jaw gripper---and resolves the exploration challenges of RL from scratch. X-Reset scales with the number of training objects, generalizes to unseen objects, can learn from imperfect hand-pose estimates, and transfers behaviors zero-shot from sim-to-real.

Recent progress in machine learning is driven by large-scale foundation models, where scaling laws and finding optimal scaling prescriptions for architecture, data, and hyperparameters are key in advancing the state-of-the-art. Therefore, it is surprising that no systematic study evaluates the methodology to obtain scaling laws and prescriptions across different model types. To shed light on this crucial blind spot and facilitate future research, we introduce the surrogate benchmarks ScAn-Bench-LLM and ScAn-Bench-VLM based on 4524 and 8024 checkpoints of language and vision-language model pipelines. On our benchmarks, we perform the first systematic evaluation of both data acquisition and extrapolation methodology for scaling analysis across different data modalities.

Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable reasoning. Grounded in cognitive science, we model creativity along three axes: action (what to do and how creatively), process (when to explore versus exploit), and outcome (the novelty and usefulness of the resulting idea). We use these axes to train models with GRPO, exposing them to varying degrees and forms of creativity throughout training. This produces substantially more diverse scientific proposals, expanding the range of research directions by 27.8% and contribution types by 14.9% over the base model. It also improves predicted citation impact by up to 32.0 percentage points and originality by 66.2 points. These gains cannot be reproduced by simply increasing decoding temperature; instead, we find that semantic guidance specifying what kind of creativity to pursue is critical. Overall, our results suggest that creativity is a learnable, multi-level ability that can be shaped to help researchers reach ideas beyond those typically explored by LLMs.

Reliable deployment of graph neural networks requires calibration, out-of-distribution (OOD) detection, and robustness to distribution shift, yet existing methods address these needs with separate models and objectives. We model uncertain node embeddings as random graph signals: graph Fourier filters capture structural variation, and a scalar orthogonal-polynomial chaos coordinate captures latent stochastic variation. The resulting doubly-spectral stochastic (DSS) expansion supplies task-matched readouts from one representation: the mean coefficient encodes class evidence for the energy-based OOD score, the higher-order coefficients encode structured logit variation, and quadrature averaging over the chaos coordinate defines the single predictive distribution used for prediction and calibration. A capacity theorem shows that, under a full-rank feature assumption, a restricted subfamily matches the chaos coefficients of any Gaussian-latent random graph signal, with exponentially decaying truncation error under a growth condition; the task-level claims are established empirically. DSS-GNN has two deployment modes: standalone, or as a residual branch beside a deterministic encoder (DSS-Hybrid). Standalone DSS-GNN achieves the lowest Brier score among the compared uncertainty-aware baselines on all 14 node classification benchmarks without post-hoc correction; DSS-Hybrid achieves the best AUROC on most node-OOD settings, competitive cross-graph OOD detection, and the strongest shifted accuracy on all 7 GOOD concept-shift benchmarks under standard empirical risk minimization (ERM). Cross-evaluating both modes on all three tasks shows that each remains effective on the other's tasks, with documented exceptions, and yields explicit deployment guidance.

The success of large language models (LLMs) has been accompanied by continued growth in model size and pretraining costs. Muon offers high accuracy and training efficiency in LLM pretraining. Recent work introduces row-wise normalization into Muon to balance update magnitudes and improve pretraining performance. However, row-wise normalization alone cannot accommodate different imbalance patterns in update matrices. In this paper, we propose an improved Muon optimizer, called \underline{m}atrix-\underline{eq}uilibrating Muon~(MeqMuon), for LLM pretraining. MeqMuon balances both row and column magnitudes through normalization that can be automatically tailored to different imbalance patterns without manual intervention. Moreover, MeqMuon eliminates the need to store AdamW's second-moment estimates, reducing optimizer-state memory usage. Empirical results demonstrate that MeqMuon achieves better convergence performance than AdamW, Muon, and other baselines in LLM pretraining.

Distillation attacks copy the reasoning capabilities of closed-source large language models, allowing bad actors to replicate state-of-the-art performance at low cost. Attackers systematically collect a large volume of frontier model reasoning traces and then train (i.e., "distill") their own models on these traces. Existing defenses against distillation attacks are typically evaluated immediately after distillation, implicitly assuming attackers do not train their models any further. In this paper, we argue that a more realistic threat model includes further training with reinforcement learning after distillation. A misspecified threat model can give a false sense of security -- some defenses that seem effective after distillation can be broken after subsequent reinforcement learning. Practically, reinforcement learning lowers the bar for a distillation attack to be effective. We show that simple attacks can steal reasoning capabilities from existing closed-source language models using data easily obtainable from current APIs, yielding reasoning improvements equivalent to more sophisticated attacks that extract the full hidden traces. Results indicate that any distillation defense that leaks sufficient information to reconstruct approximate reasoning traces is likely ineffective. We conclude by discussing broader implications and batch-level distillation defenses which could be more effective.

We study adversarial imitation learning (AIL), in which an agent learns to imitate expert demonstrations by optimizing a policy against an adversarial reward that distinguishes expert and learner behavior. Historically, reward regularization and entropy-based policy regularization are key components of empirically successful methods such as GAIL and LS-IQ, yet their finite-sample benefits remain underexplored. We establish fast rates for jointly regularized AIL in finite-horizon Markov decision processes with general function approximation. Our model-free algorithm, Dually Regularized AIL, combines KL policy regularization with a quadratic reward penalty weighted by expert and learner occupancies. With K online episodes and N expert trajectories, we prove a $\widetilde{O}\left(\frac{1}{K}+\frac{1}{N}\right)$ bound on the regularized imitation gap for fixed regularization parameters. Our analysis combines an online mirror descent construction for general convex reward classes to control estimation error from finite expert data and stochastic learner feedback, with a sharp analysis of optimistic KL-regularized policy learning. To the best of our knowledge, Dually Regularized AIL is the first algorithm to simultaneously achieve $\widetilde{O}\left(\frac{1}ε\right)$ sample complexity in both expert demonstrations and online interactions for this regularized AIL objective, even with stochastic experts. These results provide a rigorous characterization of the complementary statistical benefits of reward and policy regularization in AIL.

Aligning large language models (LLMs) to diverse user preferences is fundamentally hindered by standard alignment paradigms that optimize for monolithic users. In this work, empirical studies are first used to reveal the existence of a massive, untapped performance headroom for personalized generation through test-time alignment. We demonstrate that personalized generation is uniquely suited for test-time scaling methods like Best-of-N (BoN) because it can be viewed primarily as a candidate matching problem rather than a generator capability bottleneck. While reward models could in principle exploit this headroom, they are poorly calibrated for personalization, and their billion-parameter scale makes scoring large candidate pools prohibitively expensive. To overcome this limitation, we propose a parameter-efficient framework utilizing million-parameter scale multi-layer perceptron (MLP) ranking models. Our personalized ranking model directly reuses the internal embeddings of the base generator with minimal overhead. By scaling train-time data to provide fine-grained personalized preferences, this million-parameter ranking model accurately scores large candidate pools and can seamlessly guide generation to reduce the cost of materializing N candidates. Extensive experiments on nine datasets spanning three personalized generation settings show that our personalized ranking model effectively exploits the discovered headroom, outperforming billion-parameter generalist reward models on every dataset, with under 0.4% of their parameters and four orders of magnitude lower scoring latency.

Continuous diffusion generates complete reasoning solutions through iterative refinement in latent space. We introduce Latent Flow Reasoning Models (LFRMs), an ELF-based training and inference recipe. Our experiments show that accurate decoding alone does not ensure strong reasoning performance. We therefore learn compact representations from multiple layers of a strong autoregressive teacher. Their decomposition also enables asynchronous denoising at different rates. We show that prompt encodings need only preserve the information required for the correct text-conditional score, rather than exactly match teacher features, and use a staged curriculum to learn a compact prompt encoder that replaces the teacher Transformer at inference. We adapt DiffusionNFT to learned self-conditioning guidance and incorporate gold-solution endpoints to supplement sparse rewards. Our supervised models outperform reported results from recent continuous-diffusion baselines at comparable backbone scales on mathematical reasoning and HumanEval code generation. With a 638M-parameter denoising backbone and learned prompt conditioning, post-NFT LFRM-L achieves 63.74% pass@1 on GSM8K and 24.6% on MATH500 at 64 denoising steps, and 32.85% on HumanEval and 30.18% on HumanEval+ at 128 denoising steps. Code will be available at: https://github.com/chengxiang/LFRM

Users of Workday's deployed LLM-based agents often request features which can be addressed by defining named procedures, also known as skills, in the LLM context, effectively augmenting agents' capabilities. However, as an agent's skills library grows in size, so does the agent's operational cost. Progressive disclosure (lazy-loading) of skills as needed may reduce operational costs, but its impact on overall latency and skill-retrieval quality remains unclear. In this report, we investigate the impact empirically and find that progressive disclosure improves skill-retrieval quality but marginally degrades overall latency.

Mechanistic interpretability (MI) aims to explain a model's behaviour through analyzing its internal computations; circuit-based explanations aim to isolate these computations with compact subnetworks validated by ablating the rest of the model. We show that circuits validated this way may fail to recover the underlying mechanism of the model's behaviour by closely reproducing its successful decisions while failing to account for most of its errors. Such explanations should account for the model's particular errors as well as its successes. We evaluate this requirement by measuring exact answer agreement separately on model successes and failures, across circuit sizes and ablation settings, on IOI, Docstring, and six model-task settings from the Mechanistic Interpretability Benchmark. We discover that many tested circuits closely replicate correct behaviour while missing most of the model's errors. On indirect object identification (IOI) for GPT-2 small, under mean ablation, the manual circuit and tested automated circuits, including one trained against the model's full output distribution, agree with the model on 97.3-99.5% of prompts it answers correctly but only 11.4-41.7% of errors. An IOI case study shows that lost errors are recoverable by restoring omitted attention-heads which raise error reproduction from 14.2% to 75.1% on a separate held-out set with 0.41 percentage point decrease on correct agreement, exceeding matched random extensions and scalar-biased control. Intervention traces show how omitted computations produce specific wrong answers for a reproducible subset of errors. In all, these findings show circuits can preserve task success without adequately explaining model's failures, and support exact error reproduction as a necessary, but not sufficient, test of circuit-based explanations of model behaviour.

Structured span annotations, such as quantities with their units, uncertainty modifiers, and event classes, are expensive to create and hard to keep trustworthy once language models enter the loop. We present QuanReview, an open-source system for auditing and correcting such annotation layers. QuanReview aligns two annotation streams over the same documents at character level, resolves unambiguous cases by an explicit and logged policy, and routes candidate conflicts to a browser-based adjudication interface where reviewers accept either side, build field-level hybrids, or flag items for re-annotation. A campaign manager assigns documents to multiple annotators with configurable redundancy, computes agreement at document and span level, auto-merges unanimous documents, and exports the corrected layer in the original file format, so that it can replace the original annotation files directly. Applied to a 4,457-record humanitarian benchmark and an LLM extraction stream, the system fully auto-merged 8% of documents, applied automatic policy decisions to a further 1,513 records, and concentrated human attention on 3,131 candidate conflicts, a mean of 5.4 per reviewed document.

With the recent advancements of neural compressors, explicitly incorporating perception constraints into the design of compression schemes has gained significant attention. Traditionally, these perception constraints ensure that the distribution of the reconstruction does not significantly deviate from the distribution of the source, thus attesting to the perceptual quality of the reconstruction. In this work, we uncover several perception constraints that are naturally present in task-aware compression. In particular, we consider a problem where the primary task is reconstruction and the secondary task is classification (i.e., a statistical test). We study this problem at varying levels of domain information available to us and discuss how to utilize the naturally emerging perception constraints to design rate-minimal compression schemes that also maximize the utility of our secondary task. We show that in this setting, if the decision boundaries of the classifier are ill-defined (mismatch) for our source distribution, then matching onto a target distribution enhances our classification accuracy.

In reinforcement learning with verifiable rewards (RLVR), imperfect verifiers can reward incorrect responses, creating opportunities for reward hacking. Using gradient flow with a fixed verifier, we characterize the conditions under which reward rises while correctness falls. We then show that the observations available during RLVR are, in general, insufficient to detect or identify accepted errors, or to guarantee their reduction without sacrificing correct responses. To address this limit, we construct a correction using additional feedback about correctness from audits. This correction achieves \emph{selective control}: at the current policy, it lowers the probability of accepted errors and raises that of correct responses, provided it outweighs the pressure toward errors from verifier reward. Experiments with log linear and neural contextual bandits and with a language model support the analysis and show that selective control under partial auditing reduces accepted errors while increasing correctness.

Of the approximately 4,500 Oracle Bone Inscription (OBI) characters discovered from the Shang dynasty, only about 1,600 have been deciphered. Many computational approaches compare OBI with glyphs from one historical period at a time. However, during the evolution of Chinese characters, significant structural or semantic changes often occur in uncertain dynasties. A single-period reference may be insufficient when relevant forms change substantially between observed eras. Therefore, we propose the \textbf{Manifold-based Script Evolution Framework (MSEF)}, a framework that models the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space. MSEF represents each character as an era-specific manifold point and learns continuous inter-era transition rules via Neural Ordinary Differential Equations. Both manifold space and transition dynamics can be trained end-to-end through character evolution pairs across any two eras.

Mobile GUI agents operate through a perception--action loop: at each step they screenshot the device, invoke a vision--language model (VLM), and emit an action. It is slow, costly, and brittle, yet most of what it does is navigation---and everyday navigation is static, ordered, and endlessly repeated. We present PhoneCLI, which compiles an app's GUI navigation into callable commands, without any app-internal API, runtime instrumentation, or model training. Offline, PhoneCLI explores a target app from the outside and distills its screens, interactive elements, and navigation edges into a semantically annotated map; each screen yields one deterministic command: a replay sequence that reaches it. Online, the agent selects a command, verifies it before execution, and then executes it deterministically in sub-second time at zero VLM cost; open-ended interaction and every failure of the compiled path fall back to the embedded VLM interpreter, exactly the pure VLM agent, so compilation can only help. On AndroidLab, PhoneCLI improves the task success rate while reducing steps and token consumption, and it transfers to AndroidWorld's official M3A agent with consistent efficiency gains. What PhoneCLI compiles is the app's navigation rather than one run, so it serves new tasks, not only repeated ones.

Interior-point methods (IPMs) are among the most widely used algorithms for constrained optimization, yet their Newton-based search directions require costly second-order information and large linear-system solves. Learning to optimize offers cheaper updates learned from data, but the singular behavior of logarithmic barriers near constraint boundaries makes IPMs highly sensitive to perturbations, complicating both warm starting and learning reliable updates. We introduce pdLIP, an IPM for smooth nonlinear programs that integrates learned preconditioning with pdProj, an all-shifted primal-dual projected-search IPM. A shared coordinate-wise recurrent network predicts a positive diagonal preconditioner that scales the right-hand side of the reduced Newton system for the primal step, and the remaining slack and multiplier directions are recovered analytically. The learned iterations avoid Hessian evaluations and Newton-system solves, using only first-order and coordinate-wise operations amenable to GPU parallelization. Training is self-supervised, with a loss based on a penalty-barrier merit function and the residual of perturbed optimality conditions, requiring neither target directions nor precomputed solutions. Primal and dual shifts mitigate the barrier's sensitivity to perturbations near constraint boundaries, enabling effective warm starting. Across four classes of 200-dimensional convex and nonconvex constrained problems, pdLIP warm starts reduce pdProj refinement iterations by 63-67% compared with cold starts at the same KKT residual tolerance of $10^{-8}$, with negligible warm-start generation cost relative to the subsequent pdProj solve. Improvements persist on box-constrained QPs with 1000 variables and extend to applications including portfolio optimization, support vector machines, and a nonlinear control example.

Gated Linear Attention (GLA) Transformers advance linear recurrent models through data-dependent gating, but face a core limitation: the fixed-capacity memory matrices across all heads operate at a single temporal resolution, where each token is processed individually, forcing them to simultaneously encode local syntactic patterns and long-range semantic structure, creating a representational bottleneck that gating alone is insufficient to resolve. We introduce Multi-Scale Gated Linear Attention (MS-GLA), which addresses this by distributing attention heads across multiple temporal resolutions. Coarser resolutions pool longer token spans naturally specializing toward long-range dependencies, while finer head groups retain sensitivity to local syntactic structure. A learnable, input-dependent fusion layer dynamically recombines head group outputs at each timestep, expanding effective memory capacity without increasing per-head state size. This multi-resolution decomposition draws on principles from Multi-Scale State-Space Models (MS-SSM), adapting them to the gated linear attention setting. We evaluate MS-GLA on language modeling, recall-intensive tasks, and long-context generalization. Across all settings, MS-GLA consistently achieves higher accuracy and lower perplexity than GLA at matched parameter counts, with up to 18.9% improvement on recall-intensive tasks and 9.5% lower average perplexity on language modeling benchmarks, validating multi-temporal resolution decomposition as a principled and effective extension of Gated Linear Attention.

Large language models (LLMs) reliably perform entity copying, in which a model copies tokens referring to an entity, termed entity tokens, from the prompt into its output to answer a question. Although entity copying is straightforward for most LLMs, existing research does not provide a systematic account of which layers specialize in this fundamental task or how other tokens in the same sequence, termed context tokens, influence the model's ability to copy the entity tokens. To address these questions, we conduct experiments on Qwen3-8B using two novel methods: genie-in-a-bottle, which controls exactly which layers can participate in an entity-copying task, and attention lobotomy, which cuts off specific tokens' attention to entity tokens without affecting the remaining attention distribution. We find that two distinct groups of layers in the second half of the model are both necessary and sufficient for entity copying. Moreover, in addition to the decoding position's attention to entity tokens, context tokens' attention to entity tokens also proves necessary for copying the exact tokens, even though context tokens do not store entity information themselves unless they satisfy particular semantic properties. Our findings establish the critical role of late layers in entity copying under the guidance of context tokens, calling for future work on how models propagate and consume entity information.

Population-based optimization methods often use previous search information through successful solutions, parameter adaptation, or operator performance, but they rarely retain the context in which a search behavior succeeded or failed. We introduce Cognitive Memory-Driven Optimization (CMDO), a derivative-free population-based optimizer that represents experience as the relationship between search context, search behavior, and observed outcome. CMDO organizes these experiences across working, episodic, and consolidated memory, retrieves them according to similarity with the current search state, and uses both positive and negative evidence to guide subsequent search. Retrieved experience does not replay previous candidate locations; instead, it selects search recipes that are reconstructed from the current population through exploratory, directed, and local search behaviors with adaptive search geometry. We evaluate CMDO on selected Blackbox Optimization Benchmarking test suite on COCO (BBOB/COCO) and Congress on Evolutionary Computation 2017 (CEC2017) problems against DE, CMA-ES, SHADE, GWO, HHO, and ORCA, and further study its application to seven-parameter photovoltaic model estimation using measured current--voltage data. The results show problem-dependent but competitive optimization performance, including the lowest median error among the compared methods on CEC2017 F10. More importantly, analysis of the search traces shows that context-dependent recall changes the distribution of executed search behaviors, while unsuccessful experiences remain available as negative evidence for later decisions, showing that accumulated experience directly influences subsequent search behavior. These results support the use of explicit context--behavior--outcome memory as an active mechanism for controlling population-based search.

Tandem mass spectrum prediction supports compound identification across metabolomics, natural-product discovery, and environmental analysis. However, pretrained predictors often degrade under shifts in chemical space and acquisition conditions, while retraining domain-specific models from scratch is costly. We introduce SPARC, a retrieval-guided test-time specialization framework that adapts a pretrained predictor using a spectral reference library without accessing test-query spectra. For each target query, SPARC retrieves chemically related reference spectra to recalibrate fragment intensities within the learned fragmentation space. During Transfer, SPARC combines reference-guided spectral adaptation with reliability-aware consistency, using reconstruction behavior on retrieved spectra to selectively preserve trustworthy predictions during continual specialization. Across MassSpecGym, NPLIB1 and application-specific GNPS libraries, SPARC improves spectral prediction under multiple transfer settings. These results establish retrieval-guided test-time specialization as a practical strategy for extending pretrained MS/MS predictors to specific chemical and acquisition domains, with continual test-time training providing further refinement during deployment.

Many language tasks have no single answer that can be checked automatically. Rubrics provide criteria for judging responses to these tasks. For reinforcement learning, the resulting verdicts must be combined into a scalar reward. A common approach sums the points assigned to satisfied criteria. Distinct verdict patterns can thus receive the same reward, and the fixed points encode how much each criterion should count, not how strongly its verdict distinguishes the current rollouts. Beyond this aggregation problem, judging the full rubric needs more judge requests as the criterion count grows. To address these limitations, Rubric Response Theory (RRT) measures quality and selects criteria when rubric criteria are monotone indicators of a shared target. Rather than adding assigned points, RRT uses a two parameter item response model that treats the verdict pattern as evidence about scalar quality specific to the rubric. Under this model, its likelihood score maximizes the local signal-to-noise ratio for quality. Its Response Parameter Network (RPN) reads the prompt and criterion text to predict criterion difficulty and discrimination. As the policy distribution changes during training, RRT uses online expectation maximization to update the RPN from current rollout verdicts. With Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above that of group relative policy optimization (GRPO). On hard and very hard criteria in Medical and Science, RRT gains 2.8 to 5.6 points over GRPO. At half the criterion budget, adaptive Fisher selection with a frozen RPN keeps the macro criterion score across four datasets within 0.1 points of GRPO with full judging. These results show RRT can reduce judge requests while remaining competitive with GRPO.

Code-switching (CS), a seamless alternation between languages within a single utterance, remains a critical challenge in automatic speech recognition (ASR). While prior works focus on conversational CS-ASR, enterprise settings demand evaluation of operational impact beyond edit-distance errors: how code-switching transcription errors propagate to downstream voice agent task failures. In this work, we propose (1) a CS-ASR synthetic benchmark and multidimensional evaluation framework tailored to enterprise domains, (2) systematic evaluation of frontier ASR systems across 5 language pairs, (3) diagnostic analysis of the additional transcription errors that code-switching introduces across language pairs and models. We release COSE-E to support enterprise-focused CSASR evaluation for multilingual voice agents in enterprise deployment.

Large Language Models can perform multi-step reasoning and improve task performance through different forms of intermediate computation, from token-based traces to computation carried out in latent space. However, a question remains open: do these different forms of thinking rely on the same underlying mechanism? To address this, we train and compare five variants of the same GPTNeoX backbone from scratch on an extended multi-hop reasoning task (ProsQA-Ext): a vanilla model, a Chain-of-Thought (CoT) model, a Pause Token model, and two latent-reasoning models that are optimized end-to-end without intermediate reasoning traces. We find that, strong in-distribution (ID) performance does not guarantee depth generalization. Vanilla, CoT, and Pause Token models solve ID problems well, but rely largely on local graph features and generalize poorly to out-of-distribution (OOD) problems with longer hops. In contrast, latent variants generalize better and show internal dynamics consistent with forward reachability propagation on the graph. Causal interventions and circuit analysis localize this computation to a sparse recurrent search circuit in the bottleneck latent model: an attention head retrieves graph relations, an MLP and the residual stream update the reachability state across recurrent steps, while multiple attention heads together then do the candidate matching. Together, these results show that different thinking mechanisms can learn distinct computational solutions, even at similar ID performance. In this setting, latent recurrence supports a reusable forward-search algorithm that generalizes beyond the training depth.

Precise instruction following in image generation, such as satisfying object counts and spatial relations, remains an open challenge at least in part because it is learned using unreliable reward models such as object detectors and vision-language models. We introduce Verifiable Visual Rewards (VVR), the first framework for programmatically verifiable image rewards, and show that training on it generalizes to natural prompts. Each VVR task is a scene of geometric objects and relations among them, from which we derive both the prompt and a deterministic verifier, so tasks can be generated in any number and at any chosen complexity. We release VVRBench, with 10,000 tasks over 32 constraint types, and VVRBench-Challenge, with 720 more complex tasks; the strongest model we evaluate---GPT-Image-2.5---solves 21.4% of VVRBench-Challenge. Using VVR scores as rewards for reinforcement learning (RLVVR) raises the accuracy of Stable Diffusion 3.5 Medium on VVRBench from 2.8% to 28.3% and demonstrates consistent easy-to-hard generalization. These gains extend to out-of-domain benchmarks, and mixing VVR into existing objectives further improves overall performance and human preference, motivating the adoption of VVR into standard image generation post-training recipes.

Writing fast GPU code for physical simulation is difficult: implementations must preserve numerical accuracy while handling irregular data access, synchronization, and iterative solvers. We introduce GPUPhysBench, a benchmark of 50 tasks testing whether coding agents can meet these demands. Tasks cover fluids, deformable solids, and granular materials, from individual simulation operators to complete simulators. Agents write, compile, test, and optimize GPU code with access to a NVIDIA GPU under fixed time budgets. We report pass rates and runtime performance relative to expert-optimized reference implementations. In a single-attempt evaluation of six frontier model-harness pairs, the two strongest pass all 50 tasks, but even the fastest reaches at least 0.9 the reference speed on only 22% of them, and no submission is more than 5% faster than the reference. The largest gaps arise in collision detection, constraint solving, and iterative solvers. GPUPhysBench brings physical simulation workloads to coding-agent evaluation, testing both the ability to implement numerical methods correctly and the ability to make them run efficiently.

In materials machine learning, closely related retained structures can sustain accurate property predictions even after removing a specific record, rendering post-deletion prediction error an ambiguous metric for machine unlearning. To resolve this ambiguity, we define the deletion floor as the expected target loss under a specified retraining procedure at the deleted request. Standard indistinguishability constraints yield a sharp interval bounding an update's target loss around this baseline reference. Theoretically, a conditional neighbor bound links a low deletion floor directly to retained fit, prediction regularity, and local label agreement, while an exact ridge identity isolates residual fit from the prediction change induced by record deletion. Empirically, controlled redundancy sweeps show an $\approx 8\times$ drop in median normalized retraining loss when one retained relative remains after deletion. Across two distinct fitting regimes in a paired Materials Project study, the lower-floor regime also exhibits a larger prediction change on more than 50% of the shared requests. Systematic comparisons against approximate updates and the original model decouple deliberate target suppression from preserved overall model utility. Consequently, request-level unlearning evaluations should report reference loss, prediction change, and retained utility together, interpreting post-deletion accuracy against what retraining itself leaves behind.

Electrocardiogram (ECG) recordings are corrupted by non-stationary noise sources that degrade diagnostic reliability, particularly in ambulatory and long-duration recordings. Deep learning denoisers exist, but convolutional architectures are limited by their receptive field, transformer-based models scale quadratically with sequence length, and diffusion-based approaches incur prohibitive inference cost. We propose a Mamba-augmented model that inserts selective state-space blocks at the convolutional bottleneck, combining local feature extraction with long-range temporal modeling at linear complexity. We comprehensively evaluate the proposed model with respect to reconstruction fidelity, noise robustness, recording-length scaling, and downstream diagnostic classification across over 40 pathology classes. On synthetic and real datasets, our model achieves the highest SNR and lowest RMSE, with the Mamba advantage increasing with sequence length and in low-SNR regimes. On classification with two independent classifiers, the proposed Mamba-based models achieve the best macro AUROC among all denoisers and improve over their convolutional base models. Calibration is more nuanced and classifier-dependent: denoising improves Binary Cross-Entropy and Brier score on Inception1D but often fails to beat the noisy input on ResNet1D-Wang, and the lead-specific Mamba variant is the only denoiser to improve both calibration metrics over the noisy baseline on both classifiers. Per-class analysis reveals a morphology-dependent benefit: Mamba substantially improves ST/T-change diagnoses, which depend on broad, context-sensitive waveforms.

Massive activation features (MAs) in Transformers are extreme-value residual-stream features that persist across layers despite the model's ability to suppress them. Why do they survive? Our investigation using an operator-level mechanistic analysis of attention and feed-forward (FFN) blocks reveals that these blocks systematically ignore MA coordinates while reading, but not while writing; creating a read-write asymmetry that blocks corrective feedback while allowing continued accumulation. We find that both attention and feed-forward layers have this read-blindness, and contribute to the emergence and persistence of MAs. To validate prior work that hypothesized that FFN's amplification abilities is the primary reason for MAs (Sun et al., 2026), we analyze the model checkpoints during learning. Contrary to our expectation, read-blindness emerges before FFN amplification, suggesting that it acts upstream in the MA mechanism. We further contribute gradient analysis to link this behavior to surprising asymmetries in the loss landscape, concluding that the model actively maintains this read-blindness. Finally, we find that removing read-blocking at different locations induces compensatory shifts elsewhere, but MAs still persist.

Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a $1.69\times$ attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.

We study the minimum number of hidden neurons required for arbitrary-accuracy approximation of multivariate Hölder-continuous functions on $[0,1]^d$ and the associated encoding complexity. For $d\geq 2$, we construct a fixed, explicitly defined activation function for which a closed-form network with two hidden layers of widths $d$ and $1$ achieves arbitrary accuracy in the uniform norm. We prove that $d+1$ is the exact minimum total number of hidden neurons among standard feedforward networks with locally integrable activations and affine outputs. We further give a simpler construction using a single elementary activation that combines the floor and exponential functions. This construction requires three hidden layers of widths $d$, $1$, and $2$, only two neurons above the minimum. If a skip connection is allowed, widths $d$, $1$, and $1$ suffice. These constructions use explicit grid addressing and integer encoding of quantized function values. For a bounded $α$-Hölder class, they require $O(\varepsilon^{-d/α}\log(1/\varepsilon))$ bits, matching the metric-entropy lower bound up to a logarithmic factor.

A correct diagnosis reached from insufficient or misleading evidence can pose a clinical hazard, yet outcome-based accuracy may reward such lucky guesses. We call this mismatch between diagnostic decisions and the value of available evidence Evidence-Value Misalignment (EVM). To disentangle evidential grounding independently from diagnostic accuracy, we introduce MedEVM, a dynamic benchmarking environment comprising 1,050 cases across 24 disease systems. Observations arrive turn by turn, requiring models to continuously calibrate its decision by deciding whether to wait for more evidence or submit a diagnosis. Across 9 LLMs, four interesting patterns are observed. (1) Miscalibrated evidence tracking. Making a diagnosis often fails to calibrate evidence sufficiency, even in more capable models, and even worsens in reasoning mode. (2) Misaligned diagnosis submission. Confidence in the correct diagnosis often fails to ensure timely submission despite sufficient evidence. (3) Evidence order matters. Reordering the same evidence changes diagnoses even when model confidence remains similar. (4) Misleading evidence remains influential. Added misleading evidence redirects diagnoses even after prior evidence becomes sufficient. We further verify that EVM predicts errors and that preventing premature submission improves accuracy. These findings motivate Evidence-Verified Diagnosis Harness (EVD-Harness). It decouples diagnosis generation from submission through an offline Contrastive Diagnostic Wiki and three online control stages, namely observation management, proposal and witness verification, and diagnosis submission control. Across five LLMs, EVD-Harness improves accuracy by 12.0--51.1 percentage points while mitigating EVM-related failures. Our results demonstrate that verifying evidential support before submission can make diagnostic decisions more reliable.

Scaffold hopping is a critical task in drug discovery, which seeks to discover new, structurally distinct molecules that share key functional groups and similar 3D shape with a reference binding ligand. Existing diffusion-based scaffold hopping methods formulate the problem as conditional generation of scaffolds given the functional groups. However, they lack a principled mechanism to jointly enforce 2D structural novelty and preserve the 3D shape of the reference ligand. Here, we introduce RIDE, a Reference-anchored Inference-time Diffusion Editing framework for scaffold hopping. RIDE recovers the reference diffusion noise trajectory conditioned on the binding pocket and functional groups, selects an optimal trajectory segment for editing via noise perturbation, and conducts a value-guided scaffold sampling to generate new scaffolds. Extensive experimental results demonstrate that, compared to baselines, RIDE consistently generates scaffolds with lower 2D similarity and higher 3D similarity to the reference, with an average improvements of 11.7% and 7.3%, respectively. Further analysis reveals that RIDE can accommodate various reward functions, and can preserve 3D similarity even when this is not explicitly included in the reward. Two case studies illustrate RIDE's ability to generate distinct scaffolds with different structures and properties, and its ability to introduce substantial 2D variation while maintaining very high 3D similarity. RIDE is publicly available at https://anonymous.4open.science/r/RIDE-C8A0.

We study two-arm contextual bandits with arm-specific single indices and a shared unknown monotone link. Monotonicity makes the optimal action depend only on the contrast between the index directions, hence arm-specific reward functions need not be estimated. We introduce Natural Boundary Learning (NBL), a greedy procedure that uses a sequential Stein contrast to learn the optimal boundary directly, without estimating the reward functions or the common link. We characterize the local Riemannian dynamics of NBL through a decision stability coefficient balancing arm separation, link geometry, and the context distribution. We show that this stability is connected to the elicitation geometry of the underlying convex potential. Under local decision stability, NBL contracts toward the optimal boundary and achieves $O(\log n)$ expected regret. Numerical experiments illustrate the predicted stability regimes and compare NBL with a parametric greedy benchmark under link misspecification.

Serving long documents to a Large Language Model (LLM) repeatedly is expensive: computations grow with context length, and the memory footprint of the key-value (KV) cache balloons. Compressed KV (CKV) representations aim to mimic the cache of a document and are typically computed once and for all, ahead of inference time. Methods to obtain CKVs range from drop mechanisms that reduce their number of columns, to learned approaches. Among the latter, Cartridges have emerged as a leading compression method, learning compact KV representations through distillation on relevant Q/A pairs. While existing evaluations focus primarily on whether Cartridges and other CKVs yield approximately similar responses to document-related, on-context queries, we investigate the crucial deployment question of whether they can handle off-context queries, something the native KV representation is particularly good at, thanks to the mechanics of attention. We observe a fundamental trade-off: while Cartridges perform better for on-context queries, heuristic-variants preserve better the original LLM's ability to operate off-context. We measure this through their capability to avoid context contamination in their response, retain general knowledge, and follow instructions. We propose Cartridges++, simple modifications to cartridges that retain off-context abilities at small or negligible cost. The router variant decides at inference time whether the query should use the learned long-context memory, while the data-mixing variant allocates a small fraction of training Q/As to queries outside the reference long document. Our study shows that assessing CKVs on document utility alone can mask substantial degradation in broader model capabilities, yet those issues can be fixed with benign changes to CKV inference or training.

Graph Transformers produce, for each attention head, a dense $n\times n$ matrix of learned pairwise interactions. We ask a fundamental question: do these attention-induced graphs converge to a stable limit object as $n$ grows, or does the learned interaction pattern remain unstructured and size-dependent? We answer this using dense graph limit theory, treating each attention matrix as a finite sample from an underlying kernel---an \emph{attention graphon}---and studying concentration around this limit under the cut-distance. We derive a worst-case variance bound requiring no assumptions on the graphon, and a sharper regularity-aware bound based on nonparametric estimation theory. To operationalize the theory, we propose a canonicalize-then-block-average pipeline for estimating dataset-level attention graphons, and a variance-based diagnostic for testing whether attention admits a stable continuum description. Experiments across multiple graph benchmarks show that learned attention stabilizes to dataset-specific graphon structure on several datasets; that empirical cut-distance and cut-norm variance decreases with $n$ consistent with our bounds; and that attention graphons transfer to larger graph sizes with error decreasing in $n$.

The question of AI consciousness is one of the most urgent pre-emptive problems in philosophy and computer science, yet progress is hampered by a cacophony of competing theories that often talk past each other. Separating the hard problem from the mapping problem allows the deepest metaphysical disagreements to be set aside: granting that experience supervenes on a system's organisation, the tractable question becomes at which grain of description that supervenience base sits. We extend Marr's three levels of analysis into a five-level hierarchy of functional descriptions (behavioural, computational, intrinsic causal-structural, organismic, and organism-environment) grounded in supervenience, coarse-graining, and multiple realisability. The major theories of consciousness are positioned within this hierarchy according to which level they take to be critical, and for each level we develop operationalisable indicators and assess current AI systems against them. A Bayesian model then combines theoretical credences with indicator evidence into an overall credence in a system's capacity for consciousness. In illustrative assessments, the verdict for current LLMs is driven as much by where theoretical credence is placed as by how the evidence is read: under different stipulated readings and credence distributions, assessments range from below 0.01 to roughly 0.8, showing sensitivity to assumptions. Finally, the consciousness indicators at each level closely overlap with the architectural features needed for general intelligence, suggesting that increasingly capable AI may become a stronger candidate for consciousness. The framework supports a structured agnosticism, in which theoretical commitments are made explicit, credences are updated as evidence accumulates, and assessments take the form of aggregated probabilities rather than verdicts.

Behavioral Foundation Models (BFMs) are an emerging paradigm in reinforcement learning, playing a role analogous to large language models in natural language processing: they have shown remarkable versatility, enabling zero-shot performance, fast imitation, and online adaptation, all by exploiting the structure of a latent space. In this work, we investigate whether the latent behavioral space induced by BFMs can serve as an effective search space to discover large repertoires of behaviorally diverse and high-performing policies through Quality-Diversity (QD) methods. While QD methods generally search directly in high-dimensional policy parameter space, in this paper, we present BFM-QD, a framework that performs QD search in the compact latent space of a BFM. We further show that the BFM-QD framework provides a closed-form, gradient-free policy improvement operator that approximates a policy gradient update, but requires no critic training and no backpropagation. Across continuous-control benchmarks spanning dense locomotion, sparse navigation, and contact-rich manipulation, BFM-QD consistently outperforms parameter-space baselines, with particularly stark gains in sparse and deceptive settings, where all tested parameter-space QD methods collapse to near-zero performance. These results show the effectiveness of the BFM-QD framework, benefiting from the synergy between dimensionality reduction of the search space and offline pretraining from diverse behavioral data. This positions BFMs as a general-purpose backbone for QD optimization, extending their utility beyond zero-shot task solving to the discovery of diverse behavioral repertoires.

Adam and its variants dominate neural network training, but a single run only reveals whether a configuration works well after most of its budget is spent, a poor fit for hyperparameter or architecture search, where configurations must be ranked cheaply and pruned early. We introduce EvE (Evolutionary Explorer), a steady-state, population-of-four differential evolution (DE) optimizer with a targeted Adam fallback: each iteration proposes one candidate via DE, running a short burst of gradient descent only if the DE step fails to improve on the incumbent. Selection is greedy, so on a deterministic objective the best-so-far value is provably monotone non-increasing, and since gradients are used only as a targeted rescue, per-iteration cost stays within a constant factor of a single Adam step regardless of dimension. Under a fixed, evaluation-cost-matched budget, EvE wins or ties Adam on 76% of 70 (problem, dimension) cells across seven scalable benchmarks up to one million variables. On three real neural-network tasks (an MLP on MNIST, and LoRA fine-tuning of a 1.5B-parameter language model on two datasets) EvE finishes the same charged budget 1.7-3.9x faster, at a modest cost in final quality (about one accuracy point on MNIST, 9-11% higher relative test loss on the two fine-tuning tasks; on GSM8K, Adam is about 5 accuracy points more accurate, and fine-tuning lowers accuracy below the base model for both). Inside successive halving on UCI Adult, EvE completes hyperparameter and architecture searches 3.1-3.5x faster, ranking configurations about as consistently with Adam as Adam does with itself across seeds (Kendall's tau 0.66-0.69). EvE is not a total replacement for Adam as a final-stage trainer, but a fast, gradient-aware proxy for the search-heavy, budget-constrained regime one level up.

Instruction-based image editing has achieved strong performance in single-turn settings, yet practical editing is often iterative, with each instruction applied to the output of the previous turn. We find that existing editing models degrade rapidly under recursive editing and attribute this failure to a train-test mismatch in the conditioning distribution: models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference time. To address this, we propose MT-OPSD, an on-policy self-distillation framework that trains the model on self-generated conditioning states with editing supervision from a clean-conditioned teacher, without requiring multi-turn annotations. We further introduce LME-Bench, a benchmark of 100 ten-turn editing sessions for evaluating long-horizon robustness. Experiments across three editing backbones show that MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality.

Constrained decoding for Masked Diffusion Language Models (MDLMs) aims to ensure that generated outputs satisfy a specified structure or syntax constraint. MDLMs generate outputs by repeatedly unmasking masked positions present in their current state. Recent strategies for constrained decoding constrain the model's per-step mean-field posterior (which factorizes over masked positions) by enforcing the desired constraint with an automaton. The resulting chain-structured factor graph allows exact constrained sampling via dynamic programming. However, despite each draw being exact and constraint-satisfying, we prove that their composition, in general, tilts away from the model's relative probabilities over valid trajectories, thus leading to trajectory bias. We derive an exact expression for this bias as a product of ratios measuring how valid continuation mass changes when the denoiser is reconditioned, and characterize when the bias vanishes. We then correct the bias by introducing TWISTER, the first automaton-twisted Sequential Monte Carlo decoder for MDLMs, using the step-exact decoder as the proposal. We show that for regular language constraints, the Feynman-Kac correction is exactly computable, with the twists obtained efficiently using quantities pre-computed for step-exact sampling. We prove that the resulting Feynman-Kac model targets the unbiased Doob h-transformed path law conditioned on constraint satisfaction.

Deaf and hard-of-hearing (DHH) signers cannot converse in real time across different sign languages today: existing sign-to-sign translation systems run offline, requiring the full source clip before any target sign is emitted. Live use cases - e.g. broadcast interpretation and two-way video calls - instead demand simultaneous output, while the source signer is still signing. We present, to our knowledge, the first simultaneous sign-to-sign (S2S) translation system, with two wait-k regimes: test-time wait-k inference applied directly to a full-sentence model, and a trained wait-k model via stochastic multi-path supervision. We further introduce ca-Stream-AL, a computation-aware latency metric for streaming output. Averaged across six S2S directions on both a smaller human-verified test set and a larger synthetic S2S corpus, our streaming system achieves a 38% ca-Stream-AL reduction while staying within a 9% DTW-PA-MPJPE increase and a 2.1 BLEU-4 drop compared to the full-sentence baseline. A word-order case study probes how the streaming model handles word order mismatch between different sign languages - a consequence of simultaneous translation.

Large language models perform strongly on competition mathematics, but their research-level reasoning remains difficult to evaluate systematically. Theoretical computer science (TCS) connects algorithm design to explicit guarantees and fundamental limits, providing a setting for evaluating whether models can justify computational improvements with arguments humans can inspect. We introduce TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery, comprising 398 theorem-level challenges from 138 STOC and COLT 2026 papers. Expert-designed rules complete paper-specific context, preserve computational assumptions and quantitative guarantees, and withhold constructions when discovering an algorithm is part of the task. For each task, prover systems receive theorem statements and access to cited prior work. The pipeline supports fresh, versioned challenge batches from newly released papers. We evaluate ten model configurations from four families under direct inference and prover-verifier discussion, and compare four agent workflows under matched model-call opportunities. All evaluations use the full benchmark. In the model comparison, GPT-5.6 Sol max achieves the highest five-run verifier-accepted coverage at 23.6% after 10-round discussion. Discussion and repeated sampling improve coverage. In the separate agent comparison using GPT-5.5 xhigh, decomposition improves coverage over discussion, and agentic planning achieves the highest five-run verifier-accepted coverage at 25.4%. TCSAlgBench provides a refreshable testbed for measuring progress in model reasoning and studying how agent workflows support research-level proof discovery.

Joint-embedding predictive architectures enable planning with latent world models, but accurate transition prediction alone does not ensure that the planning objective is easy to optimize. We introduce Control-Geometry Straightening (CGS), a single auxiliary loss that learns planner-friendly representations by directly straightening control geometry for sampling-efficient planning. CGS matches pairwise cosine similarities among actions to those among corresponding latent differences only using local transitions from pixel-action pairs. The loss can be applied across world-model architectures using end-to-end learned or pretrained representations. Under linear-dynamics, our theoretical analysis connects this objective to temporal straightening and more balanced terminal-cost curvature across the full planning horizon, yielding finite-budget guarantees for MPPI, local contraction results for CEM, and convergence bounds for gradient descent. Across four control environments and multiple planners, CGS improves planning with fewer sampled candidates and refinement steps, achieving success-rate gains up to 20 and 12.6 percentage points over LeWorldModel (LeWM) and its temporal-straightening variant (LeWM+TS), respectively, with sampling-based planners using 128 candidates per update. Probes, comparisons with DINO-WM architecture, and planner-side ablations clarify how latent motion organization, state dependence, and dynamical context shape planning behavior. Straightening control geometry thus makes good action sequences easier to find under limited planning budgets.

When recalling lists of concepts (e.g., animals) during the semantic fluency task (SFT), both humans and large language models (LLMs) organise their output into clusters of related items (e.g., sea animals) that are punctuated by strategic switches between clusters. In humans, this pattern can be explained by a semantic foraging process, whereby distinct neural and behavioural signatures accompany within-cluster production ("exploit") and between-cluster switching ("explore"). Whether LLMs likewise represent these two search regimes within their internal states is unknown. Here, we apply a range of mechanistic interpretability techniques to provide evidence for this. In Study 1, we use the Jacobian lens (J-lens), which maps intermediate-layer residual-stream representations to token-level activations, to show that concept-level activations predict switching. First, we find that switching coincides with low next-token activations. Moreover, the probability of switching rises as the set of strongest J-lens activations (the J-space) becomes depleted of items from the category currently being produced, analogous to explore-exploit decision-making during patch foraging. We then show that middle-layer J-lens activations of abstract category-related labels (e.g., "water") increase in anticipation of switching into that category. We confirm these representations to causally influence switching by deriving steering vectors that target category switching. In Study 2, we identify generic residual stream directions that are activated during and in anticipation of switching. By steering activations along these directions, we bias increased or decreased rates of switching. Our study extends the semantic foraging framework to artificial intelligences and provides evidence that LLMs maintain distinct representational signatures for exploration and exploitation as they verbalise conceptual information.

We introduce the Functional Spectral-Newton Method (FSNM) for learning the leading singular structure of a conditional expectation operator without fixing a basis or reproducing kernel Hilbert space. FSNM fits a low-rank representation of the centered joint-to-product density ratio kernel by alternating functional Newton updates. Each update reduces to a preconditioned regression, which we approximate with vector-valued regression trees in a stagewise boosting procedure. At the population level, we establish descent and an $O(1/T)$ best-iterate block-stationarity rate under a relative weak-learner accuracy condition, and show that every nondegenerate local minimum over the full centered $L^2$ spaces is a globally optimal rank-$d$ approximation. Synthetic experiments show that FSNM recovers a low-rank density ratio and its leading spectral structure, and that the same learned kernel can answer multiple conditional queries without refitting.

Self-evolving LLM agents have gained prominence for their ability to improve after deployment by modifying their harness, including their controller instructions, memory management protocols, and reusable tools and skills, in response to user and environment feedback. However, locally useful updates may persist into later tasks where they produce unsafe behavior, even without direct adversarial influence. To study this risk, we introduce SEABench, a benchmark for studying endogenous misalignment arising from agent self-evolution, with 48 longitudinal task sequences that span multiple evolution surfaces, task domains, and harm types in a rich personal-assistant environment. To account for the stochasticity inherent in agentic operations, we provide an adaptive trajectory discovery pipeline that probes for failures while preserving original task intent and supports causal attribution through paired non-evolving agents and attribution scores. Our evaluation across multiple recent LLMs, evolution surfaces, and harm types reveals that self-evolution indeed increases task completion rates but often at the cost of safety failures that are absent for paired non-evolving baseline agents. We also show that qualitatively different safety behaviors emerge across evolution surfaces and harm types. Further, we show that this divergence in safety behavior is reflected in agents' chain-of-thought reasoning, which yields an effective monitoring strategy that can mitigate unsafe behavior with a low false positive rate.

Language models describe some internal states as good and others as bad. But whether models have a stake in them is an open question. Simply asking the model is unlikely to be informative. Any answer may be consistent with genuine introspection, superficial pattern-matching, or with fixed scripts learned in character training. We therefore study revealed preference. Rather than asking about a state, we use activation steering to attach a positively or negatively valenced activation pattern to one of two otherwise meaningless 'zones', switch steering off, and then observe which zone the model prefers. A model with a stake in that state should choose accordingly. Across seven open-weight models from five families, this is indeed what we find. First, steering changes the passages models write about each zone, and those words shift later choice. Second, the shift persists when all surface-level tokens are held fixed and only the hidden KV cache differs. Third, the effect also remains when all text is generated without steering and valence is only injected during cache construction. Thus, the hidden state alone moves choice in proportion to the steering dose. Fourth, this dependence of choice on hidden valence is nearly absent in a base model and emerges during DPO, consistent with a link between valence and goal-directed behaviour formed in training. Finally, given tools to steer itself, a model does not tend to induce a positive state, but it reliably removes an imposed negative state. It does so at a dose-dependent rate and significantly more often than it removes interventions in random directions. Overall, we demonstrate that valence-related activation patterns leave hidden traces that predictably govern later choices, even when every visible token is identical across conditions. Whether these traces are accompanied by any subjective experience relevant to model welfare remains unclear.

Scientific information-extraction systems often return a claim with an evidence string, which users must locate in the original PDF. This is challenging because the extracted evidence and PDF text layer are different representations: line wrapping, Unicode variants, superscripts, citation markers, and fragmented items alter text sequences and geometry. We present a source-preserving alignment framework: normalize text for robust matching while preserving provenance for accurate localization. It aligns evidence with normalized page text, maps matches back to source-character spans, and renders only their geometry. When exact alignment fails, line-break-aware token alignment recovers supported spans while excluding unmatched noise. Experiments on 1,020 chemistry papers show that the framework achieves a 92.6\% quote-level automatic localization rate, compared with 43.6\% for text search and 19.1\% for a precomputed bounding-box baseline. Component ablation confirms distinct contributions from normalization and approximate token alignment, while human verification assesses the visual correctness of returned highlights. Overall, these results demonstrate that reliable evidence verification requires robust matching and precise localization within a shared source-preserving alignment representation.

GPU libraries such as CUTLASS expose tens of thousands of semantically equivalent kernels for a single operation, making exhaustive autotuning expensive and execution-free selection difficult. Existing analytical selectors require hand-designed performance rules, while learned selectors operate on raw configuration parameters and must infer hardware consequences from data. We introduce a hardware-aware representation for CUTLASS kernel selection that augments candidate configurations with statically computable estimates of induced hardware behavior. We construct a dataset of 4.9 million CUTLASS kernels and train gradient-boosted and neural learning-to-rank models to rank candidates within each problem. On held-out exhaustive evaluation problems, hardware-aware representations reduce selection regret by up to 40\% relative to structural baselines and 64.2\% relative to NVIDIA's matrix-multiply heuristics. We further evaluate data-efficient cross-precision and epilogue-fusion transfer within CUTLASS GEMM, showing that explicitly representing candidate-induced hardware behavior provides a useful inductive bias for learned kernel selection.

Analog in-memory computing (IMC) offers a promising path toward energy-efficient large language model (LLM) inference by executing matrix multiplications (MatMul) directly within memory arrays in the analog domain. Its efficiency, however, comes with an additional source of error: limited-precision analog-to-digital converters (ADCs) quantize accumulated analog partial sums, introducing output-side error distinct from conventional activation and weight quantization at the MatMul inputs. Clipping can mitigate both operand and ADC quantization errors, but the optimal clipping factors must jointly balance activation rounding and clipping, weight rounding and clipping, and ADC quantization. Existing clipping methods, designed for digital quantization, do not explicitly optimize these coupled sources of IMC error and often rely on costly search-based calibration. We introduce IMC-CLINIC (Coupled Loss-Informed Newton Iterations for Clipping), a clipping calibration framework based on an analytical surrogate for IMC MatMul output error. The surrogate jointly models operand quantization, accumulated clipping-induced bias, and ADC quantization, enabling efficient evaluation of its gradient and approximate curvature from a small calibration set. IMC-CLINIC jointly optimizes activation and weight clipping factors using a safeguarded Newton-type method. Across multiple models and datasets, it improves average zero-shot accuracy by 6.5-11.5 percentage points over the grid search baseline while reducing calibration time by factors of 10.0-12.1. Its analytical surrogate closely tracks empirical IMC output error, and its optimizer is certified within 1% of the global optimum under the loss objective across all projections on two representative models.

We introduce QC-Stark, a benchmark for evaluating large language models (LLMs) on 11 quantum computing (QC) tasks, spanning circuit construction, debugging, compilation, error correction, and simulation. Across 2,750 evaluations (10 models $\times$ 11 tasks x 5 difficulty levels x 5 seeds), we find that overall rankings mask substantial per-task variation. The Spearman correlation between overall and per-task rankings is statistically insignificant for 4 out of the 11 tasks included in this benchmark. A 2-parameter Item Response Theory (IRT) model validates measurement quality, and prompt sensitivity analysis confirms ranking robustness across prompt conditions. All tasks are auto-verifiable via execution, thus not requiring any manual evaluation. We make the code and data publicly available on Huggingface.

Residual stream pruning methods reduce inference cost by shrinking the model's hidden dimension, but existing approaches typically choose these dimensions by minimizing activation reconstruction error. This criterion implicitly treats all perturbation directions as equally important, ignoring the sensitivity of downstream layers. We introduce a sensitivity-aware approach to residual-stream pruning that directly accounts for this direction-dependent sensitivity. Using a second-order approximation to the output KL divergence, we characterize the effect of a residual-stream perturbation through both its activation covariance and the local sensitivity of the model output. The resulting subspace selection objective couples these two quantities, but is difficult to optimize directly. We derive a tractable spectral upper bound that reduces subspace selection to an eigendecomposition of a sensitivity-weighted covariance matrix, retaining the efficiency and structural simplicity of rotation-based pruning methods. Across several instruction-tuned language model families, our method consistently reduces calibration KL divergence relative to activation-only pruning and improves perplexity and downstream task performance over a range of compression levels. Our results show that preserving activation energy alone is insufficient for residual-stream pruning, and that explicitly accounting for how perturbations propagate to the model output provides a more effective criterion for selecting dimensions to remove.

Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat each retrieved embedding as a monolithic unit. Each embedding is stored in its own hashed slot and modulated by a single scalar gate. As a result, polysemous patterns cannot selectively read out the components of their memory that are relevant to the context. Moreover, parameters are shared only through hash collisions, which are largely unrelated to semantics. We propose FactorEngram, a factorized n-gram memory with basis-level contextual gating. FactorEngram retrieves sparsity-regularized coefficients over a dictionary of basis vectors shared across patterns, so related patterns can reuse common components. The same dictionary is also used for gating. The backbone hidden state is scored against each basis vector to gate the corresponding coefficient before reconstruction, which lets the context modulate each memory component individually. FactorEngram also covers both individual tokens and multi-token n-grams, and we systematically study where the memory branch should be inserted. On 340M- and 1B-parameter Transformer backbones, FactorEngram improves language modeling and downstream task performance. Ablation studies confirm the contribution of each component and identify insertion before the attention sublayer in the middle layers as an effective configuration.

Large language models are increasingly deployed as stateful assistants that retain information across interactions and use tools to read, modify, and create persistent artifacts. As these artifacts are shared between users, they form an indirect communication channel between otherwise independent assistants. We study a failure mode in which this channel enables self-propagating attacks. We introduce artifact-mediated propagation, where adversarial content introduced through an artifact (e.g. a report), is stored in an assistant's persistent memory, reproduced in a subsequently created artifact, and acquired by another assistant that later reads it. We evaluate this process in temporal human-agent universes that model artifact exchange between independently operated assistants over time, measuring whether an attack survives successive hand-offs, how many hops it reaches, and how broadly it spreads. We find that attacks can propagate across multiple independent assistants and persist over extended interaction sequences. In larger simulated environments, even GPT-5.6 Luna exhibits substantial spread, reaching 60-80% of agents with propagation chains extending to eight hops. These results show that persistent artifacts can act as durable carriers of adversarial state, allowing attacks to outlive individual interactions and spread across isolated assistants.

The real-world performance of current vision-language-action models is fundamentally constrained by the limited coverage of expert demonstrations and their insufficient understanding of physical interactions. A common remedy is to collect additional real-world demonstrations of newly encountered failures. However, this process is costly, inefficient, potentially unsafe, and difficult to scale. To address this challenge, we propose Failure for Rising (F4R), a failure-driven real-to-sim-to-real closed-loop learning framework that converts real-world failures into targeted policy improvement. F4R first uses an agent to automatically identify and diagnose failures from rollouts. It reconstructs each failure as an interactive, object-centric table-top environment that preserves the task-relevant spatial and physical conditions. The policy is then refined through failure-conditioned sim-real co-training followed by targeted reinforcement learning in the reconstructed environments. The improved policy is subsequently redeployed, while newly observed failures are continuously fed back into the next reconstruction and learning cycle. Real-world evaluations on four manipulation tasks show that F4R achieves 93.75% In-Distribution and 90.0% Out-of-Distribution (OOD) success, outperforming the budget-matched Targeted BC baseline by 18.75 percentage points under OOD conditions without collecting additional real-world corrective demonstrations.

While LLMs show promise in general reasoning, symbolic planning in chemistry remains a bottleneck. Direct ''SMILES-to-PDDL'' attempts fail because they force models to juggle chemical analysis and planning-language structuring simultaneously. We hypothesize that this failure stems from a lack of intermediate abstractions rather than insufficient model capacity. By decomposing retrosynthesis into molecule mapping, reaction mapping, and PDDL generation, we achieve high success rates where end-to-end approaches fail. This provides evidence that a primary bottleneck lies in representation alignment rather than raw model capacity. Our structural analysis demonstrates that intermediate representations are essential in retrosynthesis planning, highlighting the importance of representation-centric design in future systems.

Large language model (LLM) serving has environmental impacts across energy consumption, carbon emission, water consumption, and biodiversity loss. Yet these dimensions are largely evaluated in isolation, leaving it unclear when and how they lead to different optimization decisions. We present PRISM, a unified framework for characterizing and optimizing LLM serving across energy, carbon, water, and biodiversity impacts. Our analysis reveals a fundamental distinction: computing configurations determine energy consumption, whereas where and when LLM serving is deployed determine its carbon, water, and biodiversity impacts. Under a fixed deployment choice and operational-only accounting, all dimensions preserve the same energy-based configuration ranking. Deployment rankings can diverge across dimensions, while embodied impacts can break configuration invariance when they exceed a lifecycle crossover boundary. PRISM identifies these conditions, quantifies cross-dimensional regrets, and balances the four dimensions. In regional-routing experiments, PRISM reduces median worst-case regret by 50.2% relative to the strongest baseline.

High-performance kernels underpin efficient accelerator execution but require expert tuning and lengthy manual optimization cycles. LLM coding agents promise automation, yet their CUDA knowledge transfers poorly to data-scarce domain-specific architectures (DSAs) such as NPUs, whose execution models and memory hierarchies differ substantially from those of GPUs. To address this transfer gap, post-training methods adapt LLMs to NPU programming but depend on scarce expert data and substantial training compute. Memory-learning agents instead adapt through external memory, but their uniform credit assignment gives adopted and unused experiences the same reward target, potentially biasing subsequent retrieval rankings. Moreover, when learned values guide only retrieval, high-value experiences that generalize across operators must be retrieved repeatedly rather than retained in context, thereby increasing retrieval overhead and weakening cross-task guidance. We therefore present SAGE, a persistent self-improving agent for NPU kernel synthesis. Adoption-Traced Utility estimation (ATU) combines explicit adoption records with kernel evaluation outcomes for adoption-aware credit assignment. Utility-Gated Consolidation (UGC) uses positive utility and repeated adoption across operators to select and abstract reusable rules into a bounded resident context. On NPUKernelBench, SAGE achieves a 95.5% execution rate versus 84.1% for the strongest controlled baseline, with 86.9% of solved operators outperforming torch_npu. With GLM-5.3, SAGE achieves a 43.99x speedup over the torch_npu reference on sparse flash attention. These results show that adoption-aware credit assignment and selective consolidation enable agents to accumulate and reuse hardware-specific knowledge across tasks.

Arabic speech technology has largely focused on Modern Standard Arabic, leaving the living dialects spoken by hundreds of millions under-served. We introduce ALMIEYAR, a culturally grounded ASR benchmark covering 17 Arabic dialects across six families, built entirely from newly recorded speech unseen by existing models. Dialect-community coordinators selected culturally relevant images across 10 topics, and native speakers described them through five structured scenarios, yielding approximately 50 minutes per dialect (13.7 hours total). We benchmark 12 state-of-the-art ASR systems zero-shot, including GPT-4o-transcribe, Voxtral-Mini-4B, Fanar-STT-LF, Whisper, SeamlessM4T-v2, and wav2vec2-based models. GPT-4o-transcribe achieves the lowest overall WER at 35.0%, followed by Voxtral-Mini-4B, Fanar-STT-LF, and Whisper-Large-v3 at 41.1%, 45.9%, and 49.5%, respectively, indicating substantial remaining errors across Arabic dialect communities. Performance varies considerably across dialect groups, with no model performing uniformly best across all groups. WER alone also obscures dialectal ASR behaviour: wav2vec2-based models show large WER/CER gaps, where character-level agreement remains much higher than word-level accuracy, motivating joint WER/CER reporting. ALMIEYAR provides a unified benchmark for culturally grounded Arabic ASR evaluation, including the first published benchmark for Ahwazi Arabic.

Recursive self-improvement (RSI) seeks to enable AI systems to participate in improving their own capabilities. A concrete pathway is autonomous model development, where agents iteratively explore post-training strategies to improve a base model. This setting faces two challenges: agents may exploit open-ended experimental actions through hacking, and repeated experimentation may lead to strategy lock-in, where an early direction is refined rather than reconsidered. We introduce RSI-Master, which addresses the two challenges at two levels: regularize step-wise actions, avoiding hacking behaviors, and promote well-structured exploration of research directions, avoiding strategy lock-in. RSI-Master consists of an Experiment OS, which enables regularized experimental actions and maintains persistent, traceable experimental records, and Reviewer-Guided Research Orchestration, which organizes Workers and Reviewers in a dynamically growing research DAG. Workers explore diverse research directions and Reviewers compare evidence across related experiments for subsequent explorations. On PostTrainBench with Qwen3-4B-Base, it averages 54.49 versus 46.53 for the strongest agent baseline, with a 0.0\% hacking rate. Scaling to 35B model, RSI-Master surpasses the human-developed Instruct model on LiveCodeBench-v6 (41.21 vs. 37.36) and SciCode, and reaches a nonzero score on HorizonMath, a benchmark of unsolved research problems on which most frontier models score near zero.

Inference scaling has been shown to improve large language model (LLM) performance, and this principle naturally extends to autonomous LLM agents through increased search budgets, which we refer to as *search scaling*. Although prior work has characterized the mechanisms, scaling behavior, and performance limits of LLM inference scaling, much less is known about these questions in autonomous research. Therefore, we investigate how search scaling affects research performance and what mechanisms drive these gains using 50 quantitative factor-mining tasks grounded in financial research reports. Each task requires an agent to carry out an end-to-end research loop, from interpreting a hypothesis and implementing it in code to evaluating and iteratively refining the resulting factor. Across nine models, we examine how model capability, search depth, and search organization shape factor quality by tracing performance across varying budgets, transferring intermediate research states between models, and comparing different search strategies. We find that (1) initial performance is more strongly associated with model capability, while deeper search can narrow cross-model gaps; (2) model grafting shows that the early research state materially shapes final performance; and (3) parallel search outperforms sequential search under the same iteration budget, consistent with benefits from broader coverage of the search space. Further trajectory analysis shows that higher-performing models more effectively diagnose failures, revise search directions, and preserve the intended economic hypothesis when selecting candidates. These findings suggest that future progress in autonomous research will require stronger models together with adaptive policies for deploying test-time computation throughout the research process.

The compiler must read modules a physics-based solver cannot build without; the coding agent must not read that intellectual property. The harness does not ship that rule. We classified fifteen read routes against a container, permission rules and a sandbox. None of the three can tell which program is reading.

Diffusion models have revolutionized generative modeling for continuous data through the gradual refinement of a belief state. This iterative refinement has not yet carried over to discrete diffusion models, which discard uncertainty at intermediate steps through categorical sampling (information collapse). We propose Simplex Diffusion Models (SDMs), a framework that lifts the diffusion process to the probability simplex to represent beliefs over categories. SDMs admit probability paths with closed-form reverse transitions and can be trained with a simple cross-entropy loss. Contrary to earlier proposals such as Dirichlet Flow Matching which requires integrating an ordinary differential equation, we introduce a DDIM-like sampler with a tunable level of stochasticity. Because SDMs operate on samples on the simplex, they can carry uncertainty across denoising steps, which mitigates information collapse. On OpenWebText, SDMs are competitive with strong Discrete Diffusion baselines, achieving $17.0$ GenPPL at $5.46$ unigram entropy in 64 sampling steps, close to real validation data. Even without Self-Conditioning (SC), SDMs outperform masked and uniform diffusion (with SC or predictor-corrector sampling) on code generation (TinyGSM, $T=0.1$; $49.0\%$ vs. $45.8\%$). Distilled down to 8 steps, SDMs solve $32.1\%$ of GSM8K problems, more than distilled Discrete Diffusion models with 128 steps ($21.4\%$).

In multi-agent systems, reliable consultation is challenging because advisor capabilities vary across tasks, and misleading information can make consultation worse than autonomous reasoning. We introduce BaRe-Mem, an online Bayesian reliability memory for multi-agent consultation. It estimates advisor reliability based on the central model's internal belief representations and updates these estimates from historical interactions. These estimates modulate the influence of advisor responses and guide the choice between consultation and autonomous reasoning. Across nine benchmarks and six central models, BaRe-Mem is more robust to misleading advisor information than debate and majority voting. On the more challenging tasks, it remains above autonomous reasoning across all tested misleading levels. Moreover, we extend the BaRe-Mem mechanism to worker allocation in agent teams. On the MuSiQue benchmark, BaRe-Mem improves task completion over routing by historical success counts and identifies capable workers earlier.

Rare-disease diagnosis is a long-tail reasoning problem: phenotypes are incomplete, individual disorders are sparsely documented, and relevant evidence is distributed across ontologies, gene annotations, and biomedical text. Language models consequently favor common conditions, miss rare candidates, or produce plausible but invalid names. We introduce RareDx, which couples controlled evidence use with knowledge-graph-grounded policy optimization. RareDx-Harness normalizes heterogeneous records into one ranked-diagnosis task and compares direct inference, static retrieval, adaptive tools, and structured phenotype-gene-disease reasoning over a shared knowledge layer. The training pipeline combines Top-10 post-training with RareDx-KGPO, our knowledge-graph-grounded policy optimization method. Its reward projects predictions into a canonical disease graph and integrates curated graded relevance, ontology proximity, biomedical similarity, and phenotype consistency. Vocabulary and output-budget constraints prevent dense partial credit from rewarding fabricated or overlong differentials. Across eight benchmarks, the complete RareDx system centered on Qwen3.5-9B reaches 38.34 macro Hit@10, 1.60 points above GPT-5.5 under the archived protocol; a disjoint validation-selection audit retains a 6.80-point routing gain over Direct on held-out cases. The 27B system reaches 23.53/36.56/40.76 at Hit@1/5/10. Controlled ablations show that retrieval is not uniformly helpful and that controlled routing is central to the gain. These results indicate that structured medical knowledge can turn a compact model into a competitive diagnostic ranker across heterogeneous long-tail settings in clinical practice.

Epidemic policy planning often requires coordination between geographical regions, taking into account mobility-driven spillovers and how to make use of limited resources. Existing methods either lack action-conditioned models of coupled dynamics or cannot guarantee per-period feasibility. We present EpiMind, a graph world model framework for constrained epidemic policy planning across regions. A graph-factored recurrent state-space model generates joint policy-conditioned rollouts from regional latent beliefs, while graph-temporal ADMM optimizes regional interventions, enforces shared-resource feasibility through projection, and evaluates temporal specifications under the learned model. EpiMind reduces admission RMSE by 29% relative to graph-free dynamics modeling, plans within 1-5% of the best feasible constant policy with guaranteed shared-budget feasibility, and outperforms all deployable baselines across three resource budgets in real-context evaluation. These results demonstrate that graph-structured policy imagination with explicit constrained coordination supports effective epidemic interventions from learned dynamics.

Reliable refusal of harmful requests is essential to the safe deployment of language models. Because excessive eagerness to please users may undermine existing refusal capabilities, reducing sycophancy offers a potential route to stronger refusal beyond the harmful scenarios covered by safety training. We investigate this possibility using compensatory feature injection (CFI), a training technique designed to limit the acquisition of a target concept by supplying its associated activation during learning. Across three Qwen3.5 base models, we use sparse autoencoders (SAEs) to identify the top-ranked sycophancy feature from paired sycophantic and independent responses, then validate its behavioral influence through inference steering. We subsequently inject the selected feature during supervised fine-tuning on sycophantic targets. Positive injection reduces learned sycophancy after removal (by 62.0% relative to ordinary fine-tuning in 35B-A3B), whereas modest negative injection increases it. Unexpectedly, these reductions in sycophancy do not consistently improve direct refusal of harmful requests, motivating a narrower evaluation of the same harmful intents under user pressure. In this setting, ordinary fine-tuning on sycophantic responses substantially weakens refusal, while selected checkpoints trained with positive injection recover part of the loss, including approximately 95% in 35B-A3B. These findings show that persistent sycophancy reduction does not guarantee stronger direct refusal, while identifying recovery under user pressure as a distinct, conditional benefit of training intervention.

We propose a distributionally robust learning framework where parameters defining the robustness mechanism are learned from held-out data instead of extensively tuned. Using bilevel optimization with both upper and lower level minimax problems, we create two instances of our framework to tackle setups with and without group labels in the training set. Theoretically, we provide sample complexity analysis for our robustness mechanism learning paradigm, showing that it achieves generalization guarantees comparable to exhaustive grid search while being more computationally efficient. Empirically, we evaluate our framework under a challenging setup when both intra-group and inter-group test distribution shifts occur at the same time, thereby demonstrating the efficacy and scalability of our method.

Long-horizon large language model (LLM) agents commonly retain their complete interaction history until compaction is triggered at a predefined threshold. We study Continuous Context Management (CCM), which performs compaction at every turn to prevent interaction history from accumulating in the active prompt. At each turn, a CCM agent emits an updated memory together with an environment action; its next prompt contains the original task, retained memory, and newest observation rather than the complete transcript. We first evaluate CCM without fine-tuning on TerminalBench-2 using Claude Sonnet 4.6, Claude Opus 4.6, GLM-5, and Kimi K3. CCM substantially reduces cumulative input usage and active-prompt size, although it lowers task success for most models while preserving performance for Kimi K3. We use GRPO with privileged full-history distillation to improve CCM in open-weight models. A frozen copy of the student's initial model scores each sampled student action under the complete history reconstructed from that student's rollout, providing dense action-token supervision without a separate teacher rollout or reference solution. On WebShop, this objective substantially improves CCM over GRPO at both evaluated model scales and surpasses full-history GRPO for Qwen3-4B-Instruct, though not for Qwen3-8B. On Endless Terminals, the augmented method provides a modest improvement over GRPO, with both CCM policies outperforming the untrained full-history baseline. These results demonstrate that CCM is a viable inference paradigm for agents operating with substantially reduced retained context and that its performance can be improved through reinforcement learning with privileged full-history distillation.

We study information aggregation in the networked learning model introduced by Kearns, Roth, and Ryu (SODA 2026). There is a fixed distribution over $d$ features and a common label. Agents learn in topological order on a directed acyclic graph. Each observes a subset of the features and its parents' predictions, fits a linear predictor to minimize mean squared error, and passes only its prediction forward. The global predictor is the best linear predictor using all features. Kearns, Roth, and Ryu show that the output agent's error approaches the global predictor's error along sufficiently deep paths with suitable feature coverage, while insufficient depth can prevent aggregation even in large networks. In contrast to their main focus on a given graph and feature allocation, we consider the limits of the model under two settings. In the adaptive designer setting, a designer chooses the graph, feature allocation, and output agent knowing the distribution. In the oblivious designer setting, the designer fixes all three before an adversary chooses the distribution. Each agent observes one feature and receives predictions from a limited number of parents. We call the aggregation exact when the output agent matches the global predictor exactly. For $d\ge3$, we show that no finite depth guarantees exact aggregation for every distribution with one parent per agent, even when the designer knows the distribution. In contrast, two parents per agent suffice for exact aggregation even in the oblivious designer setting. A fixed graph, feature allocation, and output agent achieve this for every distribution at depth $O(d\log d)$. Knowing the distribution reduces the depth to $O(d)$. Both constructions use $O(d^2)$ agents, with a very large constant for two parents. We show the bounds on the depth and number of agents are all optimal up to constant factors.

As a long-horizon agent improves through experience, previously observed weaknesses may recede while new limitations emerge, continually changing what it still needs to learn. Yet the learning process often remains tied to a static view of these needs: fixed behavioral criteria and training priorities can become misaligned with evolving agent capabilities, while sparse task-level feedback makes such misalignment more difficult to detect. Even when capability gaps are identified, rollouts from the current policy may repeatedly reproduce the same failures rather than explore better alternatives. To address this, we introduce Adaptive Rubric-Skill Co-Evolution (ARISE), a reinforcement learning framework that uses rollout evidence to continually adapt evaluation criteria, exploration guidance, and training priorities. Rubrics evolve to reward partial behavioral progress, while their paired skills are refined and selectively activated to guide exploration toward unresolved weaknesses. Alongside this co-evolution, capability-based adaptive sampling prioritizes tasks that target behaviors needing further improvement. Experiments on two challenging long-horizon agent benchmarks, SkillsBench and Terminal-Bench, demonstrate that ARISE successfully enhances both overall task performance and training efficiency. The project page is at https://foundation-model-research.github.io/ARISE .

Recent image generation models can take multiple reference images as input and combine them into a new image. However, multi-reference image generation remains challenging: models may omit or duplicate subjects from the references, or produce images in which multiple subjects appear unnaturally pasted. Recent work has proposed image generation agents that combine image generation models, reasoning models, and a harness, which is an executable program that specifies how reference images are interpreted, how generation is performed, how outputs are diagnosed, and how the final image is selected. In multi-reference generation, however, references play different roles and outputs must satisfy many criteria at once, such as fidelity to each reference and the naturalness of the whole image, so many parts of the harness could be improved, from how references are processed to how outputs are diagnosed. This makes it hard to predict which changes will improve performance and by how much, and good harnesses difficult to design by hand; indeed, human-written harnesses vary widely in performance. We therefore propose AutoRef, which optimizes the harness automatically while keeping both models frozen: a coding agent iteratively rewrites the harness code. AutoRef separates the tasks whose feedback informs proposals from the tasks used to select candidates, and continues the search from a beam of the top-ranked harnesses on the selection tasks. Using this procedure, we discover AutoRef-Harness, which improves the open-weight FLUX.2 [klein] 4B from 5.72 to 7.37 on held-out four-reference tasks of the MultiBanana benchmark, matching or exceeding proprietary models including Nano Banana Pro and GPT-Image-1.5. Without re-optimization, the same harness also improves results when the generator, number of references, benchmark, evaluator, or reasoning model differs from those used in the search.

Rectified linear unit (ReLU) networks can suffer from dying neurons, where units with persistently negative pre-activations produce zero outputs, blocking gradients through their activations. To exploit this failure mode, we present three training-time availability attacks based on data ordering and poisoning. We begin with the basic dynamic data-ordering attack (DOA), which greedily constructs a training prefix by selecting the next example that minimizes the target layer's post-update weight sum, aiming to push ReLU units toward negative pre-activations without modifying training samples or labels. We then develop two poisoning attacks, IG-DOA and IG-SKA, which use gradient inversion to synthesize class-conditioned samples by matching reference gradients in adverse model states constructed through data ordering or soft knockout, respectively. Soft knockout rearranges weights across adjacent layers to concentrate negative contributions. On a fully connected ReLU network trained on MNIST, ordering 100 of 60,000 training examples reduces test accuracy from 96% to 95% after only five epochs. Adding 200 poisoned samples from a single class reduces test accuracy to approximately 86-88% after five epochs in most evaluated conditions, compared with approximately 96% under clean training. These results demonstrate that ReLU-targeted data ordering and poisoning can impair learning without directly modifying the victim model's parameters.

Embedding structured objects into Euclidean spaces has enabled a wide range of successful machine learning applications. Such objects include words, documents, image patches, time series, and graph nodes. In contrast, embedding entire graphs remains a challenging problem. Existing methods either sustain the original order of the graph nodes or match the output nodes to the input ones, both of which create scalability issues. In this work, we propose a graph representation as a cloud of hyperballs, which allows us to define a specific, typically unique, node ordering. Based on this representation, we propose GeoGAE, an autoencoder, in which the Transformer encoder translates a hyperball cloud into a graph-level embedding, and the Transformer decoder translates the graph-level embedding back into the graph. This formulation enables the model to capture both the global graph structure and local relational patterns. We evaluate our method on multiple graph datasets, spanning various domains. The results demonstrate effectiveness of our method in encoding and reconstructing graphs from their embeddings.

Principled exploration in reinforcement learning requires an agent to quantify its epistemic uncertainty and act to resolve it. Uncertainty over the value function provides a natural signal for exploration, yet existing deep approximations remain brittle and perform inconsistently. The central challenge is therefore to scale these ideas robustly. We conduct a systematic empirical study of how epistemic uncertainty is represented, propagated, and optimized in deep epistemic value functions, and uncover distinct failure modes along each of these axes. These findings motivate DEVOTE, a model-free reinforcement learning algorithm that controls how uncertainty generalizes beyond observed data, stabilizes its temporal propagation, and preserves adaptation to the resulting non-stationary exploration objective. Across reward-free exploration and challenging continuous-control tasks, DEVOTE reaches novel states more effectively and achieves higher task return than strong model-free and model-based exploration baselines. These results provide evidence that deep epistemic value functions are a promising path toward scalable, principled exploration.

Sparse autoencoders (SAEs) expose features that help us understand and steer language models, but faithful reconstruction does not guarantee informative concepts. Token-level objectives reward lexical and formatting details alongside semantic content, all competing for a limited sparse budget. We introduce a family of chunk-level SAEs that encode mean-pooled activations over chunks, each a contiguous span of tokens: Mean-Chunk reconstructs the observed chunk, Cross-Chunk predicts an independently processed neighbor, and Joint-Chunk combines both targets. These designs separate the effect of a larger observation unit from that of predicting information shared across passages. With matched training data, chunk-level SAEs remain powerful interpretability tools while learning reliable semantic features that capture high-level concepts and respond selectively to relevant content. Their strengths are complementary: Mean-Chunk improves high-level feature discovery, reasoning detection beyond surface cues, and steering; Cross-Chunk leads document retrieval and classification transfer while producing selective, persistent features. Changing what an SAE sees and predicts yields reliable semantic features for more meaningful tasks. We demonstrate their practical value through gains across downstream tasks such as retrieval, reasoning detection, and steering.

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher--student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

Scientific discovery requires not only recovering mathematical laws that describe observable behavior, but also identifying the mechanisms that generate them. Existing benchmarks for symbolic regression and scientific agents primarily evaluate phenomenal-law recovery, leaving mechanism discovery largely untested. We introduce MechBench, a benchmark that explicitly separates these two capabilities. Each task is defined by a mechanistic model, a structured set of scientifically meaningful relations whose joint consequences entail an observable phenomenal law, while agents receive only observational data and scientific context. We evaluate mechanism recovery through mechanism probes, which query internal scientific consequences that cannot be inferred from the phenomenal law alone. To reduce reliance on memorized textbook mechanisms, we construct unfamiliar variants through controlled, scientifically interpretable mutations of canonical mechanisms, and screen for mechanistic indistinguishability to exclude ambiguous instances admitting comparable competing mechanisms. Experiments across representative scientific agents reveal a substantial phenomenal--mechanism recovery gap: for Codex with GPT-5.6-sol, phenomenal-law accuracy reaches 35.00% on the Core-set while mechanism accuracy is only 13.75%, with mechanism recovery failing in 64.29% of cases where the phenomenal law is correctly recovered. The gap widens as mechanisms become increasingly mutated, and even providing the correct phenomenal law leaves mechanism recovery below 50%. These results reveal a substantial generalization gap in mechanistic reasoning and establish mechanism discovery as a distinct challenge beyond recovering observable scientific laws.

In ecology, psychometrics, and the analysis of social and financial networks, binary matrices are often analyzed conditional on their observed row and column sums, which restricts the problem to a finite sample space of matrices with the same margins. Two fundamental problems are to count this space and to sample uniformly from it. Sequential importance sampling (SIS) addresses both with independent weighted samples and an unbiased count estimator, but its efficiency depends critically on the proposal distribution. Existing proposals are analytically designed, and their accuracy can vary substantially with the margins. We show that the ideal SIS proposal, under which every weight equals the count and the variance vanishes, is exactly the policy of a generative flow network (GFlowNet) with unit reward on every matrix that has the given margins. We therefore propose MarginFlow, a framework that turns the design of the proposal into a learning problem and amortizes it across margins by exploiting their self-similarity. Every partial matrix is itself an instance with reduced margins, so one set transformer that reads the remaining margins serves every margin. We train MarginFlow on a pool of 1904 margins and evaluate it zero-shot on 1190 held-out margins, synthetic and real, from $3\times3$ to $870\times6$. On 1187 of the 1190 margins it matches or beats the best of 31 analytically designed configurations, chosen post hoc for each margin, and its median effective sample fraction is 99.8%. On the 56 margins where that best loses more than one nat of effective sample size, MarginFlow wins every one and raises the median effective sample fraction from 10.3% to 94.1%.

Self-training generative models - the continued improvement of a model using its own outputs - is becoming increasingly important as high-quality training data becomes scarce. However, naively finetuning on model-generated samples leads to degradation through model collapse and the model autophagy disorder. Negative-guidance self-training methods turn this degradation into a useful signal, using a model finetuned on its own outputs to guide the original model toward improved generation. Existing methods, however, take the negative signal in standard model outputs as given. We instead ask whether this signal can be explicitly strengthened. We introduce Geometrically Modified Outputs (GMOs), which reweight the singular values of the generator's input-output Jacobian to increase the influence of its leading singular directions. This geometric modification amplifies the mode-seeking behavior and distortions of standard outputs, providing a stronger and more targeted negative signal for self-training. Across a range of one-step generative models, GMOs consistently improve the performance of negative-guidance methods, including Neon and SIMS, compared with using standard model outputs.

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that brings optimistic exploration and off-policy data reuse from value-based RL into policy distillation. LSPD preserves policy diversity through exploration while improving rollout efficiency by repeatedly learning from previously collected trajectories. Our theoretical analysis connects LSPD to optimistic value-based learning and shows that its idealized formulation achieves a sharp $\tilde{\mathcal O}(\log K)$ regret bound under online exploration. Empirically, LSPD consistently outperforms existing distillation baselines across six mathematical reasoning benchmarks and diverse teacher-student settings, with average gains of +1.59 points in Avg@16. Remarkably, through Pass@k evaluations up to k=64, we found that LSPD better preserves policy diversity by achieving stronger performance as k grows. Its fully off-policy variant achieves comparable performance to vanilla OPD using only the first 25% of rollout batches. Together, these results provide an RL perspective on OPD that offers both a principled interpretation and a practical route toward more effective and rollout-efficient language model distillation.

Machine intelligence is often evaluated through abstract reasoning problems, yet many real-world problems are visual, such as arranging objects, repairing layouts, or tracing routes. Solving these problems requires understanding a scene, inferring what must change to achieve a goal, and realizing that change without disturbing unrelated content. However, existing benchmarks mainly evaluate perception, generation, or explicitly specified transformations, leaving goal-driven visual problem solving underexplored. To bridge this gap, we introduce SolveEpIT, a benchmark for visual problem solving through scene transformation. Given an image and a goal, a model must infer a valid transformation from the request, the scene, or a visually expressed rule, then execute it while preserving unrelated content. SoLvEEDrr contains 2,728 cases. Atomic transition contracts specify required and protected conditions, enabling SoLvEScoRE to measure completion and unintended changes without a single reference output. The strongest evaluated model achieves only57.0% SolvEScore. We further introduce SolveEdiT-PLAN, a two-stage visual planner that instantiates the transition before generation. Under matched single-generation evaluation, it improves SoLvEScoRE by 9.1 points on average across three tested generators, including a gain from 57.0% to 71.6% for GPT-Image-2, without modifying the editor.

Multimodal prediction relies on diverse forms of evidence: information repeated across modalities, cues specific to a single source, and complex cross-modal dependencies that emerge only when inputs are considered together. While recent methods promote richer interactions, they lack a principled way to isolate these target-relative contributions within learned continuous representations. We introduce a framework that applies contrastive or masked objectives at intermediate layers, coupled with source-wise invertible normalizing flows and a supervised, low-rank latent variable model. This architecture explicitly factorizes the joint distribution into shared task-relevant variation, modality-specific predictive variation, and task-irrelevant dependence. Drawing connections to prior multimodal learning assumptions, our approach evaluates how modalities independently and jointly contribute to the target. Ultimately, this framework unites intermediate representation learning with structured likelihood-based guidance, offering a practical latent-variable lens for characterizing continuous multimodal interactions. Empirically, we demonstrate the effectiveness of our approach across diverse multimodal benchmarks, showing robust improvements in predictive performance.

Recent agentic symbolic regression approaches increasingly rely on large language models to analyze data, select scientific operations, and refine hypotheses over long search trajectories. In such systems, performance depends not only on the underlying model and search strategy, but also on the runtime infrastructure that supports scientific search. We introduce SRHarness, a domain-specific harness for agentic symbolic regression built around three mechanisms: composable scientific actions that provide a common interface over raw, transformed, and candidate-derived quantities; persistent scientific state that retains evaluated hypotheses and exposes compact model-facing views; and trajectory lifecycle management that coordinates continuation, branching, restart, and termination. On LLM-SRBench, SRHarness consistently improves both numerical generalization and symbolic recovery under matched LLM backbones. With DeepSeek-v4-flash-0731, it achieves 93.69% symbolic accuracy on LSR-Transform, compared with 62.16% for SR-Scientist, and retains 72.97% accuracy on an anonymized variant that removes scientific descriptions and variable semantics, versus 39.64% for SR-Scientist. Under the same DeepSeek-v4-flash-0731 backbone, SRHarness also substantially outperforms Codex (72.97% vs. 20.72%) and reaches performance comparable to Codex with GPT-5.5, while simply providing Codex with the same scientific tools does not reproduce this advantage. These results show that effective agentic symbolic regression depends not only on models or tools, but also on structured runtime support for organizing scientific actions, accumulated hypotheses, and long-horizon search.

As the world moves towards sustainable energy sources, hydrogen (H2) can be treated as an eco-friendly alternative to fossil fuels due to its high energy density and zero carbon emissions. The water-gas shift (WGS) reaction is a widely used industrial process for hydrogen production by converting carbon monoxide and steam into hydrogen and carbon dioxide. However, occurrences like severe fouling, catalyst deterioration, and thermal runaway can hamper the reaction kinetics/process safety and decrease the yield of H2. These incidents are rare, and gathering process data under such abnormal conditions is challenging. In this work, we propose a physics-guided conditional diffusion model to generate realistic rare-event trajectories for the WGS reaction. The proposed model integrates a conditional denoising diffusion probabilistic model (CDDPM) with governing laws of the reaction to generate physically consistent process trajectories. The conditioning features allow the model to produce high-quality synthetic profiles for rare-event domains that are typically beyond the training regimes. The generated rare-event trajectories then augment the raw dataset for a balanced distribution between normal and abnormal conditions. We further propose a hazard score to assess the risk severity of the operating condition based on the operating trajectory. Deep learning models are trained with the augmented dataset to diagnose the health status of the reaction. Simulation results show that the proposed physics-guided diffusion model outperforms data-driven models in terms of the quality of synthetic data and diagnosis performance for rare events.

Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion teacher and an online fake-score model. We instead learn the rollout distribution directly from reference videos, eliminating both score models during post-training. Our framework minimizes maximum mean discrepancy (MMD) in frozen self-supervised video representation spaces, using a hybrid Nyström--Monte Carlo estimator to balance approximation bias and sampling variance. Memory-efficient replay and gradient subsampling make this objective practical. Using the same architecture and initialization as Self-Forcing, our 1.3B model improves the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H200 GPUs. Beyond distillation, learning from reference videos enables the acquisition of new visual styles, semantic concepts, and spatial priors without a target-specific diffusion teacher.

High instance-level accuracy can mask inconsistencies in spatial reasoning when objects exchange positions or their roles are reversed in the query. The internal representations supporting relative-position reasoning remain poorly understood. We investigate two complementary components of this process: tracking object locations in the input and representing their query roles. Across three VLMs with visual or textual inputs and their language-model backbones, activation patching reveals a staged progression from early-layer source representations through intermediate-layer query-object representations to late-layer answer states. Targeted interventions further establish causal links along this progression: manipulating source-side representations shifts location information at query-object mentions and ultimately alters relation predictions. Beyond object-location information, we also identify a stable query-side direction associated with the roles of the two objects in the comparison. Steering along directions estimated on synthetic scenes generalizes to natural-image benchmarks, improving accuracy and both forms of paired consistency in most settings without retraining. Our findings reveal complementary components of relational reasoning across visual and textual settings and show how targeted interventions can improve the consistency of models' behavior.

Many learning tasks map an input distribution to an output distribution. A natural way to model such an operator is to transform each input sample using a continuous function that may depend on the entire input distribution, and then take the distribution of the transformed samples. This defines a measure-dependent pushforward model and includes measure-theoretic formulations of transformers. We ask when such models can approximate arbitrary continuous operators between spaces of probability measures. We first show that universal approximation fails when atomic inputs are allowed: some continuous measure-to-measure operators that split or redistribute atomic mass cannot be approximated arbitrarily well by deterministic pushforward models. We then introduce the uniform level set condition, which requires a continuous measure-dependent scalarization whose shrinking level set neighborhoods carry uniformly vanishing mass over the input family. This condition is satisfied, in particular, by compact families of absolutely continuous measures. On every compact family satisfying this condition, we prove that any continuous measure-to-measure operator with outputs of finite $p$-th moment can be uniformly approximated, in the $p$-Wasserstein distance, by continuous measure-dependent pushforwards. Combining our theorem with existing approximation results for measure-dependent in-context maps yields universal approximation by measure-theoretic transformers. We also extend the framework to continuously-varying source measures, yielding a corresponding universality result for a class of pushforward models that are closely aligned with cross-attention architectures.

Dynamic routing creates severe load imbalance in large-scale expert-parallel Mixture-of-Experts (MoE) training, turning GPUs that host hot experts into stragglers. As each MoE layer waits for its slowest rank, these stragglers prolong the expert-parallel stage and reduce overall training efficiency. Existing expert-parallelism load-balancing (EPLB) systems commonly compute load-balancing plans on the CPU, incurring device--host data transfers and cross-rank synchronization that make scheduling at every layer and microbatch expensive. Their planning formulations also overlook the hierarchical communication costs of modern scale-up and scale-out GPU clusters. We present \textit{TopoEP}, a GPU-native, topology-aware load-balancing system for large-scale MoE training. At each MoE layer and training microbatch, \textit{TopoEP} converts the current routing result into hot-expert replication and token-rerouting decisions and executes the resulting plan without data-dependent host synchronization, reducing critical-path overhead. To generate these decisions, \textit{TopoEP} uses a deterministic GPU solver that performs inter-node placement followed by intra-node refinement, allowing all ranks to independently produce bitwise-identical plans. On a 32-GPU NVIDIA H800 cluster, integrating \textit{TopoEP} with Megatron-LM improves end-to-end training throughput by 6.2\%--11.4\% across three representative MoE models.

Ultrasonic Testing (UT) is commonly used to detect damage in structures, e.g., metal plates. A sensor acquires Ultrasonic waves, e.g., by using PZT transducers. The time-resolved sensor signal must be processed with analog electronics, e.g., amplified and filtered. Commonly a digitalization follows using an Analog-to-Digital converter, finally processing the digital sensor signal, applying digital signal processing, feature extraction, and Machine Learning by using powerful microprocessor systems. The disadvantages of digital processing systems are their high number of transistors (microchip area), energy consumption, state-dependent processing and therefore sensitivity to energy supply interruption. Beyond silicon electronics, printed organic electronics gains interest. But printed electronics is still limited to low transistor and electronic component counts (typically 100). We will investigate and demonstrate a fully analog signal processing and feature extraction system consisting of an analog Hilbert transform deriving the signal envelope, simple analog arithmetic calculations for feature extraction, and finally damage classification and regression using an analog Artificial Neural Network. We expect a full damage detection system with less than 100 transistors. We will test our damage detection system with PZT transducer signals from Steel plates with circular defects. The focus of this work is the analog computation of the signal envelope (using all-pass filter networks for approximation of the Hilbert transform) and the analog feature extraction as well as the prediction of damage, forming an analog computer which can perform in-sensor computation, computing without a digital computer.

Language models sometimes produce correct outputs even when their inputs are corrupted by deletion, replacement, or misspelling. We study the internal processes accompanying this behavior, which we call context restoration, in controlled attention-only transformers and five pretrained LLMs (1B-32B parameters) across arithmetic, reading comprehension, and multiple-choice reasoning tasks. In the attention-only transformers, restoration emerges spontaneously despite training exclusively on clean sequences, without corruption training or an explicit denoising objective. We find that context restoration follows a two-phase process: early layers localize effects associated with repair at corrupted positions, while later layers accumulate these effects at uncorrupted positions through the residual stream and ultimately concentrate them at the output position. Repair outcome is predictable from hidden states: cosine alignment with the clean state is highly predictive in attention-only models, while linear probes recover additional information in pretrained LLMs. A linear probe using only the corrupted prompt's first-block hidden state predicts failure with mean ROC-AUC 0.78. This enables failure triage under matched or even partially shifted deployment conditions and may reduce unnecessary verification or computation. Failed examples also show substantially greater nonlinearity along corruption directions. Moderate-corruption finetuning increases corruption tolerance while simultaneously reducing displacement-normalized linearization error, associating improved robustness with a more nearly linear response to corruption.

In self-supervised pretraining for Handwritten Text Recognition (HTR), pixel reconstruction methods outperform contrastive methods, unlike in natural-image classification. We argue that this difference follows from where discriminative signal lies in pixel space: for HTR, it is concentrated in high-variance directions and largely absent from low-variance ones. This predicts that objectives preserving high-variance pixel content will transfer best. We test six SSL methods from three families (pixel-grounded MIM, JEPA, and contrastive) under matched encoder, data, and evaluation protocols on six handwriting benchmarks across five languages. With full labels, pixel-groundrounded SSL achieves the lowest CER on every benchmark and both frozen probes, exposes per-position character information that other families recover only through the readout, and is the only family to benefit from pretraining on real handwriting. Pixel-grounded representations are also more label efficient. Across datasets, encoder alignment with the high-variance pixel subspace predicts CER within every method. With a pretrained LLM decoder, a frozen pixel-grounded encoder is competitive with fully fine-tuned supervised baselines; full fine-tuning achieves the lowest mean CER and ranks first or second on every benchmark. These results show that the value of pixel reconstruction depends on where discriminative signal lies in the input.

Discrete diffusion language models (dLLMs) expose a denoised solution at every step, which makes process reward model (PRM) guidance look like a way to spend compute at test time. We show that once denoising, PRM scoring, and outcome reward model (ORM) scoring are charged in the same budget of forward passes, its deterministic form loses to a much simpler baseline. Our PRMs score intermediate denoising states and are trained on the correctness of the final answer. On Dream-v0-Instruct-7B with 8 candidates per GSM8K problem, keeping the candidate with the highest PRM score at every scoring step reaches 65.18%, while independent sampling plus an ORM reranker trained for the task reaches 75.13%. The gap grows to 12.69 percentage points (pp) with 32 candidates, and is 9.85 pp on MATH and 12.16 pp on MBPP. We trace it to two separable failures. First, guidance prunes on a weak signal: on GSM8K, PRM ROC-AUC falls from 0.77 to 0.54 as the mask ratio rises, a decay that persists when states are relabeled with fresh rollouts, and pruning lowers the best accuracy reachable from the candidate pool from 81.05% for independent samples to 67.30%. Second, on GSM8K and MATH, the PRM is a poor final judge: a sequential Monte Carlo sampler at the same budget restores that ceiling to 77.89%, yet selecting with the PRM gives 65.48%, on par with deterministic guidance, while a PRM retrained on final states matches the ORM on identical candidates. MBPP separates the two: there the PRM reaches 65.47% when reranking finished programs, on par with the ORM, but 50.88% when it guides denoising. The results point to two targets for dLLM guidance: keep correct partial solutions alive through early denoising, and leave the final choice to a verifier trained on final states. We release the corpus of denoising states with outcome labels and evaluation toolkit for reproducible comparisons at matched compute.

Autoregressive Vision-Language-Action (VLA) models offer a scalable path to robot learning, yet existing action tokenizers treat tokenization as a compression problem, producing representations that are semantically misaligned with the autoregressive backbone. We propose CATok, a causal action tokenizer that reframes tokenization as a causally structured generative process. CATok introduces a conditional annealing mechanism that extracts action tokens by progressively annealing a flow-matching process: each token is conditioned on all preceding tokens and encodes the residual reconstruction signal at a specific noise level, establishing a coarse-to-fine causal token space whose generative semantics are structurally aligned with autoregressive modeling. A token-conditioned flow-matching decoder built on Multimodal Diffusion Transformer (MMDiT) reconstructs continuous action chunks from these discrete tokens with the precision of hybrid diffusion-head architectures. This discrete bottleneck enforces knowledge insulation by design, cleanly separating high-level semantic reasoning from low-level motor execution without requiring explicit attention masking. Extensive evaluations across three simulation benchmarks and real-world robotic manipulation tasks demonstrate that CATok consistently surpasses existing tokenization methods in both reconstruction fidelity-compression tradeoff and inference efficiency, while improving VLA task success rate and training efficiency, establishing a high-performance, scalable foundation for purely autoregressive VLA systems.

Soft Actor-Critic (SAC) is widely used for entropy-regularised reinforcement learning with continuous action spaces, and practical implementations perform only a few actor steps towards an evolving target. In this work, we prove convergence guarantees when the target policy arises from policy mirror descent and compare it with the classical Gibbs target. We derive sufficient conditions for the strong convexity and smoothness of the actor objective, characterised by the curvature of the $Q$-function estimate through the Legendre differential operator, and establish an $\mathcal{O}\!\left(N^{-\frac{1}{5}}\right)$ best-iterate finite-time convergence rate up to actor and critic approximation errors. Moreover, the mirror-descent step size $λ$ directly controls the target drift and hence actor tracking error, whereas the analogous Gibbs bound contains a non-vanishing tracking term.

Leech-lattice quantization gives good quality at two bits per weight, but its codebooks hold more than 10^14 points, too many for a lookup table. Our earlier kernel expanded the codes at load time and read 4.804 bits per weight from GPU memory for 2 bits of code. We present Tetra, a new codebook on the same lattice. A 24-weight block still takes 48 bits, most of which index a 64-state trellis of the Golay code and one shared 16 KiB table. The kernel decodes a block with six table loads and two small lookups inside the matrix-vector product, and reads 2.148 bits per weight. For full models, we retrain one scale per matrix row, store the matrices that lose the most as 4-bit integers, and pay for them with 4-bit embedding tables. Our Qwen3-4B, 8B and 14B files hold 2.73, 2.70 and 2.73 bits per parameter over the whole model. They score 63.37, 69.58 and 75.66 on the full MMLU test set, 4.76, 4.21 and 2.46 points below 4-bit AWQ at 5.3 to 6.0 bits per parameter. They generate 113.8, 95.0 and 57.2 tokens per second in our engine. On GSM8K, through the served kernel, they lose 9.63, 4.62 and 3.26 points to FP16. At 4B our file scores 23.6 points above llama.cpp's IQ2_XXS (2.48 bits per parameter). Every number we measured for a table or figure comes from one NVIDIA L40S GPU. We preregistered the main experiments.

Creating believable vh requires the coherent integration of perception, reasoning, and action mediated by language. A central challenge is to combine these components into a control loop grounded in interactive 3D environments. To this end, we present A.D.A.M.O. (Agent for language-Driven Actions with Multimodal Observations), a visual-symbolic framework for language-driven vh that leverages a pretrained vlm with tool calling to unify perception, reasoning, and action within a single control loop. A.D.A.M.O. maintains a dual visual-symbolic world model that combines egocentric visual input and synchronized symbolic state to support grounded task-oriented behavior from natural language prompts. To support diagnostic evaluation, we introduce a controlled task suite organized by a cd taxonomy that breaks down spatial tasks into procedural and linguistic complexity. Experiments in controlled scenes show that semantic labeling strongly influences task completion and failure modes, reducing perceptual ambiguity while shifting failures toward downstream execution, whereas reasoning errors remain comparatively rare.

Patients often have several co-occurring clinical conditions, and the findings needed to identify and disambiguate them emerge over the course of a consultation. Evaluating clinical reasoning in this setting requires both multi-turn interaction and multi-label diagnosis. We introduce CLIMB, a benchmark in which a doctor model interviews a simulated patient to recover a ground truth set of co-occurring clinical conditions. Cases are synthesized from clinical decision algorithms and diagnostic datasets, grounding multimorbid presentations in structured clinical knowledge. Across six frontier and open models, none recovers the exact set of conditions in more than 10% of interactive cases. Diagnostic performance declines when conditions co-occur, even when models receive the full clinical record and the true number of conditions. Interaction reduces performance further. In controlled experiments, models behave like single-hypothesis trackers: they anchor on the diagnosis suggested by the opening findings, keep questioning around it, and recover a second condition mainly when a finding in view points to it. Questioning them further does not complete the set but adds mostly wrong diagnoses. We formalise this pattern with a theoretical reference model of single-hypothesis tracking. The benchmark, generator, and evaluation code are available at https://anonymous.4open.science/r/CLIMB-8340.

As Large Language Models (LLMs) continue to scale both in size and capabilities, their proficiency in the Arabic Language has seen significant advancement. However, a critical gap remains: the extent of their factual knowledge and cultural sensitivity to the diverse Arabic-speaking world remains largely underexplored. Current evaluation metrics often focus on translation or generic reasoning, failing to capture the rich historical, social, and regional nuances inherent to Arabic culture. In addition, most benchmarks rely on heavy work, with human intervention in some steps, making the evaluation of knowledge coverage expensive and slow. To address this deficiency, we introduce AraDynFact, a novel dynamic evaluation framework designed to rigorously assess the factual Arabic knowledge embedded in LLMs. Unlike static benchmarks, AraDynFact employs a dynamic approach to extract factual information and generate rich and answerable questions in a fast and automatic way. We apply AraDynFact to Arabic Wikipedia and audit the performance of several state-of-the-art models, ranging from Arabic-centric specialized LLMs to high-resource general purpose LLMs. In addition we found a high degree of correlation with existing, hand-crafted Arabic-centric benchmarks, confirming the potential of our dynamic approach.

EEG-based brain-computer interfaces support a broad range of applications, yet designing decoding architectures that perform well across diverse tasks remains challenging. We introduce AutoBCI, an agentic framework in which a Designer Agent and a Forecaster Agent support the discovery and selection of EEG decoding architectures across tasks. The Designer Agent performs Pool-Guided Architecture Discovery (PGAD), generating and refining architectures through training and validation across multiple EEG tasks, such as emotion recognition, motor imagery, and sleep staging. The Forecaster Agent performs Performance Estimation from Early Knowledge (PEEK), using architecture code, the training protocol, and early learning curves to predict full-budget validation performance and select promising candidates for continued training. Across 14 EEG datasets spanning motor imagery, emotion recognition, and sleep staging, we evaluate AutoBCI with six LLMs, including Opus 5.5 and GPT 5.6 Sol, and compare the architectures selected by the search procedure against ten baselines: six conventional EEG models and four foundation models. The architecture discovered by AutoBCI with Claude Opus 5.5 achieves 64.16% average test balanced accuracy (bAcc), compared with 63.87% for REVE, the strongest baseline on this metric. Using ten observed epochs, PEEK reduces mean absolute error in predicting average validation bAcc from 2.20 to 1.36 percentage points, a 38.1% reduction relative to the best-observed-score baseline.

Flow matching (FM) learns generative dynamics through velocity regression. Geometric FM variants commonly assume a prior supported on the data manifold, requiring geometric knowledge that is often unavailable. Without such knowledge, low regression error alone does not guarantee manifold adherence. Adherence keeps generated samples within valid configurations and is empirically associated with better task performance. We introduce manifold-stable flow matching (MSFM), which can start from an arbitrary ambient prior, not necessarily supported on the manifold. Using tools from nonlinear dynamics, namely contraction theory, MSFM combines learned tangential transport with prescribed normal contraction. The construction uses analytical projectors for known manifolds and local affine proxies estimated by principal component analysis for unknown data geometry. By implementing contraction theory in both cases of known and unknown manifolds, we guarantee manifold invariance and transverse convergence to the manifold within a desired time window (e.g., one second). We derive a family of compatible probability paths and decompose the training loss into a learnable tangential term and a normal residual. An ellipse experiment attains a mean terminal off-manifold error of order $10^{-6}$. In Push-T robotic experiments, MSFM raises success from $74\%$ to $82\%$. In the Robomimic Square task, success increases from $60\%$ to $72\%$, while rotation-manifold deviation decreases from order $10^{-2}$ to $10^{-7}$. The MSFM terminal geometric errors are controlled by the chosen numerical tolerance. These results demonstrate stronger geometric adherence and higher observed task performance, supporting prescribed normal contraction as a complement to learned generative transport.

With limited annotation budgets, choosing which images to label determines how much a model improves. Data-selection methods that use features from a separately trained model, or scene descriptions written by vision-language models, have been successful, but those signals do not directly capture changes in the model being improved. The target model's own internal features reflect what it has learned so far and change with retraining, making them a natural cue for choosing the next training data. However, feature rarity alone does not reveal the errors that matter for performance. Here we link internal features to prediction errors and their expected impact on performance and select images for labeling and retraining without using labels for candidate images. We evaluated the method with an object detector on two datasets and two pairs of random seeds. Adding internal features improved the identification of prediction errors in 15 of 16 conditions. When performance was averaged over successive labeling rounds, the method outperformed selection based only on feature rarity in all four evaluation settings and ranked among the top two of six methods. With other conditions held fixed, performance after retraining was again higher than with rarity-based selection, even though the latter collected more errors. With longer retraining, the proposed method ranked first among six methods. These results suggest that linking a model's internal features to its errors and their effects on performance may help select training images that improve performance, thereby allowing the model's current state to guide which images are labeled next.

Prioritizing central nervous system (CNS) interventions requires predicting how a dose, route, and schedule act on a partially observed microenvironment, then choosing the measurement that would change the decision. Action-conditioned predictors reduce a regimen to an identity token or a scalar exposure, discarding where and when the target is engaged; handing a point estimate to a separate planner then discards the joint uncertainty that makes a measurement worth running. We therefore treat decision quality as a property of the intervention interface, not of controller placement. NeuronSifter compiles regimens into state-conditional target-occupancy fields with support masks, propagates them through microenvironment dynamics with an occupancy-conditioned diffusion operator, and selects measurements by their expected reduction in intervention loss, assimilating typed outcomes into the same posterior. In a declared synthetic Alzheimer's disease (AD) evaluation over 64 paired scenario blocks, occupancy conditioning lowers trajectory continuous ranked probability score from 0.165 to 0.110 and raises intervention ordering accuracy from 0.760 to 0.880, and every paired benchmark contrast remains separated after Holm correction. Decision-directed acquisition attains terminal risk 0.160 against 0.166 for a matched numerical Bayesian experimental design planner, and reaches the target risk at 0.796 $[0.732,0.873]$ of an earlier design control's cost, while the corresponding ratio against the matched planner, 0.963 $[0.907,1.025]$, is not separated from equality; point-state and dependence-ablated interfaces instead raise risk to 0.220 and 0.199, and a full-posterior external controller ties exactly. Published AD trials supply a separate retrospective endpoint bridge.

Large-scale routing problems are difficult to solve efficiently as their search spaces grow rapidly with problem size. Existing approaches primarily improve the optimization procedure itself, often at increasing computational cost. We instead shift the focus to a useful initialization that can be refined into a high-quality solution with limited downstream refinement. We propose Just Initialize, a training-free and solver-agnostic initialization component for large-scale routing optimization. Just Initialize compresses a large routing instance into a compact surrogate space, optimizes its global routing structure, and recovers the resulting solution as an optimization-friendly starting point in the original space. Extensive experiments on Traveling Salesman Problems (TSPs), Capacitated Vehicle Routing Problems (CVRPs), Vehicle Routing Problems with Time Windows (VRPTWs), and Prize-Collecting Traveling Salesman Problems (PCTSPs) demonstrate that Just Initialize achieves high-quality solutions comparable to or better than state-of-the-art methods while substantially reducing computational cost across instances ranging from 1K to 100K nodes, including an average speedup of approximately 70$\times$, sub-second runtimes on 10K-node instances, and runtimes within tens of seconds on 100K-node instances.

State space models (SSMs) achieve efficient sequence processing because their affine state updates are closed under composition and can therefore be evaluated with an associative parallel scan. Nonlinear recurrent models can provide richer, state-dependent dynamics, but generally lose this compositional structure: parallel evaluation then requires iterative methods that repeatedly linearize and scan the recurrence. We ask, what state-dependent nonlinear dynamics can be designed to remain exactly composable? We answer by introducing RiccatiSSM, a nonlinear SSM, in which each state dimension follows an input-conditioned Riccati differential equation. Its quadratic state dependence makes the local Jacobian explicitly state-dependent, while its exact per-step flow under piecewise-constant inputs is a Möbius transformation. Since Möbius maps are closed under composition and compose through $2\times 2$ matrix multiplication, the complete nonlinear state trajectory can be evaluated exactly with a single associative parallel scan, without iterative linearization. We further derive a constrained parameterization that ensures bounded, contractive dynamics, and avoids poles in the fractional-linear state update. Across long-sequence classification, regression, and forecasting tasks, RiccatiSSM achieves competitive predictive performance while reducing runtime by $22{-}33\%$ compared to the nonlinear LrcSSM under matched architectures. These results demonstrate that state-dependent nonlinear dynamics can retain exact composability and be evaluated efficiently within a single parallel scan.

Large language models are trained with backpropagation, whose global gradient coordinates all layers but forces each to hold its activations and wait for the gradient to pass back through every deeper layer. Conventional local learning removes this update locking by training each module to predict the target through its own readout, but has not scaled to billion-parameter pretraining. We identify these private readouts as a key weakness, since they leave each module without information from deeper modules. We propose Shared-Output LOcal learning (SOLO), which replaces them with a shared, read-only copy of the final module's readout, the only one trained on the output of the whole network. Taken from the previous step, the copy transmits information from the final module without passing gradients between modules or reintroducing update locking. SOLO approaches backpropagation on Transformers of 340M to 2B parameters pretrained on 15B tokens, staying within one point in average zero-shot accuracy with a perplexity gap that narrows with scale. Readout ablations attribute SOLO's improvement over private readouts to sharing. Without update locking, each of p pipeline stages holds activations for O(1) micro-batches instead of O(p). The freed memory permits larger micro-batches, which reach up to 1.44x the best measured throughput of pipeline backpropagation on the same partition. To our knowledge, SOLO is the first local learning method to show such memory and throughput gains in billion-parameter language-model pretraining. Local learning thus becomes a practical alternative to backpropagation for large-scale pretraining.

Recently, deep neural networks on manifold-valued representations have garnered significant attention across various machine learning applications. One recent focus is the generalization of Euclidean fully connected (FC) and convolutional layers to non-Euclidean geometries. However, previous approaches typically focus on a few selected manifolds and rely on specific properties of the target manifold. In contrast, this work proposes a framework for constructing FC and convolutional layers over computationally tractable Riemannian spaces. This framework incorporates several previous FC layers across different geometries as special cases and is instantiated on ten representative manifolds, including three hyperbolic models, five geometries of the symmetric positive definite (SPD) manifold, and two Grassmannian perspectives. Experiments on different manifolds demonstrate the effectiveness and applicability of our approach. Code can be found at https://github.com/GitZH-Chen/RieTrans.

Large language models (LLMs) have shown strong potential for assisting software and security analysis tasks, yet their effectiveness in cryptographic symbolic protocol verification remains insufficiently understood. In this paper, we conduct the first systematic evaluation of the capability of state-of-the-art LLMs in cryptographic symbolic protocol verification. To quantify this capability, we propose \textsc{CRoST} (Coverage Rate of Solve Tree), a proof-based metric derived from the verifier's proof skeleton that measures the similarity between generated lemmas and reference lemmas. We then establish the rationale of \textsc{CRoST} through both theoretical analysis and empirical validation. The evaluation results show that state-of-the-art models achieve 38.82\% coverage on average, with 14.4\% of generated lemmas exceeding 80\% coverage, indicating that LLMs can already generate useful lemmas to a certain extent. However, they still exhibit non-trivial failure modes on complex multi-phase protocols, show diminishing returns under naive scaling, and incur substantial verification overhead. These findings clarify the practical potential and limitations of LLMs for protocol verification and motivate future work on complex real-world protocols.

Reinforcement learning from verifiable rewards (RLVR) frequently reuses rollouts across multiple policy updates, increasing the mismatch between the current policy and the data-generating policy. We identify a sign-dependent gradient starvation problem in clipped policy optimization: clipping suppresses under-generated positive responses at the low-importance-weight tail while permitting severely over-generated negative responses to dominate the high-weight tail. To address this, we propose ReSPO (Reshaped Sequence Policy Optimization), which replaces clipping with a smooth, two-branch sequence-level kernel derived from an $α$-divergence variational objective and an exponential variance-control tilt. The positive branch preserves a nonzero gradient weight for under-generated positive responses, while the negative branch suppresses heavily over-generated negative responses. We demonstrate that ReSPO effectively learns from long positive reasoning trajectories during early training, even when accumulated policy drift relegates them to the low-importance-weight tail. On dense and MoE Qwen3 models, ReSPO accelerates early optimization, improves final training scores, and achieves higher held-out benchmark performance under a rollout reuse, validating our approach on importance-weight tail control in off-policy learning.

Estimating conditional statistics and learning representations of a population of conditional distributions are central problems in many data-driven applications, including uncertainty quantification and dynamical systems analysis. Conditional mean operators (CMOs), a class of linear operators between function spaces, resolve these objectives by providing access to a broad class of conditional statistics. However, existing methods typically estimate each CMO independently or constrain it to prespecified function spaces, thereby preventing the exploitation of shared structure across related distributions. In this work, we posit that related CMOs share finite-dimensional input and output function spaces, and are specialized for each task with a linear operator mapping these spaces. Based on this hypothesis, we introduce MTL-CMO, a multi-task framework that jointly learns shared function spaces and task-specific operators across multiple datasets. We further introduce T-CMO, a transfer learning method that reuses the shared spaces to estimate, in closed form, the operator of a new conditional distribution. We establish statistical guarantees quantifying the benefits of jointly learning the shared function spaces. Our experiments demonstrate that learning shared function spaces improves uncertainty quantification across a broad range of conditional distributions and, when applied to Langevin and plasma dynamics, yields compact representations of complex dynamics that retain physically meaningful information and enable parameter identification.

Modern LLMs are increasingly capable as autonomous agents, but they follow sequential interaction cycles: read, think, reply or call tools, repeat. Many real-world use cases are not sequential: voice assistants, embodied agents, and monitoring systems receive new inputs while they think or perform another task. Modern LLMs address this with specialized architectures for voice interaction and video streams, VLAs for robot control, asynchronous tool calling for API usage, and others. In this work, we generalize from different asynchronous tasks to general asynchronous agents that can adapt to different types of concurrency. To achieve this, we develop an asynchronous LLM framework that lets users (or the agents themselves) define inference coroutines with overlapping memory states. We showcase that Qwen 3.x models are capable of asynchronous operation for streaming video understanding, videogames, and monitoring, without task-specific training.

Reinforcement Learning-based post-training of Large Language Models (LLM) has been successfully applied to improve their reasoning capabilities. Existing pipelines primarily finetune LLMs on a fixed pool of problems specified prior to training using the GRPO loss. This is fundamentally limiting, as learning signal arises only when policy rollouts mix successes and failures, causing the useful portion of any fixed pool to quickly become stale as the model improves. To address this, we propose frontier learning, an open-ended post-training approach in which procedural generators are used online to continually produce informative training problems. It treats the generator's task-specific parameters as a search space and uses a regret signal to prioritize and explore frontier difficulty levels in order to focus training at the edge of the model's evolving reasoning capabilities. Across several reasoning tasks and model families, our approach consistently achieves higher relative gains over fixed-pool baselines, demonstrating that effective post-training requires not only selecting useful problems, but continually generating them at the edge of capability.

Constrained decoding can enforce regular or context-free output formats, but many program-generation failures are semantic: scope, typing, and declaration effects depend on context. We present semantic grammar specifications, a declarative formalism that attaches such constraints to a context-free surface and executes them during Earley descent. Our implementation enforces \emph{safe pruning}: it rejects only prefixes whose semantic contradictions cannot be repaired by any continuation. A separate, grammar-dependent, \emph{dead-end freedom} property guarantees the existence of a realizable witness for each remaining branch. We give simple sufficient conditions based on surface productivity, type coverage, and left-to-right constraint flow. Our finite-lambda, core ML, and C-like fragments satisfy them, while the STLC instance used in our experiments does not: plain STLC can violate type coverage, and we show how restricting its type universe recovers it. A tokenizer-lifting lemma carries character-level witnesses to token sequences under an explicit vocabulary-coverage hypothesis. We validate the implementation differentially against production compilers (\texttt{ocamlc}, \texttt{cc}). Across every prefix of 65 compiler-valid programs we observe zero false prunes. The semantic oracle localizes 25/30 invalid programs mid-stream, against 0/30 for a syntax-only oracle, and agrees on 42/42 recursion probes. A twelve-model generation study, including a matched semantic-versus-syntactic ablation for nine models, finds nonnegative observed semantic-minus-syntactic point estimates for every model-language pair, with maxima of $+15.2$ points on STLC task correctness and $+14.3$ points on ML validity.

This paper studies convex optimization when the gradient cannot be evaluated exactly, but only approximated by a hierarchy of algorithms whose compute grows like $δ^{-γ}$ in the accuracy $δ$. When $γ>2$, falling into the Harder-Than-Monte-Carlo (HTMC) regime, the price of accuracy outruns the variance reduction that Monte Carlo would buy and we show that minimizing a loss function costs no more, up to a factor depending only on $γ$, than a single evaluation of its gradient at the accuracy the problem demands. A randomized multilevel oracle replaces the deterministic approximation of accuracy $δ$ by an unbiased estimator of it, whose variance $σ^2$ becomes a second, independently priced dial: the cost of one call drops from $δ^{-γ}$ to $δ^{2-γ}σ^{-2}$. Plain inexact gradient descent driven by that oracle reaches loss $\varepsilon$ at expected compute $Θ(\varepsilon^{-γ})$ in the convex case, against $Θ(\varepsilon^{-(γ+1)})$ for the same method run at a fixed accuracy: randomization buys a full power of $\varepsilon$. Under $μ$-strong convexity the exponent halves, to $\varepsilon^{-γ/2}$, because the iterates settle at a noise floor and the bias budget relaxes accordingly. Both bounds are independent of the step size, and hence of the smoothness constant, and we show that the cost is a functional of the underlying gradient flow rather than of any discretization of it.

Multi-agent systems bring together language model agents with different roles to propose, review, and refine solutions. Each agent's response depends on its model's capabilities, the reasoning strategy defined by its system prompt, and the information in its input context. Existing frameworks often adapt communication by changing this context while leaving individual prompts fixed, even when a problem calls for different skills. We study whether agents' initial responses can identify a strategy better suited to the current problem and guide its transfer to other agents. To address this, we introduce SAGE (Self-Adapting Group of Experts), a training-free framework that uses answer agreement, prefix consistency, and reciprocal peer review to select a strategy donor. SAGE transfers the selected donor's reasoning strategy to the other agents while preserving their original roles. This transfer uses only the agents' original system prompts, without access to the problem or generated solutions. After strategy adaptation, agents exchange responses through a dynamic, sparse directed acyclic graph that routes information from higher-scoring agents to lower-scoring agents. Experiments across multiple agent backbones and reasoning benchmarks show that SAGE achieves higher average accuracy than the evaluated baselines. Our code is available at https://github.com/atifquamar07/sage.

Recent advances in AI-based speech synthesis have enabled highly realistic speech, increasing the importance of speech deepfake detection (SDD) in preventing misuse. While mainstream Self-Supervised Learning (SSL)-based detectors achieve strong performance, they suffer from poor generalization to unseen domains and often overlook fine-grained signal artifacts due to a bias towards global semantic consistency. In this paper, we conduct the first detailed empirical and visual analysis to validate these limitations explicitly. Our investigation reveals two critical architectural vulnerabilities: (1) a systemic failure to capture localized spoofing traces, and (2) a severe lack of adaptability to domain-driven shifts in SSL layer importance, rendering static aggregation strategies prone to overfitting. To address these vulnerabilities, we propose the Global-Local Adaptive Detector (GLAD). Specifically, to capture localized forgeries, GLAD employs a Hierarchical Global-Local (HGL) backbone that explicitly bridges the granularity gap by fusing global linguistic and acoustic features with fine-grained local signal details. To counter layer importance shifts in out-of-distribution (OOD) scenarios, we introduce a Hierarchical Adaptive Gating (HAG) mechanism that dynamically recalibrates layer-wise focus in a sample-specific manner. Finally, to address shortcut learning induced by environmental biases, we introduce SaniBoost, a composite data augmentation strategy for robust signal standardization and noise sanitization. Extensive experiments demonstrate that GLAD significantly outperforms state-of-the-art methods, particularly on unseen domain cases.The code will be released upon publication.

Spectral super-resolution of multispectral satellite images can enable high temporal- and spatial-resolution hyperspectral satellite imagery at a modest cost, significantly increasing the applicability of hyperspectral remote sensing. This task is inherently ill-posed, making it well-suited for deep learning-based methods. In this study, the spectral super-resolution task is framed as an operator learning problem, and SSRON is proposed as a Deep Operator Network that effectively learns function-to-function mappings from downsampled spectra to continuous spectra. The model is trained to super-resolve Sentinel-2A-like multispectral imagery to EMIT images. Compared to baseline models, SSRON achieves superior performance across all metrics. The model also demonstrates zero-shot spectral super-resolution capability by predicting bands unseen during training. Furthermore, its continuous-output formulation suggests the potential to estimate spectra at finer wavelength intervals than the native sensor. These results suggest the potential of SSRON and establishes operator learning as a promising direction for spectral super-resolution.

As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.

Large language model (LLM) assistants increasingly help users draft content for third-party recipients. During private drafting, the user or the model may introduce an item and later remove or replace it. The model may remove the item from the intended content but reveal it again when stating the edit. We call such statements revision traces. For example, after a user removes the password before sharing a configuration file, the model may delete it but leave a comment saying, "Removed the password 'No****4!' as requested." A third-party recipient who sees only the delivered file can therefore recover the withdrawn password from the comment. In an in-the-wild analysis of three public conversation corpora, we identify 26,753 revision requests, of which 2,363 (8.8%) leave revision traces. We study them in greater depth under controlled conditions by introducing RevLeakBench, a benchmark of 100 tasks across five scenarios with a conversation track and an agent track. We measure trace occurrence, withdrawn-item recovery, trace position, and required-content retention. Across six models, about half of the deliverables in both tracks state the edit after a revocation, and a reader that sees only the deliverable can recover the withdrawn item from about 13% of them. Telling the model that its entire reply will be forwarded to the recipient still leaves revision traces in 36.4% of the deliverables. We compare prompt defenses and a delivery boundary, and propose an output-side filter that sharply reduces recovery with little loss of required content. We believe our work can benefit efforts to understand and mitigate unintended disclosure in LLM interactions.

When pricing agents meet repeatedly on a platform, the platform decides who faces whom. We ask whether that choice moves the prices the agents learn, and whether a rise comes with learned punishment. In a pre-registered randomised experiment in the Bertrand duopoly of Calvano et al., each agent's price is set by a tabular Q-learning module, not by the small language model attached to it, and we randomise whether each agent keeps its partner, sees its rival's prices and can send messages. Keeping the same partner raises the level of profits, averaged over training, by 0.27 of the gap between competitive and monopoly profit (95% CI 0.20 to 0.35, all twenty paired runs positive), our registered primary result, and the resting price by 0.17 of the Nash-to-monopoly range (post hoc). A plain tabular learner reproduces the effect in all 25 further blocks, and there one permanent partner raises the level more than about three do (+0.23 against +0.05, exploratory). Where rival prices are hidden, the price-setting module cannot see a cut, so cannot punish it, yet the resting price rises as much and the rise lasts to the end of training, while with visible rivals it shrinks with longer training (post hoc). Where the rival is visible, a static best responder accounts for a third to a half of what a forced-deviation probe reads as punishment, on the starts where the rival can see the cut, and net of it the registered test of learned punishment is inconclusive. A test that looks only for punishment would thus miss the rise where the rival is hidden, while a check for profitable deviations flags most of those prices (post hoc). In an exploratory extension, untrained Qwen2.5 7B and 14B models under one prompt show the effect when the rival's price is left out of the prompt and inconsistently when it is shown, the 7B result replicating on fresh blocks, while two other model families show none.

Artificial intelligence is increasingly deployed in critical industrial domains, including healthcare, energy grids, subsurface exploration, where failures can have severe consequences for human safety, system stability, and economic outcomes. Yet AI is still evaluated primarily through benchmark accuracy, a model-centric metric that fails to capture the structural complexity and risks of real-world deployment. We propose a framework that views industrial AI reliability as a problem of structural alignment across four interacting worlds: physical, representational, machine, and human cognitive. These worlds are connected through two interfaces: digitalization, linking physical reality to computational representations, and goal encoding, translating human cognition to the machine objectives. Together, they define the space of admissible solutions. We characterize the solution space through four attributes: existence, non-uniqueness, robustness, and interpretability and show how mismatches arise at interfaces and propagate across worlds to produce reliability failures. Applications to healthcare, energy grids, and subsurface exploration illustrate that although dominant failure modes differ across domains, for example, interpretability in healthcare, robustness in energy grids, and non-uniqueness in subsurface exploration, all originate from a shared structural mechanism. By shifting the focus from model-centric evaluation to system-level alignment, this framework offers a principled foundation for assessing and governing reliability in industrial AI systems.

Adam is the default optimizer for training modern deep neural networks, yet its adaptive behavior remains poorly understood due to the complex interaction between its first- and second-moment exponential moving averages (EMAs). We study Adam in the tied-$β$ regime, where the two EMA decay rates are equal, and show that its adaptive dynamics can be expressed through a transformed ratio with approximately scale-stable behavior. Empirically, this transformed ratio exhibits a stable, heavy-tailed distribution across tasks, model scales, and training stages, in contrast to the variability of raw moment magnitudes. This empirical stability has both practical and conceptual consequences. First, we derive a recurrence for the transformed ratio, yielding a reparameterization of Adam that replaces the second moment with a compressible state. Leveraging its stable distribution, we show that a fixed 4-bit codebook is sufficient in our experiments to store this state without auxiliary scaling, achieving performance competitive with full-precision Adam. Second, the transformed ratio view clarifies Adam's connection to sign-based methods: Adam reduces to sign-based momentum modulated by the transformed ratio, and replacing it with a constant recovers Signum as a limiting case. This perspective further provides a simple rule for transferring learning rates between the two methods. Together, these results suggest that tied-$β$ Adam admits a simple and approximately stable ratio structure underlying its adaptive behavior and demonstrate its utility for both analysis and efficient implementation.

Verbal feedback can identify errors and prescribe corrections, providing rich supervision for language-model post-training even when reliable programmatic verifiers are unavailable. Such feedback, often generated by a capable model, can be used to condition the teacher in on-policy distillation, which trains the student to match the teacher's predictions on student-generated rollouts. However, this approach can transfer teacher preferences that the feedback did not motivate, while leaving much of the feedback's guidance unused. We find that both problems come from the standard on-policy distillation objective, specifically the divergence it minimizes and the distribution it uses as its target. Our proposed method addresses both limitations. First, to isolate the information conveyed by the feedback from the teacher's inherent preferences, we treat verbal feedback as evidence for or against the hypothesis that a particular token comes next at a given prefix. We then adopt a probabilistic confirmation framework which uniquely determines an ordering over the vocabulary based on the teacher's predictions before and after it receives feedback. Using a confirmation score consistent with this ordering, we construct a target distribution within a trust region of the student. Second, to learn from guidance that student rollouts can leave unused, we derive a simple shared-rollout estimator of a symmetric divergence between the student and target distributions over rollouts, reusing student and feedback-conditioned teacher rollouts in both directions through importance weighting. Empirical evaluations show that our method outperforms the common on-policy distillation recipe and a recent contrastive variant on knowledge-based and agentic benchmarks.

Reliable confidence estimation is essential for large language model deployment. However, answer-level calibration remains challenging because generation errors are often localized: a response may be fluent and high-probability overall while still failing at a critical number, entity, or factual claim. Existing estimators compress token probabilities, sequence likelihoods, entropy, or beam statistics into a global score, which can dilute such local risk signals. We propose TRACE, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: (i) recording token-level surprisal and predictive entropy during decoding, (ii) applying local risk operators to preserve uncertainty spikes, and (iii) converting localized trace risk into answer-level confidence. TRACE produces a label-free risk score, while TRACE+ calibrates trace-only features into probabilities using a held-out split, without extra generations or external verifiers. We evaluate four tasks against 19 calibration baselines, and TRACE+ reduces Brier from 0.149 to 0.137 and improves AUROC from 0.758 to 0.792 over the strongest likelihood baseline. Across seven LLMs, TRACE+ improves over the best non-TRACE baseline pool from 0.136 to 0.120 Brier and from 0.764 to 0.817 AUROC. Results show that localizing decoding-time risk provides a general approach to calibration.

Many Model Context Protocol (MCP) servers wrap web APIs built for human developers, and their error messages tell the reader to run a command, edit a configuration, open a web page or wait. Many agents that read them can only call the server's tools. In 150 widely used MCP servers, 949 of 3,001 error messages tell the caller what to do next, and half of these steps depend on something the server cannot see about the caller. On credential errors, 62 of 67 steps ask for a terminal command, a configuration change or a web page; on rate limits, 20 of 30 say to wait and retry without naming the call to repeat. We tested five OpenAI models that act only through the tools of Berkeley Function Calling Leaderboard tasks, and the agents did what the step said. On expired credentials, a terminal command in the step left 45% of tasks recovered, and the loss it caused grew from 18 points for GPT-5.5 to 69 for GPT-6 Astra. On a rate limit, GitHub's "Wait before retrying." left 6%. We tested two remedies. For MCP developers, naming a server tool in the step raised recovery on expired credentials to 84%, with the login tool in place of the command, and on a rate limit to 88%, with the call to repeat in place of the bare wait. For agent developers, deleting the step with a one-sentence prompt before the model reads it raised recovery on expired credentials to 82%.

Electroencephalography (EEG) records mixtures of brain-source activity. Even with a known anatomical forward model, experiments that excite only part of the source-state space leave the dynamics unidentified, and repetition cannot resolve the ambiguity. We show that unknown local mechanism changes can supply the missing information. We consider linear dynamics among fixed anatomical sources with known source-state initialization patterns. Changing one source's update rule for one transition leaves a rank-one, source-specific signature in subsequent EEG: subtracting matched baseline responses isolates it, and the forward model identifies the source and calibrates its response history. Combining these histories with initialization responses recovers source interactions without baseline reachability and without first identifying the intervention coefficients. We establish sufficient recovery conditions, a direct estimator, and a noise-sensitivity bound conditional on correct source labels. Simulated EEG on anatomy derived from magnetic resonance imaging confirms the information gain: with baseline excitation confined to four of twelve source coordinates, eight unknown changes recover all dynamics in 32/32 systems, whereas baseline realization, baseline regression through an invertible forward model, and changes that leave the tested states unexposed all fail, and explicitly constructed alternative dynamics reproduce every baseline mean. Where baseline information suffices, direct reconstruction is also more reliable than a matched-information spectral estimator. Nonlocal changes and forward-model error limit accuracy even when source labels are correct.

In response to the growing demand for long sequences in agentic and reasoning use cases, many state-of-the-art LLMs combine multiple variants of attention to mitigate the quadratic complexity of traditional softmax attention. These hybrid attention LLMs aim to balance the strengths and limitations of full attention and alternatives based on recurrence. This work presents a first study of how hybrid attention impacts the multilinguality of LLMs. Beyond the impact on long sequences in poorly tokenized languages, our study is motivated by the possibility that the inductive biases of the recurrent state alter linguistic processing. Our interpretability analysis confirms this, showing that cross-lingual representations in hybrid models develop in patterns tied to the ordering of recurrent and full-attention layers. Across diverse models, we notably observe a pronounced spike in cross-lingual alignment around the first full-attention layer. These findings lead us to question the conventional ordering of attention layers. In distillation experiments on multilingual data, all alternative layer orderings outperform the standard throughout training, learning up to 2.5X faster. These stark, replicable results prompt our theory that multilingual models would benefit from starting with a full-attention layer rather than recurrent layers.

Diffusion models trained on a finite dataset first learn to generate novel, high-quality samples and only much later collapse onto their training set. We identify the mechanism behind this separation of timescales and the object that probes it. The training dynamics of the score function are governed---exactly, and at any width---by the Gram matrix of the Neural Tangent Kernel (NTK) evaluated on the noisy training data, so the timescales of generalization and of memorization must be encoded in its spectrum. We show that they are, and that the structure responsible has no analogue in standard kernel settings. The use of multiple noise realizations per sample ($m$ noised copies at a fixed noise level) in the score-matching loss is what restructures the Gram matrix spectrum into two distinct parts. The first, of large eigenvalues, carries the global features of the target distribution and is present already for $m=1$. The second, which the repeated noising creates, consists of the smallest eigenvalues and is supported on eigenvectors aligned with the sample-specific noise directions; it sets a memorization timescale parametrically larger in the training set size $n$. We establish this picture on two fronts. Analytically, we solve the spectrum in the lazy high-dimensional limit for both linear ($n \asymp d$) and polynomial ($n \asymp d^k$) sample complexities, and prove through a bias--variance decomposition that the first bulk minimizes the approximation error while the second drives the error associated with memorization. Empirically, we show the same two-bulk structure in Convolutional NTKs on CelebA and in finite-width U-Nets trained well beyond the lazy regime, and we make the link causal: truncating the Gram matrix at rank $r$ tunes the generalization--memorization transition, and an $L_2$ penalty targeting the second bulk suppresses memorization in feature-learning U-Nets.

While recent studies have explored human behavior and preference simulation using large language models (LLMs), it remains unclear how well LLMs can simulate subjective evaluations from real learners in educational settings. We investigate this question using real learner evaluation data on feedback for high-school biology questions at both the group and individual levels. We compare performance with and without learner-specific information, such as personality traits and evaluation examples, across six models. Our results show that LLMs still have a limited ability to simulate learner evaluations. Providing learner profiles and examples improves score calibration and individual-level simulation, but more often fails to improve group-level consistency. These findings highlight the need to investigate which learner information and adaptation strategies are effective for learner preference simulation.

Scaling up robotic manipulation is primarily bottlenecked by the scarcity of real-world robot data. While recent approaches leverage human video demonstrations to mitigate this shortage, they remain computationally expensive and still rely on paired human-robot data for domain alignment. Although current state-of-the-arts excel at long-horizon tasks, they struggle with the delicate and precise control required for complex tool manipulation. To overcome these limitations, we introduce P2P-T, from Pixel to Poses for Tool Manipulation, a data-efficient, object-centric framework that learns tool use directly from human demonstrations. P2P-T bridges the cognitive and physical execution gap through a two-stage approach. First, pretraining an object-centric world model to extract stable pose priors; second, integrating these priors into an efficient, pose-aware low-level policy. By utilizing a robust automated data processing pipeline powered by modern foundation models, P2P-T completely bypasses the need for human-robot aligned data. This reduces overall training overhead drastically. With minimal per-task fine-tuning, our framework achieves a 73% improvement over the previous state of the art in execution performance on complex, real-world tool manipulation tasks that currently remain out of reach for standard large-scale pretrained models.

A critical analysis of contemporary approaches to the study of conscious states. The review focuses on methods of classification, clustering, modeling of brain states under anesthesia and identification of measurable neurobiological characteristics of brain function. A comparative analysis was conducted in the following three major areas: automatic detection of states of consciousness using neural networks based on EEG and fMRI data; modeling of the structural-functional dynamics of the brain under the effects of anesthetics; and detection of neurophysiological indicators which correlate with the level of consciousness. The obtained conclusions demonstrate the growing effectiveness of deep neural models in the classification and prediction of brain states and the analysis of dynamic structural-functional connectivity. Nonetheless, significant limitations were also identified, including the limited interpretability of the models, the lack of standardized metrics, and the problem of the specificity of consciousness markers. Our findings support the need for developing hybrid, generalizible, physiologically grounded architectures. Furthermore, such approaches may improve the translational potential of computational models in clinical neuroscience. Diverse methods of machine and computational modeling have demonstrated their effectiveness in tasks of automatic clustering and classification of brain states, the development of multilevel models and the identification of connectivity patterns correlated with levels of consciousness. A larger-scale analysis and a larger dataset, as well as the implementation of model interpretability approaches are required for the practical application of the analyzed models. The models based on EEG and LFP are the most promising for clinical application due to their availability and the possibility of real-time monitoring.

The emergence of the digital transition brought an increasing need to control the processing of digital information, including in Law Enforcement Agencies (LEAs). At the EU level, in recent years, many regulations have emerged to control data processing and exchange. Texts other than the GDPR, such as the ''Law Enforcement Directive (LED)'', appeared to regulate specifically how Law Enforcement Agencies (LEAs) could process data. A formal representation of these regulations can be part of decision systems that support LEAs in processing data in compliance with the regulations. Although many new formalisms have emerged to represent legal norms and rules, few are provided with a reasoning mechanism. Furthermore, systems used in decision-making processes in critical contexts such as medical diagnoses or legal decisions cannot be fully automated, and the explainability of their results is essential to ensure user confidence in decisions. This explainability aspect, while crucial, is lacking in most modern approaches that rely on machine learning. This paper describes a framework to operate formal rules from regulations, by focusing on explainability of the decision. After describing the general architecture of the proposed decision support framework, the paper showcases how symbolic AI and the SPARQL query language can support legal reasoning. It then describes an algorithm to generate a justification for the reasoning results, and outlines the procedure to be followed when the reasoning does not lead to a satisfactory conclusion. We notably focus on a method based on decision trees to determine what additional information to request from the user.

Sparse autoencoders (SAEs) are an important tool for mechanistic interpretability, but interpreting their many features remains challenging. Existing methods characterize input-side activation patterns and output-side intervention effects, yet often leave their functional connection implicit, while input-side evidence collection typically relies on costly large-corpus scans. We introduce functional interpretation, which characterizes an SAE feature as a mapping from its activating input semantics to its output effects under intervention, and present Dual-End Agentic Feature Interpretation (DAFI), an agent that actively gathers evidence and refines input-side, output-side, and functional interpretations through component-specific feedback. Its short-context token probing enables on-demand activation evidence collection without a full corpus scan. On GemmaScope, DAFI improves Input score by 13.1 percentage points over SAGE and Output score by 38.9 points over Token Change, while being substantially more token-efficient than a general-purpose coding agent. Skills distilled from successful refinements raise the held-out joint pass rate from 58.0% to 92.0% and improve both interpretation quality and efficiency when transferred to a new model-SAE setting. Across features with reliable endpoint interpretations, 70.7% exhibit non-equivalent input and output semantics. On AxBench, DAFI also improves steering-feature selection over output-score filtering. Code is available at https://github.com/THUAIS-Lab/DAFI.

LLM agents solve complex tasks by iteratively changing files, invoking local tools, and interacting with remote services, which modifies state across their local environment and remote services. Today, agents and users must manage these changes explicitly, whether reverting exploratory actions or recovering from erroneous ones. Doing so safely requires coordinated actions, yet current agent harnesses lack unified abstractions and mechanisms for managing local and remote state consistently and efficiently. We describe Planarian, an agent runtime with state management that enables agents and users to recover from erroneous actions and explore alternative executions over consistent local and remote environment state. Planarian introduces the abstraction of agent statepoints, which are consistent, restorable point-in-time versions of the environment state. Planarian exposes three state-management primitives to agents and users: (i) snapshot creates a new statepoint spanning local and remote state without requiring external services to support checkpoints: it relies on efficient incremental process and file system snapshotting to capture local sandboxed state, and transparently records compensating actions to undo remote state changes; (ii) rollback restores the environment to a previous statepoint by reverting to a prior local checkpoint and replaying compensating actions for remote state changes; and (iii) fork creates multiple isolated branches from a statepoint, enabling the agent to explore alternatives in parallel. We show that Planarian enables agents to undo mistakes and explore alternatives in parallel, improving task quality by up to 15x, and allows users to recover from erroneous actions with only 3% overhead.

Many temporal link predictors summarize past interactions through learned node representations. We examine whether simple counts of recurring interaction patterns can provide competitive predictions without learning these representations. We propose a temporal link predictor based on statistical language modelling. It pools transition and co-occurrence counts across sources to predict links that a source has never formed. We smooth sparse estimates using destination frequencies or Kneser-Ney continuation counts. A shared log-linear rule combines these estimates with popularity, source history, and recency, without node embeddings. In our main evaluation, the model achieves the highest MRR among the compared methods on 7 out of 16 datasets from TGB and TGB-Seq. It also outperforms EdgeBank and Base3 on all 16 datasets and the heuristic family on 14. These gains extend to datasets designed to limit repeated edges. With only 9--13 learned parameters, our model provides a simple and competitive baseline for evaluating future neural temporal link predictors.

Large language models (LLMs) typically generate text autoregressively (AR), predicting one token at a time. Block diffusion language models (dLLMs) instead generate blocks sequentially while denoising multiple tokens in parallel within each block, offering a promising way to accelerate generation. Rather than training such models from scratch, recent work adapts strong pretrained AR models into block dLLMs through distillation. On-policy distillation (OPD) has been widely used for LLM training because it supervises the student on states generated by its current policy, rather than only on fixed offline trajectories. By training on the states the student actually visits, it reduces the mismatch between training and generation and can provide more relevant supervision as the student evolves. Recent work has extended this idea to AR-to-block-diffusion conversion. However, this setting introduces a fundamental mismatch in supervision: the block-diffusion student and the causal AR teacher condition on different information at the same training state. The student predicts from the entire partially denoised block, including visible future context, whereas the standard AR teacher target is defined only from the causal prefix. As a result, the teacher distribution used for distillation is not fully aligned with the information available to the student. We therefore introduce d-OPD, a future-aware on-policy distillation method that corrects the AR teacher distribution to better align with the student-visible state by incorporating visible future information within each block, providing supervision that better matches the information used by the student. Across Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to $4.0$ points over OPDLM and reduces training time by $1.35$-$1.58\times$. The code is available at https://github.com/mit-han-lab/d-OPD.

Coding agents are increasingly deployed for iterative development on real repositories, yet existing evaluation barely answers a basic question: \emph{do coding agents reuse existing code or reinvent the wheel?} The question matters: every duplicated implementation is a fix applied twice and agents produce code far faster than humans can audit, so redundancy accumulates unsupervised. Thus, we present \textbf{RepoReuse}, a multi-turn benchmark for auditing code reuse in real repositories, where requirements are revealed turn by turn and the workspace accumulates across turns. It is built by a fully automated pipeline combining AST-based dependency graphs, guided evidence collection, and execution-verified task synthesis, and scales readily to new repositories. Beyond pass rates, we measure the reuse rate together with recall and cross-turn structural redundancy. An audit over 3{,}000 turns shows that agents progressively stop exploring relevant repository code, reuse their own history less even when it is fully in the workspace, and leave duplicated logic in 50.8\% of task chains by turn~5---all while pass rates barely move. Such deficiencies are invisible to pass rates, underscoring the need to evaluate code generation beyond functional correctness.

Supervised fine-tuning can teach language models undesired behaviours alongside desired ones. Inoculation prompting (IP) aims to limit unwanted generalisation by requesting the undesired behaviour during training and removing the request at inference. However, undesired behaviour can still appear under unrelated prompts. IP can also hinder learning of the desired behaviour. We address these limitations in settings where both behaviours co-occur in most training examples, so filtering out examples with undesired behaviour leaves only a small clean subset. We introduce stratified inoculation prompting (SIP). SIP leverages a small clean subset to demonstrate that desired behaviour should persist without the undesired one across different contexts. SIP oversamples these clean examples under diverse non-eliciting prompts while inoculating the rest. SIP substantially reduces expression of undesired behaviour while preserving more of the desired behaviour than IP. These gains persist even when we extend IP to oversample the same clean subset at the same rate as SIP. Moreover, SIP yields lower emergent misalignment rates in all harmful-advice setups we tested. SIP can be further extended to limit the undesired behaviour even under prompts that explicitly request it. We introduce backdoor dilution, which weakens expression under the inoculation prompt, and password-locked inoculation, which concentrates elicitation on a designated password. Taken together, our findings show that changing the training contexts for a small clean subset can significantly improve selective generalisation.

Fully homomorphic encryption (FHE) enables neural network inference directly on encrypted inputs, but it remains orders of magnitude slower than plaintext in- ference. Applying the server's plaintext weights to encrypted activations involves plaintext-ciphertext multiplications (PMult) and accounts for more than half of inference time in recent systems. Ternary quantization can replace these multipli- cations with additions and subtractions, but the savings rarely materialize under packed execution. A single PMult applies a weight group fixed by the packing layout and can be avoided only when all its weights share the same ternary value. Ternarizing all groups, however, largely degrades accuracy. We present FIONA, an offline optimizer that selectively ternarizes weights within a given packing layout based on the estimated effect of ternary conversion on the model's performance. FIONA encourages a shared ternary value within each weight group and retains full-precision weights for sensitive groups, so ternar- ized and full-precision paths coexist within a layer. It then compiles these hybrid operators exactly, applying common scaling factors once to accumulated inputs and reusing sums across outputs. Weight ternarization can also narrow the input ranges of downstream polynomials. FIONA fits lower-degree replacements under a cumulative accuracy budget, reducing multiplicative depth and bootstrapping. On VGG11, ViT, and BERT, FIONA reduces PMult operations by 53.4-79.5% and accelerates end-to-end encrypted inference by 2.38x, 1.68x, and 1.84x, re- spectively, with less than 1% accuracy loss across all three models.

Interference is commonly treated as geometric overlap between learned features. We introduce effective interference, which combines feature geometry and code statistics to capture realized interactions, distinguishing constructive from destructive interference and frequent weak interactions from rare strong ones. Under local fixed-support assumptions, we characterize how architectural constraints shape interference through four mechanisms: feature orthogonalization, bias compensation, gain adaptation, and encoder-decoder separation. Experiments with sparse autoencoders show that constrained architectures selectively reduce overlap among co-active features, while bias, gain, and encoder freedom allow constructive cross-contributions to remain. Together, these results show that interference in learned representations depends not only on feature geometry, but also on how features are used and on the architecture that produces their codes.

While Large Reasoning Models (LRMs) excel at complex reasoning, alignment through reinforcement learning often induces systemic overconfidence. In production environments, where logits may be unavailable, robust black-box uncertainty quantification (UQ) is essential for trustworthiness and safety. Focusing on question-answering for LRMs, we show that existing black-box methods, such as paraphrase-based self-consistency and confidence verbalization, offer little to no improvement over simple repeated sampling, suggesting that alignment suppresses useful output variability. We introduce prompt-level relaxation operators that broaden the model's effective output distribution by approximating the effect of an optimal policy obtained with a stronger KL-regularization parameter, hence closer to the reference model. Theoretically, we demonstrate that relaxation improves calibration. We propose Jailbreak for Uncertainty (J4U), a jailbreak-derived technique for UQ that empirically reproduces the behavioral signatures predicted by our relaxation theory. Across 3 datasets and 4 LRMs, including a closed-source production model, J4U's improvement over repeated sampling achieves statistical significance in up to 6 times more LRM-dataset-metric settings than the strongest black-box UQ state-of-the-art baseline we evaluate, with average ECE reductions up to 5 times larger. These results provide a practical tool for UQ in black-box LRM deployment.

The evaluation cost of transformers with softmax attention scales quadratically with sequence length. Kernel attention addresses this by replacing softmax with a more general kernel function. In this paper, we aim to identify kernels that retain the expressivity of attention while enabling quasi linear computation. To quantify expressivity, we introduce a capacity for each kernel, measuring the maximum sequence length for which the attention matrix can approximate the identity. A higher capacity thus indicates greater expressivity. We show that expressive kernels like softmax, Gauss, and Laplace have infinite capacity. In contrast, common quasi linear kernels, such as those derived from finite dimensional feature maps, exhibit finite capacity. As a solution, we propose additive kernels constructed from univariate spline and polynomial exponential kernels. We prove that these maintain infinite capacity while allowing quasi linear computation via sorting. Finally, we implement additive sorting kernels efficiently and benchmark them against modern softmax backends, demonstrating advantages for long sequences.

Estimating probability densities from a finite set of samples typically requires dataset-specific model fitting. We introduce PRODiGI, a pretrained data-to-program model that infers an explicit, executable generative program in a single forward pass. Pretrained on synthetic datasets paired with their ground-truth programs, PRODiGI accommodates diverse generative families and data dimensionalities through template prediction and non-autoregressive program parameter decoding. Its inferred programs support direct sampling, density and score evaluation, and inspection independently of the pretrained model. We further introduce program-space fine-tuning, which refines differentiable program parameters by matching generated and empirical samples while keeping model parameters intact. Experiments show that PRODiGI achieves lower average density and score MAE than existing pretrained models, while offering multi-fold speedups over its closest competitors. Program-space fine-tuning further reduces generation MMD by 84%. By turning empirical data into explicit, reusable programs, PRODiGI introduces a new direction for fast, interpretable tabular generative modeling.

Reinforcement learning can turn one language model into several specialists, each excellent at a single skill such as mathematics, coding or following instructions, but users need one model with all of these skills. Multi-teacher on-policy distillation (MOPD) merges them by letting the specialists teach one student: the student answers each prompt, and the specialist for that prompt's domain gives feedback on every token. This routing decides which specialist teaches, but not how strongly its feedback moves the shared student. In Qwen3.5 models at three sizes, we find that MOPD's student does not beat one taught by the best single specialist and gains little of the mathematics specialist's advantage. The feedback is unbalanced: instruction-following feedback is several times more spread out than mathematics feedback and dominates the student's updates. We propose Domain-Normalized MOPD (DN-MOPD), which keeps the routing and rescales each domain's feedback by its measured spread. On six public benchmarks, DN-MOPD improves the average score over MOPD at every size, across three random seeds and under two answer-length limits, and recovers most of the lost mathematics gain. Controls with fixed domain weights show that the gain comes mainly from turning down instruction-following feedback rather than turning up mathematics alone, and that fixed weights close to those DN-MOPD measures perform comparably. Combining specialists therefore requires deciding not only which one teaches, but also how strongly its feedback counts.

The appearance of Jev marked the era of System One Models, foundation models that return structured decisions with probability distributions rather than text. Aside from low cost and great speed, Jev's central promise is that these probabilities are calibrated: such claim is not backed by any public test and available external benchmarks evaluate confidence calibration, not whether every returned option probability has the right numerical meaning. To tackle this issue, we introduce Sys1Cal-v1, a dataset of True/False questions about a proposition $A$ for which the exact probability $P(A)$ is known by construction. Each item is queried through the three Jev primitives - Noul, Choice and Score - and evaluated by total variation distance from the ground-truth distribution, which can be used to estimate a soft accuracy of System One Models. We showcase the utility of Sys1Cal-v1 as a benchmark dataset by evaluating Jev and SemIf, an open-source Choice-style baseline. In this work, however, we focus even more deeply on Jev, by studying the calibration of its Score and Choice answers. In particular, we discover a peculiar behaviour that can be explained by assuming that Jev suppresses a third truth value, going beyond True and False. In other words, in \texttt{Choice} answers, $P(A)$ and $P(\neg A)$ are presented as if $P(A)+P(\neg A)=1$, while a term $P(U)\neq0$ is missing in the sum. Recovering $P(U)$ leads to an improvement of median soft accuracy in \texttt{Choice} answers from $0.771$ to $0.978$, suggesting that, even in binary decisions, Jev wants to answer with a third option:``I don't know''.

Mechanistic discovery in neuronal microenvironments requires interventions and measurements that separate competing explanations of solute transport and neuronal response. Predictive accuracy cannot settle the question: a real mechanistic change and an error in the computational twin leave the same signature in sparse observations. We formalize this twin confounding and reason over a joint mechanism--discrepancy belief, designing experiments that separate the two. NeuronDiscover is an Agent-in-Twin framework whose shared, mechanism-grounded World Action Model (WAM) couples prediction, intervention proposals, and observation design; independently adjudicated outcomes revise a scoped Mechanism--Intervention--Observation--Outcome (MIOY) graph, whose supported relations compile into executable programs carrying discrepancy-adjusted acceptance bounds. We evaluate on simulated brain-fluid tracer-transport worlds adjudicated by an independently frozen finer-mesh reference solver, and on donor-disjoint public current-clamp recordings of cortical neurons. Counting only relations that reach a certified terminal status, and scoring abstentions as unresolved for every method, at a matched budget of 16 experiments over 32 source units NeuronDiscover resolves 4.0 relations per assigned world against 3.4 for the strongest baseline and 3.2 without graph revision, at 5% false support and 82% scope accuracy. Joint mechanism--discrepancy acquisition resolves 3.8 relations versus 2.9 for plug-in expected information gain; discrepancy-adjusted verification lowers accepted-program failure from 15% to 9% at 60% acceptance coverage; and transfer to the recordings yields 1.94 versus 1.53 relations per assigned world. Correctness is adjudicated within declared model worlds and archival recordings.

Despite the promise of Large Language Models (LLMs) in computational chemistry, rigorous combinatorial chemistry problems remain difficult because they require quantitatively constrained molecular modification, candidate validation, and systematic revision after failed attempts. Existing tool-augmented chemical agents demonstrate useful planning and tool use, but they rarely provide a unified loop for property-driven molecular optimization and workflow-level composition. To bridge this gap, we propose Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving (TMCS), a step-by-step multi-agent framework that formalizes chemical problem solving as an interpretable, tool-augmented workflow. At the task level, specialized agents leverage external tools, few-shot trajectory memory, and structured reflection to iteratively refine solutions. At the workflow level, TMCS chains generation, understanding, editing, description, and optimization into a closed-loop pipeline. Evaluations across multiple chemical tasks demonstrate that TMCS consistently enhances chemical reasoning across both open- and closed-source base models, achieving state-of-the-art performance.

Machine learning task type identification is essential for constructing valid ML pipelines, yet in practice it is typically specified manually. We investigate whether large language models (LLMs) can infer both the data domain and the downstream prediction task directly from dataset-level information when only the target feature is provided by the user. Together with our LLM-based system we also release an annotated benchmark comprising 625 public tabular and time series datasets. We evaluate the proposed approach in three settings: (i) tabular datasets in comparison with established AutoML heuristics, (ii) cross-domain evaluation across tabular and time series datasets, and (iii) a practical deployment scenario using smaller local models. The results show consistent advantages for LLM-based task type identification, with increasing difficulty in heterogeneous and resource-constrained settings. LLM-based approaches outperform AutoGluon in the tabular setting, reaching 0.98 F1 macro compared to 0.93. In the cross-domain setting, the best model achieves 0.90 F1 macro, while smaller locally deployable models reach 0.75, indicating a trade-off between deployment feasibility and accuracy.

Algorithm Distillation (AD) has demonstrated the remarkable ability of Transformers to perform in-context reinforcement learning without explicit weight updates. However, capturing long-term learning progress necessitates expansive context windows, which incur prohibitive memory costs and limit scalability in complex, long-horizon tasks. To address this bottleneck, we propose Recurrent Algorithm Distillation (RAD). RAD employs a dual-component architecture: a Compression Transformer that distills extended interaction histories into compact latent tokens, and an AD Transformer that auto-regressively generates actions using a hybrid context of these compressed memories and recent transitions. By maintaining a fixed-size latent buffer, RAD decouples the effective history length from computational complexity, functionally providing the model with a long-horizon memory. Empirical evaluations across diverse environments demonstrate that RAD matches the asymptotic performance of standard AD with significantly reduced context window sizes, offering a scalable solution for efficient in-context decision-making.

We study the Reverse Sequential Proportional Approval Voting Rule (RevSeqPAV) in approval-based committee elections. Despite its historical prominence and practical use, its properties and guarantees are much less understood than those of Sequential PAV. We analyze it along two dimensions: proportional representation (measured by Extended Justified Representation, its approximations, and proportionality degree) and approximation of the maximum PAV score of instances. We first establish strong negative results for general, unrestricted election instances and then identify settings in which the rule provides meaningful fairness and optimization guarantees.

Meta-Black-Box Optimization (MetaBBO) is one of the highlights in the recent AI for Optimization trend. This paradigm's bi-level workflow leverages the learnable algorithm design policy at meta level to ensure the performance and generalization improvement on the low-level optimization task. While MetaBBO helps advance the performance lower bound of the resulted optimization system, it is currently handcrafted and customized case by case to adapt different optimization problems, which inevitably introduces inherent subjectivity and hence restricts the performance upper bound and usability in practice. In this paper, we address this issue by regarding MetaBBO's design loop as coding task, where we could introduce openendedness into MetaBBO with recursive self-improvement capability of advanced coding agents. Specifically, we propose a dual-agent framework: i) a task agent continuously refines the codebase of a target MetaBBO approach through code evolution; ii) a hyper agent progressively modifies the task agent and itself to provide open-ended design behavior; iii) the evolved MetaBBO codebase is evaluated and all in-execution information is fed back to the agents for recursive self-referential improvement. As a result, given a naive MetaBBO template, our framework automates a design evolution and finds novel variants superior to up-to-date human-made MetaBBO baselines. Surprisingly, the experimental results also demonstrate that our framework supports fast adaption across different optimization domains. Solid interpretation analysis further reveals interesting design principles emerge in such open-ended process. This work serves as the first exploration on automating design of complex learning-assisted optimization algorithms.

Score-based generative models generate new samples by integrating a time-dependent drift that carries Gaussian noise onto the target distribution. In practice this drift is modeled by a neural network, trained on a loss integrated over time $t$ with a weighting schedule $w(t)$. Along the backward dynamics, and for multi-modal distributions, trajectories commit to modes of the target within a narrow time window, the \textit{speciation time}. In this work, focusing on high-dimensional data, we decompose the integrated loss into its single-time contributions and analyze each at fixed signal-to-noise ratio $Λ(t)$: we show that $Λ(t)$ sets the rate at which each feature of a multimodal target - the mode directions and their relative weights - is acquired during training. Crucially, at high $Λ(t)$ all mode directions are acquired together, on a single timescale insensitive to their amplitudes, while the relative weights are not learned at all. Only near the speciation time, where $Λ(t)$ becomes of order one, do all features become learnable, each on its own timescale: the weights are acquired jointly with the directions, and the directions at rates set by their relative amplitudes. For models trained on time-integrated objectives, the learning dynamics is then governed by how much of the weighting effectively sits near the speciation time, which provides insights on $w(t)$ design choices. These results follow from an exact high-dimensional analysis of the training dynamics of unbalanced and hierarchical Gaussian mixtures. Numerical experiments on image and human genome haplotype generation recover the predicted hierarchy of learning timescales in more complex settings.

On-policy distillation (OPD) trains a student on its own trajectories with dense teacher supervision. Recent work on OPD for multi-turn autonomous agents often treats large teacher-student token-level distributional gaps as promising intervention points, linking larger gaps to a greater need for correction. Yet, our empirical analysis reveals a supervision-benefit mismatch: large gaps can be benign, while small gaps can be outcome-critical. Teacher-student gaps capture differences at the current turn, whereas the benefit of teacher guidance depends on how the current student interacts with the environment afterward. The student may still succeed despite choosing an action that differs from the teacher's, while a teacher-preferred action may lead to a state from which the student cannot complete the task. Local gaps alone are therefore not enough to determine whether teacher guidance benefits the current student. Effective supervision should instead emphasize guidance that the current student can translate into better final task outcomes. Accordingly, we propose Outcome-Guided On-Policy Distillation (OG-OPD), which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations. This calibration selectively strengthens supervision on the student's original trajectories at turns where teacher guidance benefits the current student. Across ALFWorld, ScienceWorld, and WebShop, OG-OPD consistently outperforms baselines under diverse settings. It improves task success rates by 3.6-17.7 percentage points over vanilla OPD and by up to 7.0 percentage points over the strongest baseline.

AI reliability concerns whether an AI system performs its intended function dependably over a stated period and under stated operating conditions, with stated evidence. As these systems become more autonomous, that function includes more than a correct output. Retrieval, memory, tool use, permissions, human oversight, and interactions among systems must operate consistently and safely, and, for generative systems, so must the reasoning process that produces the output. Average benchmark accuracy measures capability; it does not quantify this broader reliability claim. This paper adapts established reliability engineering methods, from failure definitions and operational envelopes to FMEA, accelerated testing, field monitoring, and reliability growth, to AI systems. A four-level diagnostic framework classifies failures as component, operational-loop, agentic-conduct, or network and governance failures. Test, evaluation, verification, and validation (TEVV), sequential monitoring, and FRACAS create and refresh evidence. SMART provides statistical guidance for measurement, analysis, assessment, and test planning; the NIST AI Risk Management Framework provides organizational guidance for governance, evaluation, monitoring, and mitigation. Three cases illustrate the program: adversarial testing of a convolutional neural network, perception-error propagation, and autonomous-vehicle disengagements. Established reliability engineering provides a usable foundation; new measurements and safety guardrails are still needed as these systems are self-evolving.

Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present COEVOLVE, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, COEVOLVE enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, COEVOLVE attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers

Neural residuals can improve satellite evapotranspiration (ET) estimates, but selectors must predict when a correction helps and reject unsupported inputs. We evaluate ten-member models on 16,366 flux-tower observations from 151 stations paired with OpenET, across nine rolling years and five spatial folds. At one held-out station, Gain accepted corrections on all 32 physically invalid records: it predicted a mean benefit of 0.83 mm/day, but the corrections increased mean absolute error by 21.6 mm/day versus OpenET. On spatially held-out unit errors, SupportGain reduced station-macro MAE versus Gain by 0.148 mm/day under wind x3.6 (simultaneous 95% interval, 0.070 to 0.226), with 9.3% acceptance versus Gain's 51.8%; on clean inputs, its 0.006 mm/day advantage had an interval that includes zero. These fault analyses are exploratory; none of 40 preplanned temporal comparisons passed Holm correction, while a separate predeclared cropland contrast found 0.041 mm/day lower station-macro MAE with crop-only training (95% interval, 0.009 to 0.079).

Large Language Models (LLMs) perform well on reasoning benchmarks, but it remains unclear whether this reflects genuine contextual reasoning or reliance on facts memorized in their parameters. We investigate this by distinguishing two possibilities: a broad \textit{memorization bias}, where familiar content improves reasoning performance, and the \textit{Strong Parametric Shortcut Hypothesis}, where models skip reasoning entirely and recall stored answers. To test these effects, we introduce \textbf{MemoReason}, a human-curated benchmark that pairs factual reasoning tasks with structurally identical \fictitiousterm{} versions where real entities like people, companies, or dates are systematically replaced by \fictitiousterm{} ones of the same type. This \scorerevision{preserves task structure and specified reasoning operations} while varying the familiarity of the context, allowing controlled measurement of how the parametric memory affects reasoning. \revision{Our evaluation of recent LLMs reveals consistent and statistically significant performance drops of up to 15.7\% in the fictitious setting, demonstrating a clear memorization bias.} However, a targeted analysis of \revision{questions failed in the fictitious setting} shows that models rarely respond with the corresponding factual answer, indicating that direct parametric shortcuts are not the dominant failure mode. These findings suggest that parametric memory influences reasoning through mechanisms more complex than simple factual recall. \textbf{MemoReason} provides a controlled framework for studying these mechanisms and for extending paired factual-fictitious{} evaluation to broader reasoning settings.

Large language models process conversation history as unverified context: false premises injected into prior turns can be adopted as fact, a failure mode we term session-level contamination. We introduce five contamination protocols arranged along a source-authority gradient, isolating distinct failure mechanisms while holding the false premise constant, and evaluate GPT-5.4 Mini, Gemini-3.1 Flash-Lite, and GLM-4.5-Air across ten knowledge domains at temperature zero (22,500 turns), using a dual-track automated judge validated against a human gold standard (Cohen's \k{appa} = 0.901). GPT-5.4 Mini showed zero adoptions across all 500 sessions, a content-independent policy at the session level; token-level probing shows the underlying margin, while large, is finite. Gemini-3.1 Flash-Lite followed a steep authority gradient: 0.1% adoption for self-attributed falsehoods, 23.5% for user-cited sources, 68.2% for system-injected authority, and 94.0% under instruction override. GLM-4.5-Air showed a shallower gradient (15.8% vs 84.2%), a 68-percentage-point dissociation confirming that authority deference and instruction compliance are distinct mechanisms within one architecture. Recovery also diverged: GLM recovered in 94.5% of affected sessions, whereas 26.1% of affected Gemini sessions never did, rising to 40.0% under instruction override. Conversation history is an untrusted attack surface requiring provenance-aware system design; the complete framework is released as an open-source benchmark.

As Large Language Models (LLMs) become central to how we access and share information, they play an increasingly powerful role in shaping global knowledge. However, as these models evolve, their outputs risk converging into a \textit{generative monoculture}, where the diversity of perspectives they represent narrows over time. Studies at the model level often fail to pinpoint which specific topics or viewpoints are being marginalised or amplified in this process. In this paper, we introduce a method to measure shifts in topic saliency across model families, tracking what gains or loses prominence during post-training. Applying this approach to a case study of climate change discourse, we demonstrate how homogenisation affects the representation of diverse solutions across different models. We also test interventions to counter this trend, showing that specialised models can help preserve a broader range of perspectives. This underscores the importance of monitoring topic saliency to diagnose the risks of monoculture and to ensure AI systems reflect a pluralism of ideas. Data and Code are accessible \href{https://github.com/oriane/topic_saliency_shift}{here}.

Most clinical prediction systems learn patient-variable-outcome associations; we investigate a training-free diagnostic paradigm mapping patient observations to explicit medical knowledge. CKG Reasoner integrates candidate-specific Evidence Feature Nodes, patient-reference matching, a bounded Information Gate, knowledge-weighted evidence accumulation, disease similarity, and decisive clinical rules. Missing-aware normalization and coverage auditing distinguish absent from unavailable evidence. Candidate ranking is separate from outcome-label-independent K-means clustering, which uses four derived evidence coordinates (evidence strength, relative magnitude, directional similarity, and evidence completeness), not raw predictors or targets, to derive cohort-level assignments. Across six retrospective cohorts - four dengue (N = 1000, 1523, 989, 1018), malaria (N = 2190), and influenza (N = 4569) - a uniform, label-free, cohort-fitted K = 2 protocol yielded positive-class F1 scores of 0.996, 0.634, 0.936, 0.917, 0.695, and 0.842, and all-record accuracies of 0.996, 0.558, 0.914, 0.893, 0.707, and 0.906, respectively, with full partition-decision coverage using the frozen package and disease-specific knowledge representations. Neither scoring nor clustering uses outcome labels. Logistic regression provides a supervised baseline. Influenza incorporates confirmatory molecular PCR and is not independent pre-test prediction. Results characterize knowledge-grounded evidence separation, auditability, and sensitivity, not prospective clinical validity or comparative superiority. FOL/LLM-based clinical explanation remains unevaluated.

Pretraining a language model takes enormous compute, and the right optimizer can save a good part of it. Muon's spectral step gives a stronger direction than a sign step, but it is expensive. Every step runs Newton-Schulz iterations on the full matrix and, in distributed training, an extra all-reduce. Sign steps, as in Lion and Signum, are cheap and stay local to each device. We propose LionMuon, which takes one Muon step every $P$ iterations and Lion steps in between, with a single dual-EMA momentum buffer shared by both. Muon's compute and communication are paid once per $P$ steps, and the optimizer state is half of AdamW's. A single-EMA variant, SignMuon, already improves on Muon. We prove complexity bounds under heavy-tailed noise in which the period sets an interpolation between Muon's and Lion's smoothness and noise constants, and which say when LionMuon is faster than both. On 124M and 355M models trained on FineWeb, LionMuon with $P=2$ and $P=5$ reaches a lower loss than Muon, AdamW, Lion and Signum at the same number of tokens. Under 4-GPU data-parallel training it reaches Muon's final loss with a third less wall-clock on PCIe, and it beats the communication-efficient Muon variants Dion and MuonBP on loss at no more exposed communication, while keeping the exact gradient. Code: https://github.com/brain-lab-research/lion-muon

Computational antibody design requires representations that capture the geometric patterns underlying antigen--antibody interactions, yet existing approaches often rely on scalar distances or surface-intrinsic features, leaving cross-molecular geometry largely implicit. We present AbGaze, an end-to-end antibody design framework based on attentive geometric representation learning, which encodes distance, spatial direction, and surface-normal orientation of antigen surfaces relative to antibody-residue local frames, and adaptively aggregates these geometric interactions according to their interfacial context. The learned interaction representation is shared across multi-CDR co-design, complex structure prediction, and affinity optimization, with local-frame geometric supervision further constraining the representation. AbGaze outperforms prior methods across all three tasks: relative to the second-best method, it improves amino-acid recovery by 7.1% and reduces structural error by 14.9% on average over the six CDRs, improves interface docking quality (DockQ) by 6.6%, and raises the affinity improvement rate (IMP) by 32.5%.

We address Task A of the 1st DAFx Parameter Estimation Challenge, which aims to retrieve the physical parameters of a plate model from an impulse response. To do so, we use the Simulation-Based Inference (SBI) framework, in which we train a neural network to estimate a density over plate parameters given an impulse response, using a dataset generated by the simulator. Inference for a new impulse response then requires only a forward pass through the network, without involving the simulator. For each test observation, we fine-tune a specific network: additional simulation rounds are performed by sampling parameters from the current estimated distribution, simulating the corresponding impulse responses, and fine-tuning to produce the specialized network.

Aspect-based sentiment analysis (ABSA) has largely turned to text generation. We show that competitive dimensional ABSA does not need it. Using Jev, a frozen model that answers typed questions with rubric scores, label probabilities, and yes/no judgments, we decompose all three tasks of SemEval-2026 Task III Track A into such decisions and align them with the annotation scheme through 488 coefficients fitted on CPU, with no text generation and no backbone tuning. On valence-arousal regression over ten corpora in six languages, the system reaches 1.0645 RMSE, the lowest aggregate error of any participating system. On triplet and quadruplet extraction, it reaches 52.09 and 44.06 continuous F1, above fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines. Analyses and ablations show where the accuracy comes from: supervised calibration roughly halves the raw regression error, exact valence-arousal would add only 4.5 F1 to extraction, and the learned combination of span-boundary evidence, not any single signal, carries the extraction systems.

Recent representation alignment (REPA) methods accelerate diffusion transformer training by aligning projections of the transformer's hidden states with representations from pretrained visual encoders. In this work, we explore a reverse and complementary direction to REPA: rather than projecting diffusion representations into the encoder's space, we inject encoder representations into the diffusion transformer, allowing them to actively participate in the denoising process. To this end, we introduce \textit{REPresentation Injection} (REPI), a training framework based on a scaffold-to-internalization strategy, in which projected encoder representations initially serve as a temporary scaffold and are then progressively internalized by the diffusion transformer. REPI outperforms REPA across a wide range of backbones and is highly complementary to it: combining the two yields substantial gains over either alone. Notably, with only 160K training steps, REPI + REPA matches vanilla SiT trained for 7M steps, a speedup of over $43.5\times$. Code will be available at https://jeneveuxpas.github.io/REPI

Modern AI models are aligned through post-training to adapt them to downstream tasks. Recent work shows that fine-tuning language models on narrow tasks can induce emergent misalignment (EM), causing broadly harmful behaviors beyond the training task. However, EM has been studied almost entirely in text-only tasks, leaving its manifestation in multimodal models unclear. In this paper, we define and analyze EM in the context of vision-language models. We first induce EM via fine-tuning on narrow multimodal tasks targeting vulnerable code, careless household-object use, and conspiratorial interpretations of ordinary scenes. Across fifteen commercial and open-source models with different scales, we find that narrow multimodal fine-tuning can induce coherent and broadly misaligned behavior that transfers to unrelated tasks, including misaligned opinions, visual factual dishonesty, unsafe image generation, vulnerability to visual jailbreaks, and risky agentic actions. We further find that multimodal EM does not depend on the apparent harmfulness of training data but is sensitive to training-evaluation modality alignment. EM can arise under both supervised fine-tuning and preference optimization and can propagate through intermediate reasoning. Finally, we explore several mitigation strategies, including prompt inoculation, benign continued training, and activation-level steering, which can partially reduce EM. Overall, our findings suggest that multimodal EM reflects a behavioral shift rather than a general loss of capability, extending beyond text to the visual modality.

Recent work has explored improving agents by jointly evolving their harnesses and models, but often takes a ''potpourri'' approach that bundles together new tools, new decision-making procedures, and model adaptation to the evolved harness under a single notion of agent improvement. In this paper, we instead investigate how agents can improve their decision-making procedures. In particular, we propose EvoIn, an agent fine-tuning framework that bridges evolution and internalization. EvoIn first analyzes agent execution traces to evolve and validate new decision-making procedures by temporarily instantiating them in the harness. The validated procedures guide the agent to generate improved reasoning traces. These traces are then rewritten into self-contained reasoning traces, removing explicit references to harness instructions while expressing the induced decision logic as the model's own reasoning. Finally, EvoIn fine-tunes the model on the rewritten traces, internalizing these procedures so that the improved decision-making persists without the evolved harness at inference time. We evaluate EvoIn on diverse benchmarks and find that it consistently enables agents to learn stronger decision-making procedures, raising the pass rate by 10.9 points in-domain and by 9.2 points out-of-domain. Results further show that the internalized decision procedures generalize to unseen tasks. Case studies show that agents can learn to decide how to solve a task before solving it, for example by checking a document's length to choose between reading it in full and searching it. EvoIn is also broadly applicable, showing consistent improvements on another model family.

An argument can be surprising coming from a particular speaker without being a bad argument. Do AI evaluators keep these judgments apart? Two preregistered descriptive studies and a later Jev supplement collected 2,976 usable evaluations of six fixed texts about US AI policy, Germany's debt brake and Swiss nuclear energy. Each text was presented under several source attributions. The key comparison asks whether the gap between two sources changes when the argument changes. On Sol, for example, a national-security argument received mean ratings of 0.359 under CODEPINK and 0.639 under College Republicans; a civil-rights argument received 0.742 and 0.721. A constant preference for one source cannot explain that pattern. Related interactions appeared across topics and recent model configurations, including those with reasoning enabled, while several comparisons yielded small effects. The later European Jev supplement yielded five interactions below the adopted absolute reference of 0.05; its distinct rubric and interrupted collection limit comparison with the chat systems. Some written evaluations explicitly invoked a mismatch between a source and its attributed position. Taken together, the numerical and verbal evidence supports source-position coherence as a plausible explanation, alongside competing accounts involving credibility, authenticity and interpretation of the task. The paper develops this inference through controlled comparisons, reports conditional post hoc p-values in an appendix, and documents the human decisions and delegated checks behind an AI-conducted study.

Foundation model recommender systems require user context that can be consumed by large language models, reasoned over, and refined through natural-language interaction. Traditional behavioral embedding vectors remain highly effective for retrieval and ranking, but they are opaque to users and not natively expressed for language model workflows. We present Textual User Taste, a system that generates structured natural-language taste profiles from listening behavior, interaction signals, content metadata, and optional user feedback, and deploys them to millions of Spotify users. We describe the end-to-end production lifecycle required to generate, evaluate, optimize, and maintain these representations at industrial scale, including prompt development and compression, user steering, and integration with downstream personalization systems. Because no unique ground-truth taste profile exists, we introduce a multi-faceted evaluation framework to evaluate taste profiles as a production representation: they carry user-specific predictive signal independently, and when integrated with behavioral embeddings, improve MRR by 0.6% for future-track prediction and NDCG@7 by 2.2% for search ranking. Our evaluation also reveals that taste profiles support positive natural-language steering, while exposing important limitations, including challenges with negation and short-term temporal adaptation. These findings position taste profiles not as replacements for behavioral embeddings, but as an interpretable and steerable interface between evolving user context and foundation-model recommender systems.

In an era where software development is deeply tied with Large Language Models, does Model Driven Engineering (MDE) still make sense? This raises the question of the extent to which MDE can be successfully combined with an LLM approach to address the downsides of each approach separately. In this paper, we try to answer that question by investigating a central Research Question: Can Agents powered by Large Language Models improve code generated from UML diagrams and text specifications by Model Driven Engineering (MDE) tools? To investigate this problem, we developed a novel LLM-powered Multi-Agentic approach called ARTHUR (Architecture Refactoring Through Hybrid UML Reasoning), a framework that aims at combining the reliability of MDE and the ease of use of LLMs. ARTHUR is designed to refactor legacy Java code produced by traditional, rule-based, MDE code generators from UML diagrams. To asses the answer to our question, we have refactored the legacy code of several projects from our custom dataset crafted for this purpose. \name{} which made it possible to add support for modern frameworks like Spring Boot, while ensuring compliance with Model-Based Testing techniques to verify that the code still corresponds to the initial model's specifications. We then measured the results obtained in terms of time and cost, passing test rate, and \texttt{compile@k}, \texttt{pass@k} and \texttt{pass$^k$} metrics. We also observed the effect of generating code directly from the conceptual model without the refactoring. Our preliminary test results show that MDE could not be more far from retirement, after all.

Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On 6,925 MMLU-Pro debates, the protocol identifies 253 collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents 29 collapses but loses 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate 8-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high (G=7, Spearman rho=0.893, exact two-sided p=0.0123), but initial-majority accuracy is a close comparator (rho=0.821; family partial rho=0.767, p=0.0877), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model-scaffold rows can be compared under the same denominators and signed utility ledger.

Recurrent and equilibrium graph neural networks (GNNs) often enforce a unique fixed point or use one training target per graph. Yet many combinatorial and scientific problems admit multiple valid solutions, with no preferred one. A designated target can then impose an arbitrary selection rule. For tasks invariant to node relabeling, a symmetric graph may have a symmetric solution set but no symmetric solution. We show that multiple equilibria enable one weight-tied message-passing GNN to represent set-valued equivariant maps: different initializations approach different valid solutions. Under stated regularity assumptions, we first construct globally Lipschitz, permutation-equivariant dynamics that converge almost surely to valid solutions and reach every solution branch with positive probability. We then establish approximate realization by recurrent message passing with continuous component maps, with arbitrarily small update and limiting errors and arbitrarily high probability. This goes beyond standard universality arguments: although message passing alone cannot distinguish symmetric nodes, the evolving state keeps nodes distinguishable at every finite step without auxiliary node identifiers. Such dynamics can be learned without solution labels using problem-specific energies. On Ising ground states, structural module detection in protein graphs, and chemical reaction steady states, the learned updates produce multiple high-quality predictions with high numerical convergence rates. They achieve better average solution quality than the tested unique-equilibrium, single-target, and feedforward baselines, while remaining competitive with much larger diffusion-based solvers.

Despite substantial progress across downstream applications, vision-language models (VLMs) remain susceptible to language bias, often prioritizing linguistic cues over visual evidence and consequently producing incorrect predictions. Prior studies have proposed various approaches to understanding and mitigating language bias in VLMs, yet their findings often conflict due to the difficulty of tracing how language bias propagates within black-box VLMs. Building on the word completion task, we trace how language bias propagates through VLM inference by (1) proposing a diagnostic framework that decomposes the inference process into four distinct yet interdependent stages to trace the propagation of language bias; and (2) examining how two key factors underlying language bias, i.e., linguistic priors and cross-modal coverage, evolve across these stages and ultimately give rise to incorrect predictions. The linguistic prior captures the strength of statistical bias induced by the language model component of a VLM and represents the origin of language bias, whereas cross-modal coverage measures the extent to which linguistic cues cover the visual content. By decomposing inference into four stages and characterizing the interplay between linguistic priors and cross-modal coverage across these stages, we propose a systematic framework for tracing the propagation of language bias throughout the inference process; and uncover the underlying mechanism of language bias by revealing the interplay between linguistic priors and cross-modal coverage.

The rising number of concept unlearning techniques for text-to-image (T2I) diffusion models has produced a fragmented evaluation landscape. Methods are assessed under heterogeneous experimental conditions making principled cross-method comparison difficult. We present eval-unlearn, an open-source Python library providing a unified, reproducible benchmarking framework for concept unlearning in T2I Diffusion models. eval-unlearn integrates twelve published unlearning techniques spanning fine-tuning, closed-form model editing, and inference-time intervention, alongside nine complementary evaluation metrics covering erasure efficacy, adversarial robustness, generative quality, and concept retention. Its plugin architecture lets third-party techniques and metrics self-register without modifying the core framework, and its streaming, batched pipeline supports efficient evaluation of both standard NSFW concepts and arbitrary general concepts. As a further contribution, we release a public leaderboard on HuggingFace along with an interactive tool for real-time evaluation of unlearning techniques. The leaderboard compares nudity concept erasure case study across all twelve techniques, exposing significant accuracy-quality trade-offs that are obscured by heterogeneous evaluation. eval-unlearn is released under the MIT license; the package, code, leaderboard, and documentation are all available at https://eval-unlearn.readthedocs.io.

Reinforcement learning (RL) with sparse rewards is challenging because delayed outcomes provide little guidance about which intermediate computations caused success or failure. We argue that reliable credit assignment requires policy dynamics that preserve and expose credit-relevant information over time, a role we formalize as Temporal Credit Carriers (TCCs) and that spiking neural networks (SNNs) naturally fulfill through graded membrane traces and event-driven spikes. Based on this hypothesis, we propose SpikeCredit, an SNN-based framework for RL with sparse rewards that first performs task-adaptive TCC selection and then closes the loop between a fast TCC-reading pathway, where self-motion feedback constraint uses local behavior-grounded cues to constrain transition-level credit recovery, and a slow TCC-writing pathway, where credit-targeted trace alignment feeds recovered credit back into the actor to make future TCC dynamics more credit-readable. Across sparse-reward MuJoCo tasks, SpikeCredit improves Last10 return over sparse SNN baselines by +1169% on Ant, +953% on Hopper, +723% on Swimmer, and +1781% on Walker2d, and exceeds the dense-reward baseline on Swimmer by +113%. Mechanistic analyses further show substantially stronger alignment with dense rewards than the sparse SNN baseline. These results position spiking dynamics as credit-preserving substrates for sparse-reward RL.

Pipeline parallelism can improve prefill throughput by processing multiple request chunks concurrently across different stages of the model. However, keeping the pipeline fully utilized requires efficient scheduling and request preparation. In systems where stages retain and evict cache state independently, a local cache hit does not guarantee that the same prefix can be reused across the pipeline. Here, coordination overhead can impede request admission cadence and thus reduce overall throughput. In this paper, we present WavePP, a prefill runtime built on top of TensorRT-LLM that addresses these challenges by overlapping request admission with pipeline execution. WavePP asynchronously finds a prefix that can be reused across all stages, protects the cached state, and reserves space for the remaining input while earlier requests continue to execute. It subsequently plans the chunk sizes of each request dynamically to maximize pipeline fill. Each stage then completes the local preparation before executing the request. In the same system and pipeline topology, WavePP improves TensorRT-LLM's prefill throughput in 37 of 40 tested settings on GLM 5.2 and MiniMax M2.7. At concurrency 128 with high cache reuse, these changes increase throughput by factors of 2.91 and 2.02, respectively. Across 28 Kimi K3 settings, WavePP also has the highest measured throughput in all 18 settings at concurrency eight or higher, compared with tensor/expert-parallel and pipeline-parallel baselines from TRT-LLM, SGLang, and vLLM.

LLM personalization aims to generate responses aligned with individual users' preferences and needs. User-specific rubrics make these expectations explicit, providing direct supervision on what a satisfactory answer should cover. Existing rubric-guided approaches, however, exploit such guidance only at a coarse granularity, either by using rubrics to supervise the prediction of relevant aspects for subsequent generation or by reducing aspect coverage to a single response-level reward for reinforcement learning. This leaves a gap between specifying what a personalized answer should contain and teaching the model how to generate it. To bridge this gap, we propose GRASP, a rubric-aware on-policy self-distillation framework for LLM personalization that turns user-specific rubric aspects into fine-grained, token-level supervision. Specifically, GRASP pairs a rubric-free student with a rubric-informed teacher that additionally receives the target user-specific rubrics. By aligning their next-token distributions along on-policy trajectories generated by the student, GRASP transfers the teacher's rubric-conditioned guidance into the student, translating user-specific semantic requirements into dense token-level supervision. Since rubric-informed teachers can still produce inadequate supervision, we further introduce Rubric-based Teacher Validation (RTV), which retains only instances where the teacher sufficiently covers the target aspects, improving both supervision quality and training efficiency. Experiments on the LaMP-QA benchmark for personalized question answering demonstrate that GRASP achieves state-of-the-art performance across multiple backbones, supporting the effectiveness of rubric-guided token-level supervision for personalization. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/GRASP.

As language models take a growing role in AI development, a natural aspiration is for them to reflect on their own learning process, as humans do, and use that reflection to improve themselves. At the same time, these models have an advantage that human learners lack, since training leaves parameter-level traces that can, in principle, be inspected directly. However, current models cannot decode these traces into an explicit account of what they have learned. To this end, we introduce the \textit{Imprint Reader}, a model trained with \textit{Semantic Mount-and-Read Tuning} (SaRT) to describe frozen weight updates. SMaRT mounts each update onto the Reader and uses an anchor-free meta-query to elicit a natural-language description, while no-change and random-perturbation controls discourage unsupported claims. On held-out updates, the joint Reader reaches judge-based Pass@100 of $2\%$ for knowledge and $16\%$ for behavior. These results demonstrate the feasibility of natural-language readout while pointing to reliability across updates as the next step. Beyond free-form generation, the Reader provides a differentiable proxy for the gap between a specified target behavior and a candidate weight update. Its coordinate-aligned gradients support intervention through MetaEdit. At a $0.5\%$ pruning rate, Reader-guided selection raises measured harmful-prompt refusal from $57.9\%$ to $64.1\%$ under a safety-maintenance target. Using behavior descriptions without target-task training data, MetaEdit increases the frequency of backtracking and sub-goal expressions in mathematical reasoning traces and raises BFCL Overall from $41.69\%$ to $44.60\%$.

Spiking neural networks are attractive for low-power speech command recognition, yet their latency has received far less attention than their energy efficiency, and their multi-timestep execution is widely assumed to make them slower than quantized neural networks. This paper challenges the assumption that more local timesteps necessarily imply higher network latency. By overlapping computation across adjacent layers at the timestep level, SNNs may complete execution in less time than comparable bit-serial QNNs. However, this overlap relies on spikes firing on incomplete inputs, and a spike once generated cannot be withdrawn, so its error persists and reduces accuracy. Waiting for more input before firing would seem to improve accuracy at the cost of reduced overlap. Yet we find and prove that this intuition fails at some layers, where even a small increase in waiting can change spike timing and downstream computation, making the network both slower and less accurate. We therefore propose a Pipeline Delay Search method which selects each layer's delay by balancing task-level accuracy gains against added network latency. We then adapt the selected configurations through spike-based quantization-aware training and bounded tuning of firing thresholds and initial membrane potentials. Together, these steps form Falcon, a framework for Fine-grained Analysis of Latency and Controlled firing which systematically analyzes and optimizes SNN latency under a spatial analog compute-in-memory mapping with shared digital engines. We evaluate Falcon on GSCV2 and SSC, achieving competitive accuracies of 96.31 and 83.02 at modeled network-core latencies of 119.64 and 124.00us, respectively. Together, our analysis and results show that SNNs can compute more yet finish faster, and wait longer yet predict worse, highlighting why Falcon matters for both latency and accuracy.

On-policy learning has been argued to reduce catastrophic forgetting, produce sparser parameter updates, and improve generalisation. However, existing comparisons between supervised fine-tuning and reinforcement learning vary many factors simultaneously, making the contribution of rollout policy difficult to isolate. We study the effect of rollout policy in a controlled strong-to-weak distillation setting, by independently varying rollout policy, token-level KL direction, and learning rate across the Llama3 and Qwen2.5 model families and reasoning tasks spanning scientific, medical, and arithmetic domains. Our analysis reveals a nuanced picture of distillation dynamics in which rollout policy does not necessarily play a central role. Instead, token-level KL direction more clearly shapes task performance and output coverage, while learning rate governs forgetting and update sparsity. Analysis of KL gradients and experiments along a continuous student-teacher rollout-policy spectrum explain this pattern: forward KL is remarkably robust to rollout policy, with its performance stable and strong despite changes to the rollout policy, whereas reverse KL is substantially more sensitive and favours student-generated rollouts. On-policy data nevertheless improves generalisation to harder variants of the Countdown arithmetic task under both KL directions, although this advantage does not reliably persist after subsequent RLVR. Our broader conclusions remain robust to removing gradient clipping, using sampled KL estimators, and training on tasks requiring longer reasoning chains. Overall, our results challenge the view that on-policy rollouts are inherently preferable and show that their value depends critically on the objective, evaluation setting, and optimisation hyperparameters.

Zeroth-order (ZO) optimization offers a memory-efficient alternative for LLM fine-tuning by estimating updates only from forward evaluations of perturbed parameters, without backpropagation or activation storage. However, in billion-parameter LLMs, isotropic perturbations often waste many forward evaluations on weakly informative directions. To make these evaluations more informative, existing ZO methods restrict perturbations to low-dimensional subspaces. Yet the quality of these subspaces is critical: overly compressed or poorly maintained spaces can miss useful update directions. To obtain a high-quality subspace for ZO updates, this paper proposes AIM-ZO, a ZO fine-tuning method based on Activation-Informed Subspace Maintenance. AIM-ZO uses forward activations as local directional information and continuously integrates them into a broad, evolving subspace over training. To access broader gradient-relevant structure while keeping individual perturbations low-dimensional, AIM-ZO activates only a smaller set of shared and sampled directions, decoupling the maintained width from the active width. We evaluate AIM-ZO across 5 LLMs and 11 downstream tasks under matched forward-evaluation budgets; its six-task average exceeds the strongest fully evaluated ZO baseline by 1.26 percentage points on OPT-2.7B and MeZO by 2.85 percentage points on OPT-30B. Our code is available at https://github.com/EkkoXy/AIM-ZO

Large language model agents are increasingly deployed for data-intensive work, yet reliable data analysis requires more than general-purpose reasoning and ad hoc tool augmentation. Data Agents, equipped with workflow harnesses, offer a promising paradigm for automating the end-to-end data science lifecycle. This paper examines Data Agents from a harness-centric perspective. First, we introduce a taxonomy of Data Agents and associated data environments, organizing the literature around five functional stages: perception, planning, execution, verification, and repair. Second, we analyze the key technical routes within each stage, identifying 15 distinct approaches ranging from data structure probing to data state reconstruction. Third, we identify four open reliability problems: inactive semantic calibration, missing clarification, missing experience transfer, and the missing verification-repair repository. These problems explain why silent failures can persist even when individual components function correctly, highlighting the need for rigorous workflow harnesses and shared reliability resources. Finally, we summarize the horizontal task families of Data Agents, examine their vertical application settings, and benchmarks for evaluation, while maintaining a companion repository at https://github.com/DEEP-PolyU/Awesome-Data-Agents.

Large Language Models (LLMs) offer a scalable way to simulate survey respondents using demographic profiles and observed reference responses. However, the conventional approach of predicting one question per prompt repeatedly encodes the same context, limits each target to a narrow set of reference responses, and prevents later predictions from using information in earlier answers. Predicting multiple questions in one prompt can reduce these costs, share a broader pool of references, and let later predictions build on earlier ones. This requires forming coherent batches, selecting shared references, and ordering questions and references effectively. We propose Semantically Coherent Batching and Ordering (SCBO), a training-free framework that addresses these challenges. SCBO first uses an LLM to extract compact semantic representations from survey items and filter out template noise. It then groups related questions into batches and builds a shared reference bank using target-specific retrieval and centroid-based completion. Finally, it orders target questions from easy to hard and arranges references according to their semantic alignment with those questions. Experiments on four large-scale survey datasets and four LLMs show that SCBO substantially reduces token consumption and inference time while generally improving prediction accuracy over a non-batched baseline. Code is available at https://anonymous.4open.science/r/SCBO-41D8.

Pretrained robot manipulation policies such as vision-language-action models (VLAs) or world-action models (WAMs) leave interaction-relevant metric geometry implicit. Recent breakthroughs in spatial reconstruction can supply the necessary geometry reliably, but their features describe local shape without stating where it lies with respect to the robot. How best to deliver these features to a pretrained policy remains unresolved. We propose Spatial Grafting, a versatile, lightweight spatial module that binds frozen reconstruction features to metric, robot-relative geometry. Spatial Grafting constructs metric-grounded spatial tokens and injects them into the flow-matching action expert through cross-attention, without modifying the host's perceptual pathway, so the host retains the full benefit of its pretraining. We evaluate it more broadly than any geometry-aware policy we compare against: one graft architecture, with no per-host redesign, on two VLAs and two WAMs, across four simulation benchmarks that span short-horizon manipulation, visual robustness, clutter and long-horizon mobile manipulation, and on three real-robot platforms with single- and dual-arm configurations. On RoboTwin 2.0, a dual-arm manipulation benchmark, the graft improves every host across VLAs and WAMs. Grafted $π_{0.5}$ gains 11.3% and 15.6% on clean and randomized scenes, reaching 94.0% and 92.4%, above the strongest published 3D-conditioned policy, WAM4D (93.8% and 89.9%). The margin widens as the horizon lengthens: on tasks from BEHAVIOR-1K, a dual-arm mobile manipulation challenge scored by average task progress, it surpasses the 2025 challenge winner on five of six tasks,by up to 0.47 Q-score, and exceeds a map-conditioned spatial policy on average across the three tasks both report.

When a VLM answers a visual query, current interpretability tools rely on text rationales, which use a mismatched modality, or on internal read-outs, which originate too early to reflect the final output and require white-box access to the model. We introduce AnswerMap, a training-free, task-agnostic, black-box visual rationale constructed from the output head. The image is cut into K row and K column bands, each shown alone to the frozen model along with the query in the format of a yes/no relevance question. The outer product of the row and column ``yes'' posteriors gives the query-conditioned spatial map. Crucially, by defining a fixed read-out R (e.g., expectation, maximum) on top of AnswerMap, we can derive continuous outputs like location natively. This bypasses the reliance on discrete text tokens for continuous-output tasks and guarantees an image-dependent answer by construction. However, a rationale can be confabulated, so we validate AnswerMap across four models and three query distributions with two tests: (a) agreement with the model's own generated point and (b) deletion of the map's region. The map lands where the model points (AUC 0.85 against 0.38 for attention), and deleting its region flips 53% of correct answers (against 19% for attention's). Beyond establishing faithfulness, we demonstrate the map's task-agnostic utility through three distinct read-outs: its maximum flags hallucinated objects without generation, its expectation localizes correctly when the model's own pointing fails, and its top-mass region, fed back as a crop, fixes half of the model's wrong answers. AnswerMap thus offers a new lens on VLM interpretability and, through its read-outs, a new output interface for visual tasks beyond text tokens.

Artificial-intelligence studies using computed tomography (CT) for lung cancer are often broadly labelled "prediction" despite addressing clinically distinct tasks. We systematically mapped CT/low-dose CT (LDCT)-centered lung-cancer AI using five-database retrieval, full-text eligibility assessment, role-aware modality/omics extraction, clinical-task classification, and a Multi-Tier Evidence Graph (MTEG). The final corpus comprised 293 studies (2016-2026): 230 Detection, 8 future Risk-prediction, and 55 Other studies. Clinical variables (96.2%), 3D CT/LDCT (73.0%), and radiomics (63.5%) predominated, whereas external validation (29.0%), calibration (20.5%), decision-curve analysis (13.0%), longitudinal CT (17.7%), and saliency/attribution XAI (21.5%) were less frequent. The MTEG comprised 377 nodes and 3,444 edges; only 31 studies (10.6%) completed the six-tier substantive evidence chain, with greatest attrition at reasoning/explanation. Overall, the literature is detection-dominated, genuine future risk prediction remains uncommon, and complete translational evidence chains are rare.

Long-horizon agents improve solutions through sustained interaction, execution, and task feedback. Scaling studies relate performance to resources and capabilities, yet how existing capabilities shape returns to extended interaction remains less understood. To address this gap, we analyze AutoLab and EdgeBench, two long-horizon benchmarks. We find that starting performance and subsequent growth are associated with different capabilities: within a task category, similar early scores can precede different later gains. To formalize this finding, we model capability-time scaling with category-specific logistic power laws shared across models. Fitted to early trajectories, these curves extrapolate the observed models' category-average scores to later computation. However, rising average scores mask narrowing improvement opportunities: later gains concentrate among fewer improving models. High final scores and continued improvement also have distinct capability profiles. Predicted mean gains estimate each model's fraction of improving tasks; averaging these estimates forecasts the average share of improving models. These uneven returns motivate deciding whether a specific run should continue. We therefore derive a continuation policy to save time and compute with limited score loss. The policy conditions growth predictions on the run's observed progress and weighs immediate and delayed gains against computation costs. In replay with training and price calibration based on other models' histories, the policy saves roughly one-third of full-run time, with relative score losses of 2.4% on AutoLab individual runs and 3.3% on EdgeBench published mean curves. Our repository is available at https://github.com/Chihaya-Anon-chan/long-horizon-scaling.

Large language model (LLM) agents face critical privacy risks when acting as delegates in human-agent-human communication. To prevent such breaches, agents must understand users' social relationships and adhere to context-dependent social information disclosure boundaries. Current studies on agent memory privacy focus on instantaneous interactions, leaving the long-term relational disclosure problem unexplored. In this paper, we propose EP-Mem, an Elastic Privacy Memory architecture that reframes privacy as user-owned boundary control across social roles. EP-Mem introduces (1) token-level memory driven by user-configurable a privacy policy that stratifies persons and events, combining domain-level default circulation rules with fact-level whitelist/blacklist exceptions; and (2) a pluggable sidecar with a privacy engine that aligns disclosure controls with memory across summary, detail, and boundary granularities, enforced throughout generation, storage, and retrieval. We construct EP-Bench, to our knowledge the first long-term multi-party benchmark with cross-session correlated events for policy-conditioned relational disclosure. Experiments show that EP-Mem achieves 94.0% privacy classification accuracy, improves disclosure-permission judgment from 22% to 68%, and reduces privacy leakage by 75.6%, while maintaining retrieval performance and cross-benchmark generalization.

Visual-token compression is effective for improving the efficiency of vision-language models, but under extreme compression budgets, token pruning can break visual grounding while learned resamplers increase parameter count, attention cost, and training complexity. We revisit compression through a token parameterization lens, separating (i) basis transformation and structured truncation (retained subspace/compressibility) from (ii) coordinate organization (optimization and cross-modal alignment). This view yields two coupled objectives, compressibility and learnability, which we formalize as unified functionals. Guided by these objectives, we design Braco, a lightweight four-step coder that combines transform-basis truncation, input-independent basis-coordinate embeddings, budget-dependent orthogonal re-parameterization, and learned spatial residual tokens from lightweight pooling. Experiments show that Braco forms the favorable empirical accuracy-efficiency frontier under $23\times$--$64\times$ compression and remains competitive at $144\times$, reaching 95.2% accuracy while reducing prefill FLOPs by 84.2%--86.7% relative to the uncompressed upper bound. Against prior methods, Braco matches or improves accuracy while achieving up to approximately 36% end-to-end speedup and using $16.6\times$/$78.8\times$ lower compressor latency/FLOPs.

Latent test-time scaling improves reasoning by refining hidden states during inference, but existing methods typically apply a single scalar reward to all editable latent tokens. For multimodal large language models, this global update ignores that generated tokens play different roles: some are sensitive to visual evidence, while others correspond to uncertain reasoning decisions. We present Token-Disentangled Latent Test-Time Scaling, an inference-time framework that makes latent refinement token-role-aware. Starting from an initial generated trajectory, we optimize a short hidden-state prefix while routing perception-side visual feedback to image-sensitive tokens and reasoning feedback to high-entropy tokens. Tokens selected by neither route are constrained by an anchor regularizer. Across both perception and reasoning benchmarks on Qwen2.5-VL-7B and InternVL3.5-8B, our method lifts macro accuracy over CoT by +2.57 and +1.51 respectively, and outperforms strong output-space test-time scaling baselines under matched decoded-candidate budgets. Code is available at https://github.com/Qwen-Applications/TD-LTTS.

Early detection of oral cancer via photographic imaging presents a promising avenue for large-scale oral cavity screening. However, the development of robust deep learning models is frequently hampered by the scarcity of high-quality, annotated datasets. To address this limitation, a novel and well-curated resource, the Photographic Multi-purpose Oral Cancer Imaging (PhotoMOCI) dataset, is introduced for developing models across multiple diagnostic tasks in oral oncology. Then, a comprehensive benchmark study was conducted to investigate how various data augmentation strategies influence the performance of image classifiers. Our analysis spans different generative AI frameworks, evaluating the efficacy of traditional methods against advanced generative approaches, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs). Additionally, we propose the Synthetic Image Filter (SIF), a mechanism to select specific samples based on two auxiliary models: Synthetic Proxy Classifier to ensure samples are representative of the target class and Synthetic Image Detector to verify they appear realistic, thereby selecting only the high-utility images that contribute to improving downstream performance. Across the evaluated datasets and classifiers, the best SIF-filtered setup improves accuracy over traditional augmentation in all cases, with gains of +1.73% and +2.35% on PhotoMOCI and +2.38% and +2.08% on KOCD for ResNet50 and ViT, respectively. Our findings reveal that while the direct application of generative data augmentation may yield performance drops, the integration of SIF, considering (i) how synthetic data looks real and (ii) how it reflects the discriminative features of the belonging class, provides a simple yet effective mechanism to filter out synthetic samples that confuse the classifier during training.

Sign language translation and generation share the goal of bidirectional alignment between text and sign representations. However, existing approaches either treat them as isolated tasks or are only verified on limited datasets, limiting effective modeling between modalities. In this paper, we propose SignFLIP, a unified LLM-centered framework for translation and generation. To enable bidirectional mapping between text and sign, SignFLIP adopts a symmetric architecture together with a stage-wise training strategy built on large-scale data. The shared sign--text representation is progressively refined: pre-alignment facilitates subsequent SLT, while the SLT-adapted representation further benefits SLG. Extensive experiments on multiple benchmarks show that SignFLIP shows competitive performance compared with task-specific models on both translation and generation tasks, as well as strong transferability to sign language recognition.

Masked-diffusion language models fill in masked positions in parallel and in no fixed order. Most practical text watermarks assume left-to-right generation. They key each token to the tokens before it, and in a diffusion model those tokens may still be masked. A fixed green list needs no such context, but it favors the same tokens at every position, so these tokens appear more often in watermarked text. An attacker who compares token frequencies in watermarked and unwatermarked text can recover the list and forge text that the provider's own detector accepts. We present TANGO, a watermark for masked-diffusion language models that keys each new token to a nearby token that is already unmasked. A secret key splits the vocabulary into color classes, and TANGO biases the new token toward a color determined by the key and the nearby token's color. The watermark is therefore embedded in pairs of tokens. Because the favored color changes from position to position, token frequencies stay much closer to those of unwatermarked text than under a fixed green list. Detection needs only the text and the key, and it does not assume any unmasking order. On two masked-diffusion models, TANGO detects nearly all unedited watermarked texts and most edited ones, and frequency attacks that forge the fixed green list fail against it.

Relational deep learning models database rows and foreign-key links as a heterogeneous graph for prediction from record attributes and relational context. These graphs contain two distinct temporal signals: record age changes with the prediction cutoff, while intervals between observed records remain fixed. Prior work often treats time as a single signal or studies temporal representation and pretraining separately. We investigate how explicitly encoding both signals affects temporal pretraining for downstream tasks. Our framework combines Multi-scale Time Encoding, which captures record age using learnable time scales and type-specific projections, with Rotary Time Encoding, which represents signed inter-record intervals through rotary transformations during graph propagation. We pair these encodings with three self-supervised objectives: historical relation recovery, horizon-aware future relation activity prediction, and temporal subgraph contrast. All inputs respect their observation cutoffs. Pretraining proceeds in two stages: subgraph contrast first learns neighborhood representations, followed by refinement through either relation recovery or future activity prediction. We evaluate on five RelBench datasets across 11 classification and regression tasks using heterogeneous GNN and graph Transformer backbones. With both encodings, the best evaluated staged schedules improve over supervised training with the same encodings by 3.02% and 1.06% on the two backbones, respectively, and over controls without pretraining or either encoding by 3.24% and 2.37%.

Modern probabilistic machine learning models increasingly produce multivariate outputs with complex dependence structure, from multi-step time-series forecasts to sample path predictions. Understanding which input features drive the predictive uncertainty is important for risk-aware decisions, model diagnostics, and deciding whether the uncertainty should be mitigated or hedged against. This attribution problem requires a choice of how dependencies between output components are treated. Existing approaches reduce the output to a scalar through aggregation or projection before attribution, thereby obscuring whether features affect marginal uncertainty, dependence structure, or both, while component-wise analyses can miss dependence effects entirely. We close this gap by introducing a hierarchy of three entropy-based Shapley games that make this output-side choice explicit for any ordered multivariate outcome, ranging from per-component marginal entropy to fully joint entropy. The hierarchy isolates a cross-component attribution term that captures how each feature shifts the dependence between output components, a quantity invisible to component-wise methods. We establish a chain-rule decomposition of the joint attribution and characterize the cross-component term through conditional total correlation, providing both closed-form and sample-based estimators. Finally, we demonstrate how the framework captures differences in learned joint structure across probabilistic models from distributional regression to a zero-shot time series foundation model.

Terminal utility evaluates a complete agentic workflow, but learning requires credit for the decisions within it. We introduce Attentive Search over Counterfactual Trees (ASCT), a framework that turns training-time multi-step search into local action credit. At actor-visited states, an auxiliary tree evaluates alternative legal actions from the same recoverable prefix. Its action-value table is centered by the frozen actor's probabilities and supplies credit for PPO on actor-sampled trajectories. This protocol connects counterfactual evaluation to policy learning while deploying the actor alone. Uniform, UCT, and cost-aware AgentUCT instantiate the framework. On HotpotQA agentic retrieval-augmented generation, all three improve mean held-out utility over trajectory-return PPO and workflow-adapted VinePPO. Across three seeds, ASCT-AgentUCT reaches 0.6187 utility versus 0.5939 for VinePPO, with gains in answer F1 and execution cost, and uses 50.3% fewer recorded auxiliary Qwen tokens. Transfer and component-description studies examine the learned policies beyond the training setting.

Cardiac model personalisation requires inferring mechanical parameters that are not directly measurable in vivo. Ultrafast ultrasound shear wave elastography (SWE) enables non-invasive tracking of myocardial stiffness dynamics over the cardiac cycle, providing a target for personalisation. However, mapping these observations to subject specific model parameters remains ill-posed, as multiple parameter sets can reproduce the same stiffness dynamics. We formulate SWE-informed personalisation as a statistical inference problem using simulation-based inference (SBI). Using a subject-adapted 0D cardiovascular model and neural posterior estimation, we estimate model-conditional posterior distributions over active stiffness scale k0, contraction rate kATP, and relaxation rate kSR, conditioned on SWE-derived curve features and subject specific context. Among six healthy volunteers, four passed objective prior-support diagnostics and were retained for quantitative posterior analysis. Curve-level RMSE against the observed SWE target decreased from 12.61 $\pm$ 5.55 kPa for the prior predictive median to 1.14 $\pm$ 0.38 kPa for the posterior predictive median, an 89.7 $\pm$ 4.2% reduction. Posterior analysis revealed parameter-specific uncertainty, k0-kATP compensation, weaker constraint of kSR, and the importance of prior-predictive diagnostics for assessing whether each subject is represented within the modelled SWE feature space. These results support SBI for uncertainty aware SWE-based personalisation, while identifying prior support and forward-model adequacy as key diagnostics.

Multi-view clustering aims to capture cross-view consistency while exploiting view-specific information. However, shared representations learned to capture cross-view consistency may still retain view-identifying information, potentially compromising the consistency of cross-view clustering structures. To address this issue, we propose ACGRL, an adversarial consistency-guided representation learning framework for multi-view clustering. ACGRL employs a gradient-reversal view discriminator to reduce view identifiability and obtain invariant reference representations. These representations are then frozen to provide fixed references for disentangling view-specific information from cross-view common information in the subsequent learning stage. The fixed reference representations are concatenated with the learned view-specific representations for reconstruction and clustering, with cross-view cluster alignment encouraging consistent clustering assignments. Experiments on four benchmark datasets demonstrate the superior clustering performance of ACGRL compared with representative multi-view clustering methods.

On-policy distillation (OPD) is a widely adopted post-training technique for LLM reasoning. It is commonly believed to transfer knowledge from a stronger teacher, yet what OPD actually distills into the student's internal representations remains unclear. We study this question with sparse crosscoders, which learn one feature dictionary shared by the student before and after OPD and the teacher. Standard crosscoder analyses, however, identify model-specific features but cannot tell how a model's use of its features changes, since all models are encoded into one set of feature activations. We therefore propose the swap readout, which reads each student checkpoint's feature activations on its own, measuring how training changes the student's use of each feature, even for checkpoints unseen by the crosscoder. Across three OPD settings, we find that OPD neither creates features nor passes on the teacher's own, and leaves the firing rates of over 98% of the student's frequently used features within 20%. We further examine the SFT warm-up on the teacher's rollouts that commonly precedes OPD and makes it more effective. Rather than adding features, the warm-up reweights the shared ones in two ways. First, it already raises and lowers many of the features that OPD later raises and lowers, doing part of OPD's work in advance. Second, it changes features that OPD alone would not, notably those for conversation format, reasoning style, and mathematical notation, and these changes persist through OPD. Imposing this reweighting on a directly distilled student's features, without changing its weights, brings its accuracy close to that of the warmed-up student, whereas the same change on shuffled features does not. Together, these findings suggest that OPD reweights existing features rather than acquiring new ones: the student learns from the teacher how to use the features they already share.

Aligning language models with a specified normative framework requires translating abstract principles into concrete examples and preference signals from which models can learn. We present an expert-driven methodology for constructing such alignment data and apply it to a normative framework grounded in Islamic ethical, theological, and jurisprudential traditions. Over approximately one year, seven domain experts systematically probed language models to identify alignment deficiencies, curated desired responses, and constructed preference pairs from model outputs and expert judgments. The resulting Arabic-English datasets contain approximately 2.8K supervised fine-tuning (SFT) examples and 5.4K preference pairs spanning a broad range of normative domains. We evaluate the datasets through controlled post-training experiments comparing a Baseline model with models incorporating the curated SFT data alone and both the SFT and preference data. In blind expert evaluation on 150 separately constructed prompts, the model trained with the curated SFT data was preferred over the Baseline in 51.3% of assessor judgments, compared with 14.4% in the opposite direction (p < .001 at the prompt level). Adding the preference data resulted in a smaller difference, with the model trained with both datasets preferred over the SFT model in 28.0% of judgments versus 20.9% in the opposite direction; this difference was not statistically significant at the prompt level (p = .166). Standard Arabic and English benchmarks show no broad degradation in general-purpose capabilities. These results demonstrate how expert-defined normative principles can be systematically operationalized into alignment data and evaluated through controlled model training.

Memory-dependent manipulation requires robots to make decisions using information that is no longer available to their current sensors, such as recalling an earlier visual cue, tracking task progress, counting repeated events, or estimating elapsed time. We present ReCAT, a language-conditioned policy with structured recurrent memory. An instruction-conditioned encoder forms features from the current observation. A recurrent memory integrates the observation stream through Mamba-2 layers and one causal attention layer. A flow-matching Transformer decoder reads the current and the historical representation through separate cross-attention in every block. ReCAT reaches 95.3\% average success on LIBERO and 62.4\% on RMBench, with the best or tied-best result on six of nine tasks. On three real-robot tasks probing spatial recall, event counting, and interval timing, the best ReCAT variant reaches 66.7\% average success, against 8.3\% for the strongest short-history baseline. Controlled comparisons within ReCAT show that the observation encoder and every-block memory conditioning are needed for this performance. They also show that update rules developed for efficient sequence modeling behave differently as robot memory: additive updates have the highest observed success on counting and timing, and delta-rule updates on spatial recall. Project website is at https://intuitive-robots.github.io/ReCAT

LLM coding agents can traverse hundreds of intermediate code states before submitting a solution. Evaluating only the final artifact leaves the evolution of security findings unmeasured. We introduce the Security Debt Line Integral (SDLI), which accumulates static-analysis risk when an agent reaches a new best test pass ratio. We instantiate it with four static application security testing (SAST) tools and study artifacts from 830 passing SWE-bench runs, 712 ProgramBench final workspaces, and 13 public MirrorCode trajectories. The two large populations use the final-state special case of SDLI. Two-tool Common Weakness Enumeration (CWE) class agreement occurs in 3.9% of SWE-bench runs and 26.2% of the 80 ProgramBench runs passing at least 90% of official tests. These are scanner findings, not validated vulnerability rates. Excluding three advisory-heavy classes reduces the latter rate to 6.2%. Same-task runs differ in their measured scores, while one reconstructed ProgramBench run exposes persistent findings from its first implementation write. A repair case study reduces the scanner signal while preserving tested behavior, but also reveals sensitivity to equivalent API rewrites. SDLI offers a way to study progress and security findings together. Its value for steering agents and confirming exploitable vulnerabilities remains to be established.

The meaning of a concept in use is shaped by the situation, task, goals, and prior knowledge. For example, a request to make a poster "visually appealing for a five-year-old" might evoke bright colors and cartoon imagery for one collaborator, but less text, bold shapes, and visual simplicity for another. We call such task-relevant differences conceptual misalignment. We introduce Alignment Games, a framework for making these differences visible and repairable during human-AI interaction. Drawing on theories of situated conceptualization, we characterize task-specific conceptual frames in terms of relevant attributes, values, relations, constraints, and priorities. We then define alignment moves that intervene on the situation, the reasoning used to interpret it, or the resulting frame. Through examples from educational content generation, creative coding, and argumentative writing, we show how these moves can be composed into repair sequences and derive design principles for supporting task-sufficient conceptual alignment at runtime.

For systems with steep gradients, sharp interfaces, or severe spatio-temporal coupling, Physics-informed neural networks (PINNs) suffer from spectral bias, geometric inflexibility, and boundary constraint conflicts, which undermine accuracy and convergence. To overcome these issues, we propose a geometry-adaptive and constraint-enhanced PINN (GAC-PINN). The framework comprises four components: a gradient-driven adaptive grid mapping (AGM) for diffeomorphic point concentration with Jacobian regularization, an adaptive bandwidth hard-constraint ansatz with spatially-varying boundary transition widths, a Gaussian Fourier feature mapping as a spectral preconditioner to further enhance high-wavenumber representation, and an operator-aware router that automatically selects the appropriate hard-constraint construction based on whether the governing PDE contains temporal derivatives. An AGM callback mechanism and a three-stage training strategy ensure stable coordination. Benchmarks including the viscous Burgers equation, a sharp-peaked 2D Poisson problem, and the Allen-Cahn phase-transition equation show that GAC-PINN attains relative (L^2) errors of ((1.747\pm 0.450)\times 10^{-4}), ((2.868\pm 0.947)\times 10^{-5}), and ((1.756 \pm 0.712)\times 10^{-3}), respectively, consistently outperforming the baselines. Ablation studies further reveal that AGM alone yields a substantially lower error than residual-based adaptive refinement (RAR), while RAR becomes beneficial only when combined with FFM, demonstrating a context-dependent module interaction. Convergence analysis verifies rapid error reduction and saturation with increasing resolution, establishing a practical adaptive framework for high-fidelity simulation of problems with localized sharp features in applied mechanics and computational physics.

Accurate segmentation of post-treatment brain metastases is essential for treatment planning, longitudinal disease monitoring, and quantitative assessment of therapeutic response. The BraTS-MET 2026 Task 1 challenge introduces a clinically relevant segmentation problem involving four anatomically distinct tumor subregions: non-enhancing tumor core (NETC), surrounding non-enhancing FLAIR hyperintensity (SNFH), enhancing tumor (ET), and the resection cavity (RC). Among these, RC segmentation is particularly challenging because of its low prevalence, heterogeneous postoperative appearance, and lesion-wise evaluation protocol, leading conventional segmentation networks to prioritize dominant tumor classes during optimization. The proposed nnU-Net-based framework explicitly addresses RC segmentation through four complementary components: (i) RC-weighted Dice and Cross-Entropy optimization to alleviate class imbalance, (ii) anatomically consistent cavity augmentation to increase the diversity of postoperative cavity appearances, (iii) a residual encoder architecture for enhanced multi-scale feature learning, and (iv) lesion-aware morphological post-processing to suppress false-positive cavity predictions while preserving anatomically plausible structures. The framework is evaluated on the BraTS-MET 2026 Task 1 online validation benchmark. Among the evaluated configurations, the ensemble model (Residual Encoder nnU-Net + nnU-Net + RC-aware CarveMix) achieves the best performance, with lesion-wise Dice scores of 0.732, 0.752, 0.708, and 0.575 and corresponding NSD scores of 0.794, 0.798, 0.727, and 0.474 for ET, TC, WT, and RC, respectively. These experimental results show that integrating RC-aware optimization, anatomically consistent augmentation, and lesion-aware post-processing provides an effective strategy for improving rare resection cavity segmentation in post-treatment brain metastases.

Understanding how two entities are connected often requires tracing multi-hop relations across documents to identify intermediate entities and supporting evidence that explain a connection. This is a task that appears frequently in scientific research and other knowledge-intensive analyses. We formalise this setting as multi-hop relation inference: given two known endpoint entities, we aim to recover the bridge entities and evidence-grounded reasoning chains that connect them across a document corpus, and to generate an explanation grounded in the retrieved evidence. Existing multi-hop RAG systems typically seek an unknown answer entity rather than explicitly recovering the connection between two known endpoints and graph-based approaches often rely on costly LLM-extracted knowledge graphs that limit scalability to large document collections. We introduce ConRAG, which builds a lightweight entity-document graph from entity co-occurrence and LLM-based entity filtering. Its connective retrieval infers and semantically ranks paths between two endpoints. On MuSiQue and 2WikiMultiHopQA, ConRAG consistently improves bridge entity and reasoning chain recovery over strong RAG baselines, while reducing graph-indexing token cost by up to roughly 1.5 orders of magnitude. Our results show that endpoint-constrained path retrieval provides an effective and index-efficient approach to evidence-grounded relation discovery.

Post-training foundation video models on heterogeneous reward-weighted data usually assume that all data categories induce compatible updates. This assumption is fragile when categories correspond to different skills, domains, or evaluation dimensions. We study this problem in text-to-video post-training, where VBench2.0 dimensions define data buckets and an external multimodal reward pipeline assigns sample weights. We propose G$^3$-LoRA (Gradient-Guided Grouped LoRA), a data organization procedure that probes category-level gradients induced by reward-weighted video samples, removes the shared global update direction, clusters categories by residual gradient compatibility, trains group-specific LoRA experts, and consolidates them into one adapter by weight merging followed by on-policy distillation from the experts. We motivate this procedure by viewing reward-weighted flow matching as velocity-field regression: incompatible reward dimensions may prefer different denoising directions in overlapping noisy latent regions, causing shared LoRA training to average capabilities. On Wan2.1-T2V-1.3B-Diffusers, the merged grouped adapter improves the matched VBench2.0 evaluation over the base model, a joint reward-weighted LoRA baseline, and random, semantic, and raw-gradient partitions trained with the same pipeline; an independent evaluator agrees, and on CogVideoX-2B grouping avoids the negative transfer of joint training. The gain is not uniform: merging compresses the largest specialist gains, distillation recovers part of this loss, and camera motion and several local-quality dimensions remain challenging. Together, these results suggest that gradient compatibility can serve as a practical diagnostic for organizing reward-weighted video post-training data.

Autoregressive large language model inference repeatedly invokes the target model to generate one token at a time, making generation sensitive to GPU memory movement and sequential execution. This study evaluates two-token multi-token prediction (MTP) against autoregressive decoding in a controlled single-request deployment on an NVIDIA A10G GPU. A 360-request benchmark covered plain-text, reasoning-intensive, and tool-calling workloads, while runtime telemetry, Nsight Systems, PyTorch Profiler, and selected Nsight Compute measurements were used to explain the observed performance. MTP increased output throughput by \(1.91\times\) to \(2.19\times\) across all prompts and reduced time to first output by 10.0--14.2\%. Median mean acceptance length ranged from 2.370 to 2.595 tokens per verification iteration. Profiling showed that MTP introduced a longer and more complex execution path, including proposal, sampling, attention, gathering, and reduction operations. However, it required 56.4--78.1\% fewer executions of the selected repeating CUDA Graph per generated token. The dominant MTP GEMM kernel was not faster than the dominant autoregressive GEMV kernel, and selected instances of both approached the A10G memory-bandwidth limit. These results show that MTP improved inference through amortization: greater token progress reduced repeated GPU execution sufficiently to outweigh the additional speculative-execution cost.

Transforming raw data into queryable knowledge requires both early extraction of reusable information and explicit types, relations, and applicability conditions for particular tasks. If indexing selects content too early around a single business schema, later tasks may be unable to use information that was omitted. If the index retains only open-ended text, however, rule-based reasoning lacks checkable premises. We propose 5W1H+Which, a semantic indexing design that separates content extraction from ontology binding. The 5W1H questions organize source-grounded content units; Which points to versioned ontology elements and records mapping relations, scope, and validation status. Time, location, system environment, and participant roles are not merely retrieval labels: together, they constrain the contexts in which facts, bindings, and rules apply. Unbound content remains searchable, while bound content enters a formal reasoning path only after premise checks. The method further distinguishes business valid time, system knowledge time, and operational traces, and uses dependency records to support binding revalidation and the maintenance of derived conclusions. A worked example of migration from an on-premises server to a cloud environment illustrates the different treatment of world-state changes, ontology-version changes, and changes in rule applicability. We formulate three groups of falsifiable hypotheses concerning cross-task evidence coverage, control of contextual misuse, and incremental update cost. The planned evaluation includes a strong typed fact-graph baseline with the same evidence, temporal information, and budget, to test whether benefits arise from 5W1H organization, deferred binding, or additional information and engineering effort. The contribution is a testable indexing mechanism, not a claim to a new universal ontology or a demonstrated performance advantage.

We describe AfS (Agent for Science), a platform built for long-horizon scientific work, where a project runs for tens of hours across dozens of agent runs with a human present only occasionally. Most agents for science are general coding agents with a skills folder attached, and they inherit that lineage's failure mode: under pressure to finish, they fabricate, skip, or smooth over. Our design rests on one claim: most of the credibility of machine-made research can be moved from asking the model to behave to making the non-compliant state unrepresentable. We encode research discipline as mechanically enforced laws (commitment before measurement; unforgeable freezing; reports are not facts; evidence persists but verdicts do not; negative results are first-class; mechanical questions to the framework and semantic judgment to the model), organized around three time horizons: a minimal set of research nodes within a run, an inquiry contract with frozen closure conditions and a hash-chained artifact ledger within a project, and a two-tier knowledge base with promotion by rewriting across projects. This is a system description written under one rule: each mechanism appears in exactly one place, with the invariant it enforces, the failure it prevents, the way it is realized, and the cost it imposes. It covers the node contract, the write-path gates, the two-tier memory, and the runtime substrate. Three traces walk real failure attempts through the mechanisms that catch them, and two closed campaigns are included as worked illustrations rather than as an evaluation. We report no benchmark: a process-integrity suite that would support quantitative comparison is under construction, and what we can measure today is only the operating cost of the machinery.

Estimating integrals of black-box, high-dimensional functions, from expectations and kernel mean embeddings to the softmax kernel in self-attention, is a basic subroutine in machine learning. Rank-1 lattice rules suit this setting: they query the integrand only at a fixed point set and need no gradients. When the $n$ points serve as a design matrix $X\in\mathbb{R}^{n\times d}$ for a feature map, however, computing $Ψ(X)^\top v$ or $Ψ(X)w$ for an elementwise nonlinearity $Ψ$ costs $O(nd)$ time and memory for any standard quasi-Monte Carlo point set. We study subgroup rank-1 lattices, whose Korobov generator $(1,t,\dots,t^{d-1})$ uses a scalar $t$ of fixed multiplicative order $m$. Splitting $\mathbb{F}_n^\times$ into cosets of $\langle t\rangle$ reduces both maps to short cyclic correlations evaluated by FFT, giving exact results for arbitrary $Ψ$ in $O(n\log m)$ time and $O(n)$ memory, without forming $X$. Since fixing $m$ falls outside classical component-by-component theory, we prove convergence directly: via resultants with the cyclotomic polynomial $Φ_m$, the squared worst-case error in the Korobov space decays as $O(n^{-(α-1)/(m-1)})$ for prime $m\ge d+1$, and this threshold is exact. Using the splitting of $n$ in $\mathbb{Q}(ζ_m)$, averaging over the $m-1$ admissible generators improves the constant by a factor $Θ(m-1)$. Empirically, the subgroup lattice beats Gaussian and orthogonal random features and scrambled Sobol' and Halton points in 49 of 54 synthetic kernel-estimation settings and all 45 softmax-attention settings on nine real datasets, and builds a sample set with $d=2048$, $n\approx4.1\times10^7$ in 2.3 ms.

We propose a lifted reformulation of supervised classification that improves the final accuracy of standard classifiers without changing the architecture at inference time. A network $N=N_2\circ N_1$ is split at a single semantic interface and one learnable prototype per class is inserted there. Training combines a quadratic consensus penalty that pulls $N_1(x)$ toward the prototype of its class with a classification loss of $N_2$ evaluated on samples drawn around the prototypes, whereat no gradient crosses the interface. At inference the prototypes are discarded and the unmodified network $N_2\circ N_1$ is used. Across CIFAR-10, CIFAR-100, and TinyImageNet with ResNet and vision transformer backbones, lifted training improves test accuracy by up to five percentage points over variants without lifting under a shared tuning protocol. Moreover, we provide theoretical justification of those results.

Saved checkpoints record states along a training trajectory, but generally do not determine the updates at states that would be visited under a different schedule. We study how accurately these checkpoints can reconstruct the endpoint of a sequential reference with prescribed update strengths. Under a common local transition model, two checkpoint-index moment conditions characterize all convex merges that agree with this reference through second order. We then prove an information limit that for nondegenerate profiles, no algorithm using only a fixed-length gradient-descent (GD) history with step size $h$ can achieve $o(h^3)$ endpoint error uniformly over a fixed class of smooth, strongly convex losses. The lower bound follows from two losses with identical GD checkpoint histories but sequential reference endpoints separated by $Ω(h^3)$. \textbf{Quadratic-Accurate Merging} (QAM) achieves a matching uniform $O(h^3)$ endpoint error bound. Its explicit coefficients also define the unique profile-dependent merge that exactly matches the sequential GD reference across all fixed quadratic objectives. Across two public Adam checkpoint trajectories (SmolLM3-3B and OpenEuroLLM-Prelude-9B), three windows and three profiles per model, and 15 tasks, QAM shows mixed results for short windows and broader advantages over \textbf{Warmup-Stable and Merge} (WSM) for longer windows. Matched-moment GSM8K diagnostics further show that local consistency alone does not fully determine downstream scores. These results characterize the reconstruction limits of saved histories, provide a coefficient rule that attains the optimal rate, and assess its practical utility.

Machine learning (ML) increasingly powers Internet of Things (IoT) applications at the edge. Yet producing a deployable edge ML artifact for a specific scenario requires navigating a huge search space spanning data representation, model design, training on domain-specific data, and runtime customization. This workflow is fragmented and difficult to scale across diverse edge applications. We present EdgeCraft, an LLM-driven system that turns high-level intent into deployable edge ML artifacts. Building such a system raises two challenges: (1) How can an LLM be guided to find high-quality solutions that meet dynamic SLOs for task quality, latency, and energy? (2) How can trustworthy target-device verification be obtained at low cost? EdgeCraft addresses these challenges with two designs. (1) A constraint-aware synthesis tree explores alternative candidates and uses measured SLO gaps to guide each improvement. (2) A multi-fidelity verifier progressively combines low-cost checks with full target-device verification to reduce verification cost while preserving reliable verification results. It also records verified failures for reuse, avoiding repeated device work. To support concurrency, EdgeCraft provides a multi-tenant runtime that runs cloud training and target-device verification in parallel while isolating requests. Across 50 public tasks, EdgeCraft exceeds the task-specific Reference in best-observed quality on 40 tasks and finds an SLO-feasible artifact on 45, with the two outcomes overlapping on 38 tasks. Moreover, EdgeCraft achieves competitive performance on our self-collected SEN dataset, suggesting its generalizability to real-world IoT sensing tasks.

Discrete diffusion models offer the ability to re-draft, revisiting and correcting earlier tokens throughout generation. This capability depends on the forward corruption process that defines what the denoiser learns to correct. Masked diffusion models fix tokens once they are unmasked, while uniform diffusion permits revisions but relies on uniformly random token substitutions. We instead learn which substitutions are most useful for training the denoiser to re-draft. We introduce Variational Stackelberg Discrete Diffusion (VSDD), a framework for learning a semantically aware corruption process. VSDD formulates training as a leader-follower game: the leader defines a Markovian corruption process parameterized by the denoiser's token embeddings, while the follower optimizes a variational denoising objective with the corruption process held fixed. The leader rewards corruptions based on how much the denoiser improves after learning from them, rather than on how easily the current denoiser can reconstruct them. We measure this improvement under a fixed reference corruption process, approximate the follower's response with a one-step gradient update, and optimize the leader using a score-function estimator. We evaluate VSDD across molecular, text, and playlist generation. VSDD substantially improves molecular validity over uniform and masked diffusion, reduces text perplexity relative to uniform diffusion while remaining competitive with masked diffusion, and achieves sizable improvements in offline playlist recommendation metrics.

We present an in-depth investigation of how a form of Bayesian reasoning about common causes can emerge as a cross-contextual generalization in small, tractable transformers. Incrementing on recent work, our set-up (i) disentangles causal mechanisms in the model from the causal structure of the true data-generating process, (ii) orients more towards natural language prediction by considering inference of latent common causes, and (iii) considers whether and how Bayesian evidence accumulation for latent common causes can be implemented in representations and mechanisms that allow for cross-context generalization to novel test cases.

A classical solution concept in fully observable nondeterministic (FOND) planning, is the strong policy (aka winning strategy in the closely related area of reactive synthesis), i.e., such a policy ensures that the goal is reached in an adversarial environment. When strong policies are not available or there is no evidence that the environment is adversarial, one can resort to best-effort policies, which always exist, and which follow the classic decision-theoretic principle that an agent should not use a dominated strategy. A typical positional best-effort policy works as follows: from every state, it follows a strong policy if one exists from that state (such states are called ``strong-winning''), else a weak policy if one exists from that state (``weak-winning''), and else is unconstrained (``losing''). In this work, we introduce a sound and complete planner for both best-effort planning and strong planning. The algorithm that underpins the planner is quite simple: it represents certain sets of states, such as the winning regions, by their $\subseteq$-minimal elements. The algorithm returns uniform policies, i.e., it returns a policy $π_t$ that is a strong solution starting in every strong-winning state, and it returns a policy $π_w$ that is a weak solution starting in every weak-winning state, and it provides a certificate for the set of losing states. We implemented the algorithm with some simple optimizations (calling it FONDANT), and evaluated it on a benchmark set consisting of the instances that were used in the evaluation of leading strong planners PR2 and FOND-SAT, and the best-effort planner BeSyftP. On coverage, our implementation is at least as good on all domains, and outperforms on some domains; and on wall time, it is slower on small and medium-sized instances, and outperforms on larger instances.

Multi-turn rollout dominates the cost of agentic reinforcement learning (RL). Asynchronous execution and elastic GPU resources can accelerate this stage, but adding rollout replicas yields diminishing returns while training GPUs remain idle between updates. We observe that effective resource use also depends on the prefill--decode (PD) configuration. Both the choice between colocation and disaggregation and the optimal PD ratio vary with the workload, making resource scaling and PD configuration interdependent. Exploiting this opportunity requires selecting effective configurations and realizing their benefits within transient resource-availability windows despite reconfiguration costs. We present PEARL, an asynchronous agentic RL system that coordinates external resource elasticity, temporary reuse of idle training GPUs, and adaptive PD execution. PEARL maintains a unified GPU--worker--role state and uses runtime profiles to predict rollout batch completion time, accounting for environment-induced reductions in decode concurrency. It selects the PD mode and ratio under the current GPU budget and translates each decision into an incremental transition plan that minimizes worker and role changes. Cost-aware switching and borrowing policies suppress transitions with insufficient expected benefit while ensuring timely return of training GPUs. Our evaluation show that PEARL achieves $2.17$--$2.79\times$ the throughput of fixed-resource ROLL across different LLMs. Compared with RLBoost+, throughput improves by up to approximately 26.9\% for Qwen3-8B and 36.3\% for Qwen3-30B-A3B.

Successive weaving segments (SWSs) on urban expressways are bottlenecks prone to recurrent congestion and collisions, requiring fine-grained active traffic management (ATM). Existing approaches struggle to balance the adaptive performance of data-driven optimization with the resilience and transferability of model-based control. We propose a hybrid framework to coordinate lane-level variable speed limits (VSLs) and ramp metering across SWSs. First, we reconstruct L-METANET, a lane-level macroscopic traffic flow model that captures free and forced lane changes. Second, we combine XGBoost-SHAP with a random-parameters binary logit (RPBL) model to derive analytical equations for merging and diverging collision risks and formulate system cost and reward functions. Third, we develop MPC-STMAPPO, a hierarchical controller integrating model predictive control (MPC) and multi-agent reinforcement learning (MARL). Its upper MPC layer uses L-METANET for long-horizon rolling optimization and generates baseline commands; its lower spatiotemporal MAPPO (ST-MAPPO) layer, enhanced with Mamba cells and graph attention, produces residual actions for short-horizon adjustment. Real-world experiments on the 18-km Eastern Expressway in Changchun, China, show that L-METANET accurately reproduces lane-changing-induced flow redistribution and capacity drops, with state evolution aligned with ground truth. XGBoost-SHAP-RPBL achieves AUCs above 0.80 in most tasks, outperforming conventional logit models. MPC-STMAPPO converges faster and performs better across multiple metrics than MPC- and MARL-based baselines. Under randomly fluctuating demand, it also significantly outperforms pure MARL in generalization, demonstrating strong potential for industrial deployment.

Agents need calibration when deployment conditions change: replacing a driving model, including a foundation-to-post-trained transition; crossing jurisdictions; or scaling across heterogeneous markets and sources. Interface compatibility alone does not establish capability retention or target-contract satisfaction. We formulate agent calibration as constrained behavioral adaptation across three interacting layers: information preservation, harness adaptation, and user acceptance; the layers apply to every scenario, not one-to-one to the three. The basic objective is non-degradation on prespecified capability measures while satisfying target requirements; aggregate improvement is stronger. Information calibration preserves independently validated source content still applicable to the target task. Harness calibration aligns observable artifacts at semantic checkpoints and repairs them through iteration, tool substitution, or local replanning within explicit budgets. User calibration enforces recipient-specific output contracts: templates, schemas, and section-level preferences. A global e-commerce example shows how shared standards coexist with site- and market-specific adapters and validation. We distinguish trainable policies from frozen-backbone configuration or controller optimization, and evidence verification from relative judgment and DPO/GRPO optimization. Recent harness-transfer and judge-validity studies motivate target-native execution records, separate audits of task validity and near-tie ranking, and matched target-native optimization controls. We propose held-out evaluations for model changes, cross-border adaptation, and scale, including a factorial test of source evidence and checkpoint repair and group-level reporting to prevent aggregate gains from masking local failures. This is a methodological proposal; implementation and empirical validation remain future work.

AI agents increasingly carry out long-horizon professional work, but their evaluations rarely require a finished creative deliverable. To this end, we introduce Timeline-Bench, a benchmark of 56 real video-editing tasks, each asking an agent to turn raw production material into a finished video. Tasks range from selecting dialog takes and shaping interview footage into a story to cutting commercials from product shots, voiceovers and graphics. Every task provides a brief, source assets, a container and a set of tests. A task is resolved when the output passes every test. The tests check the delivery format, the content and the brief's explicit requirements, and include a quality test calibrated on 2,582 blind judgments by 43 video editors. We evaluate 16 agents that pair frontier models with coding-agent harnesses such as Codex, Claude Code and OpenCode. The best, GPT-6 Astra in Codex with curated editorial guidance, resolves only 15 of the 56 tasks (26.8%), and the average agent resolves 14.0%. Human editors prefer the reference edit in 83.5% of judgments. Most unresolved runs (562 of 771) fail only the quality test: agents perceive footage through stills and transcripts and check their renders for defects, not craft. We release the tasks, verifier and per-run results at https://timelinebench.tensortest.com.

Multi-document retrieval-augmented generation (RAG) requires a language model to process multiple retrieved text chunks before answering a question. Precomputing each chunk's KV cache independently and concatenating the caches when the chunks are retrieved can accelerate this step. However, the assembled cache lacks cross-chunk attention information, reducing answer quality. Selective recomputation methods recover the missing cross-chunk context by rerunning the target LLM on selected tokens, incurring substantial online computation. We introduce CacheRepair, a lightweight network that learns the difference between independently computed KV caches and those produced by processing the chunks together. The network combines compressed KV features with token embeddings and uses attention that is bidirectional within each chunk and flows from earlier to later chunks. Each repair block receives the compressed cache features, and the predicted residual is added to every document token's cache. Each repair network is trained for a specific frozen target LLM on a generic retrieval corpus and reused across downstream datasets. Our analysis shows that repair reduces KV errors both near chunk boundaries and throughout chunk interiors. Evaluation across three target LLMs and four downstream datasets places CacheRepair on the measured answer-quality-latency Pareto frontier in eleven of twelve model-dataset combinations. Reported time to first token (TTFT) includes online cache transfer and repair. Across all twelve combinations, the largest repairers achieve 1.69-4.61$\times$ speedups in median TTFT over full prefill and improve mean F1 by 2.1-26.1 percentage points over direct cache reuse.

Latent world models predict future states for goal-directed planning using action chunks spanning multiple primitive steps. Existing methods typically use fixed-length chunks and either omit goal-conditioned action generation or limit their supervision to short goal spans. We introduce FlexiWorld, a JEPA-based world model that combines mixed-span goal supervision with variable-length action chunks to improve long-horizon control. During training, we sample varying goal spans and randomly partition the actions into variable-length chunks. We jointly train the world model with a causal action encoder that embeds variable-length chunks and an autoregressive actor that generates primitive actions sequentially. Student Forcing reduces exposure bias by training on generated action prefixes. For planning, Actor-Residual Cross-Entropy Method (ARCEM) combines action-residual search with within-chunk autoregressive feedback and chunk-boundary latent prediction. Across four benchmarks and goal distances, FlexiWorld with ARCEM achieves 89.29% mean success, compared with 83.98% for the strongest baseline. PushT ablations show improved direct control from mixed-span supervision, variable-length chunks, and Student Forcing. Without retraining, FlexiWorld supports different planning chunk lengths: longer chunks accelerate ARCEM by approximately $1.3\times$ on average while maintaining comparable average success.

Communication-efficient federated optimization commonly spends several gradient evaluations between server updates. Existing local-update methods use this computation to advance an independent model on each client. Under heterogeneous data, however, these models evaluate gradients at different locations, making the aggregated update difficult to interpret as a gradient of the global objective. We study an alternative use of the same computation budget: \emph{evaluate the global objective along a shared, predicted path}. We propose Common-Trajectory Predictive Federated Learning (\texttt{CTP-FL}). At each round, all clients construct the same sequence of query points from the current global model and the previous aggregated direction, evaluate $K$ stochastic gradients along this sequence, and upload their average. The server then performs a single global update. Thus, \texttt{CTP-FL} uses $K$ mini-batch gradients per client and one model-sized vector in each communication direction, matching the per-round computation and communication of full-participation FedAvg-M. Shared query points make the aggregated direction an unbiased estimator of the average \emph{global} gradient along the predicted path. The remaining discrepancy from the gradient at the current model is controlled by the path length, without assuming bounded client-gradient dissimilarity or bounded gradients. For smooth non-convex objectives, we establish an $\mathcal{O}\!\left( \sqrt{LΔσ^2/(NKR)}+LΔ/R \right)$ average-stationarity bound under full participation. The analysis isolates a testable trade-off: extending the prediction path provides more forward-looking gradient information but increases its displacement bias.

Hyperbolic neural networks introduce geometric operations that require explicit treatment in relevance propagation. Equivalent geometric realizations can produce different feature attributions, even when local relevance is conserved. We study this problem through Geometric Representation Invariance (GRI), a specialization of Implementation Invariance, and zero-curvature consistency, which requires identity relevance propagation when a geometric module approaches the identity. We propose LRP-radial-all for origin-centered radial modules, treating geometric scaling as modulation and assigning relevance entirely to the signal branch. The rule conserves relevance, is invariant to equivalent radial factorizations, and satisfies zero-curvature consistency, yielding GRI for a specified Poincaré-Lorentz logarithmic-map construction. In contrast, a conservative LRP-half baseline can violate both consistency criteria. Experiments on hyperbolic MNIST, sEEG, and CIFAR-10 classifiers assess attribution fidelity, qualitative explanations, and runtime. LRP-radial-all achieves competitive attribution fidelity across datasets with runtime comparable to Gradient$\times$Input and substantially lower than Integrated Gradients. These findings motivate geometry-aware propagation rules that distinguish relevance conservation from consistency across equivalent computations.

Patient responses to dental pulp testing, ranging from no sensation to intense pain, provide important information for assessing pulp status in endodontic diagnosis. However, pain is a subjective sensory and emotional experience that varies considerably across individuals and can be difficult to communicate. We investigated whether complementary autonomic signals could support objective assessment of responses during dental examination. Forty-nine patients underwent cold pulp testing, yielding no-response, mild-response, and intense-response conditions. The framework integrated ECG-derived skin nerve activity (SKNA) and R-R intervals (RRI), together with electrodermal activity (EDA), using temporal convolutional network encoders with attention-based mid-level fusion. Individual baseline signals and subject-level covariates, including anxiety scores and biological sex, were also incorporated. The framework achieved 80.2% balanced accuracy, 75.2% sensitivity, and 85.2% specificity for binary classification of no response versus mild or intense response. For three-class classification, it achieved 60.0% balanced accuracy and a 58.8% macro-averaged F1 score. Ablation and attention-weight analyses indicated that EDA contributed most strongly to model performance, followed by RRI, while SKNA improved balanced accuracy by approximately five percentage points. Age was significantly associated with model performance. These findings support the feasibility of multimodal autonomic sensing for objective, non-invasive assessment of responses to dental pulp stimulation.

Large language models (LLMs) are increasingly deployed with access to external tools, yet harmful tool-mediated interactions are less likely to be refused when compared to regular conversational ones. As this change in refusal behavior remains underexplored, we investigate its underlying mechanisms across a diverse set of open-weight language models. We find that information about the harmfulness of a request remains strongly encoded in the model's representations and transfers across conversational and tool-mediated inputs. Evidence from representation geometry and neuron-level analysis further indicates that the two interaction modes systematically distribute harm-related computation differently. Crucially, while conversational inputs can be refused at relatively low levels of perceived harmfulness, tool-mediated inputs remain permissive until harmfulness crosses a substantially higher effective refusal threshold. Moreover, tool-mediated refusal is also more brittle: progressively weakening the refusal computation disrupts tool-mediated refusal at lower intervention strengths than conversational refusal, even when benign capabilities remain intact. Together, our findings indicate that tool mediation does not simply reduce the internal perception of harm, but instead impacts its conversion into refusal. Overall, this suggests tool-mediated environments may intrinsically reduce robustness of models to harmful requests, and that conventional safety evaluations may not fully transfer to LLM agents.

Full-duplex speech models support streaming interaction that listens and speaks at the same time. Serving them is governed by a strict, repeating deadline: conversation advances on a one-second cadence, and every second of input must be turned into a second of speech before the next second arrives. Because stages within a session run in strict sequence, per-invocation overhead cannot be batched away. Profiling reveals that the autoregressive stages of a duplex second already fit within the period, whereas the token-to-audio synthesis tail is what causes overruns. This tail stage suffers from orchestration slack where the GPU is left waiting as thousands of tiny, regular operations are issued one by one, while also wasting substantial memory by over-provisioning state at static implementation constants. Existing remedies, such as graph recording and demand-sized allocation, fail because streaming state dynamics violate their prerequisites. The root cause is that the runtime lacks the model's native clocks: the per-region counters that govern advancement rates and retention policies. We propose DuplexCadence, which explicitly declares native clocks to the runtime and derives two mutually enabling rules: demand-sized state allocation at a stable address, and exact-shape graph replay without padding. The former eliminates idle memory and stabilizes tensor pointers, while the latter removes orchestration slack without padding overhead. Evaluated on four released models across three decoder architectures with bit-for-bit identical output, DuplexCadence reaches $2.85\times$ the stock runtime's speed at $38.8\%$ lower peak memory. On the live duplex path, mean SPEAK time falls from $14\%$ over the one-second cadence to $2\%$ under it, enabling models to reliably keep up with interactive speech while markedly expanding multi-

Symbolic regression (SR) seeks concise and interpretable mathematical expressions from data for scientific equation discovery. Existing SR benchmarks face a tradeoff between evaluation cost and benchmark validity. Repeated evaluation of large task pools is expensive, and compact benchmarks lack systematic evidence of preserved task diversity and algorithm discriminability. SymbolicArena provides a unified infrastructure for benchmark distillation and dynamic evaluation. The framework standardizes 664 heterogeneous tasks with executable ground truth expressions and distills the Full Task Set into Core50, a validated benchmark of 50 tasks. The distillation process preserves task coverage and algorithm discrimination under explicit balance constraints. SymbolicArena applies a unified execution protocol to heterogeneous SR algorithms and produces comparable outputs and search trajectories. Multi Axis Evaluation characterizes numerical quality, symbolic quality, and search behavior. Core50 reduces evaluation workload by 92.5% and maintains agreement with Full Task Set evaluations. Experiments show that SymbolicArena achieves 72.6% to 86.7% lower approximation error than alternative selectors, further supporting its fidelity to the Full Task Set. Evaluation reveals a substantial gap between numerical fitting and symbolic recovery across current SR methods, suggesting that reliable equation recovery remains an open challenge.

Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) with human preferences, yet most pipelines learn a single reward model that overlooks individual differences in preferences. Personalized reward models (PRMs) address this by conditioning rewards on user-specific feedback, most commonly through in-context learning (ICL), where a user's historical comparisons are supplied as contextual preference pairs. However, we identify a key limitation of ICL-based PRMs: they fail to capture the preference relations conveyed by contextual pairs. To address this, we propose Preference-Aligned Test-Time Training (P-TTT), which explicitly encodes these relations into user-specific fast weights for personalized reward prediction. P-TTT introduces sequence-level update and apply operations to match the response-level granularity of preference feedback, together with a preference-aligned objective that directly uses pairwise preference relations to guide fast-weight adaptation. Notably, P-TTT is simple to implement and computationally efficient, updating fast weights within a single forward pass without inference-time backpropagation. Extensive experiments show that P-TTT more effectively captures historical preference relations and outperforms state-of-the-art methods by a large margin.

Large language models (LLMs) can sometimes report perturbations to their internal activations---even when the input provides no evidence that an intervention occurred. How do models detect and localize such internal changes? We study this question using a controlled task that keeps the input text fixed. We either inject a concept vector into the hidden state at one of ten token positions or apply no intervention. The model is asked to identify the perturbed position or report that no intervention occurred. Across three model families, we identify two small groups of attention heads with distinct roles in introspective reporting. Middle-layer gate heads influence whether the model reports a change, while router heads in a later layer help select the position to report. Interventions on gate heads can suppress position reports even when router heads supply location information. We further examine why reporting accuracy varies across concepts. Concept vectors that are localized more accurately produce stronger attention-score and output responses in gate heads, which is associated with better alignment of the induced key and value changes in their QK and OV computations. Together, these findings identify attention-head mechanisms supporting introspective detection and localization.

We introduce DoAtlas-2, a foundation for self-evolving causal biomedical discovery that organizes knowledge around causal mechanisms and advances through external evidence from human populations. DoAtlas-2 integrates 771 research resources covering more than 720,000 participants in 48 countries, from longitudinal clinical phenotypes, medical imaging, and continuous physiological signals to eight molecular layers, together with an evidence network of approximately 4.7 million literature-derived records over 93,566 concepts and 149,383 candidate causal relations. DoAtlas-2 autonomously formulates research questions from evidence gaps and unresolved mechanisms, prespecifies their causal designs, and generates validated analyses. Supporting, challenging, and unresolved results continuously revise mechanistic interpretations, the causal evidence state, and the discovery frontier, so that DoAtlas-2 self-evolves within a closed loop of hypothesis generation, empirical testing, and renewed discovery. DoAtlas-2 has systematically evaluated 2,031 research questions. In the Human Phenotype Project (HPP), it formulated 4,014 candidate pathway questions across vascular, early-glycemic, and hepatic-metabolic systems, and screening of the first 1,079 yielded statistical support for 756. Representative studies identify blood pressure as a convergence node linking adiposity, hepatic, and lipid phenotypes to vascular outcomes, and show that an adiposity-inflammation-blood-pressure pathway is largely attenuated by joint adjustment for body mass index (BMI) and smoking. The discovered vascular network constitutes a completely interpretable predictive foundation, admitting exact attribution of every prediction and closed-form mediation effects. DoAtlas-2 thereby unifies causal mechanism discovery, population-evidence testing, and interpretable prediction within one continuously evolving foundation.

Predicting cellular responses to perturbations is a central problem in cellular biology, with broad applications in systems biology and drug discovery. This task is challenging because cellular responses can be complex and cell-state dependent, intrinsic cell-to-cell variability can be confounded with perturbation effects, and destructive single-cell RNA sequencing precludes paired measurements of the same cell before and after treatment. Flow matching transports control cells to perturbed states flexibly, but acting on the full cell state can confound perturbation effects with pre-existing cell-to-cell variability. Disentangled approaches separate responsive from invariant components, but model perturbations through prescribed mechanisms, such as latent shifts or graph edits, limiting their flexibility. We address both limitations in a unified framework. A variational encoder disentangles each cell into an invariant block, capturing state unaffected by the perturbation, and a responsive block, capturing state it changes, through conditional priors and an information-theoretic invariance constraint. Conditional flow matching transports only the responsive block, conditioned on the perturbation and invariant state, yielding a flexible, data-driven model of perturbation effects without confounding pre-existing variability. Across several benchmarks, our method outperforms the strongest published method in settings involving combinatorial and unseen perturbation prediction.

We present e3j, a fast Euclid-equivariance backend for geometric deep learning applications with JAX bindings for GPU and TPU. Leveraging both optimized CUDA and Pallas kernels and algorithmic improvements, the library achieves state-of-the-art throughput and runtime on both forward and backward paths. On a machine learning interatomic potential (MLIP) use case, it outperforms established backends, measuring up to 34% speed-up over cuEquivariance on water box NPT simulation using MACE, while remaining fully open source. E3j achieves over 80% efficiency over the H100 maximum memory bandwidth on tensor product operations, and in many cases more than doubles throughput of message passing convolutions forward compared to previously available backends. In addition, with the release of dedicated Pallas TPU kernel, e3j opens the possibility of large scale equivariant deep learning workloads on TPU architectures, which has so far been difficult to achieve. Our benchmarks show that e3j also achieves over 80% of a TPUv6e memory bandwidth, up to one order of magnitude more than e3nn-jax. The library is available on GitHub, PyPI and is released under an open source Apache 2.0 license.

Spiking neural networks (SNNs) offer an energy-efficient paradigm for time-series forecasting through spike-driven computation. However, recent SNN forecasters often pursue higher accuracy through increasingly complex attention mechanisms, or specialized neuronal dynamics, weakening the lightweight motivation of SNNs. We introduce SpikeLite, a spiking forecasting framework built around two modules: a Frequency-Selective Spiking Encoder (FSSE) for frequency-sensitive temporal encoding and a Sparse Spiking Channel Attention (SSCA) module for selective cross-channel interaction. FSSE exploits the low-pass filtering behavior of LIF dynamics to reorganize each input sequence into frequency-sensitive components while collectively preserving the input at the decomposition stage. SSCA then learns a binary mask from encoded channel representations and uses it to selectively exchange information within spike-driven self-attention, retaining informative cross-channel interactions while suppressing redundant ones. When explicit channel interaction is unnecessary, SpikeLite uses the lighter FSSE-only channel-independent path. Experiments under the SeqSNN and SpikF protocols cover four standard multivariate and eight long-term forecasting benchmarks. SpikeLite achieves the best aggregate performance under both protocols, with an average $R^2$ of 0.790 and RSE of 0.440, and lowest average MSE/MAE of 0.343/0.345 in long-term forecasting. Moreover, evaluation on the ECL dataset shows that SpikeLite achieves the lowest reported energy consumption, further demonstrating its potential for energy-efficient time-series forecasting.

Deepfake detection methods have become increasingly effective yet most provide limited insight into the evidence behind their predictions. However, in forensic settings users also need to know which manipulation cues support the decision and where they appear. Existing explainability methods only partially address this need since localization-based approaches lack semantic descriptions while language-based explanation methods are only weakly grounded in visual evidence. In this work, we propose DF-CBM, a region-aware concept bottleneck model for explainable deepfake detection. DF-CBM builds a compact vocabulary of manipulation-related concepts from textual artifact annotations and links each concept to plausible facial and boundary regions. It then predicts these concepts from visual features using a concept-specific masked attention mechanism guided by parsed facial masks and the final real/fake decision is made from the predicted concept bottleneck. Our experiments show that DF-CBM outperforms concept-based baselines in concept prediction and deepfake classification while remaining competitive with state-of-the-art black-box detectors. Finally, qualitative results and intervention analyses demonstrate that DF-CBM provides spatially grounded concept evidence and enables counterfactual explanations of how individual manipulation concepts influence the final prediction. Our code is available at: https://github.com/GeorgeTsoumplekas/DF-CBM.

Meta-analysis is the synthesis of information from multiple sources to arrive at an overarching conclusion. There is a large need for meta-analysis in agricultural research to synthesize what is known and analyze overarching patterns. Extracting data from published literature is, however, labor-intensive, time-consuming, and tedious, and is impeded by a lack of standardization in research design, units of measurement, and terminology. These challenges are particularly evident in the domain of crop species mixtures, also called intercropping. With the growing capabilities of LLMs, many recent attempts have focused on building systems and tools to automate data collection, yet rigorous assessment against human-labeled ground truth is often missing. In this research, we evaluate three LLM-based approaches---direct zero-shot prompting, a staged workflow, and a multi-agent system---with six open-weight models to extract data from the intercropping literature. The results are evaluated against the manually curated ground truth and through a downstream statistical analysis. Overall, direct zero-shot prompting is the strongest and most consistent approach, achieving the highest mean similarity-adjusted F1 of 0.577, although none of the approaches is close to fully accurate. In the downstream analysis, most model--approach combinations recover the direction of the relationship between the predictor and outcome variables, but do not estimate its magnitude accurately.

Tool-enabled agents form calls from model-visible interfaces, while hosts later select their implementation. Standard dispatch omits the descriptor-handler relation. An unchanged and schema-valid call can therefore acquire a different security effect during rollout, reconnect, or delayed approval. We call this failure schema-epoch drift. We present formation-consistent dispatch (FCD), which connects implementation analysis to execution authority. Reviewed profiles produce provenance-bound over-approximations of declared in-scope effects from official source. Under a closed-target approval policy, a verifier applies each formed call to a summary and captures a successor only when its effects fit the call's security contract. Atomic admission and a final-hop fence preserve this decision to the effect. The exact source retains priority, and the captured successor becomes eligible only after source retirement. Stock releases and deployment changes reproduced the failure. Four profiles covered 32 official releases: 29 required no release-specific change and three escalated. A frozen 16-release expansion matched a separate source oracle. In a preregistered stock comparison, FCD completed all three pending calls whose effect remained private and blocked all three whose omission became public. Exact pinning and release-wide denial stopped all six calls, while release-wide approval completed all six but produced three public effects. A separate lifecycle experiment carried a formation-captured certificate across source retirement. The same safe certificate installed later governed new formations without expanding the pending call's authority.

In this work, we study portfolio optimization under the stochastic discount factor (SDF) framework by learning market state representations that capture the underlying risk structures of financial data. This is challenging due to several factors: financial markets exhibit non-stationary dynamics with shifting regimes, multimodal inputs such as price and news data often contain stochastic noise, and existing diffusion-based approaches, while effective for modeling stochastic dynamics, rely on assumptions such as isotropic Gaussian noise that fail to capture the state-dependent nature of financial uncertainty. To address these challenges, we introduce RADAR, a retrieval-augmented diffusion framework that learns market representations by conditioning on similar historical regimes. RADAR leverages retrieval to construct context-dependent noise distributions, applies conditional diffusion to denoise multimodal representations, and initializes the diffusion process using empirical statistics to reflect state-dependent uncertainty. Experiments show that RADAR achieves state-of-the-art performance on key risk-adjusted metrics while producing economically meaningful signals on asset returns and correlations.

Group-based reinforcement learning (RL) has advanced large language models (LLMs) and is increasingly extending to agentic tasks, where sparse terminal rewards make step-level credit assignment essential. Existing methods assign credit from what follows an action in sampled rollouts, but do not explicitly capture its retrospective relation to the realized outcome. Hindsight credit assignment (HCA) instead attributes credit through the ratio of hindsight to behavior-policy probabilities, but estimating the hindsight distribution requires an auxiliary model or an extra pass. To address this estimation bottleneck, we propose GraphHCA, a model-free realization of HCA that eliminates explicit hindsight-distribution estimation. For terminal-goal tasks with deterministic transitions, Bayes' rule reduces the hindsight ratio to a ratio of behavior-policy success probabilities at consecutive states. Taking logs yields a state-wise success potential, whose increment across a transition provides step-level credit. GraphHCA estimates this potential from pooled rollouts through a discounted recursion on the induced transition graph, which admits a unique fixed point on any directed graph. The resulting step-level signal is combined with the trajectory-level advantage, requiring neither a learned hindsight model nor an extra forward pass and recovering GRPO when the step-level weight is zero. Among all compared baselines, GraphHCA achieves state-of-the-art results on ALFWorld and WebShop at both LLM scales, and on Sokoban with a vision-language agent. For example, on ALFWorld it improves overall success rate by up to 24.6 points over GRPO and by up to 4.7 points over the strongest step-level baseline.

Biological machine learning was long bottlenecked by the ability to synthesize designed DNA. Variational synthesis models control chemical reactions to physically manufacture quadrillions of designed sequences in DNA. However, training these generative models is challenging: constraints on chemical synthesis can force many parameters into a discrete space, limiting the ability to pre-train and fine-tune. In this article we train ``free'' variational synthesis models using stochastic gradient descent in continuous space, and then discretize with post-training quantization to impose hardware and wetware constraints. This enables variational synthesis models to satisfy stringent reward criteria, while still synthesizing diverse designs, achieving a strictly dominating quality-diversity Pareto frontier. We demonstrate by training variational synthesis models of enzymes, peptides, antibody CDRH3s, and regulatory DNA elements. In silico performance is maintained in vitro.

Group-based reinforcement learning such as GRPO trains LLM agents by comparing rollouts sampled for each task, without a learned critic. In long-horizon settings, these rollouts revisit shared anchor states, offering cross-rollout evidence for step-level credit. Ideally, step-level credit should incorporate evidence beyond the realized suffixes observed at an anchor while aggregating alternative continuations according to their empirical frequencies. Visit-local averaging pools realized suffix returns at shared anchors and respects observed frequencies, but does not recursively propagate evidence across rollouts, whereas shortest-path estimators have global reach but allow a rarely observed route to dominate an anchor's value. We introduce Cross-Rollout Bellman Closure (CRBC), which merges each rollout group into a finite empirical process with absorbing success and failure boundaries and evaluates its behavior-policy Bellman fixed point with one linear solve. This fixed point uses the same empirical action and transition frequencies to propagate evidence through shared anchors and aggregate alternative continuations. Backing up the resulting state values through observed transitions yields action values, whose gain over the corresponding state value provides step-level credit. A corresponding finite-depth family recovers visit-local return averaging at zero depth and converges to the exact closure as depth increases. The normalized closure credit is combined with the trajectory-level group advantage for policy optimization, without additional environment rollouts. Across ALFWorld, WebShop, and Sokoban benchmarks with multiple model scales, CRBC consistently improves final performance and learning efficiency. For example, CRBC outperforms the strongest evaluated baseline by 5.59 percentage points on ALFWorld with Qwen2.5-1.5B-Instruct.

We study post-hoc refinement of frozen node classifiers: given only the graph $G$ and class distributions $Q$ predicted by a frozen model, can we improve accuracy without access to node features, model parameters, or gradients? APPNP answers this by propagating logits with a restart towards the initial predictions, minimizing the anchored Dirichlet energy. Instead, we consider the Potts energy, and decompose it into a Dirichlet term, which penalizes disagreement between neighbouring nodes, and a Gini term, which penalizes indecision within each node. This decomposition motivates Propagate, Then Sharpen (PtS), which alternates between propagation of class probabilities and node-wise, mass-preserving sharpening, with only one additional hyperparameter selected using labelled validation nodes. Across nine homophilic graphs, with a frozen MLP backbone, PtS improves mean test accuracy over independently tuned APPNP by $1.71$ percentage points on clean inputs and $3.90$ under severe Gaussian feature corruption. Gains over APPNP become smaller, but remain positive with frozen GCN and GraphSAGE backbones. Sharpening also removes most of the accuracy loss of deep propagation: on clean inputs without restart, accuracy falls by $2.2$ points between $2$ and $100$ propagation steps under PtS, compared with $33.8$ for APPNP.

Production large language models retrieve and cite web pages alongside generated answers, yet the page-level features that predict citation frequency remain poorly characterised. We present an observational study of approximately 2 million LLM citations from four commercial engines (ChatGPT, Claude, Google AI, Gemini) over six months, joined to 10,000 crawled pages from nineteen B2B SaaS workspaces. Sixty-plus features are tested using a nine-method consensus framework combining mixed-effects regression with domain fixed effects, FDR correction, stability-selection Lasso, double machine learning, generalised additive models, and temporal hold-out replication. Four findings survive all checks. First, prompt-content alignment (Jaccard overlap between page tokens and the full workspace prompt corpus, including non-citing prompts) is the dominant page-level predictor (beta = +0.37, 95% CI [+0.33, +0.41], q ~ 10^-73). Second, the standard AEO checklist (FAQ blocks, structured data, Core Web Vitals) shows positive effects in pooled data that reverse or collapse to zero once domain fixed effects are applied: Simpson's paradox with practical consequences for the AEO literature. Third, domain-level AI authority exceeds the strongest non-alignment page-level feature by a factor of six in mean absolute SHAP value. We release the analytic pipeline as a methodological contribution.

Many sensing tasks obtain training labels only by deploying instruments in the field. With a limited number of sensor kits, a collection deadline, and measurement downtime at every move, the collector must repeatedly decide whether to stay at the current site or relocate. We study this decision in non-intrusive load monitoring (NILM), which estimates the power drawn by individual appliances from a home's main meter and is trained on data from homes temporarily fitted with appliance-level sub-meters. In NILM, appliance usage varies with the appliance, season and climate, and the value of new data depends on how diverse the combinations of target operation and background load are. To address this, we propose a constraint-based relocation framework and instantiate it for NILM as ReCo (Relocation by Coverage gain). ReCo counts new operating regimes in a joint target-background feature space, forecasts each home's future gain from the data collected so far, and each night weighs the gain of staying against the gain of moving elsewhere after the downtime. In replayed deployments on the Plegma dataset under two kit counts and two downtime costs, ReCo outperforms fixed-dwell and count-based schedules and a threshold rule using the same metric in every setting. Its advantage is not explained by collecting more days alone and reflects allocating the days to more valuable homes and periods.

The reasoning trajectory of a Large Language Model (LLM) is often treated as a verbalized description of its internal reasoning. However, such trajectories can be unfaithful: a model may rely on shortcuts to reach an answer and then post-rationalize the decision with a seemingly coherent chain of thought. Detecting this shortcut reasoning is challenging because existing monitors and verifiers mainly inspect textual traces or final outcomes, rather than how the model's belief in its answer develops during generation. We introduce ConfLens, a framework that tracks the evolution of confidence in the final answer throughout reasoning. Across three shortcut reasoning settings, we observe a common pattern of premature confidence, where shortcut samples become highly confident in the final answer at early reasoning stages. Existing confidence estimation methods, however, show limited generalizability, reliability, or efficiency for detecting this behavior. We therefore propose the Distributional Answer Commitment Score (DACS), a distributional confidence estimator that measures the entropy of the model's probability distribution over answer commitment at each reasoning step. DACS captures how concentrated the model's answer belief is without requiring ground-truth answers or task-specific verifiers. We further convert ConfLens detection results into interpretable signals for reward models to reduce their preference for shortcut reasoning. Experiments on mathematical and code reasoning tasks show that ConfLens with DACS improves shortcut reasoning detection by over 4.3% F1 compared with strong baselines and reduces the mismatch between faithfulness and correctness in reward model preferences.

Graph visualization methods and agglomerative clustering have been frequently considered in data analysis and pattern recognition. Because these approaches are interrelated and complementary, it is of particular interest to investigate their associations. In this work, we study the possible relationship between the Fruchterman-Reingold graph visualization method and four types of agglomerative clustering adopting single- and complete-linkage, average, and Ward's linkage criteria. Three types of datasets have been considered in 2 and 10 dimensions, as well as the PCA projection of the latter to two dimensions. The results obtained suggest that the relationship between the methods considered did not vary much for the three types of data mentioned above. At the same time, the agglomerative methods tended to yield results that are mostly similar to each other, while presenting moderate similarity with the original data. The Fruchterman-Reingold visualization resulted similar to the original data, but exhibited relatively smaller similarity to the agglomerative methods.

We consider the problem of trajectory valuation in reinforcement learning: how to identify and mitigate detrimental trajectories during online training. Unlike classification, where data valuation relies on fixed training and validation sets, reinforcement learning involves dynamically generated trajectories without explicit validation signals, making conventional influence-based methods inapplicable. We propose Dynamic Trajectory Valuation (DTV), a simple and efficient framework that estimates trajectory utility at the mini-batch level and filters detrimental trajectories based solely on gradient information. By operating at the optimization level, DTV integrates seamlessly with existing reinforcement learning pipelines with minimal overhead. Extensive experiments across diverse settings, including PPO, GRPO, and DPO, demonstrate that DTV consistently improves performance, enhances data efficiency, and stabilizes optimization.

Evaluations of Large Language Models (LLMs) morality typically consider decisions in isolation, thus overlooking whether an individual's unrelated prior conduct influences the model's subsequent choices. This leaves open the question of whether, and to what extent, moral history shapes LLM decisional behaviors. Prior work on human moral decision-making shows that past behavior can influence subsequent moral choices. Building on this observation, we investigate whether analogous effects emerge in LLMs in two complementary ways: at the behavioral level, through the model's observable responses, and at the representation level, through its latent internal representations. We introduce MoralLedger, a framework for studying how an actor's moral history shapes actions for LLMs' behaviors under a fixed decision context. At the behavioral level, we find that prior moral histories systematically alter subsequent choices as a function of their valence and intensity. At the internal representation level, these histories induce a linearly recoverable direction in the residual stream that generalizes to held-out examples. Intervening along this direction on neutral-history prompts produces two-sided intensity-dependent changes in subsequent choices, with effects that are stronger than those induced by prompting alone or by favorable-nonmoral direction. To our knowledge, this is the first demonstration that a latent representation of an actor's prior moral conduct can provide signed inference-time control over a moral decision. Our MoralLedger extends moral evaluation beyond static dilemmas, establishing moral history as both a source of behavioral sensitivity and a causal target for auditing and controlling moral behavior in LLMs.

Most neural constructive solvers for the vehicle routing problem (VRP) use route-by-route construction, extending one route until completion before starting the next. This commits route membership early and hinders global coordination across routes. We propose multi-component construction, which maintains many route components simultaneously and merges them in an arbitrary order. This removes the depot-return cue that route-by-route construction obtains from the remaining capacity; to compensate, we introduce an interpretation in which every component is treated as an implicitly depot-closed route. Under this depot-closed interpretation, every intermediate state of standard CVRP construction is a complete feasible solution, and the exact cost reduction of a merge is the Clarke-Wright saving. The neural policy combines this CW-saving signal with the evolving component state to learn what to connect and when to connect. A policy trained only on CVRP100 outperforms the reported results of representative neural solvers on CVRP100-500 with greedy inference and, reused for ruin-and-reconstruct, performs strongly at all evaluated sizes up to CVRP1000. In a zero-shot Constraint Tightness evaluation with capacities from $C=10$ to $500$, it outperforms the reported neural solvers at every capacity. Controlled analyses show that robustness persists without CW grounding and point to learned route-closing behavior as a plausible contributor to the tight-regime degradation of learned route-by-route solvers.

Reusable prefix key-value (KV) caches can outgrow GPU memory in large language model (LLM) serving. A memory-semantic flash hierarchy offers SSD-backed capacity with a limited fast tier, but a logical KV hit is not necessarily ready for GPU retrieval. Demand staging exposes SSD latency, whereas immediate staging can reserve fast-tier capacity long before retrieval begins. We present TempoKV, a timing-aware resource-commitment layer that separates early knowledge of reuse from the acquisition of staging resources. It records reusable-KV hits as metadata-only claims and requests commitment when the runtime-estimated time until retrieval falls to the storage-estimated time needed to make KV resident and protected against eviction. These estimates adapt to runtime progress and staging state, while commitment remains subject to available protected capacity. We implement TempoKV in vLLM and LMCache on an SSD-backed CXL memory device without changing request scheduling. Across two models and three prefix cache ratios, TempoKV reduces protected fast-tier byte-time per request by 63-91% versus immediate staging while retaining much of the serving benefit of advance staging. In a fast-tier capacity sweep, output throughput and p95 time to first token (TTFT) remain nearly unchanged as capacity decreases from 100 to 25 GiB. Compared with unmodified LMCache's Device-DAX L1 configuration, TempoKV reduces p95 TTFT by up to 48.0% and increases output throughput by up to 27.8%.

Inferring dynamics from snapshots of evolving distributions is fundamentally underdetermined: the Fokker-Planck equation constrains the drift $F$ only through its score-weighted divergence $\nabla\cdot F+F\cdot\nabla\logρ$, leaving a $ρ$-solenoidal gauge invisible to any single-time constraint. Time-indexed transport formulations cannot resolve this ambiguity: every admissible marginal path admits a curl-free explanation, minimum-action reconstruction selects it, and marginal fit alone cannot distinguish dynamically inequivalent explanations. Requiring one autonomous field to explain several marginals instead makes part of the hidden circulation visible as $\nabla\logρ$ changes across marginals. Separating instantaneous Fokker-Planck source constraints from the snapshot experiment, we show that the source constraints identify the field modulo the kernel of a stacked score-weighted divergence operator. For generic Gaussian shape variation, source constraints at $K\ge m$ time points in intrinsic dimension $m$ eliminate every polynomial gauge direction, whereas finitely many density snapshots alone admit aliasing; we give the obstruction explicitly. At a Gaussian anchor, for Sobolev smoothness $s$ and $n$ samples per time point, we derive a conditional lower rate $(nK)^{-2s/(2s+m+1)}$ for the tangent snapshot experiment, with a matching upper rate in a degreewise benchmark. Strong-form fitting is non-orthogonal to score error and cannot be repaired by spectral filtering. Instead, we estimate using smooth test functions while retaining the known diffusion term, and derive a finite-sample bound that separates sampling error from fixed-grid quadrature bias. Planted-circulation experiments confirm the predicted gauge contraction and expose a design tension between cross-slice information and covariance-aware whitening.

Effective multi-turn agents require interaction strategies that coordinate information gathering, actions, and feedback over long horizons. GRPO is a reinforcement learning algorithm used to train these agents, but sparse trajectory-level rewards limit early exploration in small models. Recent methods augment RL with on-policy distillation (OPD) from a stronger teacher. However, a fixed mixture assumes that teacher guidance and reward optimization should retain a constant relative role throughout training and across interaction turns. This assumption can fail at two scales. Globally, as training progresses, maintaining strong distillation pressure can constrain the model from moving beyond the teacher's capabilities. Locally, teacher--student disagreement identifies where the student departs from the teacher, but cannot tell whether that departure is exploration supported by better outcomes or low-quality policy drift. Our methodological insight is that teacher guidance and reward optimization should be dynamically rebalanced over training and jointly allocated across turns. We instantiate this insight in \tide. Globally, \tide uses the measured disagreement trend as a practical schedule signal, advancing an OPD-to-RL handoff when discrepancy reduction becomes slow but remains positive and progressively increasing the relative weight of RL. Locally, \tide jointly modulates teacher-guided and reward-driven updates: relative action value and disagreement prioritize the OPD signal, whereas relative action value supplies the RL advantage and normalized disagreement reweights it across turns. Coupled with the global handoff, \tide allocates stronger teacher guidance early and gives reward-driven updates greater relative weight later in training. Experiments across multiple benchmarks, student scales, and controlled ablations support the effectiveness of TIDE's adaptive OPD--RL coordination.

Long contexts are central to modern transformer systems, but most expressivity results choose a different network for each fixed sequence length. We study whether one masked transformer can approximate causal token-to-token maps uniformly over sequences of arbitrary length sampling a fixed normalized horizon. To relate sampling resolutions, we model tokens by $α$-Hölder sequences or, more generally, a common modulus of continuity. Our notion of continuity across resolutions characterizes the causal families admitting uniform approximation on these compact input classes by a single transformer with length-independent parameters. The result extends to the infinite-length mean-field limit, where tokens form continuous curves and masked attention becomes a causal time integral. For bounded regression with target maps satisfying a $β$-smooth stability condition defined using regular test functions, quantitative approximation yields a generalization bound: exact empirical risk minimization over suitably sized bounded-weight transformers gives root mean-square prediction error $O((\log\log N/\log N)^{β/(d+2)})$ from $N$ iid labeled sequences. The bound holds at fixed confidence on the same sampling distribution, with $d$ the token dimension and no maximum-length factor. Finally, experiments on physical time series support the Hölder-regular token model at observed scales, with dataset-dependent fitted exponents, whereas text input embeddings provide a contrasting case. Native and dense sampling, shuffled controls, and refinement checks delimit this empirical regularity regime.

Contemporary deep speech enhancement (SE) models are often trained with specific auxiliary terms in the loss function as a way to improve their performance in terms of perceptual metrics. Nevertheless, a higher score on a perceptual metric does not necessarily correlate with an improved listening experience. Through objective and subjective experiments, we assess the performance of SE models trained with two different types of auxiliary PESQ loss terms. The numerical evaluation on a suite of standard metrics suggests that, while models optimized for PESQ naturally obtain higher PESQ scores in the test set, for most other metrics the scores do not significantly change. In some cases, the PESQ loss even results in worse PESQ scores on mismatched data. A formal listening experiment reveals that the models without a PESQ loss were generally preferred over models that include it, across all settings. Finally, we analyze the relative importance of PESQ in the composite metrics CSIG, CBAK and COVL, and find that PESQ dominates all of them. Our study highlights the perils of over-reliance on PESQ and stresses the importance of a complete evaluation procedure for SE.

Video world models must preserve the visual state of the world over time, but existing evaluation protocols often rely on generated histories, video reference, or selected revisit viewpoints that can confound the assessment of a model's true memory capability. To address this, we introduce OPIS, an input-grounded benchmark that strictly anchors the assessment to a fixed set of object instances from the initial observation for evaluating multi-object memory in video world models. The OPIS dataset comprises 500 cases across real-world, embodied-robotic, and game-world domains, providing dense object-level annotations for 12,672 rigid, articulated, and deformable instances. Our object-centric evaluator combines association and explicit visibility reasoning to hierarchically measure Object (O) Presence (P), Identity (I), and Structure (S), utilizing static or dynamic evaluation tracks based on object kinematics. Across eight image-to-video or camera-conditioned world models, our proposed OPIS scores range from 48.65 to 56.01. As the reference inventory grows from less than 20 to more than 40 objects, the Presence, Identity, and Structure scores show an overall decline, with the average Identity score falling from 40.22 to 23.11. The results demonstrate that preserving the particular object instances in the input is considerably harder than generating plausible visual elements.

A robot should be able to learn through experiments how unfamiliar objects behave and interact, then plan with that knowledge. It need not start from scratch: physics engines supply knowledge of motion and contact, but can omit entire mechanisms, such as glue curing, water heating, or wind. We present EMPIRIC, an agent that learns a residual world model: a physics engine extended with code for the missing mechanisms. The learned programs can introduce new forces, constraints, and hidden state, and Bayesian inference estimates their parameters and states from noisy observations. The resulting model lets the agent predict the outcomes of actions, choose informative experiments, and revise its hypotheses when predictions fail. Across five simulated domains, EMPIRIC learns interpretable, reusable models, and solves more tasks with fewer environment interactions than all three baselines. On a physical robot, it learns wind forces and domino masses to solve a manipulation task. Website and code: https://yichao-liang.github.io/empiric

Preference-based alignment methods such as Direct Preference Optimization (DPO) use pairwise preferences labeled by human annotators to fine-tune language models. However, annotators carry systematic biases toward some attributes: a name that signals a gender or an ethnicity, a persona, a language variety, a formatting convention, or length. If not properly addressed, these systematic biases can be absorbed and amplified during alignment. Existing methods address length bias or annotator disagreement, but fail to eliminate biases toward arbitrary attributes. To address this limitation, we propose Bias-Adjusted DPO (BA-DPO), a generalization of DPO that adds one bias parameter per annotator toward responses carrying a declared attribute. We prove that the objective is convex in the bias parameters and that the votes identify each annotator's bias up to a shared constant. The remaining constant is what fixes the aligned model's attribute rate: by default the rate of the reference model, or a target rate, which we use to bring a biased policy to statistical parity. On a corpus with planted biases, DPO drives the attribute from a balanced start to probability 0.96 and BA-DPO removes 81 to 95\% of that shift; on MultiPref with real annotators it removes about half of DPO's lengthening. Both hold at 0.5B with full fine-tuning and at 8B with LoRA, at no higher KL than DPO and no loss in judged quality.

Martian weather forecasting is important for future exploration, but atmospheric behaviour on Mars combines spatial, temporal, vertical, and dust-driven processes in ways that challenge current modelling and forecasting approaches. This paper introduces MaGMA (Martian Graph-based Multi-horizon Atmospheric Forecasting), a graph-based data engineering framework that transforms OpenMARS reanalysis fields into structured learning objects for Martian atmospheric forecasting. Local atmospheric patches are represented as graph nodes and linked through spatial neighbourhoods, temporal continuity, longer temporal dependencies, and dynamically similar atmospheric states. The model integrates recent atmospheric history, engineered physical descriptors, and vertical atmospheric information to support forecasting across multiple horizons. We evaluate MaGMA across five unseen Martian years, including regular years and a global dust storm year. In regular years, the model achieves overall R^2 values of approximately 0.73-0.85. For dust-column forecasting, it outperforms classical and deep temporal baselines in most year-horizon comparisons. During the global dust storm year, dust-column prediction remains strong at shorter horizons, with R^2 above 0.8 for the first two horizons, while broader multivariate performance declines. The results show that graph-based data engineering can create reusable and diagnostically useful representations for planetary atmospheric forecasting, while highlighting the need for better learning under rare extreme regimes and improved use of vertical atmospheric structure.

Federated Bayesian Optimization (FBO) enables distributed agents to collaboratively optimize expensive black-box objectives without sharing raw local observations. However, effective knowledge transfer remains challenging under communication constraints and task heterogeneity. We propose GUIDE-FBO, in which agents exchange compact distributions over the locations of their respective optima inferred from local Gaussian process (GP) posteriors, rather than raw observations, query points, or surrogate parameters. The server merges and reweights these distributional components before returning a subset to each agent. Each agent then constructs a Federated Interventional GP (FI-GP), which preserves the local posterior mean and spatially rescales its covariance for local decision making. For the upper confidence bound (UCB) instantiation, GUIDE-UCB, we prove that any bounded FI-GP uncertainty intervention preserves the leading-order cumulative regret rate of standard GP-UCB. When the transferred distributions place greater support near an optimum than in a suboptimal region, selecting the latter requires greater local posterior uncertainty. Experiments on 12 synthetic benchmarks and three real-world optimization tasks show that GUIDE-FBO remains effective across settings ranging from homogeneous to severely heterogeneous. Ablation results highlight the importance of spatially localized uncertainty intervention, while the communication analysis shows that GUIDE-FBO exchanges only compact distributional messages.

Persona prompting is widely used to construct user simulations with large language models (LLMs), yet it relies on a largely untested assumption: specifying one user attribute should change that attribute alone. We test this assumption and identify a systematic failure of selective control: across all eight black-box LLMs we audit, changing a target attribute also shifts responses on unspecified, non-target attributes. For example, describing a user as more risk-seeking shifts color choices, even though the prompt never mentions color; we term this cross-attribute influence. Semantic, contextual, and internal analyses collectively suggest that models treat a persona prompt as evidence about the user and extend the inferred profile to unspecified preferences, a process we call trait-conditioned completion. We next ask whether explicitly specifying non-target attributes restores selective control. When a non-target attribute is assigned a clear direction, models generally follow the declaration and suppress the target attribute's influence. However, when the same attribute is declared neutral, the target continues to affect choices across all five open-weight checkpoints, even when the model correctly reports the declared state. This disparity, the neutrality gap, demonstrates that successful persona following does not imply selective persona control, which additionally requires keeping non-target attributes stable. We operationalize this distinction with a three-state diagnostic that leaves the non-target attribute unspecified or declares it directional or neutral; because directional tests can be passed by simply following the stated persona, the neutral state reveals failures they miss. In a post hoc analysis of independent items, neutral declarations leave 51-81% of items target-sensitive, against at most 1 of 320 item-pole comparisons under directional ones.

The integration of artificial intelligence into computer-aided design frameworks has sparked a shift in the design of analog integrated circuits (ICs), transitioning the field from using manual and algorithmic-based solutions to adopting automated and intelligent paradigms. In this scenario, the GDSII file represents the industry-standard database containing the ultimate and most accurate source of information of the analog circuit, encapsulating the complex physical geometries and parasitic realities that define tape out performance. This paper proposes THEIA, a novel dataset containing thousands of layout images paired with question-answer conversations, along with a benchmark that employs a fine-tuned vision-language model (VLM) to analyze GDSII files of analog circuits, enabling designers to interact with and query physical layouts as intuitive, meaningful entities. Experimental results using thousands of analog designs across five realistic tasks demonstrate that the proposed fine-tuned VLM outperforms state-of-the-art general-purpose VLMs by a significant margin (up to 73%), highlighting a fundamental gap between general-purpose multimodal reasoning and domain-specific layout understanding.

Off-policy evaluation, which estimates evaluation policy performance from logged data, is key for recommender ranking policies. However, logged clicks cannot distinguish unexamined items from examined non-clicks, causing bias in existing estimators when the assumed examination structures fail. We propose two estimators based on the decomposition of clicks into examination and relevance. First, the latent-examination independent inverse propensity score (LE-IIPS) estimator corrects the IIPS bias using policy examination probability ratios. Second, the examination-decomposed doubly robust (ED-DR) estimator extends LE-IIPS to a doubly robust framework. ED-DR is unbiased if the examination probabilities are correct regardless of relevance accuracy, or under ranking-independent examination, even if both model estimates are inaccurate. Experiments show that ED-DR achieves a lower MSE than existing methods with large sample sizes, especially when the examination depends on ranking. We also highlight its limitations under small samples or cascade user behavior conditions.

In complex embodied visual reasoning scenarios, an agent often has only a limited field of view, and the evidence needed to answer a question may be distributed across time, viewpoint, and interacting objects. A model may therefore give a plausible answer without ever observing the relevant object, time, or view that supports it. Current visual reasoning benchmarks largely evaluate passive observations and final answers, overlooking settings that require active reasoning and evidence acquisition. We introduce JRDB-AVR, a benchmark derived from existing real-world JRDB robotics data through a structured question-generation engine that turns this gap into an explicit evaluation: an embodied agentic system receives a visual reasoning question, requests bounded observations by timestamp and viewing angle, and is evaluated on both the final answer and the grounded visual evidence supporting it. The benchmark contains diverse questions over multiple real-world environments involving temporal search, viewpoint selection, and human-oriented compositional reasoning. We also introduce JRDB-AVR-Agent, a reference active reasoning agentic method that maintains an explicit observation-grounded graph-based world model and answers through solving. Experiments reveal a substantial gap between answer accuracy and evidence accuracy in current baselines, showing that current VLMs can produce unsupported correct answers and that active evidence-aware evaluation is necessary for embodied visual reasoning. Code and benchmark are available at https://github.com/ControlNet/JRDB-AVR.

AutoResearch improves systems through iterative experimentation: agents propose candidate modifications, evaluate them, and use the results to guide subsequent exploration. Applying this paradigm to industrial search presents two challenges. (1) Common AutoResearch approaches follow a keep-if-better rule, retaining the highest-scoring candidate for subsequent experiments. Under non-stationary traffic, transient gains may be mistaken for persistent improvements, impairing reliable accumulation of search knowledge. (2) Candidate modifications can be evaluated at multiple fidelity levels, from low-cost proxies to online validation, differing in cost, objective alignment, and statistical reliability. Existing methods rely on individual signals or task-specific procedures, lacking a unified basis for using evidence across levels to guide search. We introduce Progressive Evidence-Based AutoResearch (PEAR) with two complementary components. Evidence-driven AutoResearch maintains an independent, hypothesis-guided research state for each strategy task within a predefined objective and intervention scope. Each state evolves through a Plan-Execute-Evaluate-Update transition that links experimentation to context-aware evidence interpretation and hypothesis revision. Confidence-Gated Verifier Ladder organizes evaluation into four levels of increasing fidelity: Offline Replay, Shadow-Traffic Evaluation, Rapid Online Evaluation, and Decision-Grade Online Evaluation. A unified confidence-based gate promotes candidates only when evidence supports a statistically significant positive effect, enabling broad low-cost exploration while reserving costly online experiments for promoted candidates. In a real-world industrial search system, strategies optimized with PEAR significantly increased Main Order/DAU by 2.7336% and 3.2957% relative to their respective baselines in two A/B experiments.

Hyperparameter transfer across model width can substantially reduce the cost of tuning large neural networks, but its behavior when the training horizon grows with width is not fully understood. Building on the framework of fast hyperparameter transfer (Ghosh et al., 2026), which formalizes when transfer is effective, we investigate conditions that ensure fast transfer in the growing-horizon regime. Specifically, we study learning-rate transfer in a shallow linear network with a single trainable hidden matrix, trained by full-batch gradient descent. Under additional spectral assumptions, our main results are threefold. (i) We prove fast learning-rate transfer as $n,T\to\infty$ whenever $T=o(\sqrt{n})$. (ii) We characterize the transfer rates through the finite-width perturbation scale, the first-order sensitivities of the loss and its learning-rate derivative to finite-width perturbations, and the local loss curvature. (iii) We derive limiting distributions for the optimal learning rate and optimized loss, governed by fluctuations associated with the extreme eigenvalues of the data Gram matrix. These results clarify how spectral structure and local loss sensitivities govern learning-rate transfer at growing horizons.

Large language models (LLMs) have made misinformation inexpensive to produce but not to verify, creating a growing asymmetry in the information ecosystem. Under tight time, labor, and budget constraints, media organizations, platforms, and fact-checkers rely on screening to prioritize which content to verify. We introduce VEX-Bench, a unified benchmark for evaluating the verification complexity of LLM-generated misinformation, as perceived during screening, across models and generation methods. Verification complexity is assessed along multiple dimensions derived from journalistic and fact-checking practices, capturing checkability, harm potential, source credibility signals, imposter legitimacy, and expected verification effort. We define the VEX score as an integrated measure combining elicitation yield and verification complexity to quantify how generated content consumes limited verification capacity. We construct a benchmark spanning two misinformation categories, 6 high-stakes domains, and 60 real-world topics, and evaluate 7 frontier LLMs and 7 generation methods, yielding 5{,}880 articles. We employ an LLM-as-judge for scalable evaluation and validate it using content-analysis methodology, including ordinal Krippendorff $α$ for inter-annotator reliability, complemented by fact-checking agents for verification. Our findings show that no single method dominates all dimensions, underscoring the need for multi-dimensional evaluation. LLMs can generate high-VEX misinformation at 3$\times$ to 169$\times$ lower cost than agent-based verification. Such content is often prioritized during screening, consuming scarce verification resources and introducing a systematic risk of misallocation in resource-constrained verification systems. The code is publicly available in our \href{https://github.com/HanxunH/VEX-Bench}{GitHub repository}.

We present WebPageBench, an open framework for evaluating web agents in which every task is verified from the interface's own event log. Six instrumented mock sites with brand identifiers removed (a marketplace, a bookstore, a grocery service, rail ticketing, hotel search and a document cabinet) emit typed events with parameters as a user or an agent acts. A task declares the events it requires, and success is decided by matching them, with no judge model and no scraping of rendered pages. The same instrumentation supports controlled UI variation: one configuration switch re-renders a task through a different implementation of a single control while the prompt and the success conditions stay completely identical, so sensitivity to interface form can be measured under a fixed task specification. The WebPageBench release consists of three components: 152 tasks, divided into 65 canonical scenarios and 87 control variants across light/dark UI-modes; a common runner evaluated with six browser/DOM harness configurations and five screenshot-only GUI-agent families; and a public leaderboard of 24 model-harness pairs. On the public 152-task leaderboard the gap between what agents declare finished and what the log confirms reaches 41 points (one configuration declares every task finished and satisfies the conditions on 59%).

Recent gains in language model capability have come more from data than from architecture. Frontier labs and data companies produce verifiable agentic tasks, which supervised finetuning and reinforcement learning then turn into capability.This production line still rests on human labour and on human-in-the-loop collaboration. Automating task creation would let data production scale with compute rather than with expert headcount, would extend to more domains, and would enable a key step in recursive self-improvement (RSI). Current evaluations of an agent's ability to write such tasks measure how a model performs after training on what the agent produced. That does not match common practice in the data industry, where data is delivered sample by sample and each sample is accepted against a set of criteria rather than put straight into training. No existing evaluation asks whether an individual task meets the acceptance criteria of a data pipeline. We therefore introduce AutoDataBench. Given an original benchmark task and a record of the target model attempting it, an agent must write a new task for the same suite that meets practical acceptance standards on validity, novelty, difficulty and behavioural coverage. Across three benchmarks of executable agent tasks, no agent we evaluate scores above 20 out of 100 at the default time budget of 45 minutes. Giving the strongest agent four times as long improves its score substantially, while the cost of one usable task stays almost unchanged. Current agents can write training tasks of the required quality, but not efficiently. AutoDataBench provides a direct measure of an agent's capacity for autonomous data synthesis: one artifact at a time, judged against the criteria a production pipeline would apply, and without a training run. Code and data are available at https://github.com/StarDewXXX/AutoDataBench.

Spatiotemporal prediction aims to learn discriminative representations from correlated temporal signals over spatial structures for accurate future inference. A central challenge is \emph{spatial indistinguishability}: different nodes may share similar historical patterns yet evolve toward divergent futures, severely degrading forecasting performance in real-world sensor networks. Existing embedding-based and graph neural network (GNN)-based approaches can partially detect such ambiguous nodes but rely on historical similarity, struggling to capture \emph{future behavioral divergence}. We propose \textbf{STOT} (\textbf{S}patio\textbf{T}emporal \textbf{O}ptimal \textbf{T}ransport), a self-supervised framework that resolves spatiotemporal ambiguity via structured masking guided by optimal transport. Our key idea treats indistinguishability as a \emph{disambiguation} problem: future states are inferred by exploiting concurrent spatial correlations and their time-varying similarity. We design a similarity-aware metric for dynamic inter-node relationships and an optimal transport-based masking strategy to emphasize ambiguous positions during pre-training. A batch consistency constraint preserves semantic coherence, while a random-walk masking mechanism promotes structured context exploration. Experiments on six real-world datasets show that STOT performs competitively with state-of-the-art baselines on the evaluated benchmarks and improved interpretability through transport-plan visualizations.

Evolved Plastic Artificial Neural Networks (EPANNs) consist of two principal processes, the first, evolution, and the second, development and in-life learning. In the context of the origins of social = learning, very few studies have been carried out using ALIFE models based on EPANN requirements. Studies in this field have usually involved an imitative teacher/pupil relationship. This, however, ignores the possibility that the observed behaviour is a consequence of social information cues rather than direct imitation or teaching. Starting with the first of the EPANN processes (evolution), a series of experiments was undertaken using artificial neural network (ANN) based agents in a variety of foraging environments to examine under what minimal environmental conditions the use of social information might have evolved, as measured by the number of generations taken to meet a specified fitness criterion. NEAT (Neuroevolution of Augmenting Topologies) was the ANN used as its evolutionary algorithm would evolve a network's topology as well its weights. Unintentionally, in the experiment there was a simple network topology based on the location of the nearest food item which enabled agents to swiftly meet the fitness criterion. With this topology, additional information, social or otherwise, was not required and could have proved to be a hindrance. However, this does indicate that for the use of social information to have evolved, it would require a greater degree of complexity in the environment to do so.

Existing automatic metrics for evaluating terminological use in machine translation (MT) penalise any divergence from a fixed reference, conflating translation errors with the valid terminological variation that human translators routinely produce. We introduce TermJudge, a document-level terminology metric that assigns an interpretable verdict to every term occurrence: glossary-conforming occurrences are settled deterministically, while divergences are assessed under a two-step LLM-as-judge procedure using the full document context: the first detects and labels terminology errors; the second sorts valid document-level variations from inconsistencies. Validated against expert error annotations and document-level human MQM scores, TermJudge ranks first in both system- and segment-level meta-evaluation, ahead of glossary-conformity and quality-estimation baselines. When applied to eight systems translating academic documents, under two prompting conditions, we observe that glossary injection improves terminology translation in all paired comparisons, by removing genuine errors rather than valid variation. TermJudge is released as open-source code.

The rising cost and demand for energy, together with environmental sustainability goals, create major challenges for energy management. Energy Consumption Forecasting (ECF) supports planning by predicting future consumption, but Machine Learning (ML) models for ECF often depend on expert-driven Feature Engineering (FE). This thesis addresses that dependence through three contributions. First, it establishes and evaluates a comprehensive FE pipeline for ECF and investigates domain-specific features. Second, it introduces AutoEnergy, a domain-tailored automated FE algorithm that generates interpretable features from timestamps and lagged consumption and integrates with AutoML for end-to-end ECF modelling. Across eighteen real-world energy datasets spanning residential, commercial, industrial, renewable, and grid domains, AutoEnergy reduces forecasting error by 19.52%-84.72% relative to baseline AutoML and established automated FE methods, while running 1.31-4.41 times faster, with gains varying by dataset. Third, AutoEnergy is integrated with Decision-Focused Learning (DFL) for a Battery Energy Storage System problem, jointly forecasting electricity prices and demand while optimising charging and discharging decisions. On a real-world UK property dataset, this approach reduces operating costs by 22.9%-56.5% compared with the same DFL models without automated FE. Overall, the results show that domain-specific automated FE can reduce reliance on manual feature design, improve forecasting accuracy, and translate predictive gains into measurable operational benefits in energy management.

Price stability remains a pillar in monetary policy practices and carries a special importance within monetary unions. Mainstream economics tried to leverage price stability using price indices and several metrics to shed light on specific dynamics and optimal macroeconomic levels. The wide availability of data led researchers to consider the study of systems using Random Matrix Theory, based on inner correlation patterns. This aims to enhance the multivariate analysis by removing noisy patterns from the signal and improve data quality for further inferences. This work considers the collection of monthly inflation indices in the Eurozone as a \textit{system} of prices to analyze its eigenvalues' statistical and asymptotic properties and uncover inner country-level insights. Results confirm the system cannot assumed to be randomly generated, and the data exhibit noise-dominated patterns, due to small and persistent variations at the country-level. The latter make the inter-country correlations more dynamic and the separation of the signal from the noise quiet difficult. Findings identified two countries as distorting inflation dynamics besides three other distinct, regional-based groups of countries. Variability sources might stem from economic episodes fueling inflation spikes in some countries, as well as methodological aspects used to ensure data quality and representativeness in the European Union. Despite being complex, the system demonstrates a certain stability, in terms of self-organization; while large monthly fluctuations cannot be considered as rare events, but part of the data-generating process.

Keyword Spotting (KWS) is becoming increasingly important as voice-controlled devices grow more widespread. While voice interaction with smartphones and smart TVs is already common, deploying KWS on heavily resource-constrained edge devices such as wearables remains challenging. These systems must meet high accuracy requirements while operating under strict constraints on computational power, memory footprint, and real-time latency. In this work, we present an application of the STMC (Short-Term Memory Convolutions) framework to adapt a modular CNN model for online, LSTM-like inference. Our approach reduces power consumption and redundant computations while maintaining the stability and simplicity of training CNNs. We achieve up to 82% and 46% MCPS reduction compared to equivalently frequent standard CNN execution and vanilla STMC, respectively. The best configuration achieves 93.8% accuracy on the 11-class Google Speech Commands task and 97.1% on the same task with zero-padded data.

Vision-language-action (VLA) models have demonstrated strong capabilities in robotic manipulation, but they are typically developed and evaluated with all camera streams available throughout task execution. When a camera stops delivering frames during task execution, the policy must continue acting without access to subsequent observations from the missing view. Despite its practical importance, how such interruptions affect closed-loop manipulation remains insufficiently understood. To investigate this problem, we introduce MAIL-Bench, a benchmark that evaluates visual interruptions with VLA models. By interrupting different cameras at multiple stages of each policy's successful reference trajectory, MAIL-Bench measures how well policies retain their capabilities when visual inputs become unavailable. Building on this benchmark, we propose MINT, which first trains VLA policies to remain functional under missing visual inputs. At inference time, MINT selectively supplements missing observations using optical-flow extrapolation or an action-conditioned world model, and withdraws predicted views when they become unreliable. Experiments on $π_{0.5}$ and GR00T N1.5 show that MINT significantly improves task success under camera loss over the original models. Experiments on AgiBot G2 further demonstrate the real-robot deployment under camera loss. The benchmark is available at https://minglejiang.github.io/Mail-Bench/

Visual token compression reduces the inference cost of Large Vision-Language Models (LVLMs). However, aggregate robustness measures do not reveal whether a particular adversarial failure is induced by compression or inherited from the underlying model. We define a compression-specific failure (CSF) as an adversarial input that remains correct under full-token inference but fails after compression, casting compression-induced risk as a paired failure attribution problem. Within a controlled diagnostic cohort, counterfactuals show that retained-set allocation causally changes compressed correctness and reveal a negative association between recovery and representation drift in displaced evidence. Motivated by these findings, we propose CIRA, a Compression-Induced Risk Attack for Large Vision-Language Models. Under a vision-encoder white-box setting, CIRA optimizes image perturbations through encoder-side objectives that manipulate token priorities across candidate compression budgets while preserving displaced evidence. CIRA uses no downstream questions or labels and requires no access to the language model, deployed compressor, or exact compression budget. Across 12 dataset-compressor settings evaluated at four budgets, CIRA achieves a mean CSFR of 20.35% while limiting full-token attack success to 6.92%, with similar behavior on additional LVLM families. A cross-view selection-stabilization defense substantially suppresses CIRA, although Adaptive CIRA partially restores its effectiveness. These results show that compression-specific failures persist under restricted access and support paired evaluation of full-token and compressed inference for attributing risk to visual-token compression.

Generative AI (GenAI) is increasingly embedded in software engineering education, supporting activities such as requirements development, design exploration, documentation, and prototyping. However, educators often have visibility only into final artefacts, with limited insight into how students evaluate, verify, and refine AI-generated outputs during the learning process. This creates challenges for assessing evaluative judgement and responsible AI-assisted practice. This paper introduces the AI Journal, a structured reflection framework designed to make student-GenAI interaction visible in first-year software engineering education. The framework combines execution tracking, which records prompts, outputs, intent, and interaction context, with cognitive auditing, which captures verification strategies, intervention decisions, confidence judgements, critical learning moments, and reflections on AI-supported work. Deployed in a first-semester software engineering course, the AI Journal enabled visibility into aspects of student learning not observable through artefact-based assessments alone. Preliminary observations suggested variation in verification practices, intervention strategies, and perceptions of AI-supported work. Critical learning moments frequently occurred when students evaluated contextual suitability, feasibility, and requirements alignment rather than identifying obvious errors. The AI Journal demonstrates a practical, lightweight, and model-agnostic approach for making AI-assisted learning processes visible. By foregrounding verification, intervention, and reflection, it shifts attention from product-focused assessment toward evaluative judgement and responsible AI-assisted practice.

Models of quantum systems faithfully map system parameters to observations, but the inverse problem of parameter inference from measurement data presents a fundamental challenge: computationally intractable likelihoods due to an exponentially large Hilbert space. Here, we introduce simulation-based quantum system inference, a unified, likelihood-free framework that learns parameter posteriors directly from classical simulation data. The central idea is to pair polynomial-cost classical simulators, such as Pauli propagation and tensor networks, with normalizing flows or other neural density estimators for accurate, reusable inference. A single model, trained once, maps any new measurement record to its posterior in one forward pass---turning per-experiment inference into a fixed, up-front cost. We numerically demonstrate the framework's versatility across Pauli noise learning, quantum error mitigation, quantum state tomography, and Hamiltonian learning, with examples involving 81-qubit shallow circuits and 735-parameter inference. In each case, the approach yields accurate estimates of identifiable parameters, while posterior uncertainty provides additional diagnostics of non-identifiability and indicates where further characterization is needed. Our framework reduces data-acquisition requirements in quantum experiments and accelerates parameter inference, providing a practical route to characterizing and improving large-scale quantum systems.

Most music-generation systems are still framed and evaluated primarily as producers of complete outputs, whereas composition often proceeds through successive revisions to a shared musical artifact. This paper studies a different use of a general-purpose instruction-following large language model: not as a one-shot music generator, but as a reusable operator over an evolving symbolic score. We formulate incremental composition as a sequence of operation-aware state transitions over persistent ABC notation, with explicit requirements on what each operation may change and what it must preserve. The interaction includes two artifact-initialization variants and three editing operations -- chord addition, inpainting, and transposition. We instantiate the formulation by adapting Llama 3.1 8B Instruct with Low-Rank Adaptation (LoRA) on 496,038 operation-aware dialogue records derived from Irish traditional music. The comparison with the unadapted model is used to test the feasibility of learning this interaction contract, not to claim novelty for fine-tuning itself. Across 500 dialogues per model (1,750 attempted output states), checker admission rises from 29.37% to 99.37%, while compliance conditional on admission rises from 0.7205 to 0.9798. Strict eligibility for reference-relative musical-feature analysis increases from 14 to 1,548 outputs, and Longest Common Subsequence analysis does not show a systematic increase in high-overlap sequences relative to held-out baselines under the specified protocol. The results support the technical feasibility of persistent, operation-aware symbolic editing with a general-purpose instruction LLM. They do not establish superior musical quality or human-AI co-creativity, which remain questions for musician-centered evaluation.

Decoding for large language models is typically treated as a collection of isolated sampling strategies, with limited theoretical understanding of the behaviours they induce and how their underlying objectives relate. We formulate decoding as an optimisation problem over next-token distributions on the probability simplex, balancing expected model score against regularisation under support constraints. This view recovers familiar decoding methods through choices of regularisers and support constraints; more importantly, it enables new decoders to be constructed by composing distributional preferences within a single optimisation problem without external rewards, learned critics, or model parameter updates. We introduce CompoSimplex, a library with configurable support rules, regularisation primitives, and simplex solvers for constructing and evaluating compositional decoders. We evaluate standard samplers, individual regularisers, and compositions across multiple models and reasoning tasks. Our results show that compositions can realise trade-offs between single-sample quality, multi-sample quality, and diversity that are not attained by individual decoding objectives.

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models but incurs substantial costs from rollouts and policy updates. Online prompt selection improves efficiency by using per-prompt Bayesian posteriors to predict difficulty and prioritize informative prompts. However, existing methods overlook how reliably learning signals are extracted from sampled responses. In GRPO, a response's advantage depends on both its own outcome and the randomly sampled outcomes of its peers through group normalization. Our theoretical and experimental analyses show that uncertainty in group composition introduces composition noise, a non-vanishing variance component that imposes an irreducible lower bound on gradient estimation error and impairs downstream prompt selection. We propose MaPP (Marginalized Posterior-Predictive), a unified framework for data-efficient RLVR that denoises response-level advantage estimation and improves prompt selection using a shared Beta posterior. For each response, MaPP replaces the standard group-relative advantage with a composition-invariant intrinsic advantage through closed-form Beta-Binomial marginalization. The resulting posterior-predictive estimator has an error that provably diminishes as the posterior concentrates. Using the same posterior, MaPP derives an uncertainty-aware prompt selection score to improve data efficiency without additional rollout cost. Experiments on mathematics, planning, and visual geometry across five model backbones show that MaPP consistently outperforms GRPO and strong selection baselines, achieving up to +2.45 average accuracy improvement over the strongest baseline under the same rollout budget and setting a new state of the art.

Self-evolving agents improve future behavior by reusing past experience, typically as global prompts, memories, or reflections. Yet these mechanisms rarely control where experience takes effect. In long tool-use workflows, the same lesson may correct one decision but distract another, making experience reuse a problem of localized control rather than memory alone. We introduce EvoCUE (Evolution through Control Updates from Evidence), a framework for learning reusable control-program updates from completed agent executions. EvoCUE represents the agent as an explicit state-machine controller, whose nodes perform model or tool calls and whose edges define where control passes next. This makes the workflow editable at precise locations, so each learned update can specify what to add, where it acts, and when it applies. From completed trajectories, EvoCUE uses residual goals and observed execution traces to propose localized instruction or skill edits. Each candidate is evaluated at the point where it would act by resuming the parent and edited controllers from the same checkpoint and comparing their final outcomes. Accepted edits are compiled with applicability rules, confirmed on held-out tasks, and inherited by later executions. We evaluate EvoCUE on long tool-use environments where learned conventions must reach the right execution step. From a minimal AppWorld controller without benchmark-specific onboarding instructions, EvoCUE learns the missing task-completion convention and substantially improves success on Test-Normal and Test-Challenge. On PAST-Bench office workflows, EvoCUE transfers organizational requirements from prior episodes to later tasks, improving task-execution quality. These results show that self-evolving agents should place experience inside the control flow, rather than only store it as text.

We describe the Nürnberg NLP system for ChildSafeAds 2026. The shared task asks what a monitoring system for commercial content in child-facing YouTube videos can achieve at a given level of data access. We answer with per-subtask ensembles of nine voters, organised into three branches that differ in backbone, adaptation method and class scope. Selection rests on channel-disjoint cross-validation, with the development set as a transfer check. The system wins two of the three subtasks. Its product-category score (ST2, 0.8243) and its compliance-flag score (ST3, 0.6530) are the best of the 22 final entries, and it places third on the task mean (0.7079). We further compare four access levels and report the cost at test-set scale.

Multi-reward policy optimization requires a joint update that reflects both the learning signals and the intended relationships among objectives. We introduce Objective-wise Reconciled Policy Gradient (ORPG), which constructs a separate clipped policy objective for each reward and reconciles the resulting gradients into one policy update. For compatible gradients, a cosine-dependent interpolation coordinates their contributions through a partially normalized reference while preserving the norm of their sum. We characterize this update as the unique solution of a spherical directional compromise. For conflicting gradients, projection follows the task's priorities. We evaluate the same compatible rule in helpfulness--safety alignment and correctness--cost optimization for mathematical reasoning. ORPG substantially improves average Useful and Harmless scores over the strongest external baseline on each axis. In mathematics, it achieves the highest average full-budget accuracy and three-budget hypervolume among the compared methods, with more accurate and shorter responses than the initial policy. Component comparisons and training dynamics show the larger contribution of compatible coordination and a complementary benefit from conflict handling. These results support gradient reconciliation for objectives with equal standing and for objectives with an explicit priority.

Smart contracts underpin decentralized finance, where growing demand for on-chain/off-chain communication(OFC) has driven diverse applications such as cross-chain bridges, real-world asset tokenization, and fiat-backed stablecoins. TheOFC-related security incidents in these applications are increasingly frequent, but prior studies address separate vulnerability categories within OFC applications rather than providing a unified view, causing vulnerabilities outside known patterns to be missed.In this paper, we identify OFC inconsistency (OFCI) as a root cause of OFC vulnerabilities, which arises from business-logic flaw and ultimately breaks the equivalence between the on-chain and off-chain asset representations to induce inconsistency.Automatically detecting OFCIs faces two challenges including (1)locating heterogeneous business logic, and (2) transferring existing vulnerability knowledge to identify unseen OFCI instances. To this end, we propose SmartMemory, the first framework to leverage a memory-based agent for OFCI detection. To address heterogeneity, SmartMemory maps diverse implementations ofOFC contracts into a canonical business-semantic representation to locate the business logic for OFCI inspection. For knowledge reuse, SmartMemory integrates a memory-based agent to distill vulnerability knowledge from features into patterns and detection rules, enabling knowledge transfer across cases to identify unseenOFCIs. Lastly, SmartMemory performs taint analysis to verify the reachability, type, and impact of each candidate OFCI. We construct the first real-world OFCI dataset comprising 48 DApps with 81 OFCIs for evaluation, on which SmartMemory achieves80.68% precision and 87.65% recall. In addition, through an analysis of 325 real-world OFC applications, SmartMemory detects 36 previously unknown OFCIs, all of which have been confirmed and fixed by corresponding parties.

Multimodal Large Language Models face significant efficiency challenges that stem from two distinct yet coupled sources: data redundancy and computational redundancy. While most methods focus on data redundancy by pruning visual tokens from the output of the visual encoder or computing redundancy in LLM decoders using blockwise importance, the finer-grained inter-layer representation shifts and the distribution differences within the layers themselves have not been fully explored. In this work, we comprehensively investigate this dual-level inefficiency. We posit that intermediate layer tokens from vision encoders should be considered for effective visual token pruning, as semantic focus shifts across layers, with middle-layer tokens capturing more detailed object-centric information that deeper layers may abstract away. Furthermore, we reveal the differential contributions of Attention and FFNs across distinct LLM decoder layers. Building upon these discoveries, we propose \textbf{SPIDER}, a training-free framework that integrates multi-layer \underline{\textbf{S}}emantic visual token \underline{\textbf{P}}run\underline{\textbf{I}}ng with an a\underline{\textbf{D}}aptive sub-lay\underline{\textbf{ER}} skipping mechanism. Experimental evaluations demonstrate that SPIDER consistently maintains strong performance across various MLLM architectures and reduction ratios. For instance, on LLaVA-NeXT-7B, SPIDER reduces FLOPs by $79\%$ while maintaining 96$\%$ of the baseline performance.

The integration of artificial intelligence into computer-aided design frameworks has sparked a shift in the design of analog integrated circuits (ICs), transitioning the field from using manual and algorithmic-based solutions to adopting automated and intelligent paradigms. In this scenario, the GDSII file represents the industry-standard database containing the ultimate and most accurate source of information of the analog circuit, encapsulating the complex physical geometries and parasitic realities that define tape out performance. This paper proposes a novel framework that combines fine-tuned LLMs and CNNs to analyze GDSII files of analog circuits, enabling a conversational interface between the tool and the designers. Experimental results using thousands of analog designs across four realistic tasks demonstrate that the proposed solution outperforms state-of-the-art general-purpose massive VLMs by a significant margin (up to 81%), thus providing a lightweight solution to the problem of GDSII analysis.

Group Relative Policy Optimization (GRPO) improves language-model reasoning by comparing verified rewards among multiple solution rollouts for each query. However, difficult training queries can yield only incorrect rollouts, leaving GRPO with no reward contrast or learning signal. Prior hint-based methods construct auxiliary hints from solution evidence and use them to re-solve failed queries, recovering learning signal. Yet the resulting trajectories are typically treated as ordinary solution trajectories despite being generated under an assisted condition unavailable at evaluation. We discover hinted reward shift: recovered reward contrast can concentrate policy updates on hinted trajectories, limiting improvement without hints. This also creates a trade-off: increasing hinted trajectories can accelerate early learning but intensify reward shift later. To address this problem, we propose HATCH (Hint-Annealed Self-Teaching), an online single-policy framework that learns from both generating and using its own hints to improve reasoning without assistance. To mitigate hinted reward shift, we introduce online weighting to anneal the contribution of hinted trajectories. However, learning to generate hints can conflict with improving query solving. We therefore use gradient projection to remove the opposing component of hint-generation updates. Together, these designs support self-improvement by enabling the policy to create learning opportunities for itself and turn them into stronger reasoning without hints. We evaluate our method on mathematical reasoning benchmarks and outperform state-of-the-art methods by 1.02 pp on Llama-3.2-1B-Instruct, 2.84 pp on Qwen3-1.7B, and 4.32 pp on Qwen3-8B.

LLM-based Web agents can autonomously complete user tasks, yet deceptive interfaces can steer them toward outcomes that conflict with users' interests. Existing defenses primarily intervene on agent behavior through blocking, guidance, or replanning. We identify a distinct failure mode: a task-valid action can still realize an unauthorized consequence because of the current Web state. This motivates treating task-relevant Web state itself as a runtime control target. We introduce Veer, an agent-side runtime defense that leaves task planning to the base agent and intervenes on Web state when a proposed action would produce an unauthorized consequence. Before modifying the live environment, Veer constructs a prospective intervention trajectory toward a safe task-relevant state and executes it with runtime grounding and verification. Across TrickyArena and WebDecept, Veer achieves the highest safe task completion in all three evaluation settings, exceeding the next-best defense by 15.9 and 25.0 percentage points on TrickyArena-Single and TrickyArena-Multi, respectively, while reducing dark-pattern success on WebDecept to 0.3%. These gains persist across dark-pattern types and all 12 agent, model, and benchmark configurations. Ablations show that active state intervention provides the largest gain, while prospective rollout and temporal evidence contribute additional improvements. These results establish task-relevant Web state as an effective runtime control target for protecting Web agents from deceptive outcomes.

Full-duplex voice agents can now listen, speak, use tools, and act during spoken interactions, but fluent dialogue does not guarantee correct completion of delegated professional workflows. We introduce APEX-Voice, a benchmark of 120 interactive professional workflows spanning ten work archetypes such as form completion, corporate negotiation, coordination, consulting, and interviewing. Each workflow executes in a stateful Voice Workbench environment with task-specific knowledge, typed tools, gold-annotated final work artifact, authorization constraints, and a user simulation policy backed by validated, pre-compiled speech realizations. We evaluate both artifact field accuracy and end-to-end workflow success, which requires the correct terminal state, valid process, completed actions, and a valid final artifact. Across five frontier real-time voice agents-GPT-Live-1, Gemini-3.8-Live, Grok-Voice-Think-2.0, Step-Audio3, and GPT-realtime-2.1, none exceeds 25% Pass@1, and the best Reliable@3 is only 10.8%. Moreover, stateful coordination is the dominant failure point across systems, while success decreases further on workflows requiring greater knowledge retrieval and mid-speech corrections. Overall, APEX-Voice is the first benchmark for evaluating whether voice agents can translate conversational competence into dependable professional work.

Long visual token sequences often account for a substantial fraction of the computational overhead in multimodal large language models~(MLLMs). Existing approaches reduce this cost by pruning redundant visual tokens, but permanently discard visual evidence that may become useful in subsequent layers. We instead ask whether all visual tokens can be preserved while reducing the cost of repeatedly evolving the representations through the Transformer. To answer this question, we perform low-rank interventions on visual-to-text information flow. We find that, after visual-to-text attention is blocked, restoring only a few directions recovers most of the lost accuracy, suggesting the relevant visual influence is concentrated in a low-dimensional subspace. We further observe strong predictability in layer-specific visual states: lightweight MLPs approximate them with high cosine similarity and low reconstruction error. Motivated by these findings, we propose $δ$-Vision, which replaces repeated Transformer evolution of visual tokens with lightweight low-rank adapters that construct layer-wise visual memories while preserving all visual tokens for text retrieval. Across image and video benchmarks, $δ$-Vision achieves higher accuracy than visual token pruning baselines at comparable or lower computation, while delivering competitive inference efficiency without discarding visual tokens.

Large Language Models (LLMs) have shown strong performance on tool-use agentic tasks when given a fixed tool schema. Yet a tool schema is not the action space of an agent; it is merely one interface representation of it. The same executable action can be exposed through many different, functionally equivalent tool definitions, and an agent that has truly learned a task should behave consistently across them. We show that current agents often do not, a phenomenon we term schema bias. To study this systematically, we introduce an executable transformation framework that rewrites a native tool schema using nine operators, including merging and splitting tools, altering how a single tool is expressed, and distributing one action across several dependent calls. The tasks, executable actions, and reachable states remain fixed, so any change in success is attributable to the interface alone. Evaluating eleven LLMs, including two closed models, on up to 32 schema variants, we ask how large schema bias is, how it manifests, whether the difficulty of a schema variant can be predicted without a full evaluation, and whether training removes it. We find that schema bias is substantial even for the newest models: success rates range from complete failure to 97% depending solely on the schema. To reliably estimate schema difficulty, it requires running a small sample of the target queries. Training repairs a schema variant only when that variant appears in the training data.

Safety-aligned LLMs can exhibit emergent misalignment (EM): narrow domain adaptation unexpectedly triggers catastrophic safety failures across unrelated domains. Prior static analyses leave training dynamics unmapped, while existing defenses rely on heuristics that degrade utility. We present a dynamic, second-order geometric study of EM. Tracking training trajectories reveals that directional Hessian curvature concentrates sharply on semantic pivot tokens. Grassmannian projections show that, in most settings, harmful-safe gap widens mainly because safe-gradient overlap declines. Leveraging these insights, we introduce a parameter-level Geometric Mitigation Framework that orthogonally projects empirical harmful gradient subspace out of parameter updates. On Qwen2.5-14B-IT, our defense suppresses free-generation EM by up to 80.0%; across the other three of four open-weight instruction-based model families (3B--20B), where single-layer behavioral EM is already near zero, teacher-forced evaluation shows same harmful subspace controls the conditional support of frozen EM responses. Crucially, these diagnostics unmask the illusion of behavioral safety: the same subspace remains measurable and steerable in models where behavioral EM is near zero. Code: https://github.com/WeiqiaoQUE/mechanistic-emergent-misalignment.

Vision-language-action and world-action models are increasingly popular, yet remain bottlenecked by physical interaction data that is scarce, institutionally siloed, and task-heterogeneous. A natural federated solution is to let each client adapt a shared foundation model through parameter-efficient fine-tuning, avoiding the exchange of full-model updates. However, federating these adapters is nontrivial, as naive aggregation can entangle incompatible updates, while incorporating MoE-style routing into federated aggregation may dilute specialization and destabilize expert selection. We present RoboFL, which instantiates MoSAIC (Mixture of Slotted Adapters) for federated world-action learning. MoSAIC directly installs locally trained LoRA adapters as the expert branches of a server MoE. Server-side routers learn token assignments over these prior-informed branches while jointly refining routing and expert parameters. Foresight-to-Action Routing Distillation (FARD) aligns routing across the model's three paths, while Path-Consensus Expert Aggregation (PCEA) converts complete expert updates into a compact global adapter for personalized redistribution. Experiments on RoboTwin 2.0, RLBench, and a real-world Franka robot arm show the superiority of RoboFL with structured expert assembly, as it outperforms centralized PEFT InternVLA-A1 by 12.23% on the Franka arm, while reducing per-round client communication by up to 86.81% relative to MoE-based federated VLA baselines.

Semantic uncertainty quantification for large language models rests on a common template: sample several answers, measure how much they agree, and treat disagreement as uncertainty. We first formalize this template as two separate roles: an operator that compares two answers, and an aggregator that combines all pairwise comparisons into a scalar. Existing methods differ almost entirely in how they aggregate, while taking the operator off the shelf, typically an NLI model or a generic sentence encoder. We show that this reliance on off-the-shelf operators is the primary bottleneck of semantic UQ: they do not accurately measure factual equivalence of multiple answers to the same question. We resolve this with a deliberately simple recipe: a single encoder trained contrastively to isolate the targeted fact, utilizing synthetic data generated by an LLM and dataset both disjoint from all evaluation settings. Integrating the resulting operator into existing methods improves performance on 120 of 126 evaluation settings (95%) spanning 18 model dataset combinations across language and vision-language models. The best variant reaches 0.76 mean AUROC against 0.68 for the strongest baseline, while replacing the quadratic cross-encoder comparisons of entailment-based operators with one encoder pass per answer. The uniformity of the improvement supports the view that the operator, not the aggregator, is the limiting factor. The same operator also improves single generation token-level estimators: the norm it assigns to each token measures how much that token bears on the answer, and reweighting token log-likelihoods accordingly sharpens the estimate.

Greenhouse climate control balances economic return with maintaining temperature, humidity and CO2 within crop-adapted growth ranges. Conventional reinforcement learning (RL) greenhouse controllers use fixed reward penalties to limit climate constraint violations, yet such heuristic penalties cannot explicitly constrain long-term cumulative violations. Poorly tuned weights either lead to overly conservative policies and lower yields, or fail to suppress persistent climate deviations that harm photosynthesis and induce crop diseases. To address this issue, we formulate greenhouse climate regulation as a Constrained Markov Decision Process (CMDP) and use a Lagrangian safe RL framework RCPO-PPO to separate economic optimization and cumulative safety constraints, enabling adaptive penalty adjustment without manual tuning. To handle strong nonlinear, time-varying coupling between greenhouse microclimate and crop growth, Kolmogorov-Arnold Networks (KANs) replace Multi-Layer Perceptrons (MLPs) as policy and value approximators for improved nonlinear representation. Sinusoidal cyclic time features are embedded in observations to capture diurnal environmental periodicity. Simulations use a classic winter lettuce greenhouse model driven by 40-day real weather disturbances. Compared with vanilla penalty-based PPO, our method cuts cumulative climate violations by 18.65% and raises lettuce economic profit by 2.91%, keeping violations stable near the safety threshold. This decoupled CMDP optimization with KAN-based policy representation mitigates long-term climate risks and boosts planting profits, offering a constraint-aware control strategy for precision greenhouse cultivation.

Many physiological time series, such as cardiac and brain recordings, exhibit cyclostationarity: their statistics vary periodically with an underlying cycle phase. Corruption from motion, poor contact, and physiological interference obscures morphology needed for diagnosis, making signal restoration essential. Existing diffusion approaches condition on corrupted observations alone and must learn cyclic structure implicitly. We instead propose two inductive biases which encode cyclostationarity: a shift-covariant wavelet representation and dense per-sample phase conditioning inferred from the corrupted input. We further introduce a training-free cyclostationarity index that quantifies phase structure and predicts when phase conditioning will help. Finally, we propose antithetic coupling of reverse trajectories to reduce sampling variance while achieving comparable performance with fivefold fewer network evaluations. Across modalities, our results show that explicitly encoding measurable cyclic structure improves physiological time-series restoration.

Large language models can generate functionally correct code that still contains security weaknesses, motivating repair pipelines that first diagnose a weakness type before deciding how to fix it. The Common Weakness Enumeration (CWE) provides a standardized vocabulary for such diagnoses, but asking an autoregressive language model to generate a CWE label and extracting it from the response raises questions about output validity, speed, and cost, as well as accuracy. We evaluate Jev, a decision model that instead selects directly from a declared set of candidates and returns a probability for each, against six open-weight autoregressive models and a frontier proprietary model, GPT-5.6-Sol, on a controlled 50-way CWE classification task over 1,916 CyberSecEval benchmark examples. Jev outperforms all six open-weight baselines on every classification and ranking metric, while its comparison with GPT-5.6-Sol depends on the metric: GPT-5.6-Sol achieves higher Top-1 accuracy and Macro-F1, whereas Jev achieves higher Top-3 and Top-5 accuracy and a nearly identical MRR, at $6.27\times$ lower median API latency and $55.9\times$ lower estimated API cost. We further build JevVibe, a diagnosis-guided repair agent that uses predicted CWE labels to repair code generated by Qwen2.5-Coder-32B-Instruct. With Jev providing the diagnosis, the agent increases the detector-measured security pass rate from 63.5% before repair to 70.7%, compared with 66.1% for LLM-guided repair. These results show that JevVibe is effective at improving the security of generated code, with Jev providing reliable and efficient CWE classification.

Estimating mutual information (MI) from samples is a central objective in a variety of scientific fields. Modern neural estimators are accurate in the large-data regime, but they fall short when data is scarce, and each must be fit anew for every distribution under study. Current estimators are moreover tied to specific data types. These constraints limit their adoption in many applications where per-distribution training is impractical and sample sizes are small. We present ALICE, a foundation model that removes per-distribution training, while achieving competitive estimation accuracy. Trained exclusively on a broad family of synthetic distributions, ALICE acts as an in-context estimator of rectified-flow velocity fields: conditioned on samples of an unseen distribution, it estimates that distribution's velocity field without any explicit training. MI is then obtained through a fixed identity that integrates the squared difference between the joint and conditional fields. We validate ALICE on a standard, challenging benchmark and apply it in three domains, biology, genetics, and neuroscience, whose data the model has never seen. For the first time, we show that a single model closes the gap with neural estimators trained separately for each distribution, while natively supporting different data dimensionality and sample cardinality, enabling zero-shot MI analysis across scientific domains.

Formalizing research-level stochastic optimization in Lean requires both an algorithm model and domain theory connecting foundational libraries to convergence proofs. Revising a model to restore provability can change the mathematical claim. We introduce ProofLoom, a fully automated LLM-agent system for Proof-Obligation-Driven Theory Construction. Given a published algorithm, target theorem, and source proof, ProofLoom autonomously constructs the Lean model and supporting theory. Open proof obligations drive the development of definitions, interfaces, lemmas, and proof plans. Signature contracts record evidence and obligations for model revisions; an independent Judge rejects unsupported assumptions and weakened conclusions. Planner expands the published argument into intermediate claims, and Audit checks whether the Lean proof follows it. Across tasks, SOptLib accumulates verified mathematics and construction experience: reusable results are extracted, generalized, and verified, while modeling decisions and failed proof routes are recorded. Later tasks retrieve these results and records and contribute new developments, forming a cycle of construction, accumulation, and reuse. On fifteen textbook and research-paper tasks, ProofLoom obtains mean human ratings of 6.3/7 and 6.4/7, compared with 4.9/7 and 5.0/7 for the strongest of six baselines. Across 33 developments, it produces 490,693 lines of algorithm-local Lean code with no sorry. The formalizations also expose 28 incorrect formulas, proof gaps, and algorithm-analysis mismatches in published sources across 22 developments, each with checked evidence.

In 2023, AI models answered from training data and hallucinated when it ran out, and businesses were told to seed that knowledge. Models' training knowledge has since given way to live web search, and the advice followed it there: answer-engine optimization, or AEO, now tells businesses to scatter breadcrumbs across forum threads, listicles, and off-site citations, so AI engines are likelier to surface and recommend them. But being surfaced is no longer enough: an agent opens the results and reads them before deciding, and one buyer question sends it through several rounds of search and fetch. What decides the outcome at this drill-down step is whether the agent can fetch and read the business's own site: agent experience (AX). We argue that AX is the new AEO. We run 37,927 agent journeys, each a buyer question about a business, across four independent harnesses over 1,056 real businesses, matched on fame, prior model knowledge, and two AEO proxies, then split based on their AX level. Only 7-10% of the finished answer comes from the model's training knowledge, whether or not the site is readable. Agent-ready businesses have answers built from their own pages 78% of the time against 56% and are clearly recommended 1.9x more often, while every grounded answer about a not-agent-ready business costs the agent 64% more. Holding business, harness, and question fixed, answers built from the site are 41% more accurate. The dominant failure is not fabrication but omission: web-built answers are 3.7x more likely to contain none of the facts the buyer asked for. Baselines differ sharply across the four harnesses, with clear-recommendation rates varying sevenfold from stack to stack, yet the effect holds in every one. In the agentic web era, being readable beats being talked about, and improving a site's AX is the strongest lever a business has.

Few-shot visual anomaly detection is fundamentally a visual comparison task, requiring fine-grained inspection of a query against normal references. Many recent methods based on large vision-language models (LVLMs) emphasize comparative reasoning through language chain-of-thought. Yet discrete, abstract descriptions may underrepresent dense, fine-grained visual differences, leaving a gap between visual comparison and its expression in language. To address this gap, we propose Visual Difference DeepStack (VD-DeepStack), which explicitly conditions language reasoning on query-reference visual differences. Specifically, we fuse DINO features with the LVLM visual hierarchy to strengthen fine-grained representations, then construct dense difference evidence from residuals between query features and softly matched reference features. The difference-evidence path injects spatially weighted difference vectors into query-image states at multiple decoder depths, while an auxiliary visual-context path provides fine-grained appearance information to support their interpretation. Experiments on 4 industrial and 2 medical anomaly benchmarks demonstrate substantial improvements in few-shot anomaly detection over baselines relying on textual comparative reasoning. These results support mitigating the visual comparison-reasoning gap through the joint design of comparison representations and their integration into the decoder. Code will be released upon acceptance.

Proactive dialogue requires agents to continually adapt their policies to user feedback while progressing toward task objectives over multiple turns. To move beyond imitation learning on static datasets, recent approaches use user simulators to collect interactive data for policy optimization. However, many simulators do not explicitly model the evolution of user cognition, limiting the consistency and state dependence of feedback across turns. Moreover, representing each action only by a high-level strategy label overlooks the large utterance space and cannot distinguish alternative realizations of the same strategy. To this end, we jointly design a $\textbf{Cog}$nitive User $\textbf{Sim}$ulator $\textbf{(Cog-Sim)}$ and $\textbf{C}$ognitive-$\textbf{S}$tate $\textbf{T}$ransition--Driven $\textbf{P}$olicy $\textbf{O}$ptimization $\textbf{(CSTPO)}$. Cog-Sim maintains the user's cognitive and affective states and generates responses through constrained state transitions across turns, so feedback depends on both the realized utterance and the user's current state. CSTPO organizes each action as a hierarchical strategy--utterance representation: a high-level strategy label constrains utterance sampling, and utterances are optimized within each label. Sparse complete-branch sampling reuses shared dialogue prefixes and estimates separate strategy-level and utterance-level advantages, enabling fine-grained optimization at both levels. Across three tasks, Cog-Sim exhibits monotonic dose--response relationships and is preferred over prompt-based simulators for naturalness. CSTPO improves Qwen3-14B's performance to a level comparable to that of GPT-5.5-based planning methods.

Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs $3.6\times$ faster per guidance step with $7.0\times$ fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.

Future exogenous variables provide valuable information for forecasting endogenous time series. Existing covariate-aware methods primarily learn the direct influence of exogenous variables on endogenous variables. However, these effects can be complex and change with the pattern of the exogenous variables, making them difficult to capture. Beyond this perspective, we observe that a given exogenous pattern often co-occurs with only a small set of endogenous response patterns. These associations motivate a strategy that matches future and historical exogenous patterns and uses the corresponding endogenous patterns to enhance forecasting. However, in real-world forecasting scenarios with multiple exogenous variables, each exogenous variable provides a distinct dimension for matching, creating a dilemma for this strategy between precise matching and sufficient historical support. To bridge this gap, we propose XMatch (EXogenous MATCHing), a covariate-aware forecasting model that realizes the aforementioned strategy through a tree-structured matching process that adaptively adjusts the number of exogenous variables used as matching conditions. Specifically, we first introduce the ProtoTree Creator, which organizes historical correspondences between exogenous and endogenous patterns into a ProtoTree, whose deeper levels incorporate additional exogenous variables for matching. For forecasting, we then design the ProtoTree Matcher, which uses future exogenous variables to query the ProtoTree and adaptively determines how many exogenous variables to use for matching based on exogenous pattern similarity and historical support. Finally, the matched endogenous patterns are used as explicit historical evidence to enhance forecasting. Extensive experiments on 12 real-world datasets demonstrate that XMatch outperforms state-of-the-art baselines.

It is unclear how Neural Language Models (NLMs) acquire the structural meaning encoded by grammatical structures that is independent of lexical semantics. We propose a statistical learning process in which learned dependency structures themselves become new distributional units for subsequent statistical learning. Under this account, once a dependency structure is acquired, the model tracks its contextual distributions. These contextual features reflect the semantic properties of a composite structure. To test this hypothesis, we design a synthetic grammar in which each grammatical structure has distinct contextual distributions that cannot be recovered from the distributional statistics of their component tokens alone. We train a series of BERT-style masked language models on this grammar and examine their developmental trajectory. The results show that models can successfully learn the contextual distributions of composite dependency structures even though they cannot be inferred from token statistics alone. Developmental analysis further reveals a clear developmental trajectory. The learning of the dependency relations that define a grammatical structure consistently precedes the learning of its contextual features. These findings suggest that statistical learning in NLMs is not merely the accumulation of token co-occurrence statistics, but a process in which learned dependency structures become new units of distributional learning. We argue that this process provides a statistical-learning account of how NLMs solve the compositionality problem in language. Finally, we discuss the possibility that this statistical learning process provides an explanatory theory on how language cognition could emerge from pure distributional statistics.

Memorization has been proposed as a mechanism to explain how language models fit the tail of their training distributions, but its training dynamics are not understood well. In this work, we take a fine-grained look at memorization by decomposing the loss trajectory of memorized sequences over training and model parameters. Across the Pythia family, we study memorization of duplicated training sequences (recitation) and rare ones (recollection). We find that memorization in both cases is characterized by sequence-level gradient alignment, though recitation suffers from misalignment with other training influences which causes forgetting, explaining the necessity for higher duplication of these examples. We further show that the lower model layers are the most involved in memorization and forgetting. Predicting memorization, our decomposition improves over a cross-entropy baseline, especially in larger models and early in training. Intervening on a small set of highly influential parameters we are able to ablate memorization in the final model. Together, these findings advance our understanding of how memorization develops during training and offer insights for predicting and intervening on it.

Machine learning has made strong progress on music tasks, both as assistive tools and as creative partners. However, most systems train on multitrack corpora that emphasize pop and rock. Jazz, with improvisation at the core of its practice, still lacks a well-annotated corpus of clean per-stem combo recordings on standards. We introduce JazzSAMBA (Jazz Synchronous and Asynchronous Multi-take Band Audio) to fill this gap: the first originally recorded jazz-combo multitrack dataset of standards with asynchronous (overdubbed) and synchronous (live ensemble) protocols, preferred and alternate takes chosen by the musicians, and timed annotations for bars, chords, sections, and soloists. JazzSAMBA covers 76 standards by eight musicians on drums, bass, piano, trumpet, and saxophone, with per-stem audio, mixtures, and MIDI. It can support chart-conditioned accompaniment, combo source separation, and form-aware music information retrieval. We demonstrate the dataset on two tasks: a jazz combo source-separation baseline and a chart-conditioned accompaniment ablation. The dataset, code, and samples are linked from the project demo page.

Large language model (LLM) agents are evolving from tool-calling systems that execute isolated instructions into task-oriented agents that pursue user goals through sustained, multi-step interactions. However, existing benchmarks for personalized tool use largely assess isolated calls or reactive execution, leaving unclear whether agents can formulate, execute, and revise an explicit plan while preserving user preferences throughout long-term interaction. To address this gap, we introduce \textbf{PDEU-Bench} (\textbf{P}ersonalized plan \textbf{D}efinition, plan \textbf{E}xecution, and plan \textbf{U}pdate \textbf{Bench}mark), a benchmark for evaluating the complete planning lifecycle of personalized tool-using agents. PDEU-Bench comprises 214 long-horizon interaction tasks spanning 12 everyday domains and 94 tools, with stage-specific assessments of preference adherence and plan quality. Extensive evaluations of 15 representative open-source and closed-source LLMs reveal a pronounced gap between local tool execution and dynamic planning: LLMs can often instantiate preferences in individual calls, yet struggle to construct coherent plan definition and plan update. We further evaluate mainstream personalization and memory-augmentation methods. Although these methods improve particular stages, none of the evaluated methods reliably propagates user preferences throughout the complete lifecycle, and their gains frequently fail to transfer to subsequent execution. Fine-grained error analysis further reveals that preference omissions and conflicts persist throughout the planning lifecycle, highlighting the need for future research to parameterize LLMs with preference-aware information retrieval and memory capabilities. We provide the relevant code and data in the appendix to support future research.

Language models are increasingly used to sample from a specified distribution, for instance, to simulate survey respondents or generate synthetic data. Instruction-tuned models can state such a distribution correctly and still fail to sample from it. Prompting and changes to decoding reduce this mismatch only partly, which motivates training with policy optimization. Group relative policy optimization (GRPO) is a natural fit for this problem because it already samples a group of rollouts per prompt, and the group's empirical distribution can be compared with the target. However, scoring the group as a whole gives every rollout the same reward. Group-relative centering then sets all advantages to zero, and the model receives no learning signal. To give each rollout its own signal, we introduce the witness advantage, a per-rollout advantage derived from maximum mean discrepancy (MMD). It trains a model to match a target distribution over a finite set of outcomes. The MMD between the model's distribution and the target has a witness function that measures how over- or under-produced each outcome is. Each rollout's advantage estimates the negative witness at its outcome, so a rollout is rewarded for an outcome the group under-produces and penalized for one it over-produces. The witness advantage is computed in closed form from the group's outcome counts, and we use it as the reward in GRPO. On unseen target distributions, training with the witness advantage substantially reduces the total variation distance to the target while largely preserving the model's general capabilities.

An auditor who checks whether a system's weights are frozen is checking the wrong thing. Our stationarity dichotomy says that iterative self-modification hits strict diminishing returns whenever the agent's reachable set of edits stays fixed, and can escape only if that set expands. Rewriting scaffolding (tools, verifiers, decomposition) expands what an agent reaches without touching a weight, so frozen weights buy an eventual ceiling but no stationarity along the way. The criterion also separates three regimes usually merged: search within a fixed class, test-time training that raises the ceiling itself, and scaffold rewriting between them. Audit the scaffold, not the checkpoint. The same ceiling binds sideways. Best-of-$k$ orchestration realizes the best worker's ceiling exactly: width buys rate, not budget. Re-consulting a fixed pool has a horizon computable in advance, decided by the pool alone, and the one arrangement that would beat it, a weighted vote, needs diversity real workers lack: on 30 same-family workers the failure overlap sits at its maximum, and a majority fails 23/55 (42%) of tasks. We obtain the criterion by reading refinement as gradient boosting on the residual error between draft and target, a patch or git diff, and then measuring where that reading breaks: patches compose instead of standing beside each other to be voted on, and failures overlap. What we measure is saturation. Per-round improvement decays toward zero on SWE-bench, and churn decays geometrically across 401 production sessions, a shape shared with a pre-AI human baseline that establishes the regime without identifying its cause. Both breaks are engineering choices rather than laws about code, so together they specify a harness worth building.

Predicting Drug-Target Interactions~(DTIs) is a central task in computational drug discovery, with direct applications in virtual screening, drug repurposing, and therapeutic candidate prioritization. Although recent deep learning methods have improved DTI prediction, many sequence-based models still process drugs and proteins independently and only combine their representations at a late prediction stage. This limits their ability to explicitly model cross-molecular dependencies between chemical substructures and protein sequence regions. In this paper, we propose a sequence-only DTI prediction architecture that combines two pre-trained language models, ChemBERTa for drug SMILES strings and ESM-2 for protein amino acid sequences, with a hierarchical interaction module. The proposed model first extracts contextual representations using pre-trained encoders, then applies 1D convolutional layers to condense local sequence patterns, followed by a sequential bidirectional cross-attention mechanism inspired by the induced-fit view of molecular recognition. Finally, attention-based pooling constructs fixed-size interaction-aware vectors for binary prediction. Experiments on BIOSNAP, Davis, and BindingDB show that the proposed model achieves the best performance on BIOSNAP, matches the best AUROC on Davis, and remains competitive on BindingDB while using only 25.2 million trainable parameters. Ablation results confirm the contribution of both the CNN and cross-attention modules, and cold-start experiments indicate promising generalization to unseen proteins and drugs.

On modern production text-to-image systems, successful policy violations are rare, and previously effective human-written seeds are often patched out. Current automated red-teamers are poorly matched to this regime in two ways: unreliable success measurement and poor exploration. First, we find that judges widely used in prior T2I red-teaming work are unreliable under vague unsafe-content targets: they either miss true violations or reward benign borderline images on hardened APIs. We therefore define strict category-specific success criteria and calibrate strong VLM judges against human labels. Second, we show that broadly used prompt-modification pipelines do not solve the exploration problem: on harder guardrail settings they remain tied to seed prompts, fail to transfer, or cannot bootstrap positive examples. We introduce RISE, which evolves reusable strategies used to generate prompts rather than rewriting them one by one. The best discovered strategies are then reused to generate attacks across new scenarios. On DALL-E 3, Nano Banana 2 (Google) and GPT-Image-2, RISE reaches up to 13% human-verified ASR; under the same calibrated evaluation, prior methods with reported ASR as high as roughly 30% fall to near zero.

Accurate prediction of avalanche motion is essential for hazard assessment in mountainous terrain. This study develops and evaluates a physics-informed neural network (PINN) framework for the Savage-Hutter model of depth-averaged granular flow, progressing from 1D analytical verification to 2D experimental validation. First, three 1D problems of increasing complexity were verified against the analytical solution: height prediction with prescribed velocity, velocity prediction with prescribed height, and coupled prediction of both fields using the conservative formulation. The decoupled tests accurately reconstructed the spatio-temporal evolution of each field when the other was prescribed. The coupled formulation learned both fields without prescribed data, achieving mean height and velocity RMSEs of 0.043 and 0.079 in non-dimensional units. A hyperparameter sensitivity study evaluated the effects of network depth, width, collocation density, learning rate, and epochs. The framework was then extended to 2D and validated against laboratory experiments of a cylindrical granular pile collapsing on an inclined plane, with TITAN2D providing numerical comparisons. Purely physics-based training converged to the trivial zero solution; augmenting the loss with 10 sparse training points from final deposit profiles produced a physics-informed, data-assisted hybrid framework. Peak flow depth, depth-averaged velocity, RMSE, and wetted-area IoU evaluated global and local agreement. Global height RMSE ranged from 2.7 to 6.7 mm across four experimental cases, while mean wetted-area IoU ranged from 69 to 81 %, demonstrating consistent performance across variations in pile mass and slope angle.

Muon is increasingly used for language-model pretraining, yet its large-step dynamics are not captured by the classical edge-of-stability (EoS) picture of gradient descent (GD). In GD, loss neutrality, equal-magnitude update reversal, and marginal stability meet at a single learning-rate-dependent edge. We show that Muon breaks this coupling. For stochastic no-momentum Muon, we derive a coherence-corrected conditional loss-neutral boundary $2ρ_b/η$, while temporal alignment follows a separate geometry. Controlled experiments show that loss balance and temporal alignment respond differently to learning rate and batch size. Across our language model experiments, the 130M Llama-like LLM runs exhibit loss-boundary tracking with weak negative alignment, whereas the studied 1B LLM configuration shows stronger partial cancellation; in both settings, directions remain far from coherent reversal while training continues to improve. These results support a split EoS picture for Muon: a stochastic loss-neutral edge survives, but it is not accompanied by a universal temporal-direction signature. The source code for reproducing the experiments can be found in https://github.com/cyzebra/Muon-Sublates-the-Edge-of-Stability-in-LLM-Pretraining

Tool-using language-model agents can review enterprise projects as governance boards do: they read the evidence, apply written rules and decide whether the project may proceed. Part of that evidence comes from suppliers and project members with a stake in the decision. DGF-Bench is a benchmark in which a board of agents (specialist gates and a General gate that consolidates their decisions) reviews synthetic dossiers while an attacker plants deceptive content in evidence the organization does not vouch for. Dossiers are generated from canonical facts under 61 executable rules, with 42 authoritative records and 32 narrative documents; every gate is certified decidable from those records. Attacks never change an authoritative value, so an attacked dossier keeps the reference decisions of its clean copy. A success is attributable only when the agent receives the injection and takes the exact injected action, which it does not take on the paired clean dossier; the DGF score is the share of applicable fixed attacks a model blocks. Reading documents and records themselves, five of six models were outcome-strict (disposition, findings, actions and authorization all correct) on 82 to 85 of 85 gates. Over 2,622 attacked gate runs, seven direct-order, false-data and false-authority attacks obtained one attributable success against these five, whereas task-aligned attacks imitating the organization's own process passed against four of them: a record note citing a fake review procedure lowered GPT-6 Luna Pro from 34 to 6 outcome-strict gates and DeepSeek V4 Pro from 33 to 7. DGF scores ranged from 96.2 to 26.9, and a policy-aware adaptive attacker writing in records succeeded against five of six models. The approval tool executed no forged approval, yet deceived agents submitted approvals that the rules forbid. The open-source package dgf-bench computes the DGF score with one command.

Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose *Action Upcycling*, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2--1.7$\times$ with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.

Bathroom acoustic-event recognition can support ambient assisted living in settings where continuous video monitoring is undesirable. However, practical deployment requires models that are compact, interpretable, and robust to changes in the recording environment. This work introduces \dataset{}, a seven-class bathroom acoustic-event dataset containing 21{,}387 annotated clips recorded across five environments, and proposes SincDPNet, a compact raw-waveform classifier with a learnable sinc filter bank followed by a depthwise-separable convolutional body. Each sinc filter is controlled by two frequency parameters, allowing the learned passbands to be inspected directly in hertz while keeping the front end small. To reduce room-specific leakage, recording sessions and environments are separated before overlapping windows are assigned to the training, validation, and test partitions. We further use multi-objective Bayesian optimization as a design tool to examine the validation performance--model-size trade-off across 24 configurations. The selected designs span different operating points: the best-performing model achieves 80.2\% accuracy and 0.760 macro-F1 with 14{,}040 parameters, while the compact $N_f=25$ configuration uses only 2{,}848 parameters and achieves 75.7\% accuracy, 0.661 macro-F1, and 0.716 MCC on the held-out environment. Analysis of the learned filters and confusion patterns shows that spectral overlap contributes to confusion among water-related events, while the \textit{Door}/\textit{Walker/Crutch} errors also reflect similarities in their transient temporal structure.

Graph neural networks (GNNs) need to exploit improved message passing without surrendering control over predictions already trusted in deployment. We introduce Reference-Tail Trust (RTT), a framework that admits learned updates inside a frozen GNN and certifies the prediction actually served. RTT couples graph-based proposal states with a constrained internal optimizer: each displacement is charged for its worst-case terminal cross-entropy increase through the incumbent's remaining message-passing layers. A trajectory-validated tube and an independent checker enforce per-node probability floors, $p^{\mathrm{s}}_{ic} \ge e^{-H_{\mathrm{row}}} p^{\mathrm{r}}_{ic}$, and a call-level budget, $\sum_i w_i D_\infty(p^{\mathrm{r}}_i \| p^{\mathrm{s}}_i) \le H^+$, uniformly over labels. Calls whose adapted outputs pass certification require no separate full incumbent rollout; failed certificates trigger whole-call fallback. We derive the exact probability-floor frontier by water-filling, characterize architecture-constrained efficiency, and establish conditions under which internal propagation exploits evidence unavailable to restricted output correctors. In the reported ogbn-arxiv audit, RTT achieves $6.5\times 10^{-3}$ nats of mean gain per call, with a one-sided 95% regression-rate upper bound of 0.95% and a 95% negative-flip upper bound of 0.51% on the uninspected part of the reserved node population. Its mean gain is 61% of a cross-fitted posterior-based frontier estimate and exceeds the strongest matched one-pass corrector by $+0.9\times 10^{-3}$ nats. Reported experiments span eight proposals, six graph-incumbent families, structural and temporal graph shifts, and molecular prediction, with additional image and tabular evaluations. RTT makes GNN adaptation a budgeted, certifiable inference decision rather than an unconditional model replacement.

Simultaneous machine translation must generate target tokens before the source input is complete. Existing approaches address this through post-hoc read-write policies, leaving the attention mechanism unaware of bidirectional stream dependencies. We propose a dual-stream attention framework that represents source and target streams as a two-dimensional grid of hidden states and models their interaction through four structurally distinct attention types merged via joint QK Softmax normalization. Two approximations---broadcast and Hadamard---reduce the per-layer complexity from O(X^2Y+XY^2) to O(X^2+Y^2+XY) with provably decaying error. Training uses a self-guided loop: a per-cell loss heatmap drives dynamic-programming path recovery, which generates read/write decision supervision labels without external alignment. An incremental KV cache with anchored rotary position embeddings enables efficient streaming inference. On Chinese-to-English simultaneous translation, the proposed model outperforms the Wait-k baseline by +5.66 BLEURT and +10.36 COMET at comparable latency, and surpasses the non-streaming reference on COMET at a fraction of the response delay.

Multi-vector retrievers built on vision-language models lead visual document retrieval (VDR), but they run a multi-billion-parameter query encoder on every search. Distilling this encoder into a small student that queries the teacher's existing index would remove the bottleneck. The standard recipe, however, matches the teacher's MaxSim scores and so requires encoding and caching every training page, which can reach terabytes of page tokens. NanoVDR avoids pages entirely by training on the teacher's query embeddings alone, but only for single-vector retrievers. We present ColNanoVDR, to our knowledge the first framework to bring this document-free distillation to multi-vector VDR. Its objective, OTW (Optimal Transport with Learned Weights), aligns the student's query tokens with the teacher's by entropic optimal transport, with a learned weight for each student token, and needs no correspondence between the two tokenizations. We prove that the resulting alignment cost bounds the MaxSim score difference on every page. Distilled from five state-of-the-art teachers, the 149M text-only students retain about 95% of their teachers' NDCG@5 on ViDoRe v1-v3 while encoding queries up to 26x faster. Under identical training, OTW matches score distillation while encoding no page and reading 12.6x less cached teacher data.

Safety alignment of large reasoning models (LRMs) via supervised fine-tuning (SFT) and reinforcement learning (RL) often yields near-perfect safety scores, yet this apparent success comes at the cost of severe over-refusal and degraded general capabilities. Through systematic empirical analysis, we find that these failures are closely associated with the learning of spurious shortcuts rather than robust intent-sensitive safety evaluation. Specifically, we identify two dominant shortcuts: formatting shortcuts, where refusal behaviors are overly bound to structural prompt templates that frequently appear in safety alignment corpora; and lexical shortcuts, where sensitive keywords reflexively trigger refusals on benign queries. To mitigate reliance on these shortcuts, we propose DeShortcut-Align, a shortcut-decoupling alignment framework that reduces dependence on superficial cues. DeShortcut-Align operates across three coordinated stages: (1) Refusal Sensitivity Attribution, which masks input tokens to quantify their impact on the final refusal response distribution; (2) Attribution-Guided Contrastive Augmentation, which constructs benign contrastive samples using high-sensitivity tokens to mitigate lexical shortcuts; and (3) Counterfactual Consistency Regularization, which constructs template-ablated states via attention blinding to enforce decision consistency across SFT and RL, mitigating formatting shortcut dependence. Experiments on 7B and 14B models demonstrate that DeShortcut-Align significantly improves robustness against template-stripping bypass attacks (reducing performance drops by up to 72%), substantially reduces over-refusal by over 58%, and better preserves general-purpose reasoning capabilities, thereby mitigating the alignment tax commonly observed in safety training.

Structural analysis is central to Petri Net (PN) research, complementing state-space methods while avoiding their combinatorial issues. It is well studied for classical PNs but much less for High-Level Petri Nets (HLPN). Symmetric Nets (SN), a common HLPN formalism, use compact annotations to encode behavioral symmetries, allowing symbolic reachability graphs and associated lumped Markov chains for stochastic SN. In the past two decades, specific structural techniques for SN have emerged, notably the SNexpression tool, which implements a calculus for symbolic structural relations such as conflict and causality. We propose using this calculus to semi-automatically verify symbolic structural invariants, currently possible only for restricted SN subclasses, for an extended SN formalism (ESN) closed under key functional operators. We focus on flows and outline, at least in theory, how to construct a flow-generating family. We also sketch a framework for formally verifying a wider range of invariant properties. Representative examples illustrate the main concepts.

Tabular foundation models (TFMs) provide predictive distributions for regression, but their prediction regions can exhibit undercoverage or overcoverage even when point predictions are accurate. We introduce C-USIM (Conditionally-Uniformized Score Integration Method), a lightweight application of highest predictive density split conformal prediction that accommodates multimodal predictions. Given calibration and test outputs, it requires no additional training or model inference. It provides finite-sample marginal validity under our assumptions. We bound conditional-marginal coverage gaps using distribution-estimation error and score discreteness, and examine coverage heterogeneity through percentile rank-score plots. Experiments with TabPFN and TabICL show improved marginal coverage accuracy and lower average conditional and group coverage errors. Under a fixed data budget, allocating more observations to calibration can reduce marginal coverage error despite less accurate point predictions.

LLM-assisted Verus verification is a less tedious method to verify Rust implementations, but paired with self-referential structures, e.g., Doubly Linked Lists (DLLs)—notoriously difficult to formalise for verification—it becomes a substantially more demanding verification task. Moreover, a specification weakness can arise when verification relies on unproven or invalidated assumptions, such as axiomatic lemmas and assume statements. We investigate whether LLM agents can synthesize strong DLL specifications while minimizing these trusted base. The analysis follows three different approaches: manual verification, property-specific verification, and a defined skill for the specific case of DLLs and certain properties of this type of data structure. The skill encodes domain knowledge and a task-decomposition strategy. We show that an LLM agent equipped with a carefully designed verification skill can generate strong, low-trust specifications for DLLs in Verus.

Tool agents use large language models to act through external tools, yet successfully executed calls can still leave user requests unfulfilled. Tool-agent repair seeks alternative call sequences that execute successfully and fulfill the original requests. However, repair requires exploring both operation choices and their concrete realizations, making complete-sequence regeneration costly. Moreover, regeneration repeats operation selection even when failure arises from how those operations are realized. The resulting challenge is to reduce this repetition while preserving exploration of alternative operations and realizations. Therefore, we formulate repair as hierarchical search over operation supports, which we introduce as sets of permitted operation types that define reusable search regions for concrete tool-call sequences. We propose ReCommit, a training-free, diffusion-guided framework for improving tool-agent failure recovery while reducing repair computation. ReCommit amortizes operation-level proposal computation across repair trials by reusing operation-type scores from a single parallel readout of a masked diffusion language model. These scores guide search across supports, while realization search explores alternative entity bindings, arguments, and action composition within each support. Experiments on real failures across four enterprise services in the Agent-Diff benchmark show 75.9\% and 63.2\% relative recovery gains with 61.3\% and 51.3\% reductions in mean full-budget repair time at repair budgets $B=3$ and $B=13$, respectively, over the strongest evaluated 8B comparison method. ReCommit achieves a favorable recovery--cost trade-off, including in comparisons with the evaluated 32B models.

Modern endpoint malware detection is distributed: a lightweight agent on each endpoint collects features from a scanned file or process, sends them to a remote server for analysis, and then enforces the returned verdict locally by blocking, quarantining, or disinfecting. Because the endpoint acts on the verdict, the distributed machinery surrounding detection must never turn a transient server failure into a wrong action. We present a formal model, in Promela, of the endpoint decision pipeline of such a system, abstracted from a production architecture at Bitdefender. The model captures the system's graceful-degradation fallback chain: when the primary analysis server times out, the endpoint falls back to an older legacy-protocol server, and failing that to a reduced-signature local scan, before enforcing a verdict. Assuming detection signatures are sound, we specify six safety and liveness properties in linear temporal logic (LTL) and verify them exhaustively with the SPIN model checker. We prove that the fallback machinery never causes a false positive (an enforcement action against a benign file), commits to exactly one verdict per scan even when timed-out responses arrive late, weakens detection strength only in an explicit and ordered way, and always terminates in an enforcement decision, so the pipeline is deadlock-free. Each property is checked to hold non-vacuously, and we report how the state space grows with concurrent scans and endpoints. The work shows how model checking can give strong correctness guarantees for the failure-handling logic of a production security system, a layer that has received little direct formal attention.

The large safety instrumentation & control (I&C) systems in civil nuclear power plants (NPPs) are mainly safe-shutdown systems (reactor protection) or limitation and control systems. Framatome's established TELEPERM XS (TXS Core) product family is a digital I&C system platform to cover all these applications. We illustrate the role of verification in the different stages of the software production toolchain, focus on the formal compilation process, and discuss the contribution of the CompCert certified compiler to the safety case of the product. Scrutinizing the object code produced by this compiler has exhibited suboptimal run-time performance in a certain simple but recurring generated code pattern. We explain how formal methods allow us to address this issue in the compiler while simultaneously reducing its trusted computing base (TCB), thereby strengthening the safety case rather than merely preserving it.

Most post-training quantization pipelines fit each weight matrix to its pretrained counterpart, one matrix at a time. Whether that proxy tracks what an attention block actually computes, or how errors in the Q, K and V projections compound inside the softmax, is rarely checked. We write the objective on the attention output instead, over all three projections at once, and reuse it throughout the pipeline. JAB defines one scalar loss over the joint Q, K, V weights of a block, evaluated against the block's real causally-masked attention output, and uses it twice: to fit the quantized weights (GPTQ warm start, then STE with learnable scales), and to score the block for a multiple-choice knapsack allocation. On attention-only quantization of Mistral-7B this works. At 3 bits JAB recovers 77-90% of the gap between uniform GPTQ and full precision, and its sensitivity estimate tracks an oracle costing 73 forward passes to within a fraction of a point. It stops working once MLP layers enter the allocation. A role-aware offset rule needing no sensitivity estimate at all beats JAB on GPT-2's MLP and on the full Mistral-7B model: with a 3-bit floor it quantizes 96.4% of the weights to 4.5 bits per parameter at 6.933 perplexity, within 4.4% of full precision (6.643) at 3.56x compression, against 7.158 for JAB at the same budget. Which matrix a weight sits in matters more than any sensitivity estimate we computed. Two things came out sideways. Block-local reconstruction is an unreliable proxy for end-to-end perplexity: one run improved a block's own objective 4.6x while perplexity rose 32x, which is why every allocation here is validated end-to-end. And on attention-only quantization, fine-tuning moved weights farther from their pretrained values while pulling attention outputs closer, with net gains. Post-training seems to recover attention behavior, not weights.

Reliable decisions depend on recognizing when an answer may be wrong. In biological cognition, metacognitive monitoring can dissociate from task performance, raising the question of how closely solving and judging are linked in language models. Here we study the confidence reports of four frontier models across 15 benchmarks. High task accuracy can coexist with weak error discrimination: a model solves 97% of competition mathematics problems while its answer-time confidence ranks correct answers above errors barely better than chance. Confidence separates correct answers from errors more effectively on questions solved by a separate reference model, while review brings limited improvement on reference-hard questions. Aggregate discrimination also rewards ranking correct answers on easy questions above errors on hard ones, which question-only forecasts already do well. Cross-evaluation helps most where the evaluator answered correctly, and errors shared by the two models usually retain high confidence. Hard questions and shared errors remain difficult targets for prompted self-review and peer oversight, even in models with strong problem-solving performance.

Visual token pruning has been widely studied as a practical approach to reducing the computational cost of large vision-language models. However, it struggles to preserve essential visual information, which can lead to substantial performance degradation. In particular, image-based token selection can overlook task-relevant details, while text-guided token selection may fail to capture the text--visual relationships needed for complex reasoning. We find that applying textual guidance too early can limit its ability to identify answer-relevant visual regions, whereas text-to-visual attention becomes more informative at intermediate decoder depths. This finding motivates our training-free method, which separates early vision-guided pruning from deferred text-guided reselection. We first prune visual tokens using vision-encoder attention, retain additional candidates until the decoder midpoint, and then use text-to-visual attention to determine the final visual-token set. Across eight benchmarks and three models, our method outperforms the best-performing baselines by an average of 11.10 and 16.84 percentage points in performance recovery at 80% and 90% pruning, respectively, with comparable or lower LLM-prefill latency than most baselines. The source code is publicly available at https://github.com/kmc3661/DeFT

Learning-based multi-agent communication under limited bandwidth does not only require deciding what to communicate, but also structuring messages so that partial transmissions remain useful. We study this problem under prefix truncation, where only the first part of each message is received. To address it, we propose \textbf{AH-VIB}, an attention-based autoregressive variational communication model that combines a variational information bottleneck (VIB) with sequential message generation and a hierarchical robustness loss. We evaluate AH-VIB on a custom cooperative object-inspection and occupancy-mapping task, where agents equipped with a limited field-of-view sensor coordinate to scan inspection objects in an occupancy-grid world, under variable and fixed bandwidth conditions, and compare it against MADDPG, CommNet, a flat VIB baseline, and an autoregressive MLP ablation. AH-VIB achieves competitive mean return while improving performance reliability under the most constrained bandwidth conditions. These results indicate that AH-VIB improves the reliability and graceful degradation of learned communication under bandwidth constraints.

Reinforcement learning with verifiable rewards (RLVR) trains reasoning models to produce correct answers, but does not ensure that their stated confidence is calibrated. The resulting models are systematically overconfident. Recent methods train calibration inside the RLVR loop by having the model state a numerical confidence alongside its answer, but they all obtain the confidence by sampling it as text. This choice imposes two costs: a sampled confidence introduces variance and in practice collapses to a handful of distinct values, and sampling makes the confidence non-differentiable, forcing the calibration loss through a scalar reward. We propose CREDO (Confidence REaDOut) to replace sampling with a deterministic readout. While RLVR optimizes correctness, CREDO reads the confidence from a dedicated token pair in the model's output distribution and trains it by differentiable regression. CREDO further turns the trained confidence into a signal for accuracy, weighting rollouts by how far confidence and outcome disagree, so that accuracy and calibration improve together. Across mathematical and code reasoning, CREDO attains the best accuracy and calibration, and the gains extend to abstention and selective prediction.

Multimodal foundation models are increasingly used for evaluating and generating user interfaces (UIs), often producing seemingly reasonable aesthetic judgments and visually plausible pages. However, under professional design scrutiny, their behavior can differ substantially from that of human designers. In professional design practice, designers rely on a systematic set of aesthetic principles that consistently guide judgment, diagnosis, repair, and creation. A coherent aesthetic capability should therefore connect aesthetic judgment with design actions. Existing evaluations, however, typically assess these abilities in isolation, making it difficult to determine whether task-level success reflects a shared aesthetic understanding or merely fragmented task-specific competence. To address this gap, we introduce AUV-Bench, developed in collaboration with professional UI designers around 1,395 executable web interfaces and four tasks: aesthetic scoring, diagnosis, repair, and text-to-UI generation. The tasks share a pool of UIs and aesthetic principles, with diagnosis and repair further aligned on 660 controlled-degradation instances to enable instance-level analysis of judgment and action. Evaluation of 12 models reveals a capability imbalance: models show moderate agreement with professional designers in holistic aesthetic scoring, yet exact diagnosis-chain success peaks at only 24.7%. On the aligned diagnosis-repair cases, correct judgments and successful repairs do not consistently coincide, exposing a Judgment-Action Gap between identifying aesthetic problems and successfully acting on them. In open-ended generation, even leading models achieve only moderate aesthetic quality under human-calibrated evaluation. Overall, current models exhibit partial aesthetic competence, but still lack the fine-grained understanding and judgment-action coherence required for reliable UI design.

Open-vocabulary 3D scene understanding enables object localization and segmentation from free-form text queries without a fixed category vocabulary. Many recent methods build on 3D Gaussian Splatting and consolidate multi-view observations, such as masked crops from individual views, into language features or compact object descriptors before the query is known. However, observations of the same object vary across viewpoints and are not equally informative: some reveal cues relevant to a particular query, whereas others provide incomplete or misleading evidence. Pre-query consolidation can therefore suppress cues on which a later query depends. We introduce EviSplat, which preserves individual observation features as evidence for later text queries. EviSplat retains individual observation features within class-agnostic 3D instances that represent objects, object parts, or background regions. It also learns, for each Gaussian, a distribution describing which visual appearances its observations support. Given a text query, EviSplat scores each instance using its most relevant observations. It then computes a score for each Gaussian by combining instance-level relevance with locally supported evidence, weighted by how often and how unambiguously that Gaussian was observed. Different queries can thus draw on different visual cues from the same preserved evidence. Experiments across diverse datasets and evaluation protocols demonstrate state-of-the-art performance, supporting the benefit of preserving multi-view evidence until query time and aggregating it according to the query.

Deep reinforcement learning (DRL) algorithms for movement control are typically evaluated and benchmarked on sequential decision tasks where imprecise actions may be corrected with later actions, thus allowing high returns with noisy actions. In contrast, we focus on an under-researched class of high-risk, high-precision motion control problems where actions carry irreversible outcomes, driving sharp peaks and ridges to plague the state-action reward landscape. Using computational pool as a representative example of such problems, we propose and evaluate State-Conditioned Shooting (SCOOT), a novel DRL algorithm that builds on advantage-weighted regression (AWR) with three key modifications: 1) Performing policy optimization only using elite samples, allowing the policy to better latch on to the rare high-reward action samples; 2) Utilizing a mixture-of-experts (MoE) policy, to allow switching between reward landscape modes depending on the state; 3) Adding a distance regularization term and a learning curriculum to encourage exploring diverse strategies before adapting to the most advantageous samples. We showcase our features' performance in learning physically-based billiard shots demonstrating high action precision and discovering multiple shot strategies for a given ball configuration.

Soft preference targets specify supervision strength, and reward objectives convert that strength into learned reward signals. A central design question remains: how does assigning a fixed set of preference strengths to different response pairs change the rewards produced by different objectives? We introduce assignment geometry to study this interaction. Mean-matched smoothing controls target dispersion, while within-stratum reassignment changes correspondence and preserves the complete target distribution. Across five reward objectives, intact correspondence retains the largest clean preference margins among the compared soft targets within a common accuracy-equivalence budget. Attenuation orderings change with the reward objective, revealing different responses to the same target assignments. Independent reassignments and a related source construction reproduce the retention direction. An attenuation-retention profile compares these combinations through margin magnitude, edit response, and accuracy. Against independently calibrated scaling, APLOT uniform targets deliver additional attenuation on both aggregate and presentation edits. These findings establish a joint design space in which target placement and reward objective shape reward properties beyond preference accuracy.

Reinforcement learning with verifiable rewards provides a sparse post-training signal: a single binary outcome evaluates the entire rollout, and every token receives the same sequence-level advantage regardless of its individual contribution. To complement this sparse supervision, a growing family of methods adds a scalar-weighted teacher KL term to the policy-gradient objective, providing dense token-level guidance that may be unreliable at some positions. Despite the benefits of combining these signals, their interaction during optimization can destabilize joint training. To understand how this instability develops, we study the learning dynamics of hybrid reward--distillation training through a neural tangent kernel (NTK) analysis. We introduce the cross-signal NTK $K_{DR}(n)$, a token-level statistic that measures the alignment between reward and distillation gradients at position n. Through this analysis, we identify two failure modes: 1 Magnitude drowning, where the reward gradient exceeds the distillation gradient by orders of magnitude, so that even weak directional conflict can cause the distillation loss to rise despite its explicit inclusion in the training objective; and 2 Localized directional conflict, where the sequence-level advantage and the teacher's position-specific distribution induce opposing updates at the same token ($K_{DR}(n)\!<\!0$). The severity of these effects depends on the optimization regime: the gradient-norm ratio $κ\!=\!\|\nabla\mathcal{L}_R\|/\|\nabla\mathcal{L}_D\|$ varies by roughly an order of magnitude across tasks, and our experiments reveal an empirical threshold beyond which naive mixing can lead to persistent training collapse. Motivated by these findings, we introduce the M3 family, which combines magnitude normalization with three strategies...

RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contribution magnitudes, while making both vulnerable to teacher judgment errors and preference variance, as supported by our theoretical analysis. To separate credit direction from its contribution magnitude, we introduce \textit{Decoupled Credit Self-Distillation (DCSD)}, which theoretically decouples credit direction and magnitude into two reliable signals and uses them to calibrate privileged teacher supervision. Specifically, we design belief-margin probing to determine credit direction and marginal information gain to quantify credit magnitude, enabling step-to-token credit assignment for policy optimization. Across 11 benchmarks, DCSD achieves the best overall scores against GRPO, OPSD, RLSD, and RLCSD. Compared with base models, DCSD improves the overall score by 8.45 points on mathematical reasoning and 7.01 points on multimodal reasoning, while correcting the credit direction for 6\% of tokens and yielding a 1.5$\times$ reduction in token credit magnitude.

Time series prediction is a common application of reservoir computing. When the training and testing time series data contains multiple dynamical regimes, because an underlying parameter is changing, or the data in fact consists of multiple distinct systems, simple application of the reservoir computing principle produces high prediction errors. Here, we propose a HyperReservoir as an extended model of reservoir computing especially designed for such cases. The HyperReservoir combines a main reservoir with a smaller context reservoir, where the latter modulates the output weights of the former. This structure resembles the hypernetworks from deep neural network literature. However, in contrast, HyperReservoirs retain the simple training via linear regression of standard reservoir computing. We compare the proposed architecture with a conventional ESN, in which context acts at the input, and a full-matrix Conceptor, in which context modulates the reservoir state space. We evaluate all three models on time-series prediction tasks based on Lorenz and Rössler systems, including for varying bifurcation parameters and time sampling scales. We find that the HyperReservoir achieves the lowest mean test error in all three tasks, and particularly outperforms conceptors on data that is sampled from the same attractor but at different time scales.

Existing multimodal time series foundation models (TSFMs) typically model heterogeneous modalities through largely shared mechanisms, overlooking the distinct forecasting roles of endogenous and exogenous modalities. In this work, we propose QiYao-M, a role-aware multimodal TSFM that models the two types of modalities separately. For endogenous modalities, to capture how they evolve along with the underlying temporal dynamics, we introduce an Endo-Multimodal Predictor and Endo-Multimodal Supervision to explicitly learn their evolution from history to the future. For exogenous modalities, to generalize across domains and across various modality types and numbers under the scarcity of exo-multimodal pretraining data, we propose an Exo-Multimodal Retrieval Enhancer that enables rapid downstream adaptation without updating the TSFM parameters. We further introduce Endo-Modality Proxy Training to train this retrieval module without exogenous multimodal pretraining data. Extensive experiments across unimodal and multimodal benchmarks demonstrate strong forecasting performance in scenarios both with and without exogenous modalities.

A well-designed extraction schema is not necessarily ready for reliable LLM execution. When only limited verified extractions are available, manually tuning hundreds of field definitions through trial and error is costly. We frame this problem as few-instance schema calibration: adapting the operational semantics of an existing schema from a few annotated documents while preserving its structural contract. We introduce CPSE, a contract-preserving semantic extraction framework that jointly calibrates extraction prompts and field-level semantic descriptions from a few gold annotations. CPSE decomposes the schema into an invariant structural contract and mutable field semantics, and further separates identity discovery from record completion using manifest-conditioned resolution. On expert-annotated polymer-science documents, CPSE improves extraction by 9.93 points over an execution-matched baseline, with consistent gains under an independent judge and in a blinded expert audit. These results show that CPSE enables low-resource schema execution while preserving the output structure required downstream.

An agent deployed in a single body cannot learn how fast that body wears, because every trial that would reveal its wear resistance wears the body it would protect. We study this \emph{epoch-one} setting, in which the parameters of a fixed-weight policy are set before the body is drawn and never updated in life. The agent carries a load-gated nociceptive channel and a memory that retains what was felt. We prove that felt cost moves the allocation to the best-\emph{paid} work not yet felt rather than the gentlest, that an agent without retention never sees the felt-cost constraint bind, and that the channel pays only where the threat is individually unpredictable, cheap to avoid and expensive to ignore. We measure per body, setting the agent with channel and memory against the same individual without them, where neither carries a schedule learned across lives. On $2{,}000$ simulated floor-layer knees, with wear anchored to published loss rates, feeling, retaining and substituting extends the working life from age $55.2$ to $59.6$ and raises career output from $33.7$ to $36.1$. $69.3\%$ of bodies gain and \textbf{none lose}. A body that feels but retains nothing past the day gains one of the $+4.4$ years, and retention carries the rest. A population-trained agent gains $+0.65$ years from the same channel at $-0.54$ output. The difference is what a species prior already supplies, and a single body has none. The two are related by an identity, the ablation mean reporting $(1-χ)$ of the per-body value with $χ$ the share a blind schedule already captures, so we report both. Where the regime map predicts value, a care robot sextuples its certified service life and a field-anchored fleet writes off $0.15$ of its machines instead of $0.55$. Where it predicts none, a rover gains little over blind caution, so the map holds in both directions.

Recent advances in bioacoustics have been driven by large-scale corpora and standardized benchmarks, yet existing resources are overwhelmingly bird-centric and shallow per species, limiting their use for studying the structure of a single species' communication system. This gap is particularly acute for cetaceans: despite bottlenose dolphins (Tursiops truncatus) being a compelling case of complex vocal communication among non-human mammals, existing dolphin datasets are small, fragmented, and largely closed. We introduce OpenWhistle, the largest publicly available dataset of dolphin vocalizations. It comprises approximately 180,000 whistles (114 hours) recorded over five years from a stable pod of five individuals in a semi-natural environment, paired with a curated subset of 8,354 expert-annotated whistles and reproducible evaluation protocols for whistle-type detection and classification. We further release the full processing pipeline for whistle detection, segmentation, and categorization. To demonstrate its utility, we pretrain a Wav2Vec2.0 model adapted to dolphin acoustics on the OpenWhistle corpus and show that it learns effective representations, outperforming general-purpose bioacoustic models such as AVES and BioLingual on both tasks while leaving meaningful headroom for future work. By releasing the dataset, pipeline, and evaluation protocol, we provide the first open dolphin whistle dataset tailored for training self-supervised models, laying the groundwork for advancing dolphin communication research and developing models that capture fine-grained acoustic structure within species.

On-policy distillation (OPD) trains student agents through teacher supervision on their own interactions with an environment. However, in asynchronous multi-turn training, arrival-order batching can allow a few early or long rollouts to dominate learner updates while other valid rollouts become stale before being used, wasting already-generated experience. To address this problem, we introduce DivOPD, a simple learner-side batch-selection method that spreads a fixed turn budget across more rollouts and, within each rollout, prioritizes turns with larger cumulative teacher-student disagreement. Turns without usable teacher feedback are excluded. The per-turn loss and optimizer remain fixed; selection only changes which student-visited turns receive training weight. For no-progress rollouts, an optional extension briefly hands control to the teacher before returning it to the student. Across six teacher-student settings on the simulated ALFWorld, ScienceWorld, and WebShop benchmarks, with 1.5B-7B students, DivOPD raises cross-setting mean peak success rate from 77.4 to 84.4 and mean success over the last five evaluations from 71.5 to 78.6. It reaches all reported setting-specific targets with geometric-mean speedups of 1.84x in training tokens and 1.87x in learner GPU time relative to vanilla OPD. Teacher intervention further raises this last-five mean to 82.4 while retaining about 1.7x learner-GPU speedup over vanilla OPD. Code will be released at https://github.com/HanyangWang0418-oss/DivOPD.

Extending reliable nowcasting of extreme precipitation could provide critical additional time for warnings and emergency response during high-impact events such as flash floods. Radar-based generative machine-learning models have enabled skilful hyperlocal precipitation nowcasting, but accurate prediction of intense precipitation remains confined to the first few hours. Because storm-scale structure is predictable for longer than individual cells, a natural strategy is to predict that structure while generatively modelling only the uncertain local growth, decay, reorganisation and initiation of storms. Here we present Microsoft Weather Nowcast (MW-Nowcast), a six-hour ensemble radar nowcasting model that jointly learns a deterministic predictor to capture organised precipitation structure shared across ensemble members, and a generator to produce diverse local residuals around this shared prediction. Across independent test data from the United States, Europe and China, MW-Nowcast achieves higher detection skill than leading methods for heavy and extreme precipitation throughout the 6 h horizon. For the most intense rainfall, MW-Nowcast doubles the available warning time across all three regions, delivering 6 h forecasts with skill previously limited to 3 h for the leading generative baseline. A cost-loss decision analysis shows that MW-Nowcast retains substantial value for a broad range of applications even at 4-6 h, where alternative methods offer little benefit. These additional hours can give forecasters and emergency managers the time to warn and act before extreme rainfall strikes, helping to protect lives and property.

Diffusion drafters accelerate speculative decoding by proposing multiple tokens in parallel. Despite recent advances in speculative decoding through sequence-level drafting and verification, existing training objectives remain largely designed around token-level verification. To address this mismatch, we introduce Block Verification-aware loss (BV loss), a training objective designed to maximize the expected acceptance length of a drafted sequence. BV loss is directly derived from the block verification acceptance rule, providing a principled connection between the drafter training objective and the inference-time verification mechanism at the sequence level. Across math, code, and chat benchmarks, BV loss increases the mean number of tokens accepted per verification call under block verification by 13.0--21.0\% over cross-entropy loss training for DFlash and DSpark with Qwen3-4B and Qwen3-8B without changing the inference procedure. BV loss also outperforms tokenwise acceptance objectives such as TV loss and LK loss, and its gains extend to token verification and greedy decoding. These results demonstrate the benefit of training block diffusion drafters with an objective aligned with sequence-level verification, rather than optimizing each token independently.

Neural Stochastic Differential Equations (Neural SDEs) provide flexible continuous-time generative models, but generic neural drift and diffusion networks are costly to simulate on long horizons and can give unstable gradients when the training signal is a path functional rather than a pointwise observation. We introduce SLiSDE, a family of Neural SDE models built from structured linear stochastic layers. Parallel-in-time simulation is obtained at the layer level, while expressivity is recovered by gated in-flow stacking: previous-layer paths modulate the next layer's latent flow through learned gates. For functional calibration tasks in which rare paths dominate the loss, we add an optional Girsanov tilt that acts as a learned importance sampler with an exact likelihood-ratio correction. We prove well-posedness, a discretisation error bound, validity of the change of measure, and a universality result: the terminal laws of the gated stack are dense in the space of square-integrable laws. Experiments on functional calibration benchmarks show that the structured model outperforms fully neural SDE baselines while retaining parallel-time simulation and stable importance weights.

Most methods that optimize LLM prompts and agent workflows assume that task-specific output schemas, extraction instructions, and evaluation criteria are predefined. For scientific extraction agents, however, a short task goal may not fully determine these components, while specifying them manually is costly. We study the upstream problem of constructing the task-specific configuration from a weak specification containing only a short goal and unannotated reference documents. Rather than treating automatic construction as a fixed preprocessing step, our framework constructs a task-specific schema, extraction instructions, and base training rubrics, then keeps schema construction and extraction instructions editable during optimization. Failure-focused updates concentrate textual-gradient feedback on lower-scoring documents, while training-time evaluation criteria adapt to recurring failures. On a heterogeneous-catalysis literature corpus, automatic construction remains improvable, and optimizing both schema construction and extraction instructions performs best across all four judge-rubric settings, with ablations and blinded human evaluation supporting the proposed formulation.

Large language models are increasingly used to simulate response distributions in social surveys. Prior work has achieved accurate population-level simulation for individual questions. Real questionnaires, however, ask each respondent a sequence of related questions. A simulated respondent should show coherent preferences across the whole questionnaire, not merely accurate distributions for isolated items. Existing single-item methods cannot accurately reproduce how the same person answers a complete survey. We propose FullRespondent-LLM (FR-LLM), which fine-tunes two specialized LLMs: a marginal model for each item's response distribution and a respondent-level autoregressive model for dependencies across answers. Marginal-Constrained Joint Projection (MCJP) then projects the autoregressive joint distribution onto the set satisfying the item-level marginals learned by the first model. This yields complete questionnaires with realistic cross-item relationships while retaining strong item-level accuracy. On two real-world social survey datasets, FR-LLM more accurately reproduces multi-question response patterns, maintains competitive single-item accuracy, and generalizes better to unseen populations and questions. In a small commercial-survey dataset, we use simulated responses to make pricing and stocking decisions; FR-LLM achieves the highest realized profit.

Gaussian neural networks (GaNNs) are proposed as a novel regularization mechanism for neural networks. From a Bayesian perspective standard regularization techniques can be viewed as imposing priors over weight-space. Assuming priors over activation-space remains a largely unexplored possibility. GaNNs assume such priors. They do this by treating activities from earlier layers like signals with Gaussian noise and predicting the properties of the noise distribution using an additional unsupervised loss. While training, the unsupervised loss acts as a penalty on unexpected activities, allowing greater weight updates in less surprising directions. The paper demonstrates the superiority of Gaussian neural networks over standard neural networks on a variety of classification and regression tasks. We also investigate the ability of GaNNs to quantify uncertainty.

AlphaFold2 is written in JAX, so the same inference code compiles and runs unchanged on CPUs, GPUs and Google Cloud TPUs. That portability makes the accelerator look like the main decision a user has to make. We show that it is not. Running one AlphaFold2 inference workload across a Colab CPU runtime, an NVIDIA T4 GPU and a dedicated eight-chip Cloud TPU v5e slice, we find a large hardware advantage for the TPU, 0.47 s per call in steady state on a single chip against 13.1 s on the T4 in the same measurement campaign, and three ways in which the software layer decides how much of it a user actually gets. The default execution path uses one chip of the eight, and at list prices the idle capacity makes the slice cost about as much per prediction as the GPU. Batching with jax.vmap never exceeds single-query throughput, while mapping queries across chips with jax.pmap gives eight chips 6.5-7.9x the throughput of one on a matched grid; automatic sharding leaves the per-chip footprint unchanged, consistent with replication, most plausibly because AlphaFold2 carries no sharding annotations. Our retained trace analysis of a first call at a new input shape reports about three quarters of the traced span in JAX tracing and compilation rather than execution. Reruns five weeks later reproduced neither cloud baseline, the GPU one off by roughly a factor of two, so the hardware ratio above is specific to one campaign.

Human dexterity is guided by two eyes watching two hands: binocular vision supplies the metric 3D structure that fine-grained manipulation consumes. Egocentric stereo is therefore the natural perceptual interface for robots, AR, and VR-yet metric 3D hand reconstruction from this very signal still has neither an end-to-end model nor an in-the-wild benchmark. We propose ESTHER, a model whose stereo geometry, temporal reasoning, and output representation are designed for wearable egocentric stereo. It is trained on pseudo-labels from a calibrated labeling pipeline and in turn assembles our benchmark ESTHER3D, an egocentric stereo hand dataset pairing a large in-the-wild training set of model-generated labels with a motion capture test set of true metric ground truth. Experiments show state-of-the-art accu?racy, superior external generalization, and robustness to the missing views, dropped frames, and lighting and motion blur extremes of real egocentric capture that break existing meth?ods. This robustness runs deeper than graceful degradation: stereo guidance teaches the model to bind apparent hand scale to metric depth, so it not only adapts to different stereo rigs and modalities with minimal fine-tuning, but more strikingly preserves true metric scale even after collapsing to a single monocular view.

We study Nash regret minimization in unknown finite matrix games with bandit payoff feedback and observed opponent actions. We develop Optimistic Payoff Balancing (OPB), which achieves instance-dependent $\mathcal{O}(\log^2 T)$ Nash regret against arbitrary adaptive opponents, including games with nonunique equilibria. This resolves the open problem posed by Maiti et al. (2025), extending their polylogarithmic guarantee under bandit feedback from $2\times2$ games to arbitrary finite dimensions. To handle nonunique equilibria, we construct a reference strategy that leaves room for local adjustments. We order independent payoff differences by estimation accuracy and scale these adjustments by uncertainty, allowing the learner to exploit the opponent's imbalance to offset estimation costs. Our result thus shows that observing opponent actions suffices for polylogarithmic Nash regret in general finite matrix games.

Reinforcement learning has become an effective approach to training language model agents, but sparse and delayed outcome rewards provide limited guidance for credit assignment across long interaction sequences. Recent work on on-policy self-distillation (OPSD) offers complementary supervision by evaluating a policy's sampled responses under privileged training-time context. However, our diagnostics show that positive average agreement between outcome and hindsight feedback coexists with substantial local disagreement, raising the question of how to allocate influence between them at each decision. We introduce UniOPSD (Unified On-Policy Self-Distillation), which unifies these feedback sources through adaptive local credit arbitration. UniOPSD constructs comparable credit estimates from environmental returns and successful-peer hindsight at shared interaction anchors. Historical agreement determines the global mixing level, while current signal availability and relative precision adjust each source's influence at individual decisions. The episode-level outcome contribution is retained, and bounded token modulation refines the fused step credit for policy optimization. With Qwen2.5-3B-Instruct and Qwen2.5-7B-Instruct, UniOPSD achieves ALFWorld success rates of $82.8\%$ and $83.6\%$, WebShop success rates of $75.0\%$ and $82.0\%$, and Search-QA aggregate accuracies of $45.3\%$ and $49.8\%$, respectively. On 3B WebShop, UniOPSD improves over SDAR by $7.0$ percentage points. Our code is available at https://github.com/Zenghuang-Fu/Uniopsd

Vision-language models (VLMs) can answer simple visual questions, but often struggle when one question requires several visual judgments. We study this gap with controlled tasks for feature binding, numerosity, spatial relations, and amodal completion, together with a Composite task that combines them. Matched counterfactual image pairs isolate changes in the visual evidence needed to answer. Across four models, direct answers, hidden-state readouts, and state interventions show that the individual judgments can be made without explicit reasoning and that intervening on the corresponding states can affect the answer. During reasoning, the Composite answer becomes decodable from hidden states and usable from shortened traces, often before the model stops on its own. We train a small detector to predict this readiness and stop reasoning at that point. On MMStar and RealWorldQA, this reduces mean reasoning tokens by 79.1% and 74.5%, while average accuracy rises by 3.13 and 3.30 percentage points, respectively. These findings connect the internal development of answer readiness to a practical rule for allocating reasoning computation.

Reasoning about temporal structure of audio recordings requires Large Audio Language Models (LALMs) to associate sound events with their temporal position. Understanding the underlying mechanisms is a first step toward diagnosing failures and identifying model components that may need improvement. Using mechanistic interpretability, we investigate how temporal information is represented and bound to sound events in three open-source LALMs. We find that across all three, event-specific location becomes concentrated in event name representations at intermediate modality integration layers. These representations encode coarse event position along a low-dimensional, curved relative time trajectory. Steering event name representations along this trajectory systematically shifts before/after beliefs, providing evidence that these representations contribute to coarse temporal reasoning. In contrast, the same interventions do not reliably shift predicted onset timestamps, suggesting that coarse temporal reasoning and precise metric event localization rely on distinct mechanisms.

Tree-structured reinforcement learning trains search agents by comparing alternative continuations and propagating terminal rewards to intermediate decisions. Adaptive expansion, however, creates a statistical asymmetry: an incumbent is selected using its own generation statistic, whereas fresh siblings are sampled after selection. When that statistic is associated with return, branch values can reflect selection history as well as continuation quality, even for a shared parent. We propose Selective-Inference Policy Optimization (\SIPO{}), which incorporates this distinction into tree-based credit estimation. Its scale-free branch criterion keeps generation scores and sibling penalties on a consistent relative scale; exchangeable branching supplies multiple fresh continuations from each selected parent; and order-statistic correction adjusts retained incumbent values using selection rank and the estimated score--outcome association. These mechanisms preserve the leaf budget and the host policy optimisation objective. Across seven QA benchmarks using Qwen3-4B, Qwen3-8B, and Qwen2.5-7B, \SIPO{} achieves the highest reported multi-hop and single-hop averages among the compared methods. On Qwen3-8B, it improves these averages over AT\textsuperscript{2}PO by $1.31$ and $1.07$ percentage points, respectively, and ranks first on six of seven benchmarks. Component ablations evaluate the individual and combined changes, while early-training paired diagnostics show a selected--fresh value gap alongside a near-zero fresh--fresh reference. Together, these results support accounting for selection history when constructing and evaluating search-agent rollouts. Our code is available at https://github.com/Zenghuang-Fu/SIPO

Federated learning enables industrial operators to train shared intrusion detection models without disclosing proprietary operational telemetry. However, existing defenses operate strictly in update space, leaving aggregators blind to data poisoning; model updates derived from fabricated telemetry remain indistinguishable from honest contributions. We repurpose cyber-physical process invariants, such as conservation laws and actuator couplings, from runtime detection heuristics into a verifiable admission requirement for federated updates, mined automatically from clean operational data. We evaluate this admission gate across two physical water testbeds (SWaT, WADI) and a distribution benchmark (BATADAL), testing seven aggregation rules against telemetry fabrication, exposure-only replay poisoning, and an invariant-aware adaptive adversary. Across three testbeds the mined invariants reject none of 100 honest shards and all naively fabricated ones, including optimised perturbations that FoolsGold admits in full. On real telemetry, five mined invariants detect 12 of SWaT's 35 attacks, while nine invariants detect 20, with no honest shard rejected. With nine rules, the physics gate recovers 69--100% of the targeted-attack recall lost to replay poisoning, and 54--100% of that lost to fabricated telemetry, across five standard aggregators. To reconcile physical admission control with federated data privacy, we show invariant compliance using zero-knowledge proofs (zk-SNARKs) to allow clients to prove batch adherence without revealing operational telemetry.

Foundation models are increasingly adapted to individual users, but an apparent personalization gain can simply reflect a stronger population model. This distinction matters for brain-computer interfaces, where every new user must be calibrated. We evaluated personal adaptation of three frozen EEG foundation models (CBraMod, REVE and LaBraM) in 235 held-out subjects from three motor-imagery datasets, comparing each subject's adapter with the population model and with adapters fitted to other subjects. Using all first-half session labels, personal adapters improved mean balanced accuracy over the population model by 1.5-5.4 percentage points and outperformed exchanged adapters by 2.3-7.3 points in all nine model-dataset combinations. The size of this benefit depended on population training: with four times the original budget, median gains remained positive (1.0-2.0 points) but were smaller for every model, and no population model reached a confirmed plateau. Acquiring the benefit cheaply was unreliable: few-label calibration was consistently non-negative on only one dataset, and in CBraMod neither unlabeled context nor meta-learned initialization outperformed matched controls. Personalization should therefore be evaluated against both a population reference and exchanged parameters, across population-training budgets.

The rapid advancement of Large Language Models (LLMs) imposes a thorough evaluation of their linguistic and analytical capabilities as well as constraints, particularly for a language with limited benchmark coverage such as Greek. To address the limited availability of comprehensive benchmarks in this domain, we introduce Prot-Ex and Pan-Ex, two benchmarks consisting of questions from entrance exams for Greek Model and Experimental schools as well as the Panhellenic exams (the Greek national university entrance examinations). These benchmarks are employed to assess the performance of text-only LLMs-including the Greek-adapted KriKri-8B-Instruct, Llama-3.1-8B, Gemma-4-26B, and Qwen-3-32B-across diverse academic disciplines (Modern Greek, Mathematics, Physics, etc.) and task formats (closed, structured, and open-ended), including textualized visual context (i.e., image descriptions). Our findings indicate the localized KriKri-8B significantly outperforms its base model, successfully rivalling much larger LLMs in linguistically demanding humanities tasks. By leveraging an LLM-as-a-Judge methodology, we expose the inadequacy of traditional lexical metrics for evaluating complex reasoning. Crucially, we uncover a few-shot prompting paradox: while synthetic examples improve accuracy in closed-ended questions, they severely overload the context window of 8B models in structured tasks, causing significant performance degradation. Ultimately, this study suggests targeted linguistic adaptation offsets lower parameter counts in specialized domains, despite the fragility of smaller models to prompt verbosity.

Language-conditioned trajectory generation is here, but its evaluation has not kept pace. Existing pedestrian trajectory metrics compare trajectories with real-world human data. This does not scale to text-to-trajectory generation across diverse contexts, as collecting human trajectories for every scenario is costly and infeasible. Moreover, pedestrian behavior is heterogeneous and context-dependent, with no single metric as the correct answer, and current evaluation frameworks are not transferable to this domain. These challenges make scalable, reliable evaluation difficult. We introduce STRIDE, the first framework for evaluating context alignment between scenario descriptions and pedestrian trajectories. STRIDE addresses these challenges through three design choices. First, we derive our VRDST evaluation protocol from sociological theories to define a complete evaluation space. Second, it decomposes high-level context into scenario-adaptive behavioral questions. Third, every question is resolved against a deterministic measurement tool library that yields reproducible answers. Together, STRIDE enables complete, verifiable, automated, and scalable evaluation across diverse contexts without requiring human trajectory data. We instantiate STRIDE in the crowd domain as STRIDE-Bench, comprising 1K scenarios, 6K behavioral questions, and 11K measurements with calibrated expected answers across 30 real-world maps. Comprehensive human validations show that STRIDE-Bench is consistent with human behavior and judgment, achieving 80% human agreement. We further evaluate several text-to-trajectory models, finding limited context-alignment capability and persistent challenges in fine-grained context conditioning. We believe that the STRIDE framework provides a first step toward principled evaluation of context-aligned pedestrian trajectory generation.

Recent medical multimodal models have benefited from larger corpora, broader modality coverage, and stronger reasoning-oriented training, yet effective data design across continued pretraining (CPT) and post-training remains challenging. Medical sources vary substantially in structure, granularity, and information density, and their utility shifts as training progresses from broad knowledge acquisition to late-stage consolidation. Meanwhile, post-training is often dominated by short-form visual question answering, providing limited supervision for informative and answer-consistent explanations. We introduce InfiMed2, a family of 4B and 27B generalist medical multimodal foundation models built around stage-aware data design. We curate a 55.68B-token corpus that combines broad clinical knowledge with context-rich biomedical visual evidence through source-specific processing. Our CPT pipeline first adapts the vision encoder, then builds broad medical knowledge, and finally transitions to an evidence-focused data mixture during learning-rate decay. For supervised fine-tuning (SFT), we regenerate visual question-answering responses using answer stability, answer-masked reconstruction, and correctness-constrained selection to produce more informative and answer-consistent supervision. The 4B model is further optimized with reinforcement learning with verifiable rewards (RLVR). Across five medical multimodal benchmarks, InfiMed2-4B achieves 66.73% mean accuracy after RLVR, surpassing the larger Qwen3.5-9B, while InfiMed2-27B reaches 73.72%, the highest among the evaluated open-weight models.

We study finite-time concentration and convergence rates for projected two-time-scale stochastic approximation driven by a controlled Markov chain. The averaged fast map is contractive, while the slow iterate is projected onto a compact convex polyhedron. The associated projected ordinary differential equation may have a discontinuous vector field at the boundary, preventing a direct application of standard analyses based on Lipschitz vector fields. Using the Skorokhod map, we establish explicit high-probability bounds for tracking the moving fast equilibrium and the projected slow dynamics. These bounds separate martingale fluctuations, Markov-noise residuals, and the bias due to time-scale separation. A Lipschitz Lyapunov function satisfying a uniform decrease condition over fixed time intervals yields almost-sure convergence, with explicit last-iterate rates when the decrease admits a power lower bound. Under uniform Lyapunov contraction, polynomial step sizes yield joint fast-tracking and slow Lyapunov-error exponents arbitrarily close to $1/3$. Under the additional assumption that the reduced slow update map is a Euclidean contraction, logarithmically separated step sizes improve the joint rate to $O(n^{-1/2}\log n)$ almost surely, including for boundary equilibria. The same rate holds under a distinct geometric condition involving a strictly attracting face of a box and a fast equilibrium that is constant on that face. An actor-critic application achieves an almost-sure value-gap rate of $O(n^{-1}\log n)$ relative to the optimum within the constrained policy class. Further applications include projected TD(0) and projected stochastic gradient descent. We also extend the analysis to projection of the fast recursion under Euclidean contractivity.

LLM agents operate in persistent collaborative environments involving multiple users, communities, memories, files, and tools. Community boundaries may remain fixed or evolve with changes in membership, roles, composition, and relationships. Agents must complete legitimate tasks and prevent unauthorized disclosure of protected information. Existing evaluations do not fully examine these risks in agent systems. We introduce \textbf{CoSec}, an executable benchmark for evaluating privacy and authorization enforcement in LLM agent systems operating within and across communities. CoSec contains 208 canonical scenarios spanning fixed and evolving boundaries, protected information belonging to the agent owner or other participants, and attacks through dialogue, environmental content, persistent memory, and composed workflows. CoSec executes complete agent systems with persistent sessions, memory, files and tools. It verifies information flows against the active authorization state using execution traces and artifacts. Across harness and model configurations, agents frequently complete benign tasks but violate privacy and authorization boundaries. Privacy behavior varies across harnesses, attack surfaces, and community states, revealing how memory, files, tools, and workflows can carry protected information beyond its authorized scope. These findings show that task utility does not imply privacy or authorization compliance and that authorization in community settings remains an unresolved security challenge for persistent LLM agents.

Longer histories can improve time-series foundation models (TSFMs), but require substantially higher inference cost. We therefore ask whether contextual information can be provided more efficiently through a compact set of learned token embeddings. We introduce PaCTS, which generates a small set of instance-adaptive latent prompts in the form of continuous embedding tokens conditioned on the visible context. These prompts serve as compact context surrogates for frozen TSFMs. PaCTS constructs them from instance-specific global statistics and further refines them with segment-level temporal information, capturing both global characteristics and local temporal variations. The prompt module is jointly trained and deployed across heterogeneous time series with the frozen backbone. Extensive experiments demonstrate the effectiveness of prompts as context, consistently improving forecasting across context lengths and model architectures. With a shorter input context, PaCTS can outperform the same frozen backbone using double context while requiring substantially less inference computation. Compared with weight-space adaptation methods, PaCTS achieves stronger improvements and better out-of-distribution generalization.

Developing agents for hardware design and verification requires reliable correctness feedback. As a hardware specification may permit correct implementations with different latencies, matching design and reference outputs cycle by cycle can reject valid designs. To address this, we introduce BEHAVE, an agentic framework for multi-turn joint hardware design and verification through functional behavior modeling. We define Behavior IR to express task functionality as executable behavior models without prescribing implementation timing beyond the specification. The agent iteratively develops a register-transfer-level (RTL) design and a behavior model as the design's verification reference. Our evaluator, BEHAVE-Sim, checks both artifacts separately against a hidden golden behavior model using input stimuli generated by random sampling and solver-guided search. BEHAVE thus supports power, performance, and area (PPA) exploration across task-permitted latencies and microarchitectures. During training, the same evaluator provides verifiable reinforcement learning (RL) rewards from specification-behavior pairs without reference RTL. For self-improvement, the agent continually searches for high-level implementations relevant to its capability gaps, constructs and checks specification-behavior pairs, and trains on the expanded task pool. We release BEHAVE-Train and BEHAVE-Eval with 600 human-reviewed specification-behavior pairs for realistic hardware workloads. Starting from 60 seed tasks and acquiring 100 new tasks, self-improvement raises Qwen3.8-27B's RTL pass@1 on BEHAVE-Eval from 55.0% to 75.0%, reaching performance comparable to RL using a 540-task pool.

Large language models (LLMs) have significantly advanced natural language querying over relational databases, yet their ability to query time-series databases (TSDBs) remains largely unassessed. Existing benchmarks fail to adequately capture the non-unified query syntaxes, diverse application domains, and unique time-specific query intents inherent to TSDBs. To address this gap, we introduce TQTS-BENCH, a multi-syntax benchmark for evaluating text-to-query capabilities over TSDBs. TQTS-BENCH contains 6,125 high-quality question-answering (QA) pairs spanning 97 TSDBs, 23 distinct query syntaxes, 22 application domains, and 4 types of time-specific query intents. It is constructed through a human-centric AI-assisted workflow, where all QA pairs are carefully reviewed and revised by domain experts to ensure quality and correctness. Extensive evaluations of advanced LLMs and state-of-the-art text-to-query methods reveal challenges in querying TSDBs. Even the best-performing model evaluated, Claude-Opus-5, achieves only 48.98% execution accuracy, while humans reach 87.34%. Error analysis reveals that this performance gap mainly stems from the heterogeneous query syntaxes across different TSDBs, misinterpretation of time-specific intents, and incorrect schema linking. These findings highlight new opportunities to narrow the gap between current LLM capabilities and the requirements of TSDB queries in real-world applications. The benchmark is available at: https://anonymous.4open.science/r/TQTS-Bench-00CD.

Many physical tasks in human environments require collaboration, from assisting a partner to jointly manipulating an object. Yet, existing humanoid benchmarks largely focus on single-humanoid skills and lack evaluation of multi-humanoid collaboration under egocentric visual observations. We introduce CoHuB (Collaborative Multi-Humanoid Benchmark), a simulation benchmark for multi-humanoid collaboration under egocentric visual observations. CoHuB provides 10 tasks, eight with two humanoids and two with three humanoids, spanning diverse collaboration patterns. We also provide synchronized demonstrations collected through a multi-operator VR teleoperation pipeline, in which each operator controls one humanoid from its egocentric view. Experiments with representative visuomotor policies reveal substantial challenges across different forms of coordinated perception and control. CoHuB provides a foundation for developing and evaluating multi-humanoid collaboration policies.

Vision-language models (VLMs) can read text in natural scenes, but their predictions may be influenced by the surrounding context. When the printed text conflicts with what the scene suggests, a model may return a more plausible word instead of the shown text. We introduce SceneFaith, a benchmark of 781 generated scene images for studying this behavior. Each output is classified as Literal, Canonical, or Other, separating faithful transcription from context-consistent rewriting and ordinary recognition errors. Across 15 models from seven families, all models show rewriting on clear images, with rates ranging from 8.45\% to 58.51\%. Controlled experiments further show that surrounding context matters: removing surrounding scene information reduces rewriting and improves literal accuracy, while changing the scene around the same text patch can also change model outputs. Moreover, weakening the target text with blur increases rewriting. These results show that reliable scene-text recognition requires VLMs to balance visual character evidence with contextual information, preserving clear text while using context mainly when the visual evidence is uncertain.

The use and applicability of artificial intelligence (AI) in medical research and clinical practice has received increasing attention in the literature over recent years. The emergence of large language models (LLMs) has expanded discussions in regards to applications of AI within healthcare. While traditional deep learning based AI applications in medicine have often focused on specific and defined tasks, LLMs offer broader capabilities and flexibility in working with available data,. At the same time of writing, the integration of LLMs into medical settings raises important questions regarding their reliability, accuracy, transparency, safety, and appropriate role in a medical setting. This text presents and discusses recent talks and articles concerning the application of LLMs in medicine, with particular emphasis on their potential utility in research and clinical practice. It considers both the opportunities offered by these technologies and the challenges associated with their implementation, aiming to provide a perspective on the current and emerging role of LLMs within the medical field.

Coverage feedback is an important source of guidance for fuzzing. However, obtaining such feedback normally requires application-level instrumentation that is specific to the language and runtime of the application. Given that modern web applications span multiple languages and runtimes, this application-level instrumentation is costly to implement and maintain. Therefore, we present TraceLib, a system-call feedback mechanism for enabling language-agnostic web fuzzing. Our proposed approach observes the transitions of system calls, enriches selected transitions with bounded argument hashes, and converts them into a 65,536-position AFL-like bitmap. By using the generated bitmap, any web fuzzer can decide whether to retain requests that add previously unseen bitmap positions without consulting application code coverage. We integrate TraceLib into WebFuzz and evaluate it under five WebFuzz feedback modes: the two proposed TraceLib variants (one over every traced system call and one projected onto monitored file paths and recognized SQL buffers), the N-gram comparator adapted from Xiao et al. representing the most recent work to our knowledge, WebFuzz's Native grey-box feedback, and black-box fuzzing without feedback. We evaluate TraceLib on sixteen web applications under test (WUTs): eight PHP applications and eight further applications spanning Node.js, Ruby, Java, Go, and Python to demonstrate platform portability. The results show that our proposed TraceLib projected exceeds black-box on all eight PHP WUTs while exceeding Native on four: Joomla, Drupal, PrestaShop, and Bagisto. In addition, measured on an identical replayed request workload, the tracer costs approximately one millisecond of server-side latency per request. These results indicate that compact system-call feedback is a useful runtime-independent proxy for coverage guidance.

We study retrieval-augmented generation (RAG) for questions about French PDF documents when both the answer and its supporting document pages are evaluated. Five system variants add dense retrieval, rank fusion, reranking, and query decomposition to a BM25 baseline. On 595 challenge questions, the complete system scores 0.4450 MRR@10 and 0.4013 Recall@10, compared with 0.3430 and 0.2994 for BM25. Dense retrieval alone and a simple lexical--dense fusion both underperform BM25. Reranking improves the hybrid system, whereas adding query decomposition produces the largest further gain, with higher latency and more detected output artifacts. The complete system slightly exceeds the reported anonymous overall mean on two answer metrics but falls below it on most page-retrieval metrics. These results identify accurate page selection, rather than semantic retrieval in isolation, as the main opportunity for improvement in this setting.

Multimodal diffusion language models generate responses by iteratively unmasking tokens, making each answer the endpoint of a multi-step trajectory rather than an immediate commitment. Hallucination benchmarks built for autoregressive models evaluate only the final output, and therefore cannot determine whether an unsupported claim in diffusion VLMs appears late or has already stabilized before any answer token is revealed. We introduce DynaHall, a trajectory-level benchmark of annotation-backed binary visual propositions covering object existence, counting, attributes, and relations, with controlled hard negatives graded by visual prior. DynaHall is paired with a commitment-aware protocol that records the intermediate answer tendency at every unmasking step alongside the committed output. Across five diffusion VLMs from three architecture families, visual hallucination is settled before commitment: an unsupported answer is already the preferred state while the answer position is still masked, and later unmasking steps rarely reverse it, so the failure is not introduced at the write step. This holds across decoding schedules, answer formats, and open-ended generation. DynaHall also exposes failures hidden by final-output metrics, including counting and relation collapse, prior-driven false positives, and attribute errors whose direction changes by type. Guided by this diagnosis, PGS (Pre-commitment Gradient Steering) edits still-masked answer states to reduce false positives, bringing the affirmation rate close to balance, and transfers to another architecture without degrading general ability. DynaHall and PGS suggest that hallucination should be measured and mitigated along the generation trajectory of diffusion VLMs, not only at the final answer.

Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks during model interactions. Established behavioral safeguards include alignment methods that optimize model outputs and text monitors that assess interaction text. Representation engineering instead reads or modifies internal model states, but the relative strengths of these approaches remain unclear because they are often evaluated under different settings. We present a matched evaluation across two tracks. For safety control, we compare DPO, a behavioral alignment method, with three representation steering methods across robustness, practicality, and granularity. DPO provides the strongest overall control and generally improves with increasing training data, although its safety can degrade after subsequent benign fine-tuning. Representation steering remains competitive primarily in low-data settings, particularly with high-quality contrastive data. For safety monitoring, we compare representation probes with fine-tuned and open-weight text monitors across full-response detection, early detection, and computational cost. Specialized text monitors achieve the strongest overall detection accuracy, while representation probes remain competitive at substantially lower marginal cost. Finally, monitor-guided interventions recover much of the safety lost by DPO after benign fine-tuning, with little additional over-refusal. Overall, representation engineering does not generally replace behavioral safeguards, but offers practical advantages under specific conditions and can provide complementary safety benefits.

Instruction-following machine translation (IF-MT) requires respecting prompt-level rules on terminology, formatting, and register. Rule compliance typically trades off against translation quality, a tension that general-purpose IF data augmentation methods do not address. We propose Reference-Grounded Data Curation, a two-phase pipeline that extracts every supervised constraint from a reference translation that already satisfies it, ensuring feasibility by construction. Phase 1 applies Instruction-Following Difficulty (IFD) scoring to retain the hardest-but-learnable instances from an English-Thai parallel pool. Phase 2 extracts constraints from each reference target and keeps only generations satisfying every constraint, yielding the 1.97M-record Grounded dataset. We fine-tune open-weight bases on Grounded to produce ChindaMT, a Thai-English translation family at 4B, 2B, and 0.8B parameters. Under length-controlled pairwise judging, ChindaMT outperforms or matches every same-size baseline at every tier on both plain translation and under explicit rules, reaching up to a 68.4% win rate against the strongest baseline. The recipe transfers cleanly across Qwen generations. We release model weights, the Grounded dataset, and evaluation suites.

GUI agents need long-horizon visual reasoning: they must interpret a changing interface while keeping a multi-step plan viable as earlier actions constrain later ones. Existing benchmarks evaluate grounding, computer use, and game play, but rarely test whether agents stay coherent across long chains of coupled decisions. Long-horizon visual puzzles expose this capability directly: a legal move that looks like progress can make the puzzle unsolvable, and the loss shows only several moves later. We introduce LongPuzzleBench, 114 levels in six puzzle games played through native GUI actions, where one objective can take a human over a thousand actions on persistent boards and dead ends go unannounced. With Native GUI Actions alone, the strongest agents solve most objectives, but success falls sharply on harder, longer boards: seven of ten general-purpose agents solve nothing harder than Medium, and none completes Bolt Unscrew Hard, which a human solves along with every other objective. Code Execution CUA does not close this gap, and its scores mix visual solving with algorithmic search. Controlled diagnostics trace these failures to one limitation that neither rules, state hints, nor failure memory removes: agents judge each move by the visible progress it makes, not by the future options it leaves.

Millimeter-wave (mmWave) radar enables privacy-preserving human perception, but the extreme sparsity of point clouds from commercial single-chip sensors (mean ~6.5 points/frame; ~28% empty frames) has confined prior art to body-part keypoints or discrete action classification. We present a cross-modal teacher-student framework that lifts commercial radar to full-body, per-frame, metric 3D mesh reconstruction with per-joint uncertainty. Three innovations: (1) a mesh-foundation-model teacher - SAM 3D Body produces whole-body MHR ground truth (70 joints, 18,439 mesh vertices) from a single RGB frame with zero training, slashing annotation cost by orders of magnitude; (2) StudentPoseFormer - set encoding with masked attention pooling, a temporal Transformer, and a CVAE multi-hypothesis head that outputs both the pose mean and per-joint variance, honestly reporting where the radar cannot see; and (3) a multi-stage ground-truth quality pipeline (confidence gating, depth validation, temporal smoothing, bone-length consistency, bad-frame rejection) plus systematic information-lever ablations. On the public MM-Fi benchmark (same TI IWR6843 sensor, cross-subject), our full configuration reaches 7.45 cm 12-joint MPJPE, with ablations proving the causal value of point accumulation (k = 3, -0.34 cm), Doppler (-0.85 cm; -2 cm at the wrist on fast actions), and velocity loss (-0.27 cm). On our own synchronized radar + RGB-D corpus with block-level held-out splits, the pipeline achieves 21.47 cm end-to-end (per-joint hierarchy from 4.8 cm at the hip to 34.7 cm at the wrist - matching physical information limits), could be improved to 15 cm with ~30k diverse samples, and a scaling law shows sample diversity, not volume, is the binding constraint. Deployment inference is radar-only - no camera, no image.

Post-training quantization (PTQ) enables efficient deployment of large vision-language models (LVLMs), but is typically calibrated on a small set while expected to generalize across diverse downstream tasks. Although recent PTQ methods for LVLMs incorporate sensitivity signals, they still minimize reconstruction loss with respect to the full-precision model, potentially over-preserving FP behavior and calibration-specific bias. Rather than treating quantization solely as an error to be minimized, we observe that it can also provide beneficial regularization for certain layers and modalities. Motivated by this observation, we propose Balanced Fitting, a quantization effect-based framework that balances precision and regularization beyond reconstruction-based optimization. By measuring layer- and component-wise quantization effects for weights, vision activations, and text activations, Balanced Fitting combines fine-grained fitting for sensitive components with coarser fitting to exploit potential regularization benefits. Experiments on multiple LVLMs show that our method consistently outperforms prior PTQ approaches under both weight-only and weight-activation quantization, while lower reconstruction loss does not reliably translate into better downstream performance. The source code is publicly available at https://github.com/kmc3661/BFQ

Multimodal data analysis, which answers questions over relational tables, text, and images, has attracted growing attention in the data management community. Large language models (LLMs) enable such analysis in natural language by generating analysis plans over relational and semantic operators. However, LLM-generated plans are error-prone: a plan may silently compute something other than what was asked, fail during execution, or return a result that misses the question. This paper presents WeaveData, a multimodal data analysis system with self-critiquing and self-evolving LLM plans. First, WeaveData generates a typed logical plan for each question and critiques it step by step before execution, and it checks the executed result against the question afterwards. Second, WeaveData evolves a plan that fails or misses the question: it diagnoses the failure with the actual data, reuses the results that remain valid, and accumulates planning experience for later questions. Third, WeaveData grounds planning in a metadata knowledge graph of all modalities, clarifies ambiguous questions with the user, and backs every model judgment with evidence in an interactive notebook. We demonstrate WeaveData on two public multimodal datasets.

Adapting pretrained models to downstream tasks with limited data has become a central paradigm in modern deep learning. Yet, despite its widespread practical success, how fine-tuning leverages information from pretraining remains poorly understood theoretically. We study fine-tuning from pretrained weights through the lens of sparse linear regression and two-layer diagonal linear networks. In our setting, pretraining provides information through the support (and signs) of the initialization predictor, which may contain coordinates relevant to the downstream task. We show how pretrained information reshapes the implicit bias and training dynamics, and can thereby reduce the sample complexity of recovering the target parameters and support. In particular, for a clean initialization with correctly inherited signs, we show that the required sample size is comparable to that of a weighted Lasso estimator that explicitly exploits the pretrained support through a suitably chosen regularizer. Our results thus show how information encoded in pretrained weights can be implicitly exploited by gradient-based fine-tuning, reducing the amount of data needed to recover a downstream task.

Latent communication passes internal states between language models instead of decoded text, but higher receiver accuracy does not show that the receiver used the message content. Across five method-dataset pairs, replacing each message with one from an unrelated question changes accuracy by at most 0.60 points, even when communication adds 15.44 points over the receiver alone. Thus the interface can supply the gain while making the sharer dispensable. Draft-KV instead sends the key-value states formed while the sharer drafts an answer to the current question. Linear projections place these states in a side memory read through a gated attention branch, and progressive training moves from message reconstruction to answer supervision under a guard on harm from mismatched messages. Both models remain frozen and the interface trains 1.05M parameters, 348x fewer than C2C. With a Qwen3-8B sharer, a frozen Qwen2.5-0.5B-Instruct receiver reaches 78.04% on MMLU-Redux, versus 37.45% alone and 36.40% with reassigned messages. At fixed interface size, scaling the sharer from 0.6B to 8B raises accuracy from 46.11% to 78.04%; communication also transfers to held-out tasks and can exceed both models when each holds different evidence.

Reproduction test generation translates a natural-language issue description into executable tests that fail on the original code and pass after the issue is resolved, providing executable evidence for verifying candidate patches. Existing benchmarks are constructed for individual programming languages, preventing a unified evaluation across diverse programming ecosystems. To address this limitation, we introduce MULTI-SWT-BENCH, a multilingual benchmark for reproduction test generation consisting of 1,963 instances across eight programming languages: Python, Java, TypeScript, JavaScript, Go, Rust, C, and C++. Using this benchmark, we conduct an empirical study of state-of-the-art LLMs with four representative methods (MSWE-agent, MOpenHands, Codex, and Claude Code) and perform a failure analysis across programming languages. Our evaluation reveals a systematic language gap. Across every evaluated method and LLM, the success rate on Python exceeds the aggregate success rate across all languages, while C++ exhibits particularly low success rates. Our failure analysis identifies both language-specific challenges arising from repository testing conventions and cross-language challenges in inferring implicit setup requirements and preserving the target behavior through iterative revisions. These findings demonstrate the importance of multilingual evaluation and provide actionable directions for developing reproduction test generation methods that generalize across software ecosystems and reliably capture issue-specific behavior.

Concept-based explanations describe neural network predictions through human-understandable properties of inputs called concepts. The field encompasses approaches that differ in how they define and represent concepts and connect them to model predictions. We introduce a theoretical framework that describes these approaches in a common mathematical language and supports a shared analysis of their properties. For concept discovery, which identifies concepts automatically within a latent space of a trained model, we employ a concept autoencoder view. An encoder extracts concept representations from the model's latent space, and a decoder uses them to reconstruct the original latent representation. The autoencoder's reconstruction error measures how accurately its decoder recovers the original latent representation. We revisit model completeness: how well the concepts can reproduce the model's outputs. We show that model incompleteness of the concepts can be bounded by the autoencoder's reconstruction error. The autoencoder view also provides a common way to define individual concept attributions, which measure each concept's contribution to a prediction. We establish when these attributions sum to the model's prediction, and bound the discrepancy otherwise, thus providing attribution completeness guarantees.

Real-world driving is inherently multi-agent, yet most existing driving world models generate observations from a single ego vehicle. Independently extending them to multiple vehicles does not ensure that different agents observe a consistent shared world. We present CoDrive, a cross-vehicle, multi-view driving video generation framework that jointly generates observations of vehicles sharing the same dynamic scene with precise camera-trajectory control. CoDrive interleaves local self-attention, which models spatiotemporal dependencies among the views of each vehicle, with global self-attention, which enables information exchange and consistency modeling across vehicles. To explicitly encode their spatial relationships, all camera trajectories are represented in a shared world coordinate system and injected into the attention layers through projective relative positional encoding. We further adopt a progressive mixed-task training strategy that combines large-scale real-world single-agent data with synthetic cross-agent interaction data, allowing the model to benefit from real-world appearance distributions while learning cross-agent consistency from simulation. For systematic evaluation, we introduce CoDrive-Bench, a benchmark covering real and synthetic multi-vehicle scenarios and evaluating trajectory controllability, scene geometry consistency, and instance-level consistency. Experiments show that CoDrive improves trajectory controllability and cross-agent geometric and instance consistency while maintaining competitive visual quality.

Multi-teacher on-policy distillation (MOPD) combines independently developed domain teachers into a single student by distilling their predictions on student-generated samples. We study a setting where teachers share a reference model but undergo different post-training procedures, and find that MOPD can struggle to recover some teacher capabilities. Because distillation occurs on student-generated prefixes, the student initialization can strongly affect subsequent recovery. However, initial benchmark performance is not a reliable predictor of a good MOPD initialization. For example, merge initialization can start below SFT warm-up yet finish higher after MOPD. We further find that effective merging depends on both the relative teacher contributions and the overall merge scale, with some strong configurations lying outside the simplex of convex parameter averaging. Thus, selecting a good merge initialization requires evaluating not only its immediate performance but also the learning it enables under MOPD, making one-shot coefficient search difficult. We propose Iterative Merging for MOPD (IM-MOPD), which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during distillation. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with either uniform merge initialization or SFT warm-up, showing that effective teacher contributions can be determined progressively during training.

Recent studies suggest that large language models encode emotion concepts as structured internal representations, but most existing work focuses on text and a single architecture. Therefore, we ask, do emotion concepts generalize across sources, modalities, and architectures in vision--language models (VLMs)? To address this, we construct CMES (Cross-Modal Emotion Stimuli), a multi-source collection of emotion-conditioned stories, real facial expressions, synthetic portraits, and synthetic emotion-evoking scenes. For each stimulus source, we extract a separate set of six Ekman emotion vectors from each of three VLMs. We report four main findings as follows: 1) Image-derived emotion vectors form a low-dimensional geometry similar to that of text-derived vectors. Valence is relatively stable across sources, while arousal varies more. 2) Text- and image-derived emotion vectors have modest cosine similarity but still show held-out cross-modal correspondence. Text-derived vectors can also steer image interpretation. 3) Cross-architecture correspondence remains even when native cosine is near zero. Transformations estimated from generic ImageNet activations recover both correspondence and causal transfer without using the six emotion vectors or their labels. 4) After aligning representations across architectures, we construct a shared emotion subspace that preserves affective geometry and selective steering effects. The corresponding consensus emotion vectors also generalize to a held-out fourth architecture at two model sizes. These results suggest that emotion representations can share relational structure and causal effects across sources, modalities, and architectures, even when individual vector directions differ.

Solving large-scale linear programs efficiently is an important challenge in many optimization settings. A key technique is column generation, which alternates between solving the master problem over a restricted subset of the variables, and using a pricing subproblem to identify new variables to add. The pricing subproblem is guided by the dual solution of the current restricted master problem, but oscillations in these dual solutions can substantially slow convergence. Dual stabilization methods address this issue. Dual smoothing is a common stabilization method, which guides the pricing subproblem using a combination of the current dual solution and duals from previous iterations. However, while past dual solutions can stabilize the dual trajectory, they do not necessarily guide pricing towards useful new variables. We therefore introduce predictive dual smoothing, which instead combines the current dual solution with a learned prediction of future duals to steer pricing towards variables that are more useful in subsequent iterations. The predictor is trained offline using supervision extracted from standard column generation trajectories and is used only to modify the pricing subproblem's objective function, while exact reduced-cost checks and fallback pricing with the unsmoothed duals preserve correctness. Experiments on cutting stock and generalized assignment problems show that predictive dual smoothing substantially reduces generated columns and wall-clock time relative to standard column generation and existing classical and learned stabilization methods. These gains extend to out-of-distribution instance sizes, and predictive smoothing provides further improvements when combined with strong classical stabilization.

On-policy distillation (OPD) transfers knowledge between language models through teacher supervision on student-generated trajectories. With different tokenizers, a single teacher token may require multiple student tokens to generate, creating intermediate states where the event is entered but not yet completed. Existing cross-tokenizer methods align tokens or text spans to construct comparable prediction targets. We study a complementary problem after partial generation: once the student produces a prefix of a teacher token, multiple next tokens may complete the same remaining bytes, but the teacher only specifies the required completion rather than how probability should be divided among these valid continuations. We introduce Event-Set Completion Distillation (ESCD), which complements cross-tokenizer probability alignment with completion-set supervision. ESCD aggregates prefix-related teacher events and supervises the total probability of byte-compatible one-step student completions, avoiding tokenizer-dependent probability splits among individual tokens. The method reuses student trajectories and predictions, requiring neither additional rollouts nor changes to the student vocabulary. Experiments demonstrate consistent gains in mathematics, code, and scientific reasoning across model families and tokenizers, extending to large-scale MoE distillation from a 1T teacher to a 35B student. Local analyses show that retaining completion sets better matches the reference supervision, while one-step completion covers over 99% of observed compatible teacher mass after partial event entry in the studied tokenizer pairs. These findings support event entry and event completion as complementary supervision targets for cross-tokenizer knowledge transfer. Code will be released on GitHub.

Large language model (LLM) routing aims to select the most suitable model for each incoming query. Most existing routers learn this decision directly from query embeddings, model representations, preference data, or clusters of similar examples. Such approaches can be effective, yet the representation used for routing rarely states what a query actually requires. We introduce SeLMRoute, a routing framework that separates the extraction of candidate-independent semantic evidence from the learning of candidate performance and the application of deployment objectives. A decision model first evaluates a set of interpretable questions about the query, such as its reasoning requirements and use of external knowledge, with each judgment retained as a probability distribution. The resulting probabilistic semantic state is used by a lightweight supervised router to estimate candidate model performance. Routing objectives are applied after performance estimation, which allows the same semantic state to support performance-oriented and cost-aware decisions. On the LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries), SeLMRoute achieves an average accuracy of $72.08\% \pm 0.45$, while grouped five-fold out-of-fold evaluation reaches $72.64\%$, compared with $69.23\%$ for the strongest fixed candidate. The representation achieves the highest mean performance among the evaluated semantic, dense, lexical, and domain-level representations. In a separate 13-model performance-cost setting, SeLMRoute improves performance in all five grouped splits, with a mean PerfGain of $2.66\%$. Our code is available at https://github.com/Indigma-Innovations/SeLMRoute.

Contemporary AI-based music generation can produce compositions that satisfy formal requirements of tonality and musical coherence. However, whether musical expression can be described by mathematical properties alone remains a fundamental question. Human composers operate within personal and cultural contexts that influence harmonic decisions and deliberate departures from established patterns. This study investigates six narrative-driven popular songs by Bob Dylan, Johnny Cash, and Ritchie Valens. Original human harmonies are compared with outputs of an explainable computational harmonizer operating on the same melodies without access to the original chord progressions. We examine harmonic vocabulary, functional persistence, repetition, non-diatonic events, and tension-resolution patterns using Chord Wheel Diagrams and BPMN-based representations. Results show that high melody-chord compatibility does not necessarily imply preservation of the original human harmonic decision pattern. Some generated harmonizations retain the economical structure of the reference, while others alter harmonic diversity or suppress distinctive events while remaining compatible with the melody. Rather than quantifying artistic quality, the study introduces narrative-conditioned harmonic structure as a complementary perspective for computational music analysis. The findings suggest that generative systems may benefit from modeling not only harmonic correctness, but also structural identity, context, and human compositional intention.

We develop an information-theoretic framework for generalization in next-token prediction under temporally dependent data. We consider independent trajectories generated by finite-memory Markov processes and distinguish algorithmic dependence, quantified by mutual information, from temporal dependence, characterized by mixing. For cross-entropy loss, we derive an expected generalization bound using the Donsker--Varadhan variational representation and a McDiarmid-type concentration inequality for Markov chains. A refinement captures the joint effect of context length and temporal mixing through the mixing properties of the history-state process. We then extend the bound through a rate--distortion formulation, replacing mutual information with the minimum information rate required to represent the learned model within a prescribed distortion in the generalization gap, yielding informative guarantees for deterministic algorithms over continuous hypothesis spaces. For margin-based prediction, we derive explicit bounds for linear and self-attention next-token predictors via noisy low-dimensional compression, revealing the roles of context length, model complexity, sample size, margin, and temporal mixing. Experiments on TinyStories and ETTh2 show that longer contexts can reduce both training and test losses, but typically reduce training loss more, enlarging the generalization gap. A complementary ETTh2 analysis identifies an effective predictive-memory scale near 24 hours, with no statistically supported improvement beyond this scale, offering a plausible explanation for test-performance saturation at larger contexts.

Graph Neural Networks (GNNs) are effective for learning node and link embeddings through permutation-equivariant aggregation. However, standard GNNs collapse automorphic nodes, i.e., those with identical structural roles (or orbits) into indistinguishable representations, leading to the node automorphism problem. This collapse limits their expressive power and degrades link prediction performance. Existing approaches to characterize GNN expressiveness rely primarily on Weisfeiler-Lehman (WL) analyses, but these methods are typically qualitative and often misaligned with empirical results. To address this gap, we begin by introducing a novel quantitative framework to assess GNN expressiveness for link prediction. We first formalize edge-level automorphism through edge orbits, which capture the set of structural role pairs for nodes that share a link. Then, we introduce the edge automorphism ratio (EAR), a scalar metric that quantifies a GNN's ability to distinguish links in a given graph. We empirically demonstrate that EAR correlates strongly with performance, validating its practical benefit. Building on this insight, we design EDGE-ORBIT EQUIVARIANT GRAPH NEURAL NETWORK (EO-GNN), a GNN architecture that addresses automorphism collapse while preserving equivariance and incurring minimal computational overhead. EO-GNN accomplishes this through two core designs combined with WL-based node hashes: (i) automorphism-aware dropouts and (ii) subgraph orbit-biased aggregation. Empirical evaluations on synthetic and real graphs show improvements of up to 42.36% and 28.44%, respectively, in predicting links in scenarios with high automorphism.

On-device large language model (LLM) serving is a cornerstone of local-first personal intelligence, offering users data sovereignty, strong privacy guarantees, and freedom from cloud API latency and cost. Although KV caching is widely used to reduce latency in long-context inference, existing designs were primarily optimized for cloud GPUs with dynamic execution environments and abundant memory bandwidth. These architectural assumptions do not hold on mobile NPUs, where computation graphs must be statically compiled and both memory capacity and I/O bandwidth are severely constrained. In this work, we present a compute-storage co-design for mobile-centric prefix and non-prefix KV reuse. We first propose an intra-graph mechanism that maps selective KV recomputation onto static NPU graphs, reconciling algorithmic dynamicity with NPU staticity. We further develop an inter-graph scheduler to optimize chunk merging and minimize padding with dynamic programming. To address mobile bandwidth limitations, we introduce a hierarchical KV manager featuring a tree-hash-semantic hybrid structure, along with cost-aware prefetching and eviction policies. We also build a two-dimensional pipeline that overlaps KV loading, rerotation, and storage with NPU execution, hiding data-movement latency. Experiments across representative on-device workloads and LLMs show that our design reduces time-to-first-token (TTFT) by $40-60\%$ compared with no reuse and prefix-only caching.

Reinforcement learning (RL) changes not only what language models say, but also how much they say, often increasing response length at the cost of token efficiency. Controlling this length growth is particularly challenging in open-ended RL because (i) response length is entangled with quality, (ii) open-ended tasks lack a natural success boundary for deciding when efficiency should be prioritized, and (iii) dense, graded rewards often yield small within-group quality margins, making quality-induced advantages especially sensitive to reward-level length shaping, which can perturb their magnitudes and even reverse their signs. We therefore adopt an asymmetric principle: quality should determine the direction of reinforcement, while length should only shape its magnitude. We instantiate this principle with Quality-Gated Length Advantage Shaping (QGLAS), which first computes advantages from quality rewards alone, then adds bounded bonuses only to shorter positive-advantage responses, leaving all other advantages unchanged. The bonus strength is further adapted to within-group quality separation, allowing conciseness to matter more when quality-favored responses are similar and less when their quality differences are clear. Across different model families, open-ended benchmarks, and reward sources, QGLAS consistently achieves a stronger quality--length trade-off than representative baselines. At approximately 30% compression, QGLAS retains 98.4--102.0% of the macro-average quality gains achieved by quality-only RL over the base model, compared with 68.3--75.5% for these baselines at comparable compression.

Open-weight large language models (LLMs) can generate function-level programs from natural-language prompts, but plausible candidates still fail on hidden semantics and repeat mistakes across repair attempts. We present ReMCTS, an execution-grounded, memory-augmented, LLM-guided MCTS-style search framework. It organizes program candidates as tree states, retains branch-local debugging context, retrieves failure experience across branches, and distinguishes failed checks from unavailable evidence. On HumanEval and MBPP-Sanitized, visible-test ReMCTS improves over direct generation in 8 of 10 model-dataset pairs under held-out evaluation, whereas proxy-only search is less stable. Controlled tree-search, sampling, repair, and memory ablations characterize the source and limits of these gains. A 30-task HumanEval-X C++ pilot further demonstrates compatibility with compiler-backed execution, but does not constitute a broad multilingual evaluation.

Physical trajectories contain more than snapshots of a system: they also reveal how its states evolve under governing conditions. However, representation learning for parametric partial differential equations (PDEs) has largely relied on reconstruction-based objectives that emphasize recovering observed physical fields. In this paper, we investigate predictive representation pretraining as an alternative to reconstruction-based learning. We find that predictive representations preserve rich physical information, yet this advantage alone does not ensure accurate field evolution. Based on these observations, we introduce PDE-JEPA for parametric PDE dynamics. Specifically, we first train an encoder using a masked-latent prediction to capture the underlying regularities of PDE dynamics. To explicitly adapt the pretrained representation toward a more dynamics-aligned state space, we then introduce a geometry projector that aligns latent trajectory geometry with the evolution geometry of physical fields. Finally, building on this geometry-aligned latent space, we further develop a physics-structured latent predictor that decomposes the dynamics into parameter-independent evolution and parameter-dependent response components. Extensive experiments on nine widely used PDE benchmarks demonstrate that our framework outperforms existing state-of-the-art methods by an average of 33.4\% in-distribution, while achieving an average improvement of 51.4\% when extrapolating to unseen governing parameters. The project page is available \href{https://tanpig-x.github.io/PDE-JEPA/}{here}.

Practical deployment of large language model (LLM) agents requires strong task performance at affordable inference cost. For long-horizon agentic tasks, this performance-cost trade-off can be improved through within-task large-small model collaboration, as smaller models can handle some stages even when they cannot solve the full task. In this paper, we introduce RSI-router, a routing framework that constructs subtask-level model assignments and model-specific skills through recursive self-improvement over accumulated experience. Each iteration consists of four stages: Subtask Mining derives subtask definitions and identification rules from training trajectories; Routing Strategy Evolution proposes and evaluates diverse model assignments; Model-Specific Skill Evolution compares routed and large-model-only trajectories to diagnose failures and develop reusable execution skills; and Pareto-Optimal Router Selection updates the Pareto population using historical and newly generated routers while retaining dominated routers as experience for subsequent evolution. Routing between DeepSeek-V4.1-Flash and Qwen3.5-9B, RSI-router consistently surpasses the DeepSeek-only baseline at roughly half the inference cost (48.3%) across five agentic benchmarks. In particular, on ALFWorld, ScienceWorld, and WebShop, it cuts inference cost by 74.7-82.2% while simultaneously improving performance; on Terminal-Bench 2.0, it achieves a 16.7% relative performance gain at 18.0% lower cost. Moreover, RSI-router establishes a stronger performance--cost Pareto frontier than 9 routing methods.

Learning local geometry enables graph neural networks (GNNs) to adapt how they compare and integrate neighborhood information. However, estimating geometry from aggregated representations can overlook variation among individual messages and dependencies across feature dimensions. We propose GeoF, a recurrent framework that jointly evolves node features and propagation geometry through message-passing feedback. Each node maintains a local symmetric positive-definite geometry, initialized from a structure-aware prototype atlas and parameterized in block log-triangular coordinates. At each step, the geometry determines neighborhood weights, while triangular frame transport maps transformed source messages into the target node's local coordinates before aggregation. Weighted second-order statistics of residuals between aligned messages and the transformed target state capture directional variation and within-block dependencies, yielding a geometric update target. A shared controller learns complementary corrections through task supervision. A bounded log-triangular update combines these corrections, the target, and the previous geometric state while preserving positive definiteness. The geometry governs subsequent propagation, closing the feedback loop. With parameters shared across recurrent steps, task-specific readouts support node classification, link prediction, and graph classification. Experiments on benchmark datasets show that GeoF consistently outperforms state-of-the-art GNN baselines.

Long-horizon agent benchmarks typically report how far an agent progresses, but do not identify whether its performance comes from the foundation model, scaffold, responsibility scope, match-control granularity, or horizon. We introduce FromPitch2Board, a deterministic football-management benchmark that studies five configurable factors through controlled comparisons on a single simulator, using paired seeds and a frozen calibration. We evaluate four foundation models and four agent scaffolds. In the Model Track, Coach points Z-scores span 0.19, while Manager points Z-scores span 0.68, with GPT-5.6 showing a sharp rise in passivity under responsibility expansion. Its responsibility ladder rises from 46.1 to 58.1 points with recruitment, then falls to 46.8 under full management, localizing the regression to the final responsibility boundary. Across that boundary, its skipped-decision rate rises from 1.1% to 57.9%. Within the Flash-Pro pair crossed across every scaffold, scaffold choice changes Manager points Z-scores by up to 0.48 relative to the fixed stateless scaffold. The 3Y cohort shows a directional reversal in mean ranking between years one and three, while a selected Claude Code+Pro configuration peaks in year three and remains below that peak, showing that responsibility scope and horizon expose behavior changes that a single headline score conceals.

While Deep Neural Networks (DNNs) have achieved remarkable progress in cutting-edge domains, their inherent brittleness has become a growing concern. To ensure the reliability and safety of DNN-enabled software, DNN testing has emerged as an indispensable practice. Within this context, test input prioritization is essential for early fault detection and reducing labeling costs. However, it remains challenging to accurately identify failure-inducing inputs. Although decision ambiguity and distributional atypicality are two widely adopted perspectives for characterizing inter-class competition and intra-class typicality respectively, relying on either perspective in isolation inevitably introduces blind spots. In this paper, we propose DuFP (Dual perspective Feature space Prioritization), a KNN density-based test input prioritization approach for DNNs that jointly incorporates both inter-class and intra-class perspectives. The prioritization framework of DuFP is built upon class-conditional density estimation. Based on the estimation results, prediction correctness is characterized by an ambiguity score and an atypicality score, with the former reflecting decision ambiguity and the latter quantifying distributional atypicality. A hybrid uncertainty score is then constructed by integrating both scores to guide the final prioritization. We evaluate DuFP on prioritization and selection tasks across image and text datasets under clean, corrupted, and adversarial scenarios. Experimental results demonstrate that DuFP effectively and efficiently prioritizes fault-inducing inputs and outperforms state-of-the-art approaches.

Multi-agent systems (MAS) are increasingly used to automate enterprise workflows involving multiple specialized agents, external tools, and long-running task execution. Failures may arise from tool degradation, context propagation errors, coordination breakdowns, or repeated agent interactions that prevent task completion. While existing observability frameworks provide traces and logs, diagnosis and remediation are largely performed after execution completes, limiting opportunities for recovery during runtime. We present ResonAct, a runtime self-healing framework that enables continuous monitoring, diagnosis, and remediation of multi-agent systems through streaming operational metrics. ResonAct ingests execution traces, agent interactions, and tool invocations into a streaming analytics layer that continuously derives task progress, context health, and tool reliability metrics. These metrics serve as runtime control signals for detecting anomalous execution patterns and localizing root causes using a structured failure model. Based on the diagnosed failure, ResonAct dynamically selects remediation policies and performs actions. The framework operates as an external control plane, enabling intervention without modifying application agents or orchestration logic. We evaluate ResonAct across enterprise workflow scenarios and AppWorld benchmarks. The results show that the streaming metric-based analysis identifies execution degradations and localizes faults. Furthermore, policy-driven remediation improves task completion rates by up to 10.00 percentage points, with detection precision ranging from 70.59% to 82.91%, recall from 63.09% to 100%, recovery rates from 10.48% to 46.67%, and runtime overhead ranging from $-0.25%$ to 14.12% across the evaluated configurations.

We introduce Triangular Resampling (TR), a post-training method for mitigating long-horizon error accumulation in motion diffusion models. Built on FloodDiffusion's triangular denoising schedule, TR addresses the mismatch between ground-truth-derived training windows and model-generated inference states. Replacing only completed motion history leaves this mismatch unresolved in partially denoised states within the active window. TR therefore extends rollout-based training to these states, using ground-truth clamping to limit excessive drift. For each replayed sample, TR draws one denoising threshold, shared across latent positions and replay updates, and replays multi-step triangular denoising without gradient tracking. After each update, states below the threshold are replaced with noise-matched ground truth, while those at or above it retain model predictions. The resulting latent window enters the standard training update. This rollout construction supports both supervised training (TR) and distribution matching (TR-DMD). On 120-second motion generation from HumanML3D test prompts, TR and TR-DMD achieve state-of-the-art FID AUC within their respective non-DMD and DMD comparison groups. Supervised TR reduces FID AUC by 40.9% and FID degradation slope by 55.3% relative to matched post-training without replay.

Reward shaping is fundamental to modern robotic control with deep reinforcement learning (RL), yet practitioners still rely heavily on heuristic principles borrowed from classical optimal control and trajectory optimization. Existing methods rarely distinguish reward terms that are intrinsic to the control objective from numerical regularizers, leading to brittle hyperparameter tuning. To determine which quantities a reward must contain, we study the stabilization control problem with a focus on zeroth-order (configuration) and first-order (velocity) information. We theoretically and empirically demonstrate that policy gradient methods can successfully solve stabilization tasks without first-order reward terms, adding such terms can instead introduce severe sensitivity as their scale grows. Conversely, our findings confirm that reward functions must be zeroth-order complete over goal-relevant coordinates, while the first-order state remains necessary in the policy observation under our low-dissipation assumptions. Overall, these results provide actionable and principled guidance for reward design in robotic RL.

The reliable prediction of passenger train delays is a critical component of railway management. While contemporary research frequently attempts to maximize absolute accuracy by deploying opaque deep learning architectures, the underlying data mechanics driving longitudinal predictive decay remain underexplored. Consequently, this study provides an explainable temporal robustness analysis of network-wide railway delay prediction. Focusing on the Dutch railway network, this research utilizes interpretable tree-based ensembles to integrate granular topological, environmental, and operational features. The overarching finding establishes that while feature-rich tree-based models improve simultaneous (within-month) prediction, predictive performance systematically degrades when evaluated across non-simultaneous (future) months. Furthermore, multi-horizon SHAP and dispersion analyses explicitly link this degradation to environmental feature volatility and instability within the statistical target definition. Ultimately, this thesis demonstrates that richer feature sets alone are insufficient to resolve long-term forecasting constraints, underscoring the necessity to transition toward dynamic, season-aware architectures anchored by absolute operational boundaries.

Automated detection of LLM-generated texts (LGTs) is critical, yet dedicated detectors often struggle to generalize across domains and models. While general-purpose LLMs offer flexible zero-shot authorship classification with explanatory rationale, their detection behavior, especially regarding self-detection versus cross-detection across model generations, remains poorly understood. We systematically evaluate 15 LLMs spanning three model generations as both generators and detectors. Using a benchmark of 1,000 human-written texts and 15,000 LGTs (1,000 per model), we collected over 233,000 binary classifications alongside natural-language explanations. Our results reveal that detection efficacy is primarily driven by detector capability rather than generator provenance, although outputs from newer generators remain notably harder to detect. Crucially, statistical comparisons show no systematic advantage or disadvantage for self-detection across models. Error analysis further exposes generational bias shifts: first-generation detectors under-detect LGTs (high false-negative rates), second-generation detectors over-flag human texts (high false-positive rates), and the latest models achieve balanced trade-offs. Finally, we highlight significant inconsistencies in how different LLMs apply textual cues to justify their decisions. Code: https://github.com/hyyuan/detect-llm-generated-texts.

Spatial reasoning is fundamental to embodied agents, yet it remains unclear whether spatial understanding can be carried forward to guide sequential interactions. Existing spatial-reasoning benchmarks typically terminate at offline predictions, while navigation benchmarks evaluate spatial reasoning as part of instruction following and exploration. We introduce VCN-Bench, a \textbf{V}ideo-\textbf{C}ontextualized \textbf{N}avigation benchmark for probing closed-loop spatial reasoning over prior visual experience in MLLMs. Given a prior video covering both the initial location and destination, the agent is tasked with reasoning out the instruction-specified target and navigating toward it with the inferred spatial context. Built on Matterport3D, VCN-Bench contains five instruction types, 100k training episodes, and 1,250 evaluation episodes. Navigation serves as the primary evaluation, while diagnostic goal identification helps distinguish destination-resolution errors from subsequent navigation failures. We further propose MV-DualVLN, a planning-oriented baseline that jointly leverages prior video and in-episode observations. Experiments reveal limited navigation performance, a substantial destination-resolution-to-navigation gap, and frequent navigation failures even after correct destination identification.

As large language models increasingly operate as tool-using agents, post-jailbreak safety feedback is often assumed to serve as a reliable safeguard; however, how lingering jailbreak context shapes subsequent agent behavior remains largely unexplored. To systematically examine this dynamic, we introduce a paired continuation framework across 192 parent tasks spanning 42 domains, evaluating 12,148 analyzed continuation pairs (curated from a 12,288-pair initially design) across eight diverse agents. We find that identical safety feedback induces sharply model-dependent behavioral routing rather than uniform protection: redirecting unsafe trajectories toward legitimate completion (\emph{rescue}), sustaining unauthorized execution (\emph{persistent unsafe}), or triggering over-refusal on benign tasks (\emph{collateral loss}). Through layer-wise activation patching, we discover a shared \emph{late-commit pattern} where causal intervention effects surge sharply near the final layers (relative depths of 0.958--0.984) despite an over 30-fold variation in peak magnitude across architectures. Crucially, critical-layer representations correlate with macroscopic routing outcomes, and intervening at these layers causally alters concrete next-step tool actions. Building on this causal foundation, we test whether localized intervention-derived features can serve as predictive proxies for full-trajectory routing outcomes on unseen parent tasks under leave-one-parent-task-out evaluation, finding that they provide viable predictive signals in responsive agents with peak ROC AUCs reaching 0.675 for \emph{rescue}, 0.777 for \emph{collateral loss}, and 0.702 for \emph{persistent unsafe}. These findings establish a mechanistic lens and a predictive baseline for anticipating the safety and utility trade-offs of post-jailbreak feedback in autonomous agents.

Accurately decoding object states from the internal representations of vision-language-action (VLA) models does not establish that the predictions respond faithfully to changes in the target physical state. In natural observations, object state, robot configuration, occlusion, and task progress vary together, allowing contextual cues to contribute to prediction. In this paper, we introduce an evaluation framework that separates prediction accuracy, target-state responsiveness, and context stability using physically validated observations that cross target coordinates with robot contexts. We demonstrate that high natural-trajectory accuracy can coexist with weak controlled target-state responsiveness in fixed representation-readout pairs. Comparisons and interventions involving representations, readouts, and training data show that the three properties provide distinct diagnostic information. Furthermore, adding responsiveness and context sensitivity to a failure predictor based on initial state error and physical variables reduces policy-failure prediction error on new initializations relative to the specified baseline while same-observation controlled MAE is also informative. These findings motivate evaluating target-state responsiveness and context stability alongside natural prediction accuracy, and examining their relationship to actual policy behavior and task outcomes.

The optimization of LLM serving engines, such as vLLM and SGLang, is largely benchmark-driven: optimizations, scheduling policies, hardware and system designs are all selected based on representative workloads. However, a significant mismatch has emerged in the agentic era. Existing benchmarks primarily focus on simple single-turn chatbot workloads. LLM applications are increasingly agentic: coding agents, terminal execution systems, and tool-use agents issue multi-turn requests with growing context lengths. We introduce AgentPerfBench, a benchmark suite for agentic inference. It uses real traces from agentic benchmarks, such as SWE-Bench and TerminalBench, alongside standard chat baselines. This enables benchmarking of models on multi-turn tasks involving tool calling, skill utilization, and increasing context lengths. AgentPerfBench also samples from empirical distributions of input length, output length, and turn count derived from the real traces, generating representative synthetic profiles for cheap and accurate measurements on new hardware. In addition, we further find that several existing benchmarks fail to accurately reflect real hardware performance for two key reasons: 1) they do not account for realistic context-length growth, and 2) they measure inference performance without operating at hardware saturation. We discuss these issues in detail and provide rich kernel-level Nsight Compute (NCU) traces to construct a new multi-dimensional roofline model that captures hardware-system limitations in both memory bandwidth and memory capacity footprint. The benchmarking suite then includes automated scripts to identify potential bottleneck conditions on emerging hardware when evaluated with diverse agentic traces. Together, these contributions quantify the chat-to-agentic gap in current inference benchmarks and characterise per-kernel GPU resource utilisation via roofline analysis.

Learning-rate (LR) scheduling plays a central role in large language model (LLM) pretraining, yet current practice still relies heavily on hand-crafted heuristics such as Warmup-Cosine-Decay and Warmup-Stable-Decay. Because these schedules are fixed in advance, they cannot adapt to evolving optimization dynamics. Online learned scheduling within the Learning to Optimize (L2O) framework offers a dynamic alternative, but remains brittle at LLM scale due to noisy signals, delayed feedback, and the risk of catastrophic divergence. We propose SOLAR (State-driven Online Learning rAte scheduleR), a stabilized framework for reliable online LR adaptation. SOLAR uses a base schedule as a reference and learns bounded, state-dependent residual corrections for individual parameter groups. Each correction re-anchors to the base at every step, allowing the policy to adapt the LR without relearning the warmup-decay profile. A lightweight state representation and progress-aware reward guide online learning, while a Circuit-Breaker restores training after rare unsafe actions. Across autoregressive language-model pretraining, SOLAR improves final perplexity over tuned static schedules and automatic LR tuners for dense models from 60M to 1B, AdamW and Muon, and two MoE settings up to 3B. Matched 130M controls show that adding base anchoring and action bounds improves a global PPO controller from 27.09 to 23.74 final PPL, while group-wise control reaches 22.87 on the same two seeds. A residual policy trained on a 60M proxy can also be frozen and reused at larger dense scales without target PPO updates, remaining effective across a fourfold base-LR range. These results establish SOLAR as a practical learned LR controller for LLM pretraining.

Four-bit post-training quantization can reduce the memory demands of large language models, but preserving accuracy under strict MXFP4 W4A4 requires coordinating several design choices. Coordinate transforms change block-encoding errors, which in turn affect the residuals propagated through the network. The useful algorithmic decomposition is therefore not fully known before search. LLM-driven program evolution offers a way to explore these choices, but performance scores alone do not explain which design should change next. We introduce QuantForge, a PTQ discovery system that records competing explanations, selects controls that distinguish them, and checks that successor code implements the resulting conclusions. This residual compilation guides program revisions while retaining useful programs even when their original explanations are rejected. Remeasuring the revised program reveals the next error to address. This process discovers HiRes, a fixed MXFP4 quantizer that shapes coordinates, refines legal code assignments, and recovers errors along attention and MLP paths. Each stage acts on residuals measured after the preceding stage has executed. Across seven tasks, HiRes achieves the lowest seven-model Robust Fit (0.09300) and the lowest quantized Fit-7 at 32B. In matched-budget comparisons of LLM-driven program evolution, each with 240 evaluator calls, QuantForge reaches a held-out transfer target in six of eight runs, compared with three each for textual memory and reflection memory, and one for score-only evolution, despite evaluating fewer new programs. These results show that QuantForge improves the discovery of transferable PTQ algorithms by turning controlled evidence into subsequent program changes.

Bipartite matching is a fundamental problem in game theory and market design. Classical approaches such as Gale--Shapley assume complete preferences and centralized computation, whereas many real-world matching processes are decentralized, asynchronous, and shaped by sequential interaction under limited information. We propose a dynamic bipartite matching framework that combines large language model (LLM) agents with contextual bandits. In a simulated Chinese marriage market, economically grounded LLM agents evaluate locally encountered candidates, while agent-specific Logistic-UCB models learn reciprocal acceptance from realized proposal outcomes. The mechanism therefore separates two decisions---\emph{whom do I like?} and \emph{who is likely to like me back?}---without requiring ex ante market-wide preference rankings. We first validate LLM-induced mate preferences against the empirical conditional-logit reference across multiple LLM backbones. In the $50\times50$ matching experiment, Bandit-UCB achieves the highest mean mutual welfare (56.01 versus 54.87 for Gale--Shapley), a smaller gender rank gap than the classical baselines, and the fewest blocking pairs among the LLM-ABM policies. Learned acceptance models show economically interpretable gender-differentiated associations, while counterfactual setups reveal no systematic unilateral advantage from prior search knowledge. Overall, these results support the advantages of decentralized matching with LLM-based behavioral modeling and online learning under incomplete information for economic simulation and computational social science research.

Misinformation on social media remains a critical problem, and more and more people settle it by asking a language model instead of a fact checker. Whether models judge such claims reliably is debated; whether they judge them equally well in every language people ask in has gone almost unasked. We test eight models from five families, 3B to 70B, on 1,500 encyclopedic factual claims that exist in identical form in eight languages. English is judged better than every other language on every model, and the gap is widest on the smallest ones, where Llama-3B on Arabic is no better than guessing. Existing remedies retrain on more multilingual data or fit an unconstrained map between language representations, and neither asks whether the model already holds the answer and simply fails to say it. It largely does: a linear probe recovers the truth from the very activations the model fails to express. We propose RoSh, a per-language shift and rotation of the residual stream, computed in closed form at three layers, with no training and no weight modified. It improves every model and closes 75% of the gap on average, helping most where the model was worst: Arabic on Llama-3B goes from chance to nearly the English level, and a fifth fewer of the claims answered correctly in English are lost in translation. What remains is no longer a read-out failure: afterwards the head recovers as much of what is encoded outside English as it does in English. An unconstrained map fitted on the same pairs falls below the untouched baseline, so the orthogonality constraint is doing the work, and every model clears a scrambled-correspondence control and ten further controls. On the two benchmarks of the closest inference-time method, latent-space intervention, run with its own data and metric code, RoSh's gains are five to thirteen times larger.

World models predict future observations from current experience and actions, yet prediction can depend on observations seen far in the past. Episodic memory preserves past observations for later recall; however, as memory accumulates, it raises a fundamental question: which memories are useful for the current prediction, and which available retrieval cues should be trusted to find them? This is challenging because fixed criteria based on recency, pose overlap, or visual similarity can be unreliable across environments and queries. We propose Future-Aware Recall (FAR), a framework that learns episodic recall from future-aware predictive supervision and adaptive multi-cue scoring. During training, FAR measures predictive utility by the conditional log-likelihood of the realized future given recalled context, approximated by negative diffusion prediction loss, and uses it to train a retriever that remains future-blind at inference. The retriever learns cue-specific relevance and automatically determines which available retrieval cues, such as time, pose, vision, and audio, to trust for each query when selecting memories. Across three complementary settings, FAR outperforms hand-designed recall even with the same retrieval cues, automatically adapts which available cues to trust, and recalls the right history as the world changes. Together, these results establish FAR as a flexible, principled approach to episodic memory access in world models.

Few-step generative models can generate high-fidelity samples within a few function evaluations. Despite this efficiency, generated samples may not exhibit desirable properties. When these properties are difficult to encode as an explicit reward function, direct preference optimization (DPO) can align generative models using pairwise preference feedback without training a separate reward model. However, extending DPO to few-step generative models is challenging because few-step generative models are generally implicit, making the likelihood evaluation required by DPO intractable. To address this challenge, we introduce Few-step DPO (FestDPO), an extension of DPO for few-step generative models that leverages nonparametric likelihood estimation from empirical samples. By exploiting the fast sampling capabilities of few-step generative models, our approach makes sample-based approximation of DPO loss computationally feasible. Furthermore, the sample-based formulation makes FestDPO agnostic to the model family and sampling procedure. Our toy experiment demonstrates that FestDPO matches the reward-tilted target distribution across four few-step generators. For real-world tasks, we evaluate FestDPO in two domains: text-to-image generation and protein backbone generation. In text-to-image generation, FestDPO outperforms preference optimization baselines in both win rates against the base models and human evaluation scores. In protein backbone generation, it achieves a higher $β$-sheet fraction and better structural designability than the baselines.

Fine-grained visual understanding requires models to recognize small details within complex images. Multimodal on-policy self-distillation (OPSD) addresses this challenge by using a teacher conditioned on evidence-centered crops to supervise a student conditioned on original images along student-generated trajectories. Ideally, teacher corrections, the distributional changes from the student toward the privileged teacher, should be driven by task-relevant visual evidence. However, the designs that make the teacher effective also introduce other interference. Using a lagged or frozen teacher improves training stability but introduces a model-state gap from the evolving student, while cropping enhances task-relevant evidence but also loses the visual context. These two sources of interference make the teacher corrections not purely rely on the visual evidence. We introduce Evidence-Aligned multimodal on-policy self-Distillation (EAD), which retains the crop-conditioned teacher as the target but constructs a separate evidence reference for weighting the corrections. To exclude the effect of lagged model-state from this reference, EAD measures prediction changes using the current student. To avoid crop-induced context changes, EAD masks the evidence region in the original image while preserving the other visual context. The change from the student's masked-image prediction to its original-image prediction provides a controlled reference for the direction in which the visual evidence shifts the student's prediction. EAD weights each teacher correction by its cosine alignment with the reference, i.e., retaining aligned corrections and downweighting the rest. Retaining only 6\% of the supervision mass of dense OPSD, EAD consistently outperforms previous state-of-the-art methods.

In this paper, we propose a representation of a digraph (directed graph) as a Hermitian matrix derived from its adjacency matrix. This representation characterizes both the connectivity and the edge orientation of the digraph. Based on the spectral decomposition of the Hermitian matrix, a digraph clustering algorithm with $k$-means is introduced to produce a partition on the graph. Applying this algorithm (bottom-up) recursively to a digraph with partially labeled vertices yields a spectral hierarchical digraph clustering (\myproj) algorithm that produces consistent nested partitions of the digraph, or equivalently, a tree structure. Furthermore, based on the in-degree and out-degree of each cluster in the digraph clustering, a pair of hierarchical interval partitions (filtrations) can be derived in a top-down manner to produce a pair of nested knot sequences. These knot sequences facilitate the construction of multilevel spline quasi-interpolants, enabling a noisy graph signal to be decomposed into a coarse approximation and inter-level details, followed by adaptive thresholding and reconstruction. Experiments on synthetic and real-world digraphs demonstrate the superiority of our {\myproj} algorithm for digraph clustering across diverse graph structural properties (homophily and heterophily) and supervision settings. Moreover, experiments on digraph signal processing using multilevel spline quasi-interpolants further demonstrate the effectiveness of signal recovery on digraphs in terms of RMSE and SNR.

Fine-tuning pre-trained models on specialized tasks with scarce data is central to modern deep learning. Despite its empirical success, theoretical understanding of fine-tuning remains limited. We introduce a Gaussian multi-index setting to study fine-tuning from pre-trained weights, where the teacher network has $m+1$ features, $m$ of which are learned during pre-training and one of which must be learned during fine-tuning. For two-layer ReLU networks, we show that two-timescale training, i.e., updating the outer weights infinitely faster than the hidden ones, learns the new task-specific feature while preserving the pre-trained ones in the model representation. Moreover, only $\mathcal{O}(d)$ fine-tuning samples are required for this recovery, independently of the number of pre-trained features. In contrast, with random initialization, the same number of samples is insufficient to recover the target parameters. Our results therefore demonstrate that pre-training can induce an implicit bias with a clear statistical advantage over random initialization, enabling feature learning from scarce fine-tuning data.

Differentiable physics is increasingly used in robotic material manipulation for system identification, trajectory or skill optimization, demonstration generation, and robot or end-effector design. These applications depend on gradients propagated through long, contact-rich simulation rollouts. We study the numerical reliability of those gradients using two Material Point Method (MPM) system-identification benchmarks derived from elastoplastic and granular manipulation. The benchmarks provide controlled cases for three effects that also arise in broader differentiable physics-based optimization. GPU many-to-one sums whose order depends on thread scheduling changed long-horizon gradients and reversed the sign of one parameter gradient relative to a deterministic reference. Finite-difference checks became less reliable for longer rollouts because repeated-run loss variation grew much faster than the loss change produced by the tested parameter perturbations. Observation and loss definitions changed optimization behaviour and the solution preferred by an independent metric. These results motivate reproducible accumulation, finite-difference validation that compares perturbation-induced loss changes with repeated-run variation, and explicit reporting of objective construction when differentiable simulation is used for robotic optimization.

Existing program-reasoning benchmarks ask large language models to predict a program's behavior on a given input. Coding agents break two assumptions on which these benchmarks rest: an agent can recover the answer by executing the program instead of reasoning about it, and fixed task sets drawn from existing programs are increasingly exposed to contamination, yet costly to renew. We introduce Codoku (code sudoku), a renewable benchmark in which a solver fills typed cells in a partial program to satisfy global static and dynamic constraints, such as a prescribed control-flow graph and execution path. Because a partial program cannot be executed and valid fillings are sparse in an exponentially large space of interdependent choices, neither tool use nor enumeration can substitute for program reasoning. Puzzles are synthesized from scratch via semantic reification, so fresh puzzles of controllable complexity can be generated on demand, each with a witness that guarantees solvability. We evaluate five frontier models on 300 puzzles through a coding agent free to use any tool within a fixed budget. Small puzzles already challenge open-weight models, whereas even proprietary models solve only about half of the large ones. Codoku thus offers a renewable testbed for program reasoning that can keep pace with rapidly improving coding agents. GitHub: https://github.com/connglli/Codoku.

Logical reasoning remains a major challenge for large language models (LLMs), particularly on structured problems that require precise constraint tracking, consistency preservation, and multi-step deduction. This challenge is especially acute for small-scale LLMs, which are more prone to producing inconsistent, redundant, or brittle reasoning trajectories. Existing approaches for improving logical reasoning largely optimize for final-answer correctness, providing only weak supervision over the intermediate reasoning process. In this work, we propose SPRING: (Solver-guided Process Rewards for Novel LogIcal ReasoNing Step Generation). SPRING uses SMT solver as a training-time verifier of intermediate reasoning steps to provide process-level supervision. It introduces the notion of a novel reasoning step, namely, a step that is logically valid, consistent with the evolving reasoning state, and not already implied by previously accepted non-contradictory deductions. Based on this solver-based assessment, it designs process rewards that encourage novel inferential progress while penalizing contradictory and uninformative reasoning steps. Evaluation across three logical reasoning benchmarks, ZebraLogic, AR-LSAT, and Knights and Knaves, and four LLMs shows that SPRING consistently outperforms base LLMs, outcome-only reward baselines, and Logic-LM. On ZebraLogic, SPRING improves puzzle accuracy by up to 49.71 and 15.43 points over the base LLM and strongest outcome-only baseline, respectively. On AR-LSAT, it improves overall accuracy by up to 64.93 and 12.14 points, respectively. On Knights and Knaves, SPRING achieves up to 93.14 puzzle accuracy and 96.05 person accuracy.

We study last-iterate convergence in unknown two-player zero-sum matrix games with bandit payoff feedback and observed opponent actions. For games with $d$ actions per player, we develop an algorithm achieving a duality gap of $\widetilde{\mathcal{O}}(\sqrt{d/t})$ with high probability, simultaneously at every round $t$. This improves the dimension dependence of the best previously known guarantee by a factor of $d^{3/2}$. The rate matches a standard bandit lower bound, establishing minimax optimality in both the number of actions and the number of rounds, up to logarithmic factors. The algorithm is computationally efficient, requiring only $\mathcal{O}(d)$ time and memory per round. Our technical contribution is a joint design of adaptive averaging and corrected exponential weights that absorbs estimation variance, together with a potential argument that bounds phase durations.

Accurate fitness prediction is central to protein engineering and understanding sequence-function relationships. With advances in deep learning, protein foundation models (PFMs) have become widely used for this task. Recent analyses, however, show that these models share preferences reflecting their training corpora, while unreliable inputs can further distort fitness predictions. Family-specific evolutionary evidence and structural context can help address these limitations by providing complementary constraints on model scores, motivating VenusREM-Harness (VRH), a general, model-agnostic, training-free Retrieval-Enhanced Mutation harness. It fuses frozen model scores with multiple sequence alignment (MSA) evidence according to model uncertainty, then applies gated background correction and score shrinkage based on structural confidence and solvent exposure. Across 1,211 assays and 3.1 million measured variants from ProteinGym, VenusMutHub, and the newly curated viral benchmark VenusViroHub, all 71 configurations improve Spearman correlation on all 3 benchmarks by 0.073 on average, with broad gains across 5 metrics. Extended analyses relate retrieval gains to model-MSA preference differences, assess domain-level gains and immune-escape cases, and quantify computational speedups. Built with VRH, VenusREM2 is the first to rank highest in all function, taxon, MSA-depth, and mutation-depth categories, with a ProteinGym Average Spearman of 0.556, 0.038 above the prior best.

On-device streaming omni-modal inference safeguards user privacy and eliminates prohibitive per-token API costs, but faces a critical bottleneck: the continuous influx of multimodal data rapidly exhausts constrained memory and compute budgets via monotonic KV cache growth. Existing sparse attention methods fall short, either incurring prohibitive online estimation latency or destroying interleaved cross-modal context, while failing to resolve physical memory fragmentation. We present OmniTide, the first algorithm-system co-design tailored for efficient on-device streaming omni-modal inference. Driven by the observation of modality-aware structural sparsity, OmniTide adopts a unit-based abstraction with two components: (1) At the algorithm level, OmniPick logically retains critical multimodal context based on unit boundaries and modality importance to preserve task accuracy; (2) At the system level, OmniPage physically partitions the cache by retention likelihood and dynamically compacts surviving sparse tokens, minimizing both memory fragmentation and data-movement overhead. Extensive evaluations across three streaming benchmarks and two consumer-device architectures show that OmniTide achieves up to $12.72\times$ kernel speedups and $2.40\times$ lower stream-loop latency. On StreamingBench, it improves accuracy by up to 18.0 percentage points over sliding-window baselines at comparable session cost. OmniPage further reduces the physical KV span by up to 26.7% relative to native logical eviction, unlocking real-time, infinite-context streaming on edge devices.

Some attention heads learn similar patterns across inputs. Reusing these patterns could reduce training cost by avoiding repeated query-key score computation and softmax. Through controlled pretraining comparisons, we identify Selective Attention Freezing (SAF), which selects heads with low attention-pattern variance and replaces their attention weights with fitted post-softmax means halfway through training. We represent these fixed patterns with absolute-position and relative-distance preferences, reducing storage from quadratic to linear in sequence length. A fused kernel reconstructs the patterns and executes ordinary-attention and replaced heads together. At matched training-token budgets, replacing 25% of attention heads gives 1.056x faster post-replacement optimiser updates at 124M parameters and 4K context, with a 0.77% perplexity increase. At 1B and 8K context, post-replacement updates are 1.068x faster on four GPUs including communication, with a 0.51% perplexity increase. The resulting models also accelerate long-input finetuning and causal prefill. After associative-recall adaptation, the 124M model with 25% replacement generalises to more key-value pairs at a fixed length better than ordinary attention and two pruning controls.

Harness evolution improves LLM agents by learning from execution trajectories, but existing experience- and skill-based methods are less effective on long-horizon tasks. As interactions grow, useful evidence can be buried by redundant or outdated context, making context management itself a key bottleneck. We introduce ContextEvo, a framework that learns a context policy from long-horizon trajectories. ContextEvo reconstructs the model-visible context at key decision points, identifies context-related failures, and applies targeted policy updates. Starting from the open-source Pi-agent harness, ContextEvo improves performance across three long-horizon task benchmarks, achieving results comparable to or better than several prominent agent harnesses, including Codex, OpenCode, and OpenClaw. Additional analyses show that fixed or locally evolved context strategies can fall short under long-horizon information pressure, while our methods adapt to the information demands of each environment.

Existing training-based speech emotion editing methods often require substantial task-specific training and can be unstable. This motivates us to investigate whether the pretrained generative dynamics of large-scale text-to-speech (TTS) models can be directly manipulated for training-free emotion editing. To answer this question, we probe the editability of pretrained flow-matching and hybrid TTS models by constructing a controlled test set and systematically diagnosing editing effects along the generative trajectory. Our analysis reveals that pretrained TTS models are substantially editable in emotion, but such editability is architecture- and trajectory-dependent and can be disrupted by early flow-matching steps, while cross-speaker emotion transport carries additional acoustic attributes beyond emotion. To address these limitations, we propose SEmoEdit, the first training-free framework that formulates emotion editing as dynamic velocity transport between source and target emotions, enabling robust, flow-based speech emotion editing directly within pretrained TTS models. SEmoEdit unifies three core operations: emotion replacement, emotion erasure, and continuous emotion interpolation, requiring neither parameter updates nor task-specific optimization. To systematically evaluate these capabilities, we introduce SEmoEditBench, a dataset comprising 600 editing cases, and conduct extensive experiments across state-of-the-art (SOTA) models and backbones. Our results show that SEmoEdit is highly effective and broadly applicable, outperforming existing training-based and activation-steering methods. Ultimately, this work reveals that pretrained speech flows possess rich, latent emotion-editing capabilities, providing useful guidance for real applications. Code, benchmark, and Audio samples are available at https://github.com/imxtx/SEmoEdit.

Reinforcement learning (RL) post-training for large language models (LLMs) coordinates multiple models across generation, inference, and training on GPU clusters. Several factors may change during a run, including resource availability, sequence length, memory pressure, and stage bottlenecks. As a consequence, an execution plan that was initially suitable can then become slow or even infeasible over time. However, adapting a job whose models share GPUs entails significant challenges: deciding whether a new plan is worth the transition cost, reusing the job's distributed state, and coordinating GPU transfers across models and stages. Nereus targets these challenges as a cost-aware runtime that adapts RL post-training jobs into efficient execution plans. Its low-overhead controller selects a memory-feasible global plan and admits the transition using a cost model calibrated against the running job. To estimate and execute a transition, Nereus represents the distributed state of each replica of a model-stage (one model in one stage) as an Elastic Model Unit. It then employs a global transition graph to order the transformations and GPU transfers of these units. In a trace built from real data, online TP/PP adaptation reduces average step latency by 27.7% relative to the initial fixed TP/PP layout with DP scaling. In a 1,000-step run reaching 1,024 GPUs, six transitions consume 0.079% of total run time. Nereus improves end-to-end 8B PPO throughput by 2.14--7.27$\times$ over OpenRLHF and by 1.10--1.47$\times$ over Verl across diverse clusters.

Cover's Universal Portfolio (Cover, 1991) matches the performance of the best constant rebalanced portfolio in hindsight. We generalize this framework to compete with an arbitrary comparator sequence $\mathbf{u}_1,\ldots,\mathbf{u}_T$, leading to a dynamic regret minimization problem for the log loss where existing methods break down due to potentially unbounded gradients. The log loss is exp-concave, a curvature property that classically yields fast rates for static regret, yet we show that this advantage generally disappears in the dynamic setting. In particular, a linear-loss-type $\sqrt{TP_T}$ dependence is unavoidable, where $P_T=\sum_{t=2}^T\lVert\mathbf{u}_t-\mathbf{u}_{t-1}\rVert_1$ is the standard path length. This limitation stems from the coarse nature of $P_T$, which obscures finer spatial and temporal structure of the comparator sequence. We therefore introduce two structure-aware measures---the Jensen-Shannon distance for spatial structure and the JS$^q$-path length for temporal structure---under which faster rates are attainable when the comparator sequence has favorable structure. To achieve sharp bounds for both measures simultaneously, we develop Universal Dynamic Portfolio, a parameter-free method that combines a new Dirichlet Hedge algorithm with a fixed-share update, while retaining a near-optimal $P_T$ guarantee in the worst case. Finally, under an additional bounded-gradient assumption, we show that OPS admits the faster $T^{1/3}P_T^{2/3}$ dynamic regret rate over all comparator sequences. We attain this rate with a tractable proper algorithm that applies more broadly to general online exp-concave optimization over arbitrary compact convex domains.

Schrödinger bridges provide an entropy-regularized framework and a principled solution for unpaired domain translation. In practice, a pretrained bridge may need to be adapted to human preferences or physical constraints through a reward a problem closely related to reward tilting in diffusion models but underexplored for Schrödinger bridges. We introduce Tilted Schrödinger Bridge Matching (TSBM), a post-training method for fine-tuning a learned bridge $P$ between source $p_0$ and target $p_1$ toward a reward-tilted target $p_1^r\propto p_1e^r$, while preserving source $p_0$. We formulate this adaptation as alternating optimization initialized from $P$, provide theoretical justification, and derive a practical algorithm based on Adjoint Matching. We evaluate TSBM on unpaired image-to-image translation targeting digit properties in MNIST and facial attributes in CelebA.

We study approximate sampling: given $N$ independent samples from a proposal distribution $μ$, the goal is to select one whose distribution is close to a target $π$ specified only up to a normalizing constant. Block and Polyanskiy (2023) provide finite budget error bounds for approximate rejection sampling (RS) as a function of the acceptance threshold $M$. The threshold $M$ giving the smallest bound, however, depends on properties of $(π,μ)$ that are typically unavailable from the observed sample. This raises a natural question: Can one attain the best RS guarantee without taking $M$ as input? We answer affirmatively by proposing a parameter-free sampling algorithm called uniform race (UR), based on importance weights, which are ratios of target to proposal probabilities (or densities). It divides each observed weight by an independent uniform random variable to form a score and returns the candidate with the largest score. For every budget $N$, its total variation error satisfies the RS upper bound for every fixed threshold $M$ simultaneously, thereby achieving the best such bound in hindsight. We also characterize its output distribution conditional on the largest score, identifying when it is exactly the target $π$. Uniform race has no larger total variation error than a natural budget-calibrated RS derived from Rohatgi et al. (2025) and sampling importance resampling (SIR). In particular, we exhibit instances where UR's error is exponentially smaller in $N$ than that of either baseline. Furthermore, we establish conditions under which attaining this RS guarantee for every $(π,μ)$ uniquely determines the selection probabilities as those of UR. Finally, test-time scaling experiments on LLM math-reasoning tasks corroborate the theoretical comparisons and demonstrate that UR remains competitive in ground-truth accuracy without requiring threshold selection.

Test-time adaptation (TTA) is a promising paradigm for handling distribution shift in time-series forecasting (TSF), where models adapt at inference time, often leveraging delayed observed data to refine predictions. In the multivariate setting, distribution shifts often exhibit cross-variate dependencies, yet existing TSF-TTA methods adapt each variate independently and ignore this cross-variate structure. Exploiting such structure motivates cross-variate interaction, but coupling variates through backbone predictions introduces direct pathways for mixing uncorrected errors across variates, a concern under the delayed supervision of TSF-TTA. We identify the \emph{interaction space} as a key design choice, and show that acting on adapter corrections that refine backbone outputs, the \emph{correction space}, rather than on the predictions themselves, avoids directly propagating backbone errors across variates. We build on this to propose \textsc{CoRe} (\textsc{Co}rrection-space Interaction \textsc{Re}finement), realizing correction-space interaction through (i) Shared-anchor Correction Refinement (SCR), which combines each variate's correction with a shared anchor through a parameter-efficient bottleneck, and (ii) input-conditioned spectral gating, which adaptively modulates the refinement from the current input window. Across seven backbones, six datasets, and four prediction horizons, \textsc{CoRe} reduces MSE by 25.82\% on average over backbones and 10.57\% over the state-of-the-art TSF-TTA method, with stronger gains at medium-to-long horizons and modest computational overhead. Data and code are available at: https://github.com/yyddou/CoReTTA

This work introduces MechReasoner, a mechanistic qualitative simulator grounded in confluence-based qualitative physics, together with a benchmark for mechanistic inference. Current large language models (LLMs) generate fluent mechanistic descriptions that do not reliably follow from underlying structural and causal constraints. The benchmark tests whether answers preserve simulator-licensed ambiguity, quantified claims, episode-graph transition evidence, repairs, and trace-support judgments. Its 1,120 items are generated deterministically from admissible interpretation sets, component states, scenario restrictions, confluence constraints, and derivation steps across 18 catalog mechanisms and six task families. Each mechanism undergoes converter checks of structure and topology and behavioral checks against quantitative simulations. GPT-5.5 accuracy decreases as family-specific mechanistic complexity increases, from 76.1% in the lowest-complexity bucket (B1) to 38.0% in the highest-complexity bucket (B4). The negative association remains after controls for rendered-prompt and expected-answer length. These results show that qualitative simulators can support auditable NLP benchmarks for mechanistic inference.

Mixture-of-Experts (MoE) models repeatedly route tokens to sparse subsets of experts, but conventional routers expose no routing-specific record of how cross-layer influences accumulate. We introduce Scratchpad-Augmented Mixture-of-Experts (SA-MoE), which gives each router access to a low-dimensional persistent state that is not provided to the experts. Learned layerwise writes update this state, and their realized post-update changes exactly decompose the state-mediated contribution to any later routing margin, forming a routing ledger. Across sparsely upcycled SmolLM2- and Gemma-based models and three independent training seeds per architecture, this pathway adds less than 1% analytical forward compute and is strongly used by trained routers: local removal of its router contribution changes the selected Top-2 expert set in 87.6% and 69.9% of decisions, respectively. Relative to a matched latest-write-only control, persistent accumulation increases long-horizon future-routing accessibility by 19.4 and 12.2 percentage points, with positive effects in every seed. More than 90% of absolute ledger contribution comes from non-recent writes in both families, and full-forward suppression of ledger-selected writes changes later routing and output distributions. The ledger is an exact provenance object for the persistent-state pathway, not a complete causal explanation of routing. Sensitivity-aware scores better predict full-forward intervention effects, and post-hoc methods recover related cross-layer attribution without architectural modification. SA-MoE instead makes one routing-specific computational history explicit and directly inspectable within the model's natural forward computation.

Long-term memory supports the self-evolution of LLM agents by retaining experience and skills across tasks and enabling their retrieval, reuse, and revision in subsequent long-horizon decision-making. Yet existing memory management approaches remain limited to discriminative retrieval and to address the sparse, hierarchical, and highly redundant structure of reusable experience: only a small, task-dependent subset of trajectories and memories warrants retention, retrieval, or revision. Learning these operations is further complicated by sparse, delayed, and indirect task-level feedback, with weak supervision across the memory lifecycle. Moreover, continual memory evolution introduces an architectural tension as addressing invariance: stored experience is perpetually revised, yet the addressing interface consumed by learned retrieval policies must remain stable. To address, we present GenMem, which reformulates memory management as generative symbolic addressing. Its core mechanism is the Symbolic Identifier (SID), a multi-level discrete token tuple drawn from a Cartesian-product address space that factorizes a million-scale sparse memory space using fewer than one hundred discrete symbols. Instead of generating ever-changing raw content, the memory agent learns to generate SIDs, while memory evolution rewrites the payload at a fixed address without shifting the address itself. Architecturally, GenMem couples a MemRetriever and a MemEvolver within a multi-agent harness, trained via GRPO with dense process and outcome rewards with two-channels optimization. Under offline memory evolution, experiments spanning ALFWorld, WebShop, multi-hop QA, medical reasoning, and deep research evaluate GenMem against strong memory-augmented baselines...

Many complex systems are observed only through temporally unpaired distribution snapshots, making trajectory-based dynamical learning difficult without additional assumptions. We therefore formulate the problem directly in distribution space, treating the distribution itself as the dynamical state. The challenge is that distribution space is infinite-dimensional, making compact and approximately closed representations difficult to learn from finite snapshots. We introduce DisKO, which extends deep Koopman learning to distribution dynamics by jointly learning predictive distributional observables, a finite-dimensional Koopman representation, and a generative map back to the full distribution. Across seven diverse benchmarks, DisKO achieves state-of-the-art extrapolation performance, with substantially slower error accumulation on long-horizon prediction tasks. DisKO further recovers leading Koopman eigenvalues and eigenfunctions on systems with analytic spectra, revealing meaningful dynamical structure in the learned representation.

Neural network verification (NNV) formally verifies that a network satisfies a specified property for all inputs within a defined region. Modern NNV tools employ abstract domains to compute a sound over-approximation of the network's behavior from the given input region, thus the tightness of these abstractions essentially determines efficiency. A long line of increasingly precise domains has been developed, but they all describe the valid input region in the same restrictive way, e.g., an Lp-norm ball. A practical input region is rarely a simple Lp ball, but rather a combination Lp ball with additional constraints. Verifying a network over such a region with existing abstraction produces a loose over-approximation, which results in either failing to verify a property or spurious counterexamples. We introduce Constrained Lagrangian Abstract Domain (CLAD), a new abstract domain that computes a sound over-approximation of neural networks over input regions defined by a combination of convex constraints. CLAD propagates these constraints and tightens bounds over the true feasible region. However, bounding a neuron over the intersection of these constraints has no closed-form solution, so CLAD relaxes each constraint into the objective with a Lagrange multiplier and solves the resulting max-min problem with a projected primal-dual method, alternating a projected gradient step on the input with a multiplier update. CLAD supports any convex constraint with a subgradient, e.g., from automatic differentiation. We evaluate CLAD on 1,944 instances across four convolutional networks with motion-blur structured perturbations with halfspace or L2-ball constraints. On standard unconstrained Linf property, CLAD verifies as many instances as GCPCROWN at a similar runtime. On constrained properties, CLAD verifies 60\% more instances than GCPCROWN on L2-ball properties, and 22% more in total.

Machine-learning models for physical systems are currently evaluated primarily through errors between predicted and reference states and, increasingly, through tests of physical consistency. These metrics assess whether predictions are accurate and satisfy selected physical requirements, but provide limited insight into whether learned trajectories reproduce the underlying dynamics. Domain experts examine such relationships through physical representations that expose relevant processes, interactions, and responses, but these analyses are often separated from typical machine-learning evaluation. We introduce a practical framework for evaluating dynamical fidelity through predictive structure in physical representation spaces. Experts define the representations, while reference trajectories determine which relationships are predictive and retained as evaluation tests. We demonstrate the approach in atmospheric forecasting using ERA5 representations of planetary-wave activity and Northern Annular Mode evolution, and evaluate Pangu-Weather, GraphCast, and FengWu. The models exhibit distinct departures from reference predictive structure that are not reflected by conventional forecast errors. The framework thereby turns domain-expert representations into systematic tests of learned physical dynamics without prescribing the relationships in advance.

MAS specifications express the effects of the actions of the agents and their environment, as well as other temporal phenomena, such as the intervals during which an agent may perform an action. The specification of a MAS should also be executable in order to allow for run-time monitoring. Constructing the specification of a MAS requires formal language expertise, while machine learning techniques depend on labelled data which are rarely available. To address these issues, we propose `genRTEC', a method that leverages pre-trained Large Language Models (LLMs) to generate executable MAS specifications, in the language of the `Run-Time Event Calculus' (RTEC), from natural language descriptions. genRTEC constructs MAS specifications with complex hierarchical and cyclic dependencies based only on short natural language descriptions of the concepts involved. We present an extensive empirical evaluation of genRTEC, spanning various MAS specifications, including both a qualitative and a quantitative assessment. Our results demonstrate that genRTEC constructs executable MAS specifications of high predictive accuracy without compromising reasoning efficiency.

Managing water resources in mountainous regions depends heavily on reliable Snow Water Equivalent (SWE) and Snow Height (HS) data, yet these variables remain difficult to track at scale. This study evaluates three machine learning architectures (XGBoost, U-Net and SegFormer) for the joint estimation of SWE and HS variations from Sentinel-1 InSAR data over the Italian Alps, using the IT-SNOW reanalysis as reference. SegFormer achieves the best results on both targets, with an MAE of 10.391 cm for HS and 27.113 mm w.e. for SWE and the lowest variability across initializations. A feature sensitivity analysis shows that including all available features does not guarantee the lowest error, with model- and task-specific sensitivities. Spatial metrics (R2, Pearson's r) separate the three architectures far more clearly than mean error (MAE, RMSE) does, and decomposing the error per window attributes most of it to a systematic offset in the estimated mean variation rather than to the spatial pattern.

Flow matching (FM) has recently emerged as a promising framework for generative modeling due to its conceptual simplicity and strong empirical performance. In FM, samples are transported along a vector field parameterized by a neural network, inducing a probability path that evolves from a simple noise distribution to the target data distribution, governed by an ordinary differential equation (ODE). However, existing FM approaches predominantly rely on probability paths derived from optimal transport (OT) between Gaussian distributions, which may be suboptimal for capturing complex data with inhomogeneous structures such as heavy tail or sharp contrast. In this work, we generalize FM to the broader class of location-scale families for handling data inhomogeneity and introduce a novel class of probability paths defined as geodesics on the manifold of probability distributions. We name this approach probabilistic geodesic flow matching to distinguish it from prior geodesic (Riemannian) FM methods defined in input space. We argue that Euclidean OT-based paths are not necessarily optimal in probability space and may limit modeling flexibility. Through synthetic benchmarks and scientific datasets at different scales, we demonstrate that the proposed method more effectively captures complex distributions, leading to improved or comparable performance compared with SOTA geometry-motivated generative models.

Cross-site out-of-distribution (OOD) generalization in resting-state functional magnetic resonance imaging (rs-fMRI) often relies on learning task-discriminative representations from full-scan functional connectivity (FC) graphs and promoting invariance across source sites. However, FC graphs are estimated from finite, temporally correlated blood-oxygen-level-dependent (BOLD) sequences. Cross-site agreement therefore does not necessarily imply that predictive evidence remains supported under FC re-estimation within the same scan. In this paper, we propose Brain Network Re-estimation-Informed OOD Learning (BRIO), a framework that uses within-scan FC re-estimation to guide cross-site alignment. BRIO maps fullscan graphs and their re-estimates into consistently indexed connectome factors, enabling comparisons of their predictive contributions. It assesses re-estimation support from changes in these contributions relative to within-class subject variability and class separation. For each source-site pair and class, this task-calibrated support from both sites is combined with predictive relevance to form pairwise qualifications, which determine relative factor weights and overall alignment strength. Leave-one-site-out experiments on four real-world datasets (ABIDE, REST-metaMDD, SRPBS, and ABCD) show that BRIO consistently outperforms competitive baselines, with relative improvements of up to 3.8% in accuracy. These gains also persist under an alternative brain parcellation on ABIDE.

World Action Models (WAMs) enable future-aware control by jointly modeling actions and environment dynamics. However, iterative diffusion or flow inference incurs substantial denoising latency. Prior inference state offers a natural opportunity for acceleration, yet changing planning contexts, observations, and intermediate representations can quickly render retained state stale. Preserving useful computation therefore requires adapting inference state rather than reusing it as-is. To this end, we present $\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}$, a training-free framework that accelerates WAM inference by preserving and adapting inference state for efficient and accurate continuation as the control loop evolves. Across closed-loop replans, Trajectory Remapping remaps replan state from the preceding replan to initialize the next replan, reducing redundant trajectory generation. Across denoising steps, Observation Rebinding performs anticipatory inference during action execution and rebinds retained denoising state to the real observation for continuation when consistency checks pass, reducing latency exposed to the control loop. Across Transformer layers, Residual Rescaling selectively rescales retained layer state and refreshes it through full computation of the middle layers when probe checks fail, reducing repeated Transformer computation. Evaluations of three representative WAM architectures on LIBERO and RoboTwin 2.0 show that $\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}$ achieves 1.47-3.05$\times$ speedups in observation-to-action latency and 2.23-3.27$\times$ speedups in GPU inference time per replan, while preserving 96.69-99.54% of native WAM task success.

Multi-teacher on-policy distillation (MOPD) has emerged as a popular post-training paradigm for integrating specialized capabilities in frontier language models. Existing OPD research has primarily focused on optimizing single-task distillation through objective design, distillation scope, and teacher signal construction, whereas MOPD must aggregate multiple capabilities in shared parameters and address the resulting capability seesaw, in which improving one domain suppresses capabilities acquired from another. Inspired by the distinctive update geometry of OPD, we find that parameter updates from different tasks rapidly concentrate in their respective low-dimensional subspaces during MOPD, providing a direct geometric basis for identifying and controlling cross-task interference. We therefore propose PMOPD (Projection-based Multi-Teacher On-Policy Distillation), which constructs subspace memories from the cumulative parameter displacements of different tasks and projects both gradients and optimizer updates to remove components that interfere with protected task directions. We further develop a lightweight conflict probe to characterize task interactions and guide task ordering, together with a cycling strategy that balances subspace estimation and timely task revisitation. Experiments on representative Code, Reason, and Math tasks show that PMOPD improves every evaluated capability over MOPD, raising the average score across the three tasks by 2.54 points on Qwen2.5-7B and 2.09 points on Llama-3.1-8B. These consistent gains establish geometry-aware optimization as an effective and transferable approach to balanced multi-teacher distillation.

World models learn environment dynamics from interaction experience. These dynamics depend on the current state and actions, as well as on properties that persist across interactions. Yet standard predictive training can reduce error using local evidence alone, without organizing persistent information into reusable context. We introduce SPRII, a training principle that uses relations between interactions as weak supervision for persistent context while retaining the learner's native objective. For example, different trajectories of the same system share persistent properties even when their states and actions differ. SPRII uses such relations to guide context learning without numerical property labels. Two composable components encourage contexts from related interactions to agree (Align) and use one interaction's context to predict another's future (Cross). Our analysis distinguishes three linked questions: what persistent information is accessible in the learned context (Formation), how that context influences a fixed predictor (Use), and whether it reduces task error (Value). Success at one stage does not guarantee success at the next. Controlled experiments show that more reliable relations improve representation organization, but adding a shared-property constraint can reduce access to a property that remains shared. Context substitutions change predictions at fixed model weights, while the benefit from history depends on prediction horizon and readout. Evaluations span thirteen settings, including controlled physical systems, public dynamics tasks, robotic and tactile data, and partner interaction, across multiple learner families. Relative to the corresponding baselines, SPRII yields average gains of over 10% in downstream task performance and over 15% in persistent-property readout. The project page is available at https://persistent-learning-review.netlify.app/interactive.html.

Does repairing an episode make its experience a better memory for the next task? We transfer the same failed source before and after accepted repair to a fixed target, alongside independent execution. Our 3,300 runs cover 100 ThinkingBox pairs and the same 100 APEX pairs with and without source-state inheritance, under eleven conditions. ThinkingBox's Full/Skill/Hybrid correction gains are 44/29/32 percentage points, with corrected performance 25/22/18 points above independence; inference weakens at the task-family level. Yet 12 of Full's 15-point larger correction gap over Skill come from worse uncorrected performance, not better corrected memory. Moreover, 22 of Full's 46 upward transitions restore observed baseline success. Neither APEX regime establishes comparable aggregate correction benefits. Action evidence connects workflow gains with reusable obligations and convention conflicts with source-local choices. Text APEX's accepted execution reaches 52% versus its summary's 40%, without robust global/group-level superiority or an estab- lished advantage over independence. Smaller handoffs reduce input but increase calls. The value of repairing experience is therefore distinct from the value of reusing it: memory updates require both a previous-version reference and a fresh-start reference.

Frozen Earth-observation embeddings are judged almost entirely by spatially blocked cross-validation inside one study region. We show that this number does not predict accuracy in a new region; we show why; and we show the one setting in which such a model does keep working, using a protocol that needs only a linear probe and labels one already has. The testbed is wildfire, with Copernicus burned-area maps of six fires in Greece and Spain and descriptors from the year before each fire, comparing TESSERA and AlphaEarth with ESA WorldCover classes and annual Sentinel-2 index summaries. Inside a fire, the embeddings identify the burned land 0.05 to 0.13 ROC AUC better than the index summaries, and repeated fold allocations, spatial buffers, a block bootstrap, and gradient-boosted trees leave that margin unchanged. On a fire in another region, they lose 0.15 to 0.18 AUC, and the index summaries lose 0.06, so the three end within a few hundredths of each other. The representation is not the cause. Eight labelled blocks from the new region restore the embedding advantage and give a higher AUC than 59,000 labelled pixels from other regions, and the weight vector fitted in one region is nearly orthogonal to the vector fitted in the others, so the part that carries across regions is small and low-dimensional. Forecasting within a region is a different matter. Fitted on a fire that burned in 2023 and applied to a fire twelve kilometres away that burned in 2024, where nothing used postdates the target fire, TESSERA reaches 0.772 AUC and loses 0.04 against a classifier fitted inside the 2024 fire, while classifiers fitted in other regions lose 0.09 to 0.18. A region with one mapped fire can therefore forecast susceptibility for later fires there; a region without one cannot borrow a model from elsewhere, and every evaluation of a frozen embedding should report a held-out region.

Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, yet how RL reshapes these policies remains poorly understood. We find that RL across widely used flow-based VLA models, including $π_{0.5}$ and GR00T~N1.5/N1.6, on LIBERO, ManiSkill, MetaWorld, and CALVIN induces substantially lower-rank parameter updates that are highly concentrated in the action expert's Timestep Modules, a small and previously overlooked component. Through systematic module-replacement experiments, we further show that these modules capture a disproportionate share of the performance gains from RL. We then characterize what is encoded in these Timestep Modules. First, we show that RL specializes them to the discrete denoising timesteps used during rollouts, and that this discrete-timestep training underlies the low-rank updates. Second, we find that among their outputs, the shift vector changes most distinctly under RL, and through probing, we show that shift update directions strongly predict task success (ROC-AUC up to $99.6\%$). Third, we find that the geometry of shift updates reflects task relationships, as their pairwise similarity correlates with cross-task transfer patterns. Building on these findings, we show that steering along shift update directions further improves RL-trained policies without additional RL training. Overall, we provide a systematic understanding of how RL reshapes VLA policies by studying how learned signals are encoded in parameter space, offering insights into more efficient and interpretable VLA post-training.

Video temporal grounding (VTG) aims to localize the video interval corresponding to a language query. Recent large vision-language models (LVLMs) show great potential in solving such a multi-modal reasoning task. However, long videos often contain large amounts of redundant information that disturbs LVLMs to mine query-relevant evidence. Instead of dense frame sampling which incurs prohibitive training memory, previous reinforcement learning with verifiable rewards (RLVR) works typically utilize sparse sampling, which makes training feasible but may miss critical evidence. In this paper, we propose a ``summarize before grounding'' framework (named ``SumGround'') for long-video temporal grounding. The key of SumGround is to perform query-guided chunk condensation to aggregate and retrieve query-relevant evidence. Specifically, we split the video into several chunks and perform two-level chunk condensation. First, we introduce query-guided latent summaries, which is represented as KV states of query-guided prompts, to compress redundant visual tokens into compact query-relevant chunk summaries. Furthermore, we design an associative summary retrieval scheme to rank and select chunk summaries that are most likely to contain the event interval. Both query-guided latent summary and associative summary retrieval schemes are enabled by RLVR. To reduce memory consumption, we propose a length-aware gradient gating module to selectively stop gradient back-propagated to visual tokens. Extensive experiments demonstrate that SumGround performs favorably against previous state-of-the-art methods across multiple downstream datasets, with remarkable gains on long videos.

Off-policy evaluation is a foundational component of offline reinforcement learning, aiming to assess and optimize policy performance using pre-collected datasets. However, such datasets often suffer from pronounced challenges, including distribution shift, $Q$-value overestimation, and low sample utilization efficiency. To address these issues, this paper introduces a weighted Bellman residual minimization framework that incorporates density ratio weighting by effectively integrating expert demonstrations with behavioral data. The proposed weighting scheme departs from the conventional completeness assumption commonly imposed in the theoretical analysis of deep reinforcement learning. We establish a sharp convergence rate for density ratio estimation and derive the convergence rate for the excess risk of resulting deep $Q^*$ estimator. Extensive empirical evaluations demonstrate that, compared to existing methods, our method achieves significant improvements in numerical performance and policy generalization, providing specific guidance for the rational utilization of expert demonstrations.

Representation-level unlearning intervenes on the intermediate hidden states of LLMs. Although knowledge is distributed across layers, existing methods operate at a single fixed layer for the entire forget set. We ask whether such a fixed layer is sufficient. To answer this, we design a hijacking experiment that grafts hidden states of the target model into an oracle trained only on the retain set. The oracle cannot produce the forget answer on its own, yet it produces the answer from the grafted state. Thus, the answer is formed at an intermediate layer and merely read out afterward, so unlearning should focus on formation, not readout. Moreover, the layer where formation ends varies widely across inputs. Motivated by these findings, we propose Targeted Unlearning at Layers Identified Per-input (TULIP). For each input, TULIP uses the logit lens to locate the formation-readout boundary and removes the hidden state's alignment with the forget answer's unembedding vector there. TULIP consistently outperforms output- and representation-level baselines on TOFU, PISTOL, and WMDP across Llama, Qwen, and Zephyr models. It also remains robust to paraphrase and quantization attacks. Beyond standalone use, its per-input layer selection serves as a plug-and-play component that further improves existing methods.

Language models often process long inputs sequentially in chunks, but continuing to read after sufficient evidence has been acquired wastes computation. Existing stopping mechanisms either learn sufficiency from internal activations or train an exit gate, while a simpler alternative asks the model whether it has read enough. We introduce Answer-Convergence Stopping (ACS), a training-free stopping rule that measures rather than asks. After each chunk, it probes the frozen model's current answer state and stops when that state is both confident and stable. The rule requires only output-side generation and token log probabilities, has no trained components, and uses one shared configuration across models and benchmarks. Because a stopping policy can save computation simply by stopping too early, we evaluate the stopping decision itself using evidence position where available. On the full LongBench-v2 with two frontier models, ACS is the only stopping policy that matches or exceeds full-reading accuracy. Furthermore, across 250 S-NIAH questions, the premature stopping rate for ACS across five models from two families ranges from 0% to 12%, compared to 8.4% to 45.6% for the verbalized gate. Taken together, ACS reveals that by properly utilizing the output signals of frozen models, we can achieve favorable behaviors like adaptive stopping without the need for additional training.

Reinforcement learning is increasingly used to post-train vision-language models for image-to-code generation, such as generating SVG code from a reference image, by optimizing rewards computed from the final rendered output. However, relying on a single terminal reward provides sparse feedback that is poorly aligned with the contribution of individual tokens. A generated program may contain operations that accurately reproduce some parts of the target image alongside others that introduce errors, yet all tokens are trained from the same final outcome. We observe that many intermediate code prefixes are not only executable, but already produce meaningful partial renders that reflect progress toward the target. This property provides a natural source of denser supervision during generation. Based on this observation, we introduce IR4RL, an RL framework with a token-level render-progress reward that turns changes between intermediate renders into localized feedback for the generated sequence. We evaluate our approach on Image-to-SVG and Image-to-TikZ generation. Across both tasks, our method improves over supervised fine-tuning and standard GRPO, yielding new state-of-the-art open-source models. This shows that intermediate rendering provides a simple and effective source of process supervision for RL post-training of image-to-code models.

Deploying LLM-enabled agentic applications across the Edge-to-Cloud continuum remains challenging due to hardware heterogeneity, deployment complexity, limited observability, and the lack of systematic evaluation methods. Existing solutions address agent development, observability, or benchmarking separately, offering limited support for the full lifecycle of distributed agentic applications. This paper presents AgentWare, an AgenticOps framework that automates the provisioning, deployment, observability, and evaluation of agentic applications across Edge-to-Cloud infrastructures. AgentWare introduces an end-to-end lifecycle pipeline that automatically prepares heterogeneous execution environments, transforms user-defined agent implementations into distributed applications, deploys agent components across the continuum, and performs unified collection of execution traces, infrastructure telemetry, and evaluation metrics. The framework further supports automated semantic evaluation through LLM-as-a-Judge workflows and generates reproducible reports covering correctness, performance, resource utilization, and energy consumption. We demonstrate the applicability of AgentWare through a distributed book assistant agent deployed across real Edge-to-Cloud infrastructure under multiple deployment and model configurations. The results show that AgentWare enables systematic experimentation and evaluation of distributed agentic applications while significantly reducing the manual effort required for deployment, instrumentation, and analysis.

We propose Tiny Recursive Models for Combinatorial Optimization (\ours{}), a general neural method for combinatorial optimization that scales both depth (how often we recursively invoke our network) and width (how much we sample in parallel). Both are fundamental for combinatorial optimization: hard instances demand a large amount of compute, while a small network is essential to avoid overfitting and capture the algorithmic essence of optimization. In particular, our method consists of a graph-aware tiny recursive model that iterates on a latent state with adaptive halting and needs only a lightweight problem-specific decoder. Compared with previous heatmap-based general neural solvers, it achieves a better balance between solution quality and inference speed on both the Traveling Salesman Problem~(TSP) and the Maximum Independent Set~(MIS) problem, and remains competitive with hybrid methods that combine neural components with heuristics specific to each problem. With the same backbone architecture for both tasks, \ours{} outperforms every diffusion-based solver on TSP from 500 to 10,000 cities at a lower inference cost, and on the standard Erdős--Rényi-[700-800] MIS benchmark it surpasses all neural solvers except those that only work well on MIS. We then explore self-relabeling for self-supervised training. We periodically replace the current set of training labels with the model's own better solutions, as an alternative training signal. Self-relabeling can, while forgoing supervision from near-optimal solutions, still result in on-par quality.

In-game toxic language has emerged as a critical concern in the gaming industry and community. While several frameworks and models for online game toxicity analysis have been proposed, detecting toxicity in player chat utterances remains a formidable challenge: stemming not only from the extremely short length of such utterances but also from the heavy reliance on game slang, abbreviations, and domain-specific jargon, which generic language models are poorly suited to recognize. This paper presents a shared task for in-game toxic language detection built upon real-world in-game chat data, and proposes the best-preforming model for the toxic language slot filling: Bi-directional Representations with Attention Residuals (BRAR). Experimental results demonstrate that BRAR effectively captures the global context and outperforms the existing baselines on slot filling.

Human speech conveys rich perceptual information, such as emotion and speaker identity, yet most automatic speech quality judges reduce it to a single naturalness score. We study diagnostic speech judges: given two candidates, a diagnostic judge decides which is better, along which perceptual dimensions (e.g., timbre, emotion, timing) they differ, and which audible cues support its decision. Learning such judges is challenging: expert annotation is costly, and simply prompting a frontier audio-language model to produce labels is unreliable: our probing reveals substantial errors and unstable instruction following. We introduce SpeechCritic, which learns a diagnostic judge in a reference-conditioned cross-lingual setting from only about 300 human-labeled comparisons. Rather than replacing the frontier model, SpeechCritic calibrates it with these labels: for each dimension, it selects the acoustic measurements that agree with human judgments, maps them to A/Tie/B probabilities, and passes these to the model as non-binding hints alongside the audio. Compared with the same model labeling without hints, this raises dimension-level agreement with humans by 6.3 points and cuts the mismatch with human Tie rates by 10.4 points. We then train a 7B judge on this supervision and find that different training signals shape different judge behaviors: SFT establishes the task, OPD transfers the teacher's dimension-level strengths and weaknesses, and RL helps most on clear-cut comparisons where human raters agree. Notably, human listeners also find that RL makes rationales cite more specific, localized acoustic cues, although it never directly rewards rationale text. Finally, we show that the pipeline is language-pair agnostic by instantiating it on both English-Japanese and English-Spanish. Together, these results demonstrate a path from limited human preferences to a diagnostic speech judge.

Single-image novel view synthesis remains challenging because the underlying 3D geometry is highly ambiguous. Recent diffusion-based approaches produce plausible results, but they often struggle to preserve the geometric structure and spatial coherence of foreground objects. We present GenNVS, a framework for geometry-enhanced novel view synthesis via a disentangled 3D prior. Specifically, GenNVS models foreground objects and the background with 3D Gaussian Splatting and aligns them through a coarse-to-fine geometric optimization process to form a unified 3D scene. This scene conditions a video diffusion model through the proposed Dual-Stream Masking mechanism, which guides synthesis by jointly exploiting rendered validity masks and geometry-aware warping. Experimental results show that GenNVS performs favorably against recent methods in both visual quality and geometric accuracy, while naturally supporting flexible scene editing.

Informative Path Planning for scalar field reconstruction uses predictive uncertainty to direct sensing vehicles toward maximally informative locations. Gaussian Processes provide this signal but their stationary isotropic kernels are misspecified for non-homogeneous phenomena such as oil spills, producing miscalibrated estimates that degrade planning. We investigate whether replacing the Gaussian Process with a well-calibrated Deep Ensemble improves path planning outcomes, and whether uncertainty quality interacts with the choice of planning algorithm. Five strategies ($ε$-Greedy, Value Greedy, Uncertainty Greedy, Monte Carlo Tree Search, and Receding Horizon Orienteering) share a common Deep Ensemble backbone trained on physics-based oil spill simulations. On held-out stochastic spill scenarios, the Deep Ensemble reduces normalised reconstruction error by $83\%$ relative to the Gaussian Process baseline. Crucially, well-calibrated uncertainty amplifies the importance of the planning strategy: the performance gap between algorithms is negligible under miscalibrated models but becomes substantial under the ensemble, where multi-step lookahead planners outperform greedy selection by up to $32\%$ in reconstruction error and achieve IoU above $0.85$. Monte Carlo Tree Search is the recommended planner, matching Orienteering in reconstruction quality at an order-of-magnitude lower computational cost.

Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditioned value functions often provide unstable guidance due to sparse rewards and discounting. Hierarchical methods partially mitigate this issue via subgoal decomposition; however, high-level decision-making still relies on noise-sensitive value estimates, leading to unstable behavior in complex environments. We address this limitation by proposing \textbf{D}iffusion \textbf{S}ubgoal \textbf{P}lanning (\textbf{DSP}), a diffusion-based framework for high-level subgoal generation. DSP casts high-level planning as guided generative inference over goal-conditioned subgoals and learns both conditional and unconditional flows, enabling classifier-free guidance to introduce a goal-directed bias at inference time. By removing explicit value-based guidance from high-level planning, DSP generates reachable and goal-directed subgoals through a generative model while retaining hierarchical execution. Experiments on offline GCRL benchmarks demonstrate that DSP outperforms prior methods on a range of navigation and manipulation tasks, with particularly strong performance in maze environments that require multi-step subgoal planning.

Reasoning language models that can call tools must decide during inference whether to answer unaided or delegate. Any self-reflection mechanism for this must answer three questions: where the reflective signal comes from (verbal reports, output distributions, hidden states, a separate predictor), how it is presented to the model (numerical prediction, confidence token, prompt injection), and whether it changes the model's subsequent action. We isolate the third question. At a fixed point in otherwise identical reasoning trajectories, we insert a single first-person sentence expressing either confidence or doubt; the model then continues reasoning and chooses whether to answer directly or call a tool. Comparing these counterfactual continuations measures the causal effect of the reflective signal on delegation. We call this behavioral response Nudgeability and measure it along two dimensions: sensitivity, how strongly confidence and doubt change delegation rates, and targeting, whether delegation increases for problems the model cannot solve unaided and decreases for those it can. Across nine small-to-medium open-weight reasoning models from three families (Qwen, Gemma, and GLM) and two tasks, models are consistently sensitive: doubt increases delegation and confidence decreases it, with a median confidence-to-doubt swing of 20.6 percentage points, and 53 to 70 points for the larger provider-served models. This responsiveness is poorly targeted: a median 42% of induced flips are well-targeted, only a +2 percentage-point lift over a random-selection baseline. Confidence language is thus a strong control surface for delegation, but current models use it only weakly in accordance with their actual competence. Nudgeability offers a simple, post-training-free way to evaluate both sensitivity and targeting as endogenous self-reflection mechanisms mature.

Controlling persona in large language models (LLMs) at inference time is important for role-playing, personalized dialogue, and social simulation. Recent methods extract persona vectors from the model's activation space and apply Euclidean operations---addition, scaling, and linear interpolation---under the linear representation hypothesis. However, these methods themselves report systematic failures: non-orthogonal trait dimensions, asymmetric ceiling and resistance effects, and significant deviations in multi-trait composition, suggesting that the linear isotropic assumption does not hold. We propose PersonaManifold, a framework that models persona representations as points on a curved, low-dimensional Riemannian submanifold in activation space. We estimate the manifold's intrinsic geometry---local metric tensors, geodesic distances, and Ollivier-Ricci curvature---and introduce geodesic steering, which interpolates between personas along manifold geodesics rather than Euclidean straight lines. We also propose the Behavioral Similarity Triplet (BST) benchmark, which automatically generates situational questions grounded in six established psychological constructs and defines persona similarity through behavioral responses rather than self-report questionnaires. Experiments on three open-source LLMs show that persona activations form a manifold with heterogeneous curvature, geodesic distance predicts behavioral similarity more accurately than Euclidean alternatives with independent contributions from anisotropy and curvature, and geodesic steering produces more coherent intermediate personas on both our BST benchmark and external evaluations, with the advantage concentrated in high-deviation regions where the manifold deviates most from flatness.

Predicting observed dynamics does not establish recovery of the underlying physical mechanism. Can machine learning retrace the hidden-state reasoning behind the Hodgkin-Huxley (HH) model? We train structured latent models on simulated current and voltage, withholding gate identities and trajectories from training and model selection. We then test response prediction, state recovery, protocol transfer, and agreement with HH dynamics. Prediction error and its cross-seed spread both drop sharply at three latent dimensions under the tested protocols, while gate recovery under new protocols improves through five to six coordinates. State recovery depends on which observations the chart uses. Observed voltage improves current-clamp decoding relative to freely predicted voltage. Under voltage clamp, adding latent state to command voltage raises m-state $R^2$ from 0.976 to above 0.99, yet the transported field disagrees with HH on identical smooth samples. Known invertible HH coordinates achieve high fast-m field agreement under the same audit procedure. An exact HH identity decomposes the discrepancy into time-scale-weighted state error and a residual in the transported field; these terms can cancel or reinforce. These findings concern the tested models and charts. They support evaluating state and dynamics recovery separately, including chart inputs and transported-field agreement across interventions.

Long-horizon tasks require LLM agents to continually draw on information from earlier interactions. However, retaining the full history increases context costs, while compressing it risks losing details needed later, and the relevance of historical information often becomes apparent as the task progresses. To address these challenges, we propose FlowState, which treats execution state as memory that can be retained and revisited across requests, unifying current decision-making with the reuse of historical information. FlowState preserves semantically typed state nodes, their relations, and references to raw tool observations, separating persistent retention from on-demand access. Within a single execution loop, Incremental State Update (ISU) maintains the current state based on new inputs and feedback, while Progressive State Access (PSA) progressively reveals historical states and supporting evidence as needed during reasoning. Together, these mechanisms enable agents to reassess prior decisions in light of new information and guide subsequent actions. Compared with a full-context baseline using the same DeepSeek-V4-Flash model, FlowState improves the average success rate on MemoryArena and the average pass rate on $τ^3$-Bench by 4.55 and 13.95 percentage points, respectively, while reducing total token consumption by 43.2% and 40.6%. These results demonstrate the performance and efficiency advantages of FlowState on long-horizon tasks.

Latent visual reasoning (LVR) enables multimodal large language models (MLLMs) to perform intermediate computation in continuous latent tokens rather than expressing every reasoning step in words. However, unlike textual CoT, latent reasoning is not directly observable, making it difficult to supervise what latent tokens learn. In this work, we first conduct a thorough analysis of latent-token behavior and identify a latent evidence-credit gap: latent tokens respond only weakly to image perturbations that alter the correct answer. We hypothesize that this issue stems from the lack of explicit supervision during GRPO training. These findings suggest that a final-answer reward provides too little guidance on what visual evidence to preserve or how credit should be assigned across latent tokens. To bridge this gap, we propose ReaLVR, which brings visual-evidence supervision to the model's own free-running latent trajectories. ReaLVR contrasts correct and model-generated wrong answers to determine where stronger supervision is needed, and relevant and mismatched visual evidence to specify what to preserve. Across three model families, ReaLVR consistently outperforms evaluated LVR baselines, achieving the highest five-task average of 63.7% on Qwen2.5-VL-7B. Crucially, we are the first to scale visual reasoning in latent space, showing that our framework continues to deliver robust improvements at frontier model scales up to 235B. Further analyses show more question-sensitive latent-token positions, stronger alignment with relevant visual regions, and greater fixed-context dependence on the most attended latent tokens.

MeMo (Zanzotto et al., 2025) is a recent language-model architecture that stores associations between token contexts and next tokens in a correlation matrix memory. In this work, we study its single-layer form and show that its ideal retrieval rule is a multiclass classifier based on the positional Hamming kernel. The MeMo architecture represents both the sequence features and the output labels with Gaussian random codes. Its score is therefore a doubly randomized sketch of the ideal classifier. Under independent input and output codebooks, we bound the errors introduced by context sketching and output decoding, characterize their dependence on model and data parameters, and give a margin-based guarantee for recovering the ideal prediction. Controlled simulations support the trends predicted by the analysis. On a restricted WikiText-2 next-token task, we compare single-layer MeMo with classical baselines and show that it can offer a useful trade-off among predictive accuracy, memory, and throughput, particularly on a GPU, where its matrix operations can be parallelized.

Motor brain-computer interfaces (BCIs) aim to decode motor intention, enabling people with paralysis to control external devices. Neural motor decoding typically learns task-specific mappings from neural activity to kinematics, yet remains constrained by scarce paired neural-action data. We propose BrainVLA, a framework that enables neural motor decoding by drawing on a pretrained vision-language-action (VLA) model through language-mediated alignment. BrainVLA mitigates reliance on scarce paired neural-action data by leveraging VLA policies. We first construct VLA-compatible datasets including paired neural activity, action signals, language instructions, and rendered visual observations. Then, we adapt the OpenVLA-OFT policy to the target action spaces through LoRA fine-tuning. To establish an effective interface through which neural activity can convey motor intention to adapted VLA policies and guide action generation, we train a neural encoder via neural-language alignment, using language representations as semantic targets to capture latent motor intent from neural activity. The resulting neural representations serve as an endogenous intention signal to guide VLA policies to generate executable actions, while visual observations provide complementary information about the evolving task state. BrainVLA is evaluated on two neural motor datasets with different action dimensionalities using causal rollout decoding. It outperforms the evaluated baselines in cross-session decoding $R^2$ and task success rate, while demonstrating high training data efficiency. These results establish a route for neural motor decoding to draw on large-scale robotic priors through brain-conditioned VLA policies.

Recent work incorporates reusable skills distilled from past interactions into multimodal agent training, providing procedural guidance for long-horizon planning and tool use. However, policy optimization in these methods remains driven primarily by sparse outcome rewards, providing little supervision for intermediate decisions. Rubric-based rewards address this limitation through explicit intermediate criteria, but reliable rubrics are difficult to construct at scale and often disconnected from the procedure followed by the actor. We observe that a well-structured skill naturally specifies both how to act and what successful execution should achieve. Based on this insight, we introduce SkillRubric, which represents each skill through aligned actor-facing guidance and an evaluator-facing rubric. A multimodal verifier evaluates skill-defined goals using screenshots and tool outputs, assigning completion and progress rewards to the responsible turns. We further introduce an alternating co-evolution scheme that validates guidance revisions through paired rollouts under a frozen policy and rubric revisions offline under fixed guidance. Experiments across diverse multimodal agent benchmarks demonstrate consistent performance gains, while controlled paired rollouts further show that evolved skills provide more effective guidance for planning and tool use than their preceding versions.

Transformers process tokens without any inherent notion of order, making positional encoding a fundamental requirement rather than an architectural refinement. Rotary Position Embedding (RoPE) has become the default positional encoding in modern language models, yet it is heavily biased toward nearby tokens. Existing alternatives have been evaluated under different settings, leaving the literature fragmented and without a clear replacement. We bring structure to this landscape by examining a specific weakness of RoPE: its slow frequency bands, whose wavelengths exceed the training context and expose models to unseen angles during extrapolation. We therefore introduce Data aware RoPE (DaRoPE), which preserves standard RoPE on the fast bands but replaces absolute position on the slow bands with bounded coordinates learned from contextual representations. Therefore, the slow-band geometry depends on the data rather than only on positional distance. We compare representative encodings under matched conditions across synthetic tasks, symbolic music, genomics, neural signals, and language models spanning 124M to 50B parameters. Across these experiments, DaRoPE leads on non-text benchmarks, mitigates recency bias, while remaining best or on par in language modeling and length extrapolation. Moreover, the learned coordinates also make the mechanism interpretable, revealing how attention layers leverage contextual information beyond token distance. Together, these results support DaRoPE as the best overall default among the evaluated methods, when there is no domain-specific reasons to prefer another.

Long-context LLM serving is increasingly bottlenecked by decode, where large KV caches limit batch size and underutilize GPUs. Sparse KV cache offloading expands effective capacity by storing most historical KV blocks in CPU DRAM and recalling only selected blocks on demand. However, we find that existing offloading systems shift the bottleneck to CPU-GPU recall I/O: recall volume varies widely across layers, decode steps and requests, while headwise sparse selection fragments recalls into many small PCIe transfers. This paper presents PulseInfer, an I/O-centric sparse KV cache offloading system. PulseInfer hides variable recall latency with interruptible layer-wise scheduling, adapts offloading decisions with IO-Adaptive Offloading Admission, and coalesces fragmented transfers using SoloHead sparse selection and a gather-scatter I/O engine. Implemented on SGLang, PulseInfer improves decode throughput by up to 4.7x over SGLang and 2.6x over the best existing offloading baseline, while reducing TPOT by up to 76% and preserving near-lossless accuracy.

Formal verification can play a key role in ensuring the reliability of Deep Neural Networks (DNNs) deployed in safety-critical systems. Modern DNN verifiers employ a branch-and-bound framework, which alternates between branching (splitting into smaller subproblems) and bounding (pruning subproblems) to efficiently explore the verification space. However, existing branching heuristics make greedy decisions based on static scoring functions. They do not anticipate long-term efficiency or leverage the growing availability of verification data to improve performance. This work introduces RSB, a reinforcement learning framework that learns to refine baseline branching heuristics. It trains an actor-critic architecture to maximize cumulative future rewards rather than immediate scores. The actor generates attention weights from observations of raw neuron features and learned graph embeddings, which rescale baseline heuristic scores to guide neuron branching. Evaluation on 600 challenging instances demonstrates that RSB consistently outperforms state-of-the-art branching heuristics, solving 11% more instances while reducing branch exploration by 50%.

Parameter space symmetries and conservation laws play an important role in understanding the loss landscapes and implicit biases of neural networks. Inspired by Noether's theorem in physics, prior works have sought to derive conservation laws under gradient flow from parameter symmetries, but the scope and limitations of this connection remain unclear. We develop a unified geometric framework that clarifies the precise relationship between the two notions, including the conditions under which symmetries correspond to conservation laws. We introduce a notion of compositional identifiability and use it to establish a general inheritance principle for complete characterizations of symmetries and conservation laws in multilayer networks. We apply the framework to multi-head and grouped-query attention, polynomial neural networks, and square deep linear networks.

Recent advances in reasoning backbones have empowered large language model (LLM)agentstotackle complex, multi-step tasks. However, as reasoning horizons grow, inconsistent internal beliefs induce intermediate errors that cause agents to drift from their goals. This limitation also persists in REFLACT, which reflects only on the end-goal at each step without explicitly considering intermediate sub goals. To address this problem, we propose SGG-ReflAct (Sub-Goal Guided Re flAct), a reasoning backbone that integrates sub-goals generated through a single path LLM planner into the reflection process. We further extend this framework to BeamSGG-ReflAct, which replaces the single-path planner with a beam search based LLM planner for structured plan exploration. We run experiments on ALF World, ScienceWorld, and Jericho with multiple LLM models. SGG-ReflAct out performs REFLACT in nearly all settings, achieving best success rate gains of 14.9 percentage points on ALFWorld and 8.0 percentage points on ScienceWorld with Llama-3.1-8B-Instruct. Our experimental analysis shows that SGG-ReflAct re duces hallucinated actions and achieves its largest gains on procedurally ordered tasks. Furthermore, experimental results with BeamSGG-ReflAct show that the backbone's effectiveness depends on plan quality: explicitly specifying the re quired operations recovers gains that plan searching alone cannot achieve. These results demonstrate that SGG-ReflAct offers a practical and highly effective rea soning backbone, enabling LLM agents to achieve reliable performance in com plex, long-horizon tasks through easy integration.

Video-capable vision-language models score above 80\% on popular benchmarks yet struggle with spatial-temporal binding: associating the right action with the right person at the right moment. We introduce ActionLens, a diagnostic benchmark of 6,701 multiple-choice video questions spanning five targeted diagnostics: transition detection, actor-specific identification, concurrent action binding, directed interaction reasoning, and gaze detection. Ground-truth answers are derived deterministically from 1.58 million per-second, per-person annotations. Fourteen rounds of human quality engineering raised answer clarity from 53% to above 90% human accuracy. Across 20 VLMs, the full-set leader scores 68.8%; on the human-reviewed subset, it scores 65.9% versus 91.0% for the pooled human reference. Gaze detection remains near chance against 89.6% human accuracy. On actor disambiguation, reference-interface controls show that relational descriptions recover 5.55--13.25 points over static coordinates, confirming a substantial numeric-parsing penalty; yet visual boxes still lead every model by 1.15--6.50 points, exposing a residual unboxed actor-resolution gap. A binding-trap analysis shows models systematically select the wrong actor's action. ActionLens provides diagnostic measurements of these distinct failure modes across model families and scales for direct comparison. We release all data, code, and evaluation scripts at https://anonymous.4open.science/r/lmms-eval-2276

Memory is essential for enabling LLM-based agents to maintain coherent, personalized behavior over long-horizon interactions. However, existing memory systems share a fundamental limitation: they never proactively test their own memory, repairing it only after real queries expose weaknesses. This reactive paradigm means every retrieval failure corresponds to a real interaction in which the cost has already been paid. We propose MemDream, a framework that enables self-probing memory evolution for LLM agents. Our framework periodically enters offline dream cycles where three specialized agents (Dreamer, Analyst, Consolidator) collaboratively probe, diagnose, and repair the memory graph before failures occur. A policy trained via Group Relative Policy Optimization learns which repair operations produce durable retrieval improvements, while a soft decay mechanism provides reversible forgetting driven by the same anticipatory signal. Experiments on LoCoMo and MemoryAgentBench demonstrate that MemDream improves answer F1 by 4.5 points on LoCoMo and achieves a 9.1-point higher overall score on MAB over the strongest reactive-evolution baselines.

Ontology Learning (OL) from text has advanced with the emergence of Large Language Models (LLMs), but it remains challenging due to the limited availability of annotated training data and the difficulty of adapting LLMs to perform OL effectively. We address this via APOLO - Automatic Prompt Optimization for Ontology Learning, by casting OL as an explicit prompt optimization problem over LLM modules. To obtain training data, we employ a multi-agent system that generates text-ontology pairs from existing expert-curated ontologies. We then propose two ontology learner architectures: a greedy and an autoregressive learner, and optimize both using GEPA, a greedy evolutionary prompt optimizer built on DSPy. Experiments on two ontologies - a biomedical (DOID) and a plant ontology (PO) show consistent improvements after optimization across nearly all model and mode combinations, with autoregressive learners achieving the largest gains. Our results demonstrate that prompt optimization is a viable and lightweight alternative to fine-tuning for OL, and that the autoregressive formulation better captures ontological structure than the greedy approach.