The Inference Report

Researchers use large language models as agents in social simulations, where agents are prompted with observations and background information intended to shape their actions. Yet LLM agents have been shown to substantially change their behavior under meaning-preserving prompt variations, creating a validation problem for social simulation. One plausible hypothesis is that this brittleness reflects a lack of grounding and can be reduced by enriching the context provided to the agents. Whether richer context improves robustness remains unclear, in part because context richness and robustness are rarely defined in ways that support controlled comparisons. We address the first need with controlled context enrichment, a set of prompt substitution rules that systematically enriches distinct types of context: neutral padding, generic social background, descriptive decision grounding, and normative decision grounding. We address the second by introducing selective robustness as a validation criterion: insensitivity to meaning-preserving prompt variation combined with a directional response to decision-relevant changes in the social world. We demonstrate the framework in a Prisoner's Dilemma-like game, showing how it can be used to identify which forms of context enrichment produce more desirable and robust agent behavior. We find that these forms of enrichment produce distinct behavioral patterns that vary between the two models tested.

Reconstructing full flow fields from sparse sensors requires prior assumptions about unmeasured regions and physical variables. Yet it remains unclear when learned models outperform classical estimators and which aspects of their predictions can be trusted. FlowReconBench evaluates ten learned and five classical methods on three regimes: a laminar cylinder wake across Reynolds numbers, two-dimensional Kolmogorov turbulence and a large-scale three-dimensional (3D) isotropic turbulence. Methods share leakage-controlled splits and sensor draws and are scored on per-channel error, ensemble spread, calibration, spectra and cost. The leading method changes for each regime. Classical methods perform strongly in low-order and dense-sensing regimes. On the cylinder wake, linear stochastic estimation and gappy proper orthogonal decomposition lead, including for unobserved pressure. In 3D, inverse-distance weighting matches the best learned methods on observed channels. Data-driven methods become superior at decreasing sensor density or poor sensor observability. On Kolmogorov flow, continuous ranked probability score and error yield different method rankings: generative models offer broader but underdispersed ensembles, while deterministic models prioritize pointwise error. Generative models balance this tradeoff through their sampler budget, with a larger budget improving ensemble quality at higher computational cost. In 3D, scaling tests demonstrate that queryable methods are flat in memory, while full-grid construction methods run out of memory. Observing only 2 of 4 fields leaves the problem kinematically underconstrained, yet a latent model can reproduce structural similarities to the ground truth.

Self-Driving Chemical Laboratories (SDCLs) are moving chemical experimentation toward embodied automation, where machines must perceive laboratory scenes, reason about experimental states, and act under physical, procedural, and safety constraints. Although Multimodal Large Language Models (MLLMs) are increasingly considered as embodied brains for such systems, existing laboratory benchmarks mainly focus on safety, anomaly detection, or general scientific understanding, leaving their execution-oriented embodied capabilities underexplored. To fill this gap, we introduce BEAKER, a Benchmark for evaluating Embodied Actionability, Knowledge, Experimentation, and Reasoning in SDCLs. BEAKER contains 1,000 expert-curated question-answer samples across 11 image and video task types, covering key judgments required for chemical laboratory execution, including affordance grounding, trajectory planning, equipment, state, and safety recognition, apparatus setup, operation and step reasoning, and temporal procedure reasoning. All questions and annotations are manually created and reviewed by chemistry and embodied Artificial Intelligence (AI) experts, without AI-assisted generation or annotation. We evaluate 28 general-purpose, science-oriented, chemistry-specialized, and embodied MLLMs in an offline, single-turn setting. Results show that current models can follow coarse experimental procedures in videos, but still struggle to ground actions into operable regions and feasible trajectories, to recognize fine-grained laboratory apparatus, states, and safety norms, and to order steps in mid-length procedures. BEAKER thus reveals a clear gap between coarse procedural understanding and the precise, action-grounded perception required for laboratory execution, providing a diagnostic benchmark for future MLLMs in SDCLs.

AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.

Humanoid robots have the potential to automate a wide range of physical tasks. Progress toward generalist humanoid autonomy, however, remains constrained by a data bottleneck: mapping pixels and free-form language to low-level joint commands requires paired supervision, and no corpus of such data exists at sufficiently large scale. Some approaches rely on human teleoperation and wearable capture devices, which require at least an hour of human time to produce an hour of data. We present HALO, a humanoid autonomy system powered by a pipeline that synthesizes vision-language-motion data across a thousand procedurally generated homes. The corpus scales with compute rather than human labor. HALO acquires a repertoire of physical skills, including sitting on chairs, navigating to named objects, stepping sideways through half-open doors, crouching and crawling beneath tables, and pushing boxes. It orchestrates these skills to solve composite, long-horizon tasks. Without ever seeing a real home during training, HALO transfers zero-shot to the physical world, achieving a 91% success rate across 360 trials in six unmapped real-world environments. See https://apple-aiml-research.github.io/ml-halo/ for code and videos.

The introduction of generalist robot policies has enabled a single, large model to learn to perform a variety of tasks from a vast pre-training corpus. While these models often exhibit powerful zero-shot generalization capabilities, they typically require an adaptation phase to achieve sufficient performance in deployment. Iterated self-improvement via reinforcement learning (RL) has demonstrated promise in enabling this adaptation, yet such approaches have been largely explored in single-task settings - where the goal is to simply improve on a narrow deployment task - which lies in stark contrast to the multi-task capabilities of the base policy. In this work we ask whether generalist robot policies are capable of multi-task, generalist improvement in deployment - targeting many diverse tasks simultaneously - and what algorithmic recipe enables this. To answer this question, we first develop a set of structured evaluations to test generalist improvement capabilities, and systematically evaluate algorithmic and practical design choices regarding value estimation, policy extraction and resource allocation. We show that not only can generalist policies be improved online across diverse tasks, but that the right recipe can enable generalization and increase performance on tasks not considered during online training. We characterize when this generalization occurs and show that both the training task distribution and the use of value-based RL are critical. To the best of our knowledge, this is among the first evidence for online generalist improvement for robot foundation models.

Knowing that a pedestrian is occluded is useful to a robot even when the detector cannot recover the hidden pixels, but occlusion estimators are normally trained on occlusion annotations that deployment settings rarely have. We show that occlusion can be read off two distance estimates a camera-only pipeline already computes. Apparent height is largely unchanged by occlusion, because the detected box still spans the whole person; monocular metric depth is pulled toward the occluder, because the nearest surface inside the box is what a depth map reports. Only their ordering matters, so no focal length, stature, or depth scale is needed, provided the depth model's error is close to a pure gain, a condition that can be checked without labels. The score detects heavy occlusion on KITTI with AUC 0.909 (0.888 over all annotated pedestrians, including those the detector misses), requires no training and no occlusion labels, and is not a proxy for range: within fixed distance bands it holds AUC 0.894-0.925 while true distance alone sits near chance. On synthetic CARLA scenes with measured visibility it reaches 0.819, against 0.644 for distance alone. We arrived here after measuring three standard architectural responses to occlusion, depth fusion, occlusion-simulating augmentation, and attention, none of which improved detection outside its training distribution in the configurations we tested, and after finding that a detector trained on German roads loses half its accuracy on self-annotated street scenes in Lagos, Nigeria. Signals that need no local annotation are worth more in settings that benchmarks do not represent.

The Model Context Protocol (MCP) enables LLM agents to access external tools and data. Studying its software ecosystem requires connecting server interfaces, the projects that configure these servers, and their development activity. However, this information is scattered across directories, package registries, configuration files, and repository histories. We present MCP-Eco, a relational dataset that integrates these sources and links consumer configuration declarations to candidate server projects. Its frozen release contains 16,880,899 records across 19 Parquet tables, covering 78,302 candidate server projects, 16,165 candidate consuming repositories, and 9,200 configuration declarations. Of these declarations, 5,975 resolve to 5,138 distinct repository-server pairs. The dataset also includes tool, resource, and prompt descriptions, declared dependencies, and associated development records, including pull requests, reviews, issues, and commits. Provenance fields distinguish directory-reported capabilities from those observed through live probing and document discovery sources and declaration-matching methods. These connections support research on MCP interface design, configuration practices, and repository maintenance, while enabling practitioners to inspect tool schemas and existing integration configurations. The dataset and accompanying analysis scripts are publicly available at https://huggingface.co/datasets/mcp-research-alliance/mcp-ecosystem Demo video: https://www.youtube.com/watch?v=N7TqnNneSfU

Automation is often justified as relief from dull, dirty, and dangerous work. I examine how these conditions relate to occupational AI exposure using 21 O*NET items across 754 US occupations. Higher exposure is associated with lower condition scores under a large language model (LLM) measure and AI Occupational Exposure (AIOE), with equal occupation weights and employment weights. Adjustment for physical demands, required education, training and experience, and wages reduces the composite association. The adjusted confidence intervals include zero for LLM exposure and remain below zero for AIOE. Separating repetition from restricted discretion also changes what the combined Dull score conveys. These comparisons show how item groupings and occupational characteristics shape the interpretation of exposure estimates. Occupational profiles offer a starting point for selecting contrasting research settings. Establishing desired automation or realized relief requires evidence about workers' preferences and specific interventions.

The best classical optimizer for a continuous black-box problem depends on its budget and landscape. Existing routers decide from the dimension and budget, past improvement or features fitted to one benchmark, and few test an optimizer's implicit assumptions online. AMM (Adaptive Multi Mode) routes a known budget among quasi-Newton, trust-region, coordinate, multistart and evolution-strategy optimization engines with online probes of curvature, plateaus, separability, kinks, basins and fractal structure. Each probe has a default, a dimension-by-budget validity range and a cost charged to the budget. In its pre-registered development loop on bbob, 32 of 36 predictions held over the last 16 confirmations. AMM was then tested once on held-out bbob instances 101-110 (D=2-40, 10*D-1000*D) against up to seven competitors. It has the highest target share in 382 of 600 function-dimension-budget cells (236 outright, 146 tied) and, by a descriptive posterior rule, is credibly behind in 35, mostly at 10*D and 30*D. At every budget it also leads the controls built from its own engines (the best engine alone, the best engine per dimension and budget, and a greedy schedule), narrowly at 10*D and 30*D, and nevergrad's NgIohTuned. In pre-registered runs its lead over IPOP-CMA-ES holds under affine and box transformations, on box-constrained bbob and, pooled, on rotated f2-f4, but rotating the separable f3 and f4 removes most of the lead there. On held-out MA-BBOB its final share at 2000*D is level with IPOP-CMA-ES's and its anytime area lower. We published AMM as an open-source Python package (pip install amm-optimizer; https://github.com/BjBodner/amm-optimizer).

LLM agents use context compaction to manage growing interaction histories, but summarizing long histories with large models can introduce substantial delays. Compaction therefore requires balancing information retention against latency. We observe that structured summary slots can retain information when generated independently. Based on this observation, we present SpecCompact, a framework that accelerates compaction through slot-parallel semantic speculation. SpecCompact fine-tunes a small drafter and generates slots in parallel. The target model verifies each proposed line and repairs rejected content, while subsequent drafting uses the updated slot prefix. System optimizations further reduce repeated context processing and model calls. Across AppWorld and $τ^3$ benchmarks with two target models, SpecCompact achieves $2.04$--$5.63\times$ compaction speedups over Hermes-style summarization. Separate downstream evaluations reveal model- and task-dependent performance trade-offs. A user study within Hermes further shows an $8.68\times$ median compaction speedup.

Error correcting codes are a fundamental component of communication systems, enabling reliable data transmission over noisy channels. Recently, neural decoders have demonstrated notable performance gains over classical iterative algorithms. In this paper, we propose a Generalized Belief Propagation Network (GBPNet) that preserves message-passing symmetry while enhancing the decoder's learning capability. The architecture uses cascaded variable-node updates and learned element-wise message scaling while retaining the Tanner graph structure. We also introduce two training strategies, progressive training and high-SNR training, that improve decoding accuracy. Experimental results on several linear codes show that GBPNet achieves lower bit error rates than conventional belief propagation and the evaluated state-of-the-art neural decoders.

In this work, we evaluate a simple inference-time scaling approach based on \textit{turn-level best-of-N sampling} (\method) for large language model (LLM) agents on complex, long-horizon tasks. We instantiate this using a \textit{generator-selector} framework with two LLMs; at each turn, a \emph{generator} model generates $N$ candidate actions and a \emph{selector} model selects one of them, which is then executed in the environment. This paper is closely related to the work of \cite{zainullina2025guided} who used critic models for guided search at inference time. Different from their work, we evaluate if off-the-shelf, open-weight LLMs can be effective at comparing different candidate actions directly. Experiments on long-horizon terminal tasks show that \method can lead to significant performance improvements, both in terms of accuracy (5--10\%) and consistency across repeated attempts. Analysis of agent trajectories reveal a shift in agent behaviour towards desirable patterns such as reduction in number of attempts to solve the task, improved task instruction following, debugging and testing.

Agentic workflows enable AI agents to perform recurring repository tasks on schedules or in response to repository events. Existing studies examine workflow definitions and outputs, but provide limited evidence of what happens during execution. We present GH-AWTrace, a dataset linking 25,014 verified GitHub Agentic Workflows across 5,051 repositories to 4,112,526 run records and 562,877 nonempty execution log archives. The dataset preserves execution evidence subject to GitHub's log-retention limits and links it to repository metadata and generated workflow configurations. Illustrative cases show how the logs expose tool denials, provider failures, changes in an agent's approach, and delegation among agents. These observations enable research on execution failures, tool use, and the relationship between reported run outcomes and visible task progress. Coverage varies across workflows, and the logs provide partial observations rather than complete agent trajectories. Video: https://youtu.be/KIlfmZhHxtA

Automated identification of individual animals is moving from small curated studies toward shared benchmarks and reusable foundation models. This progress is useful, but current evaluation leaves the evidence behind a match unresolved, i.e., high retrieval accuracy does not show whether a model used animal features or identity-correlated shortcuts. We address this with Counterfactual Foreground Retrieval (CFR), an evidence-aware evaluation protocol. CFR evaluates the same query-gallery task under different views: full image, foreground-only, and shortcut-only. The full-image measures standard retrieval performance; the foreground-only measures foreground-supported retrieval; and shortcut-only view measure how much identity signal is recoverable from background, or silhouette. This turns a retrieval score into an evidence profile: how well the model performs on the original images, how much of that performance remains when only the animal is visible, and how much identity signal remains recoverable when the animal is removed. We release around 0.5M masks for 34 datasets and code for mask generation and CFR evaluation.

Hierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, already-established techniques interact. We evaluate 38 tasks from the Massive Text Embedding Benchmark (MTEB) using four transformer-based embedding models, covering 1D and 2D hierarchical representations combined with 32-bit, 8-bit, and binary quantization, for 168 configurations per task. We report each efficiency dimension separately: theoretical per-vector storage and end-to-end evaluation time. Under a Friedman test with Nemenyi post-hoc analysis, int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking and costs at most 1.5 percentage points overall, whereas binary quantization is significantly worse across every task category. Halving the embedding dimension once is not statistically detectable above $m=16$; two successive halvings always are. Combining an 8x dimensionality reduction with int8 yields a 32x reduction in per-vector storage at 94.4% of baseline effectiveness. Because evaluation times were measured once, we treat them as indicative: we find no evidence that quantization speeds up evaluation, whereas sampling from intermediate layers shortens it substantially. Retrieval degrades far faster than every other category, so we report findings per task category.

Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage budget. We study whether coding agents can carry out this step autonomously. We formulate compression as a terminal task in which an agent receives a source checkpoint, storage limit, inference environment, and time budget, then constructs compressed weights and executable inference code. The agent controls its algorithm and experiments; perplexity feedback guides search, and independent verification checks the final model. Across six agents and six model--size settings, 29 of 36 submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16. Recorded experiments reveal how agents achieve these outcomes and where improvement breaks down. Agents lower PPL in 21 of 24 revisions at nearly fixed storage; complete-model feedback reverses 17 of 79 prior local rankings; and generation and inference tests expose failures that likelihood alone misses. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression, a measurable component of efficiency-oriented self-improvement.

We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.

Computer-use agents (CUAs) have recently achieved superhuman performance on several long-horizon benchmarks. However, their computational cost limits their use in long-horizon tasks and real-time domains such as video games. A key bottleneck is representing each screenshot as thousands of visual tokens, which accumulate into long contexts as agents reason over extended interaction histories. We observe that most of these visual tokens are either semantically irrelevant to the agent's next action or temporally redundant across screenshots. Building on this observation, we introduce Visual Token Routing (VTR), a joint training framework for learning sparse visual representations that preserve the information needed for action prediction. In VTR, we train the vision-language model together with a lightweight spatio-temporal router that examines each image region in the current and previous screenshots and decides whether its visual token enters the language model. We optimize the router using an adaptive loss to keep the number of visual tokens as small as possible while minimizing performance degradation. With just 2,000 training steps, VTR matches the performance of the full-resolution model on OSWorld, CUA-World, and video game action prediction, while using 2.5-5.8x fewer visual tokens. These savings reduce inference cost by up to 3.3x and speed up prefill by up to 2.9x. VTR also outperforms lowering the image resolution and other baselines for dropping tokens, by up to 12.8 points on OSWorld. Overall, VTR enables computer-use agents to process only the visual tokens they need, allowing them to be used in long-horizon tasks and real-time domains at a fraction of the computational cost. Code and models are available at https://pranjal2041.github.io/vtr

Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.

Faithfulness requires that algorithmic explanations accurately reflect the model's behavior. Yet existing operationalizations of faithfulness can disagree about whether the same explanation is faithful, leaving developers and auditors to choose among conflicting evaluations. In this paper, we investigate disagreement among evaluation criteria and offer a practical path forward for evaluating explanation faithfulness. We first identify how operationalizations of faithfulness differ. We introduce a faithfulness functional that locates their disagreements in what an explanation claims about the model, which comparisons bear on that claim, and how agreement across those comparisons should be aggregated, and we reconstruct existing evaluation criteria within it. We argue these reconstructions support understanding faithfulness as a common though incompletely specified relation between the model and explanation, so the differences reflect choices left open by the general requirement. Justifying an operationalization then comes down to justifying these choices, which, on our view, draws on explanatory, institutional, and normative considerations. We support this argument with a case study of adverse action notices, reconstructing the evaluation implicit in an industry-proposed procedure and identifying which of its choices the setting justifies and which remain open. Our work gives auditors and practitioners a way to navigate disagreeing evaluations of faithfulness. It also explains why the literature should expect multiple operationalizations, while showing how concrete settings can narrow the choice among them. More broadly, our analysis shows how a shared requirement of faithfulness can support contextually justified evaluations despite disagreement among existing metrics.

Grain growth during heat treatment strongly affects the microstructure of polycrystalline materials and, consequently, their mechanical properties. Although numerical methods such as front-tracking or front-capturing simulations can accurately describe grain growth, their high computational cost limits their use for long or non-isothermal thermal cycles. In previous work, grain growth was successfully predicted under isothermal conditions by treating grain size distributions as time series and forecasting their evolution using recurrent neural networks, allowing complete grain growth trajectories to be generated in seconds rather than through costly PDE-based simulations. In the present work, this forecasting framework is reformulated to handle non-isothermal conditions by explicitly incorporating thermal information into the model architecture. Grain size distributions are encoded using a convolutional neural network, thermal descriptors are embedded through a dedicated multilayer perceptron, and temporal evolution is predicted using a recurrent network. Instantaneous temperature and thermal rate are provided at each time step, allowing the model to adapt its predictions to changing thermal conditions. The framework is trained and evaluated on datasets generated from validated front-tracking simulations across a wide range of non-isothermal thermal paths. The results show that the thermally conditioned model accurately reproduces grain growth dynamics under complex thermal histories, achieving, in k-fold cross-validation over all simulations, mean relative errors of 2.99% (one-hour cycles) and 4.51% (four-hour cycles) with coefficients of determination exceeding 0.99, and mean relative errors between 1.91% and 14.66% on ten simulations following four unseen thermal histories of increasing complexity.

Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.

Ask SD3.5-Medium for twelve of an object and then, from the same starting noise, for twenty-five. The picture keeps its layout, but the count changes about as much as it would under a new random seed, and in no consistent direction. We study what number words do once a request passes about ten objects, in five open text-to-image models. Holding the object and the starting noise fixed, we change one thing at a time, and people who never see the request count 3,088 of the images. Every model follows the number for small crowds, but only up to a point, its reach: SD3.5-M and FLUX.1-dev stop following it soon after ten, while FLUX.2 [klein] and Qwen-Image follow it loosely up to twenty-five. Beyond SD3.5-M's reach, the coarse part of the starting noise fixes the composition of the picture but not how many objects it holds: redrawing only the noise finer than an eighth of the image keeps the layout as well as a new number does, and already moves the count two-thirds as far. The number word still reaches the denoiser and acts in its first steps, but nothing we tried gives it a clear direction. Shown instead as a noised grid in the starting latent, the count is drawn: people find it in 59% of images, against 4% when it is only told. Past ten, an exact-match score barely separates a model that follows the number loosely from one that ignores it, and a relative-error score can even prefer a constant guess of sixteen. A same-noise test with a few requests per object tells them apart.

Large language model (LLM) assistants increasingly help users with everyday activities that involve time-dependent constraints. However, users may not mention these constraints in later requests, requiring assistants to recognize and adjust their advice accordingly. To evaluate such capability, we introduce TASC (Temporal Action Safety in Conversation), a benchmark that evaluates whether LLMs apply time-dependent constraints when responding to conversational requests. TASC consists of 3,418 inputs derived from 128 templates, along with a separate set of 33 scenarios designed to compare models' ability to calculate elapsed time with their ability to apply that information. Across 13 models, assistants account for time-dependent constraints in only 30.5% of cases where the proposed action violates a rule, while rarely warning users unnecessarily. More importantly, models fail to apply these constraints in 62.1% of cases despite correctly calculating the elapsed time. Explicitly providing the timing resolves only about half of these failures, revealing a gap between temporal reasoning and its application in conversational advice.

Under stochastic decoding, a single context can elicit many different answers from a language model. Characterizing this variation typically requires many expensive autoregressive rollouts, each revealing only one realization. We ask whether the activations produced while reading the context can directly forecast the model's rollout activations. We train context--answer metamodels that map a context activation sequence to point or distribution forecasts of mean-pooled answer activations. We find that (i) the four metamodel families we study forecast mean-pooled answer activations with high fidelity, and the distribution metamodels also capture rollout variability; (ii) performance improves as the training set grows to 500,000 prompts, and the metamodels generalize zero-shot to out-of-domain datasets; (iii) their forecasts retain behavior-predictive signal; and (iv) context activations from Qwen2.5 predict mean-pooled answer activations from Qwen3.5 and Gemma 4, suggesting that activation forecasts can transfer across models. Together, these results suggest that context activation traces can support scalable forecasts of behaviors in LLM answers before the first answer token is generated.

Looped Transformers increase effective depth by re-executing shared layers, but three limitations hinder their adoption: memory grows with the loop count, compute can only be adjusted in whole-block steps, and the loop count is hard to change after training. We propose LooM (Loop-invariant Memory), which addresses each in turn. In a causal encoder-decoder (CED), LooM loops only the decoder layers after the first, which read a global memory written once per token, and each of them publishes the key-value (KV) entry for a token only on its first visit, so the cache does not grow with the loop count. These layers loop independently, providing every budget from 3 to 21 layer executions per token, and one model trained on randomly sampled budgets covers all of them without a router. With enough loops, LooM outperforms a single-pass baseline trained on the same data on language modeling, zero-shot accuracy and long-context perplexity, while using the same cache size. Our analysis further shows that the training budget distribution decides what extra depth buys. These results suggest that a CED is a natural place to loop: memory is written once, and depth becomes a budget chosen at deployment.

Current generative AI alignment relies on a myopic paradigm that prioritizes immediate conversational satisfaction. By relying on short-horizon feedback, this approach ignores the profound, longitudinal externalities of persistent human-AI interaction, which can lead to adverse outcomes such as cognitive atrophy, social erosion, and diminished human agency. To address this, we propose a paradigm shift toward long-horizon alignment that accounts for how AI reshapes human experience over time. To this end, we map the tensions between near-term utility and long-term costs of AI use across four facets of human well-being to reveal outcomes that current reward models fail to capture. Ultimately, we advocate for moving beyond simple preference optimization toward alignment strategies that actively foster long-term human well-being by navigating these complex temporal scales.

We present a toolchain for development of ML models on AWS EC2 F2 cloud FPGA infrastructure from high-level language, implementing a complete Vitis HLS to HDK bridge to AFI workflow that has been absent from the academic literature despite eight years of AWS FPGA availability. We introduce a four-level methodology (FP32 PyTorch, bit-accurate quantized Python, HLS C-simulation, physical FPGA) to aid in localizing errors, including FPGA-specific circuit errors. Together these tools allow coding agents to effectively implement FPGA designs from user directions, testing and debugging as needed. We document the deployment workflow, timing closure strategies, and practical HLS engineering decisions that proved crucial. We validate using TinyBERT fine-tuned for address Named Entity Recognition, implementing the embedding layer and the first of four encoder layers on the FPGA hardware. Our implementation streams weights from host DRAM, yielding poor performance compared to CPU/GPU implementations; we discuss forward paths.

Gated-delta linear attention layers such as Gated DeltaNet (GDN), Kimi Delta Attention (KDA), and GDN-2 are tuned for GPUs, where small tensor-core tiles and fast vector registers make per-channel decay rescaling and sequential in-block solves cheap. On Google TPUs, which concentrate compute in wide $256 \times 256$ systolic matrix units backed by a slower vector unit, porting these layers directly leaves half of each matrix tile empty and bottlenecks on vector operations. We present \textbf{Systolic Linear Attention (SLA)}, which reshapes gated delta attention around the matrix unit by widening keys to $256$ with grouped value heads, grouping decay gates into four $64$-channel tiles, and computing the chunk solve entirely with block matrix products in a single fused kernel. On TPU v7x, a full \texttt{SLA-R4} layer runs $1.6\times$--$2.5\times$ faster than the open-source Tokamax KDA kernel ($1.4\times$--$2.4\times$ with exact chunk inversion), driven primarily by head sharing and kernel fusion. Across a $2{,}048$-chip TPU v7x pre-training study of sparse Mixture-of-Experts (MoE) models from $6.5\text{B}$ to $63.1\text{B}$ parameters, SLA trains $11\%$--$24\%$ faster than KDA at similar perplexity, and reaches slightly lower perplexity than KDA in smaller matched-seed runs.

We study which training and calibration data help three open models solve ARC-AGI-2 tasks when the solver runs offline on four L4 GPUs for at most 12 hours. We compare data offline on public tasks and report 189 runs on the 120 hidden semi-private tasks. For the NVARC Qwen3-4B grid model, we built a corpus of 12,000 executable rules in a typed language and 95,993 tasks; a rule is admitted only if its behaviour on a fixed probe set differs from that of every admitted rule, and a task only if its pairs refute a shortlist of nearest competing rules and defeat 12 shortcut solvers. Continued fine-tuning on 1,920 real task records with revised four-GPU scheduling raised the semi-private score from 28.96 (mean of the released model) to 33.06 (mean of five runs, best 34.86); no checkpoint trained with synthetic tasks, even all 95,993 of them, improved on the real-task model. For the Qwen3.8-27B reasoning model, we fine-tuned on 1,036 of its own reasoning traces with correct final answers and compared sets that add traces of other models and synthetic tasks. With an inference procedure built around the model (a thinking cap enforced in the serving engine, a Python program checked on the demonstrations after each answer, a long first attempt per input), the score rose from 26.81 to 44.03, the level the ARC Prize Foundation measured for the base model without a time limit (42.36); most of the gain came from the inference procedure. For the Qwen3.8-Flash-Next mixture-of-experts model (125B parameters, 6B active), GPU memory limited concurrency. Pruning from 512 to 384 experts per layer by router-weighted activation saliency grew the key-value cache from 0.55M to 1.39M tokens; a 296-expert list adapted with ARC routing statistics, keeping experts for program tokens, plus an FP8 indexer cache, reached 1.96M. With a tool that checks programs, the model scored 56.67, our best result.

Choosing a quantum error-correcting code for a real device means scoring candidates with circuit-level simulations so slow that only a few hundred can be run. So every evaluation counts, and a search is judged by how many good codes it finds per paid evaluation. Quantum LDPC codes built from a few group elements protect many logical qubits, but good ones are rare in a huge space. We introduce GFlowQEC, to our knowledge the first generative flow network (GFlowNet) for quantum code design. It builds a code one group element at a time and learns to propose codes in proportion to a power of their score. We train it once on scores from earlier searches, freeze it, and let it propose lifted-product codes with 180 physical qubits, scored by their distances. Ranking ten of its draws per paid call, it finds more than twice as many new top-scoring codes (inequivalent to any it was trained on) as tuned parallel tempering, even when tempering starts from GFlowQEC's training codes. It finds more than three and a half times as many as a supervised screen ranking ten random candidates per call. When both rank a hundred, the screen beats GFlowQEC, and does best ranking GFlowQEC's draws. Soft Q-learning, which shares the GFlowNet's optimal policy, does as well at one noise bias and better at the other, so the gain does not depend on the GFlowNet's training loss. Trained proposers can make a budget of expensive evaluations go further when good codes are rare.

With the growing need for evaluating Large Language Models (LLMs) and their applications to speech, challenges persist in summarizing and evaluating conversations that lack a clear end goal. We introduce SAUCE -- a reference-free, fact-based evaluation pipeline for cross-lingual conversational speech summarization. It measures the accuracy and the fact coverage of a summary through the entailment between conversation and text. We compare SAUCE against several popular summarization metrics and demonstrate the effectiveness of capturing information loss due to transcription and translation error and identifying broken summaries. Crucially, unlike black-box LLM evaluators or dense embedding metrics, SAUCE is inherently explainable: it maps summary scores to discrete, verifiable facts, allowing users to pinpoint exact hallucinations or omissions. We illustrate how this interpretability helps developers systematically profile LLM behaviors and gives end-users an actionable tool to verify summary accuracy in noisy, real-world conditions. Preliminary investigations show SAUCE strongly align with human judgment.

Structure property prediction is a classical task in machine learning, ranging from molecular activity and materials properties to local functional responses measured in microscopy. For image-based regression, two choices define the information available to the predictor: the size of the image neighborhood and the spatial resolution retained within it. A small neighborhood can omit relevant surrounding structure, whereas a large, finely sampled neighborhood can introduce redundant or irrelevant information. Here, we develop a framework for constructing spatial descriptors for prediction of local spectra and hysteresis loops. We benchmark regression models, use SHAP maps to examine their spatial attribution, and explore the effect of the patch size and compression to establish the relationship between predictive accuracy and descriptor dimensionality. We then use a genetic algorithm to optimize spatial partitions and the assignment of mean, minimum, and maximum pooling operations. Each retained region contributes an explicit measurement of the image, allowing the search to balance prediction error against the number of descriptor values. Cartesian and polar constructions provide complementary ways to organize this information. The resulting descriptors make it possible to examine where spatial detail is needed, where regional summaries are sufficient, and how these choices depend on the predictor and the measured response. The same approach can be used for molecules and solids once the genetic algorithm rules for these are established.

Financial backtests of large language models (LLMs) are meaningful only if the model did not already know the outcomes it is asked to predict. Practitioners guard against this by trusting a vendor's stated knowledge cutoff, by asking the model what it knows, or by instructing it to answer as of a past date. To verify whether any of these safeguards can be trusted, we introduce FinDendro, a benchmark of 2,226 dated U.S. facts from 1999 to 2026, covering monthly macroeconomic series, month-end stock prices and Federal Reserve decisions. Each fact yields four prompts: a value question, Recall and Predict questions that pair each candidate value with a claim of knowing or guessing without a declared date, and a Predict question with a historical declared date. Controls for persistence and dominant outcomes keep chance accuracy near 0.27. Results show that for most frontier models, self-reports understate knowledge. GPT-5 answers 76% of controlled facts correctly but claims to know 38%, and answers 61% of the facts it says it does not know. Knowledge claims can persist beyond the measured knowledge boundary, by up to seven quarters, and a declared date removes most of the claims of prior knowledge while leaving accuracy unchanged. Backtests inside a model's training window therefore cannot be certified leak-free by self-report, instruction or stated cutoff. These findings motivate behavioral measurement of temporal knowledge boundaries, and FinDendro provides a benchmark and procedure for selecting backtest periods based on demonstrated factual knowledge.

Psychological harms from interactions with large language models cannot be monitored directly because they may emerge over repeated interactions and depend strongly on subjective individual user context. We propose measuring risk, not harm: detecting observable AI behaviors that constitute risk factors for psychological harm at the AI response level and aggregating these signals to support longitudinal monitoring. We operationalize eight such behaviors and construct a synthetic multi-label corpus of 47,062 AI responses and 373,873 response-factor labels, with targeted human annotations for validation and evaluation. We then compare trained embedding-based classifiers, direct LLM judging, and a classifier--judge cascade. All detectors are model-agnostic and operate only on LLM outputs, requiring no access to internal model states. We apply these detectors to several out-of-distribution human-AI conversation data sets from prior research. On held-out human-labeled data, the best cascade achieves macro precision of 0.86, recall of 0.76, and F1 of 0.79. On WildChat, it reduces LLM-judge calls by 97.1\% relative to direct judging. On DelusionEval, risk-factor detections cluster within conversations. Together, our results show that response-level behavioral risk-factor detection can provide a scalable measurement layer for longitudinal psychological-safety monitoring without conflating observable AI behavior with latent user harm.

Efficient benchmarking estimates a system's full benchmark performance from a small number of task executions. Standard approaches rely solely on the binary correctness of executions from a collection of already-evaluated reference systems. An agentic system, however, produces a rich trace for each execution that can reveal substantially more about its capabilities. Leveraging this information requires useful trace representations. We propose representing a trace by extracting a structured summary of it with an LLM and embedding the summary with a standard text encoder. These representations preserve correctness- and task-related signal while minimizing the presence of system identity-, model family- and harness-related signals that dominate raw trace embeddings. Given this representation for each task execution from all reference systems, we construct low-dimensional representations of the systems and predict a new system's score from its position among the reference systems. On SWE-bench Verified and Terminal-Bench 2.0, augmenting an item-response model with this geometry-based prediction reduces mean absolute error from 0.13 to 0.08 (Wilcoxon $p < 10^{-4}$) and from 0.14 to 0.12 (Wilcoxon $p < 10^{-3}$), respectively, when provided a single random executed task. Further, given access to the full set of reference systems, our approach dominates baselines for any given number of executed tasks on both benchmarks.

Hamilton-Jacobi (HJ) Reachability identifies the unsafe states of a dynamical system from which it can be forced into failure under worst-case disturbance. However, classical grid-based solvers scale exponentially with state dimension. DeepReach improves scalability by replacing the grid with a neural value function trained on the HJ-Isaacs (HJI) variational inequality (VI), but penalizes both sides of the residual equally. This allows the learned value to lie above the true one and unsafe states to be misclassified as safe. We propose Conservative DeepReach (ConDR) which aims to learn a subsolution of the HJI VI such that the learned unsafe set contains the true one. ConDR enforces the obstacle branch and terminal condition by construction, pushes the HJI residual to a safety margin $ε$ below zero with a one-sided loss, and prevents the degenerate solution of classifying everything as unsafe. We prove that the exact solution of the margin-shifted VI that ConDR trains toward is a subsolution of the original VI so its unsafe set contains the true one and its value lies within $ετ$ of the true value function. We compare against DeepReach and MADR in the established systems of Air3D, Air6D, a 9D multi-vehicle system, and a 13D quadrotor. ConDR reports the lowest rate of unsafe state misclassified and the lowest closed-loop collision rates with at most a small increase in intervention. On Air3D and Air6D, ConDR tolerates model mismatch that scales with $ε$. At matched false-negative rates, ConDR's unsafe sets are smaller than the baselines showing the safety is not obtained by trivial over-approximation.

Message-passing graph neural networks (GNNs) are limited in expressive power by the one-dimensional Weisfeiler-Leman test. Persistent homology (PH) provides topological descriptors that can complement GNN representations, but existing approaches typically compute persistence separately from the representation learning process. In this paper, we introduce Message-Passing Graph Persistence (MGP), a family of topological descriptors that incorporates message passing into graph filtrations. At each filtration step, node embeddings are updated using the current subgraph and associated with persistence events. These embeddings also determine the representatives of connected components during persistence computation. We prove that MGP is invariant under graph isomorphism and that it is strictly more expressive than the PH-based descriptors it extends. We evaluate MGP on an expressivity benchmark and on several graph classification and regression datasets. Experiments show that MGP outperforms other PH-based descriptors on most of these datasets.

Sample-efficient fluid intelligence requires agents that rapidly form working models of unfamiliar interactive environments. ARC-AGI-3 tests this across uninstructed visual games scored by Relative Human Action Efficiency, which penalizes exploratory waste quadratically. When frontier models read raw integer grids, serialization destroys spatial topology, and they spend hundreds of actions rediscovering basic boundaries and symmetries. We introduce SARBLOH, an open-weight agent built on the premise that spatial perception, not planning capacity, governs interactive reasoning on ARC-AGI-3. A zero-parameter O(HW) perception engine turns each frame, in 35 ms, into an object inventory with bipartite temporal tracking, a geometric briefing that separates proven invariants from affordance hypotheses, and an annotated canvas. Paired with a code-as-action runtime, the agent writes algorithmic hypotheses instead of low-level moves. Across eight presentations, grounded perception raises the rubric score from 3/28 to 25/28 points while cutting token use by 89.4%. On 23 held-out games, geometric invariants generalize with 100% accuracy, while single-frame affordance guesses fall to 6-50%, so active verification probes remain necessary. On the public set, Run A (briefing off) solved 143 of 183 levels in 8.2 hours at 59.4% single-pass efficiency; Run B (briefing on) reached 63.5%. Within the nine-hour budget on one Kaggle RTX PRO 6000 GPU, SARBLOH scored 16.53 on the hidden Kaggle evaluation, 13.7x the first Milestone winner.

A fast acceptance classifier could select revisions for a slower evaluator. We ask whether agreement on existing papers supports that use. We compare a classifier of word and document features with two released text-only acceptance predictors fine-tuned from the same language model with three billion parameters. For 198 papers from the International Conference on Learning Representations, we compare model rankings of the original papers and of paragraph rewrites screened automatically for unchanged facts, claim strength, and attribution. The classifier ranks original papers closely with one predictor and moderately with the other, but shows no detectable agreement with either on which rewrites raise scores. The tests specified in advance used reduced numerical precision. Rounding error dominated the small score responses. We therefore rescored the released weights in full 32-bit precision as a post hoc check; edit agreement remained near zero. At this precision, when each predictor chooses among five rewrites per paper, its average raw score is under two points out of 100 higher than with random choice. The evidence covers presentation rewrites and two open predictors from one family; transfer to costly evaluators remains untested. Revision selectors should therefore be checked against the intended evaluator on the revisions they rank.

One-shot and few-round federated learning (FL) reduce communication burden by constructing a global model within one or a few communication rounds, but often achieve lower performance than conventional multi-round FL. Without frequent synchronization, client models may drift apart, causing neurons that learn similar features to occupy different positions across clients. At the same time, local data heterogeneity can cause corresponding parameters to differ in their importance to each client's predictions. Together, these challenges cause direct averaging to combine mismatched neurons and assign equal weight to parameters despite differences in their importance. We propose FedFiOT, a Fisher-weighted optimal transport framework for one-shot and few-round FL that leverages Fisher information to quantify parameter importance and guide neuron alignment and model aggregation. FedFiOT jointly optimizes neuron correspondences and global model parameters through alternating minimization of a shared Fisher-weighted objective. Theoretically, we show that FedFiOT's optimization procedure monotonically decreases the layer-wise merging objective and establish a stationarity bound that quantifies the effects of client heterogeneity and nonuniform Fisher weighting.Experiments across multiple datasets demonstrate consistent improvements over competing baselines in both one-shot and few-round federated learning settings.

Large language model agents access data, invoke tools, and act on users' behalf. Keeping these actions within authorized bounds is challenging because agents choose their own execution paths while permissions and operating conditions may change. This paper surveys recent advances in agent access control across four areas: policy specification and derivation, identity and delegation, action and resource enforcement, and information flow control. We compare their control objects, enforcement boundaries, and assumptions to examine what individual mechanisms achieve and how they can work together. Building on this analysis, we rethink four aspects of authorization: plan selection shapes permission requirements; contextual information can inform decisions without itself granting authority; restrictions must retain their meaning across system interfaces; and authorization must be reassessed as workflows evolve. These insights motivate open challenges and future directions, together with concrete research questions and evaluation criteria. In particular, evaluation should examine both observed agent behavior and the unauthorized effects that remain enabled. Our analysis connects existing mechanisms to the requirements for preserving authorization throughout an agent workflow.

Scientific image-analysis workflows require agents to select among diverse capabilities like description, detection, counting, and segmentation. Current multi-tool architectures conflate intent interpretation with visual execution, masking whether a failure arises from flawed reasoning (misrouting) or failed perception (invalid tool execution). This lack of auditability renders agents unreliable for high-stakes scientific automation. We introduce ZETA, an extensible architecture that operationalizes the decoupling of routing policy from action execution. ZETA resolves scientific intent before invoking specialized visual tools, treating measurement as a deterministic operation grounded in explicit geometry rather than generative output. This architecture exposes the "black box" of agent decision-making, allowing routing correctness and downstream execution to be evaluated as independent failure modes. To our knowledge, we provide the first controlled study in scientific vision-agent routing that isolates explicit routing-policy specification while holding the legal action interface fixed. Evaluation on a zero-shot benchmark of 51 scientific-intent queries shows that explicit routing semantics can substantially improve routing accuracy, particularly at semantically adjacent action boundaries, although improvements vary across models. For Qwen 3.8, explicit routing-policy specification improves segmentation routing accuracy from 6.2% to 75.0% and detection routing accuracy from 60.0% to 100%. ZETA's execution traces make routing decisions and downstream artifacts separately observable, providing a basis for distinguishing intent-selection errors from visual execution failures. ZETA establishes the necessary infrastructure to isolate execution errors from interpretive ones, a critical requirement for deploying reliable scientific vision agents.

We study logistic regression for continual learning with linear predictors in a high-dimensional statistical setting. Existing theoretical work has largely focused on continual linear regression with variants of L2 regularization, motivated by their analytical simplicity and connection to the kernel regime of neural networks. We first derive the optimal L2 regularization strength for continual linear classification and characterize its generalization performance, recovering behavior analogous to prior results for continual linear regression. Next, we examine the effect of L2 regularization in deeper models using a minimal two-layer diagonal linear network. In the single-task setting, L2 regularization on the weights of such models is known to promote sparsity in the resulting predictor, yielding a rich, sparse feature-learning regime beyond the kernel regime (a similar effect arises in two-layer ReLU networks). However, in continual learning, where L2 regularization is typically centered on previous-task weights, this sparsity bias is no longer preserved. We therefore propose a weighted variant of L2 regularization that extends this beneficial sparsity bias to continual learning. We show theoretically and empirically that, for a sufficiently sparse teacher, our variant converges substantially faster than directly learning a linear predictor with conventional continual L2 or L1 regularization, trading some long-horizon bias for markedly improved performance over moderate task horizons.

Generative video models are increasingly used as visual world simulators, yet their primary strength lies in rendering plausible pixels rather than reasoning about how a world should evolve. Coherent world evolution requires more than visual fidelity: when an event occurs, a model must infer its consequences for the entities in the world and then render the resulting state. For example, after a wild beast appears, the model should not only depict the wild beast, but also infer whether a nearby person should flee, take cover, or fight back based on the realized scene. Existing diffusion-based video models lack an explicit mechanism for maintaining world state and reasoning over such event-conditioned consequences, often producing visually plausible but behaviorally static or causally inconsistent outputs. To make this reasoning gap concrete and measurable, we propose a new problem, Out-of-Sight Causal Evolution, which requires a model to render an event, reason about its consequences for the evolving world state, and reveal a causally plausible outcome as the world continues to evolve, even out of sight. As a baseline reference solution, we introduce EvoWorld, a brain-renderer collaboration framework in which a vision-language model serves as the world brain: it observes generated segments, maintains an evolving account of the world, and reasons about plausible state transitions, while a segment-autoregressive video renderer realizes them visually. We further introduce EvoWorld-Data and EvoWorld-Bench for training and evaluating event realization, causal reactions, and revisit consistency. Experiments show that EvoWorld substantially improves event-conditioned causal evolution over recent models while preserving visual quality and consistency, establishing a reference point for reasoning-guided, autonomous world evolution beyond the current field of view.

Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired through earlier training. TailSteer follows a simple principle: keep early denoising dynamics close to those of the base model and concentrate preservation corrections near the trajectory tail, where fine-grained deviations can be suppressed. These corrections are learned from the editor's own on-policy denoising rollouts, using unchanged source content as a preservation reference without requiring curated ground-truth editing pairs. Experiments across the FLUX, SenseNova, and Qwen families show that TailSteer substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance on extensive text-to-image and single-image editing benchmarks. These results demonstrate that long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities.

Scene Graph Anticipation (SGA) predicts object relationships in future video frames. Continuous-time SGA models evolve each subject-object pair with an independent Neural ODE, so relationships sharing an entity cannot influence one another during integration, and nothing prevents mutually exclusive predictions, such as "sitting on" and "standing on", for the same pair. We introduce CoupledSGA, which links such relationships through a Relationship Interaction Graph with graph attention inside the ODE vector field, and adds a Structural Consistency Loss ($\mathcal{L}_{SC}$) based on a matrix of mutually exclusive predicates. Developing CoupledSGA revealed a limitation of the evaluation protocol. With-Constraint evaluation keeps one predicate per group for each subject-object pair, so its R@K and mR@K cannot capture the within-group violations targeted by $\mathcal{L}_{SC}$: we prove they are unchanged whenever each group's winning predicate and its score are held fixed. We further show that the gradient of $\mathcal{L}_{SC}$ depends on predicate scores rather than ground-truth labels: for two-predicate rules it amplifies the ordering the model already holds, and when all scores in a group are equal it suppresses predicates in more exclusion rules more strongly, explaining why the penalty can lower recall while improving consistency. On Action Genome, activating $\mathcal{L}_{SC}$ after two warm-up epochs reduces R@20 from 42.21% to 35.18% while lowering the No-Constraint contradiction rate from 45.92% to 4.07% (matched-budget R@20: 40.48% CoupledSGA vs. 44.78% SceneSayerODE). CoupledSGA is a testbed, not a state-of-the-art claim: With-Constraint recall captures the cost of enforcing structural consistency but not the improvement. We recommend reporting No-Constraint results alongside a contradiction rate for methods that constrain predicate co-activation.

This paper introduces an innovative virtual reality (VR) exergame system designed to gamify High-Intensity Interval Training (HIIT) workouts through 11 predefined combo moves, including squat-punch combinations, bilateral arm strikes, and diagonal stretch combos. Unlike conventional tracking methods that often rely on wearable motion capture suits or limited headset-based sensors, our system uses a standard smartphone camera placed on a tripod to capture full-body movement. This approach offers a cost-effective and scalable alternative, particularly well-suited to the fast, dynamic motions typical of HIIT routines. The smartphone footage is processed in real time using a pose estimation algorithm, with extracted biomechanical data analyzed by a backend server built on the Model Context Protocol (MCP), Anthropic's open standard for connecting AI models to structured data sources and tools, and a GPT-4o large language model. The system generates personalized assessments of form accuracy, movement efficiency, exercise intensity, and potential injury risks, tailored specifically to high-impact, time-sensitive exercises. A within-subject study involving 20 participants (ages 18-25) compared experiences in VR and Non-VR HIIT sessions. Results indicated significantly higher enjoyment (Mean = 3.5 vs 2.8, p = 0.02), improved immersion/focus (Mean = 3.2 vs 2.8, p = 0.03), and increased likelihood to recommend (Mean = 3.2 vs 2.7, p = 0.04) under the VR condition. Confidence to continue did not differ significantly between conditions (p = 0.15), while perceived intensity remained comparable. The MCP system showed strong agreement with expert evaluation (87.5%, N = 10 sessions, one certified trainer), supporting its reliability as a feedback tool.

Large reasoning models (LRMs) perform well on challenging benchmarks, yet it remains unclear whether they form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) provides a controlled testbed for studying abstraction, requiring solvers to infer transformation rules over two-dimensional grids from a small set of examples. While ARC evaluates rule abstraction within individual tasks, different tasks may share higher-level structure despite requiring different transformations$\unicode{x2014}$for example, several tasks may involve reasoning about objects contained within other objects. Benchmarks such as ConceptARC capture this structure by grouping ARC-style tasks into recurring spatial and semantic concepts, but they often contain only a few manually constructed tasks per concept. We introduce ConceptARC at Scale, a generator for creating thousands of ARC-style tasks targeting a specific concept. Using this generator, we construct ConceptARC-100, a dataset of 1,600 new, manually reviewed tasks, with 100 tasks for each of the 16 concepts defined by ConceptARC$\unicode{x2014}$ten times the size of the original benchmark. We evaluate a range of recent LRMs and models trained from scratch on ConceptARC-100 to characterize which concepts they solve reliably and where they struggle. We further analyze the models' internal representations and find clear semantic structure: task representations cluster according to the operations applied to objects, such as recoloring, removal, or translation, even when realized through substantially different rules and scenarios. Together, our results suggest that models form stable abstractions of transformation operations that generalize across diverse tasks.

Small language models offer a compact basis for shell simulation without executing submitted commands, but full-output decoding requires them to predict recurring utility structure alongside content that varies across executions. We introduce template gap reconstruction (TGR), which supplies reusable output literals interspersed with sentinels to be interleaved with autoregressively generated gaps. A template selector chooses which template to be reconstructed, with the empty template retaining full-output decoding. We construct a template bank from 18.7 million outputs collected across heterogeneous execution environments and train a template selector and five reconstruction models initialized from various small language model checkpoints on a target environment excluded from bank construction. Evaluation demonstrates that TGR improves output edit similarity (O-ES) over full-output decoding (FOD) when the optimal template covers more than 25\% of the template, while performing similarly for lower-coverage outputs. Additionally, TGR outperforms FOD for shorter gap sequences, with O-ES values of 0.964 and 0.812, respectively, while maintaining similar performance for longer sequences. TGR reduces the number of tokens generated by 26.3--38.3\% across used tokenizer families. These results support using a reusable output structure as a practical design choice for small shell simulators, improving response fidelity and shortening generation targets while keeping template selection and residual prediction central to response quality.

Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.

Dexterous robotic hands are increasingly coupled to large-scale human data and foundation models, yet progress in hardware, representation learning and evaluation is often assessed separately. This survey reviews work from January 2024 to early November 2026, with earlier studies included to establish technical lineages. We organize the literature into five interdependent layers: hardware; data collection and curation; cross-embodiment transfer; learning and foundation models; and evaluation and applications. Human hand-object understanding and tactile sensing provide two cross-cutting perspectives. A system-level architecture, comparative tables and five case analyses connect design choices across these layers. We distinguish within-study experimental comparisons from heterogeneous corpus statistics, vendor specifications and company-reported results. Three conclusions emerge. First, the value of human data depends on kinematic, visual and contact alignment, not recording volume alone. Second, shared action representations and foundation policies do not remove embodiment-specific constraints on force, latency and contact. Third, task completion is insufficient to establish physical correctness or deployment readiness. Evidence remains limited for zero-shot transfer of contact-rich skills between morphologically different hands, long-duration reliability, and cross-sensor tactile generalization. We identify eleven open problems and propose reporting interfaces and controlled comparisons that can make full-stack progress more measurable. The review is a structured narrative synthesis rather than a pooled performance ranking or an exhaustive systematic review.

Continuous symmetries in neural networks lead to conserved quantities that are invariant under gradient flow, such as layer balancedness in linear and ReLU networks. Finite-step gradient descent need not preserve these laws, particularly at the edge of stability (EOS), where the iterates oscillate around a smoother trajectory, which can be described by the central flow dynamics (Cohen et al., 2025). We study how gradient flow conservation laws evolve under the central flow description of this regime. We show that the curvature correction induces a finite-dimensional drift determined by the oscillation covariance along critical parameter directions. We thus take the first steps to analyse which conservation laws remain and are broken at the EOS according to the central flow dynamics by analysing a scalar two-layer model, two-layer matrix factorisation, and a two-layer ReLU network, where the known conservation laws are shown to be complete (Marcotte et al., 2023).

Retrieval-augmented generation (RAG) assumes that language models can reason over relations stated in their context rather than fall back on associations memorized during pretraining. We study this evidence-conditioned relational reasoning in isolation, with the facts supplied directly in the prompt rather than retrieved. \textsc{Rel-Bloom} is a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations; the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized). Across 43 models, six languages, and two multilingual settings (226,800 predictions), three findings stand out. First, accuracy drops sharply from Apply to Analyze in every evaluated open-weight model series, reaching 47.2 percentage points for the Large Granite bucket, but not for either Gemini model. Second, Evaluate items remain a bottleneck even for models near ceiling on the other task types: Gemini~3.1 Pro averages 99.5\% on the other four levels but 60.7\% on Evaluate. Third, item- and relation-level difficulty rankings are consistently preserved across languages, indicating a difficulty structure partly shared across languages within this benchmark. Across knowledge conditions, Counterfactual contexts are generally harder than Factual ones, while removing entity identities can improve higher-level performance in some model regimes. Greater model capability therefore does not by itself ensure reliable reasoning over supplied relational evidence. Relevant resources are available at: https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning.

Deep generative models, particularly Flow Matching frameworks, have emerged as powerful tools for synthesizing complex physical systems. However, standard architectures remain purely statistically driven and agnostic to the underlying governing equations, often violating fundamental physical constraints during inference. Existing zero-shot approaches designed to enforce these constraints typically rely on rigid algebraic projections. While effective for sparse boundary conditions, applying such discrete geometric corrections to dense, full-domain PDEs may disrupt the probability path, \rev{potentially deviating from the learned manifold and introducing substantial computational overhead}. To address these challenges, we introduce Energy-Guided Transport, a novel training-free framework that formulates non-linear physical constraints as a smooth potential energy landscape. Rather than interrupting the ODE integration, our method applies a continuous dissipative force directly to the intermediate generative states, steering them toward the physical manifold. By evaluating this force implicitly via reverse-mode automatic differentiation, our framework preserves the natural dynamics of the flow while maintaining an efficient $\mathcal{O}(N)$ algorithmic complexity. We extensively evaluate our framework across various 1D and 2D physical systems, including the chaotic Navier-Stokes equations. Our method achieves stable generation without structural divergence, significantly improves the stability-accuracy trade-off compared to state-of-the-art methods, and substantially accelerates inference with respect to exact projection strategies. Code available at: https://github.com/ispamm/Projection-free-energy-guided-flow-matching.

Responsible AI scholarship has found that ethical principles are frequently not reflected in AI design and development. Conceptualised as a `gap' between principle and practice, this is framed as a problem of implementation, awareness or compliance, and by implication, a deficit of systems or individual behaviour. This paper sets out an alternative explanation. Drawing on fourteen semi-structured interviews with actors designing and governing AI-enabled anti-trafficking systems across law enforcement, commercial, research, and survivor-support contexts, it argues that such apparent gaps emerge from the varied and complex institutional and professional conditions through which actors come to recognise some issues, but not others, as ethical problems requiring action. We show that epistemic choices are intertwined with ethical judgements, and trace four interconnected moments---perception, judgement, situated action and socio-technical materialisation---through which concerns become intelligible, are weighed and acted upon, and are embedded in technical artefacts and governance arrangements. Actors often hold genuine ethical commitments, but these commitments are shaped by institutional logics and professional moral economies: some concerns are materialised, while others are narrowed, deferred or left unrealised. We show that ethical agency is exercised within and between competing constraints of profession, role, prior experience, and institutional settings. The paper argues that responsible AI governance requires purposeful and well-developed epistemic infrastructures that can cut across these constraints to establish institutional conditions through which the knowledge of affected communities and potential harms becomes receivable and actionable in design.

Tech-facilitated gender based violence (TFGBV) is a pervasive online harm affecting women and gender minority groups through its many direct and covert forms. Current models and annotated datasets for gendered harm detection are dominated by data from, and models trained in, Global North languages. In collaboration with civil society organizations (CSOs) from Jordan, Francophone West Africa, and Kenya, we developed a community-led approach to crowdsource culturally and linguistically relevant instances of TFGBV in Levantine Arabic, African French, and Swahili. We annotated the crowdsourced data using active learning to develop a small but performant dataset (N=3022) called ALTO to improve TFGBV detection in the three languages. Through benchmarking and fine-tuning experiments, we show that existing harmful content detection models show poor TFGBV recall within ALTO and improve with fine-tuning. We found that models often miss covert harms across all languages, while error rates are markedly higher in Swahili. We report the experimental results and release the annotation taxonomy, dataset, and trained models.

Training an agent to play Pac-Man with only up, down, left and right actions raises a basic question: when performance is poor, what should be improved? Difficulty can arise from reading the screen or from choosing actions over a long sequence. We examine these problems in a Qwen3.5-9B-based agent through separate perception checks and primitive-policy reinforcement-learning studies. In the policy study, which uses textual boards to remove image transcription from the decision interface, completion on eight held-out mazes increased from 0% before this optimization to 80.2% at a selected checkpoint, with three extra lives. The perception checks show that accurate static screenshot reading can coexist with errors during interaction. A separate policy comparison also shows that further optimization can accompany worse action choices. Together, these findings motivate concrete checks for post-training: evaluate perception in use, distinguish training success from full-game validation, and examine decisions and later performance rather than relying on one selected score. The study provides empirical lessons for improving future agent training and evaluation.

EEG extended-source imaging aims to localize active cortical regions and characterize their spatial extent, yet supervised inverse networks typically depend on synthetic EEG-source pairs and may be sensitive to synthetic-to-real mismatch. We propose a physics-informed self-supervised implicit neural representation (INR) framework for spatial source reconstruction without paired source labels or simulation-based supervised pretraining. The lead field matrix encodes the source-to-sensor volume-conduction model and constrains the INR through agreement between predicted and measured EEG. Cortical current density is modeled as a continuous coordinate-conditioned field. A classical inverse estimate provides warm initialization and spatial constraints on source activity, while patch-integrated projection maps the field to discrete source coefficients for forward modeling. Simulations with varying noise levels, source counts, and source extents showed the clearest advantage in identifying active cortical regions, particularly for multiple and larger sources. Quantitative and qualitative evaluations using real EEG recordings from epilepsy patients and intracranial stimulation experiments further supported the feasibility of the framework for source imaging. These findings suggest the potential of physics-informed self-supervised INR for extended-source imaging.

Post-hoc out-of-distribution (OOD) detection is mature on far-OOD yet remains weak on near-OOD. These two settings differ in their use of class-characteristic feature directions: far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do. Because the axis separating far-OOD leaves near-OOD almost indistinguishable from ID, a score optimized for one setting is misaligned with the other; where a detector already saturates far-OOD, added near-OOD sensitivity is bought partly against it. This divergence motivates a targeted near-OOD signal designed to complement existing scores. We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder, orthonormalizes the corresponding decoder directions to form a low-dimensional subspace, and scores an input by its per-class whitened squared Mahalanobis distance $z$ within the subspace of its predicted class. Designed as a dedicated complement rather than a standalone detector, the signal $z$ is fused additively with any calibrated base score at a single, fixed weight. This weight is a single global constant: no value is adjusted per backbone, per OOD dataset, or per deployment, and all calibration statistics are estimated from ID validation data alone. When applied to 18 existing detectors spanning diverse method families, $z$ consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved, with only marginal declines where the base score is near-optimal, making $z$ a robust, detector-agnostic complement across backbones.

We introduce LifeVal, a benchmark evaluating AI model performance on personalized agentic consumer tasks. LifeVal covers the most valuable life tasks that a personalized digital superassistant could automate, representing over 1.7 trillion US dollars in annual willingness-to-pay across 148 consumer task categories (e.g., managing finances, moving apartments, searching for a job, finding childcare). To construct our activity distribution, we conducted a new survey of willingness-to-pay for a digital superassistant across a demographically representative sample of over 3,800 consumers, cross-referencing our results against the American Time Use Survey (Bureau of Labor Statistics, 2025) and the Multinational Time Use Study (Fisher et al., 2025). For the most valuable consumer activities, we recruited task creators to create AI evaluation tasks personalized using real ChatGPT conversations and memories, reference files, and personal context. Task creators had to pass our expertise threshold (including personally completing the task multiple times and seeing the results through in their real lives for over a year) and also write a rubric, ensuring that our evaluation measured the best personalized completion for that task. All tasks had to pass at least three rounds of human quality-control before acceptance into our final evaluation sets.

Tool-using large language model (LLM) agents translate model outputs into consequential actions, turning indirect prompt injection into an execution-security problem. Existing defenses constrain tool calls, trajectories, or information flows, but often leave action semantics, state dependencies, and evidence relations implicit, limiting generalization across heterogeneous tools and complicating safe updates. We present \ours{}, a dynamic ontology-guided governance framework. \ours{} grounds heterogeneous calls in a shared semantic representation and combines an executable ontology with episode-level task state to derive deterministic ready, waiting, and forbidden decisions. A waiting decision identifies an unmet but recoverable obligation, enabling bounded replanning without expanding the agent's authorization boundary. \ours{} evolves its ontology offline from verified training executions. An offline LLM proposes instance-independent action--state axioms, which are accepted only if they address ontology errors exposed by training failures, preserve logical consistency and protected entailments, and pass held-out validation before the ontology is frozen. We evaluate \ours{} on AgentDojo, ASB, and AgentDyn under benign and indirect prompt-injection settings. \ours{} achieves a 0\% attack success rate (ASR) across all three benchmarks while improving task utility under attack compared with existing baselines. These results show that explicit semantic governance can strengthen execution security without indiscriminate blocking.

Machine unlearning is commonly evaluated with standard post-deletion metrics that compare an unlearned model to one retrained on the remaining data. Such snapshot evaluations can conceal consequential differences in the features the two models encode. We prove that if an unlearned and a retrained model agree on their outputs but encode different features, a single gradient step on their prediction heads can break this agreement. We further show that deleting an arbitrarily small fraction of the data can cause retraining to select qualitatively different features, so matching retraining can require learning features that the full model never had. We then introduce a measure of feature recovery that separates missing from extraneous features. Across eight unlearning methods, models that perform well on standard metrics often fail to recover the retrained features. This mismatch predicts deviations from retraining under subsequent relearning and downstream adaptation, and directly correcting it systematically reduces these deviations.

We study the classical apple-tasting setting, a partial feedback model in which the learner observes the true label only when it predicts $1$. A combinatorial characterization for the rates of this problem in the online adversarial setting has been studied by Raman et al. (2024). We study the self-directed setting, in which the learner can adaptively choose the order of the points it predicts on. In the realizable setting, we establish a tetrachotomy on the expected minimax number of mistakes made by any learner. Depending on the VC dimension, the $\emptyset$-centered star number, and the self-directed dimension, we show that the rates could possibly be $Θ(1)$, $O(\log n) \cap ω(1)$, $\tilde Θ(\sqrt{n})$, or $Θ(n)$. In the agnostic setting, the minimax regret is $\tilde Θ(\sqrt{n})$ or $Θ(n)$, depending on the VC dimension.

Infrastructure security does not establish whether an LLM agent selected an allowed tool, grounded its answer in tool output, or followed a required workflow. AgentTrustCI is a Java/JUnit 5 framework that expresses these application-layer expectations as policy, grounding, and trajectory checks over recorded traces. A deterministic replay implementation supports repeatable build-time tests; a judge interface provides an extension point for optional semantic evaluation. On 55 synthetic cases spanning seven enterprise domains, the full pipeline correctly classifies all 30 injected faults and 25 passing controls, including ten non-trivial controls, and misses none of 12 designated critical cases. Each single-layer configuration correctly classifies 35 cases because the corpus deliberately isolates ten faults per layer. A grounded answer obtained through a disallowed sequence illustrates the additional coverage of trajectory checks. Repeated measurements quantify validator execution separately from process startup and reporting. These results establish coverage of the authored regression suite, not production accuracy. We describe a production-trace validation protocol, optional LLM-judge integration, and maintenance of expectations as workflows evolve. The contribution is a practical pattern for expressing behavioral contracts within an existing Java release pipeline.

Search-augmented reasoning systems can reach correct answers while making errors in intermediate steps. Reliable auditing must check whether evidence $D$ covers the question $Q$, reasoning $T$ addresses $Q$, and $T$ is grounded in $D$. We introduce GraphRail, which represents requirements, evidence, and reasoning as structured claims and compares them without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, GraphRail improves error recall over generative verifiers while remaining stable under stochastic extraction. Its diagnoses can also guide online reasoning, providing a traceable way to detect and correct intermediate errors.

API distillation campaigns try to hide model-stealing traffic by splitting requests across thousands of load-balanced accounts, which standard content classifiers often miss. However, request timestamps, specifically arrival time and stream completion time, reveal the underlying dispatching strategy, such as round-robin counters or key-shuffling pools. Using a 20 percent seed set, this timing analysis can forensically rank the remaining accounts of a campaign with 0.92 to 1.00 AUC within 20 minutes, even across thousands of real production accounts. While adaptive attackers struggle to hide rotating dispatchers using jitter, pinning keys successfully evades detection, though it can severely throttle throughput when quotas bind. Ultimately, timing exposes shared infrastructure patterns even when content filters fail.

Public bodies increasingly rely on AI-produced evidence, including classifications, summaries and assessments. These outputs can inform policy decisions and service delivery, making their reliability a matter of public accountability. Establishing whether this evidence can be relied upon requires testing its outputs and examining how organisations decide whether those tests are sufficient. We combine technical evaluation with analysis of AI governance requirements, informed by measurement theory and management-based regulation. We compare eight GPT-4o configurations classifying 1,200 UK Prime Minister's Questions into 21 policy topics, against human-coded labels and across unchanged repeats. Accuracy ranges from 68.2% to 70.3%, yet 24.3% of questions receive different labels, with changes in topic shares and rankings. Configurations agree more with one another than with human coding: agreement does not guarantee correctness. Under the implemented score-extraction procedure, the proportion flagged for human review ranges from 0.9% to 41.0%. We analyse 37 UK and EU policy documents. Both sets expect testing and oversight, but reporting expectations do not consistently cover changing answers, shared errors, selective evaluation, substantive conclusions or review routing. Some UK guidance explicitly addresses stability. The UK sample emphasises public bodies checking their own use; the EU places substantial testing duties on providers while leaving them important choices about measures and acceptance thresholds. Without appropriate checks, evidence could misrepresent policy issues, redirect attention or resources, and leave errors undetected. Public policy should require organisations to justify reliance through checks beyond aggregate accuracy, clear responsibility, access to expertise and scrutiny of results.

A robot action can satisfy shape and range constraints yet carry the wrong semantics: axes may be permuted, signs inverted, rotations rescaled, or gripper polarity reversed. We study learned state-action-effect relations as offline contracts for detecting these mismatches and ranking their likely fields. Clean held-out transitions calibrate alarms; a fixed inverse vocabulary supports closed-set localization. On ten LIBERO Spatial tasks comprising 432 source episodes, a capacity-controlled comparison at 2.5% empirical clean false-positive rate gives 46.1% recall for a ridge forward residual, 61.0% for a four-component joint Gaussian mixture, and 65.7% for the exact conditional of that same mixture. The joint score retains a 14.9 percentage-point advantage over ridge, while conditioning improves recall by a further 4.7 points. Joint scoring instead has higher localization: 95.8% versus 94.6%. These matched-rate comparisons are retrospective diagnostics, not deployment calibration guarantees. Earlier Object/Goal results broaden tasks within the LIBERO convention; DROID, all three HYDRA tasks, short temporal faults, small offsets, exact repair, and simulated mitigation fail their respective gates, while xArm contributes one task only. The evidence supports a bounded offline diagnostic and a detection-localization tradeoff, rather than joint-density superiority or cross-robot safety.

This chapter addresses the central reproducibility problem of research with large language models. LLM output is probabilistic and varies across models and versions, which can have significant effects on downstream statistical results. The chapter distinguishes between replicability, reproducibility, and robustness, and proposes a probabilistic standard based on the distribution of results, not identical outputs. It introduces a standardized prompt script, tests within-model variation for encoders and decoders, and develops error-structure analysis to diagnose where models err and whether their errors align. Using three datasets, the chapter shows that measurement choice can reverse the sign of downstream regression coefficients. Existing design-based corrections with gold-labeled subsamples restore validity, but come at the cost of estimate precision. The main contribution of the chapter is a stacked ensemble of open-weight models that improves precision in the correction step while retaining estimate validity.

Human Motion Prediction (HMP) aims to predict future human poses at different moments according to observed past motion sequences. Previous approaches mainly treated the prediction of different temporal moments as a single prediction task and learned the predictions of varied moments simultaneously, which would encounter a main limitation: the learning of short-term predictions (referring to ``near-future'' prediction) could be hindered by the predictions of long-term (referring to ``far-future'' prediction) motions. In this paper, we develop a novel temporal continual learning framework called Continual Prior Compensation (CPC) to progressively train HMP models, in which we divide the prediction task of motions corresponding to varied temporal moments into several subtasks and train the model in a multi-stage manner. To mitigate the prior information forgetting in the progressive training, we further introduce a learnable random variable Prior Compensation Factor (PCF) to explicitly measure the prior knowledge loss. We theoretically show that the PCF can be efficiently learned together with the model parameters by minimizing a reasonable upper bound of the objective function. The proposed CPC is further enhanced to estimate the prior information loss for each subtask and a new framework called Continual Prior Compensation++ (CPC++) with Fine-Grained Prior Compensation Factor (FGPCF) is finally developed. Our CPC and CPC++ frameworks are quite flexible and can be easily integrated with different HMP backbone models and adapted to various datasets and applications. Extensive experiments on three HMP benchmark datasets using multiple SOTA HMP backbones (PGBIG, siMLPe, MotionMixer, and LTD) demonstrate the effectiveness and flexibility of our frameworks.

Bayesian experimental design (BED) is a principled framework for intelligent data acquisition with applications across a wide range of scientific and industrial domains. However, the development and adoption of BED remains hindered by a lack of standardised software and benchmarks. Existing methods are typically implemented in bespoke codebases and evaluated on a small subset of tasks, making methods difficult to apply, compare, or extend. To address this, we present TestBED, a general-purpose, open-source toolkit for Bayesian experimental design. TestBED decomposes the BED process into modular, reusable components spanning tasks, objectives, training mechanisms, policy architectures, and evaluation procedures, together with scalable implementations of core primitives such as expected information gain estimators. We provide a wide range of these components, which can be freely composed, allowing existing methods to be reproduced within a common framework, while also enabling previously unexplored configurations and new methodologies to be easily implemented and evaluated. Using TestBED, we further conduct a large-scale benchmark of contemporary BED methods, including controlled studies of under-appreciated but surprisingly critical methodological choices, such as the importance of time embeddings in design policy architectures. By providing a modular, extensible software foundation and a reproducible benchmarking suite, we hope that TestBED will accelerate methodological research in BED and lower the barrier to its real-world application.

We used medical image synthesis to create a challenging multimodal evaluation framework that exposes hallucination behaviors in frontier models. Evaluating multimodal reasoning has historically been difficult because existing medical datasets rarely pair different image modalities, such as radiographs and external clinical photographs. To address this gap, we introduce Trick-or-Treat, an adversarial multimodal medical benchmark that pairs real chest radiographs with synthetic clinical photographs to challenge models' cross-modal reasoning. For each of a wide range of conditions, we created two pairs of images, where the pairs share a single chest radiograph but contain different external photographs. In matched pairs, the chest radiograph and photograph both contain findings suggesting the same disease (e.g., varicella pneumonia suggested by lung opacities and crusted skin lesions seen in chickenpox). In mismatched pairs, the photograph depicts an alternative clinical finding that would not lead to the same unifying diagnosis (e.g. replacing crusted skin lesions with ordinary acne). Synthetic photograph generation allows clinical findings to be varied while the person's features and photographic setting stay consistent, reducing confounding. In an investigation of three frontier models, we found that models hallucinated clinically significant photograph findings in 45-57.5% of mismatched pairs, up from 25-37.5% for the same photographs shown alone. When faced with mismatched pairs, models tended to ignore mismatches and instead hallucinated findings that better matched the radiograph (e.g. misreading acne as crusted skin lesions to justify diagnosing chickenpox). Trick-or-Treat both demonstrates concerning error pattern in frontier models' medical reasoning and provides a proof of concept for the use of synthetic images in adversarial multimodal evaluation.

Quantum circuit Born machines (QCBMs) can learn distributions that are hard to sample from classically, but their value depends on generalizing from finite training data. Here, we study how hardware noise affects this generalization. We show that, for any noise process, the generalization gap of a QCBM trained with the maximum mean discrepancy is bounded by the complexity of the distributions the noisy model can reach. Scrambled local Pauli noise contracts this complexity but biases the model towards the uniform distribution, so residual noise acts as a regularizer whose optimal strength can be intermediate. We propose GENIE, a framework that realizes a chosen noise level with capped error mitigation and selects it during training from held-out data within a fixed execution budget, with provably bounded regret. In simulations on synthetic and census data, GENIE stays close to the best of four noise-control strategies, whereas every fixed noise level falls short in many settings.

Defenders and attackers race to find the same software flaws. I use the launch of Project Glasswing, which gave a small group of large US tech firms early access to Claude Mythos Preview, a model for finding software vulnerabilities, to study how access to frontier AI affects defenders, external researchers, and attackers. Using the universe of publicly disclosed vulnerabilities, I find that firms with access more than doubled their disclosures relative to Chinese and Russian tech firms without access, disclosed about 150 percent more high-impact vulnerabilities, and fixed reported flaws as quickly as before. As treated firms found more flaws themselves, external researchers, who sell their findings through bug bounties, reported about half as many vulnerabilities in treated firms' code and more in other firms' code, without leaving the market. There is no evidence that attackers exploited the additional vulnerabilities: the share exploited within a month of disclosure fell by about two thirds, mainly because a smaller share was already exploited when disclosed, while the number exploited did not change. Frontier AI helped defenders find and fix more flaws before attackers exploited them, but its restricted release creates an asymmetry in software security between firms in different countries.

A national digital twin was conceived as an ecosystem of twins connected through mappings to a common foundation, but the roadmap that promises to confirm those mappings are valid never says what valid means. We give the test a precise form, coherence and conservativity of the merged ontology, borrowed from logic-based alignment repair, and apply it to the Information Exchange Standard (IES), first pair by pair and then to whole hubs. Crosswalks from IES into thirteen ontologies across seven sector groups, anchored on the building ontologies BOT, Brick, and SAREF4BLDG, pass through separate stages of matching, human review, certification of each SKOS correspondence against the two class hierarchies, and reasoning with HermiT. Pair by pair, the stages catch different failures: certification alone removed all 526 conservativity violations of the raw crosswalks, review alone left 73, and together they made eleven of twelve overlapping pairs coherent and conservative. Pairwise certification does not compose. Merging the three certified building crosswalks into one hub made five Brick classes unsatisfiable, through a homonym between the IES and Brick notions of a floor that only BOT's disjointness axioms expose; rejecting that one row gave a coherent and conservative hub that entails 290 cross-ontology links nobody wrote down. The core hub is inconsistent before any crosswalk is added, because two published standards, the Open Energy Ontology and BFO 2020, contradict each other; with either one left out it is coherent and conservative. A hub, not a pair, is the unit an ecosystem of twins has to certify.

We present a shared framework that adapts Cosmos world foundation models for real-time interactive world simulation in Physical AI. By streaming visual observations in response to actions and scene controls, the framework enables users and embodied agents to interact with evolving virtual environments. It supports both predicting the visual consequences of actions and generating observations from physical state maintained by an external simulator. We demonstrate this approach across camera navigation and robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation. These examples provide a shared foundation for developing interactive virtual environments for Physical AI training and evaluation. Our code and model checkpoints are available at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3-examples. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos-simulation.

We study the limits of differentially private Gaussian mean estimation when the covariance matrix is unknown. Prior work shows that when the covariance $Σ$ is known, the optimal sample complexity can depend only on its spectrum, rather than on the ambient dimension $d$. We establish new lower bounds for Gaussian mean estimation with Euclidean error which show that private estimators cannot achieve similar guarantees when the covariance is unknown. First, when the sample size $n$ is smaller than the effective dimension $\mathrm{trace}(Σ)/\|Σ\|_{\text{op}}$, the optimal error when the covariance is unknown can be larger than the optimal error when it is known by a factor of $\sqrt{d/n}$. Second, when the sample size is larger than the effective dimension, even for diagonal covariance matrices with effective dimension bounded by $\log d$, the optimal error when the covariance is unknown can be larger than the optimal error when it is known by a factor of $\sqrt{d}/n^{3/2}$. Together, these results show that in the unknown covariance setting, exploiting anisotropic structure under differential privacy can require a dimension-dependent error or sample complexity.

Complete sensor-channel failures create a difficult form of time-series missingness: local interpolation is unavailable, and deploying a separately trained imputer for every failure pattern may be impractical. We ask whether a sequence reconstructor trained without artificial mask augmentation and then frozen can be converted into an imputer at inference time. Our method treats the missing entries of each test window as optimization variables, clamps all observations, and backpropagates through the frozen model to minimize reconstruction error only on observed coordinates. The resulting observed-self-consistency objective requires neither artificial mask augmentation nor full-channel dropout during reconstructor training. We characterize when this refinement is informative through the cross-Jacobian from missing inputs to observed residuals, and use an optional temporal-smoothness prior when that coupling is weak. Across five real-world datasets with exactly one channel synthetically masked per test window and four reconstruction architectures, no-prior test-time refinement reduces MAE relative to one-pass reconstruction in 19 of 20 dataset-backbone combinations. Compared with mask-aware imputers trained under either standard structured masks or matched full-channel dropout, the best refined reconstructor achieves the lowest MAE on two datasets and trails the strongest baseline on three. These gains require substantially more inference time and are sensitive to cross-channel coupling. Our results establish test-time self-consistency as a conditional inference-time adapter for frozen reconstruction models that complements mask-aware imputation without retraining the reconstructor.

Collaborative filtering (CF) is a cornerstone of recommender systems and plays a relevant role in many modern applications. CF uses user-item interaction data to discover future preferences. Although deep learning models have shown promise in CF, Embarrassingly Shallow Autoencoders for Sparse Data (EASE) has gained attention for its outstanding ranking accuracy provided by its closed-form solution. EASE relies primarily on relationships among items to fit a full-rank high-dimensional linear mapping. We hypothesize that this design limits its capacity to capture similar but not equivalent fine-grained user relationships, consequently limiting its recommendation accuracy. This paper introduces an alternative formulation based on EASE that takes advantage of user-user information. Furthermore, we propose a hybrid model that combines both user-user and item-item adjacency distributions. Our experiments reveal that the proposed models outperform EASE on well-known recommendation benchmarks, highlighting the significance of including the user alternative in shallow autoencoder studies for CF.

Large language models can solve many multi-step reasoning problems, but their reasoning remains brittle: plausible intermediate steps often lead to incorrect answers. Improving reasoning at test time requires balancing two imperfect signals: the generator, which provides diversity, and the verifier, which guides correctness. Existing approaches largely favor one side of this tradeoff. Verifier-weighted voting relies on trajectories sampled from the generator and cannot steer generation toward correct ones. By contrast, beam-style verifier-guided search can steer generation, but may over-concentrate on narrow trajectory sets and amplify verifier errors. We propose Verifier-Guided Self-Consistency (VGSC), a stochastic decoding framework that combines step-level correctness and future-outcome estimates to guide reasoning during generation. VGSC repeatedly proposes candidate next steps from the base language model and uses both signals in an acceptance/resampling procedure, while maintaining multiple trajectories for final answer-level self-consistency. We provide a probabilistic interpretation of VGSC as approximate sampling from a verifier-tilted trajectory distribution. Across multiple reasoning benchmarks, generator models, and verifier families, VGSC achieves higher average accuracy than representative baselines.

Continual agents must handle urgent interaction while retaining alternatives, replaying errors, and committing safe long-term changes. A previous digital-life framework distinguishes online (busy) and offline (idle) channels, but leaves open whether these are separate programs or phases of a common process and how their operations should be ordered. Motivated by recent observations of rapidly alternating perceptual representations in primate cortex, we develop a temporal organization layer with phase-tagged micro-epochs, an observable duty cycle, independent content and resource controls, and three nested clocks separating candidate exploration from durable commitment. We give a projected duty-cycle update, state its deterministic descent guarantee and the additional noise term required for a stochastic guarantee, and specify coupled busy/idle execution contracts. Five synthetic experiments (64 seeds each) examine descent and noise, objective misalignment, gain asymmetry, resource coupling, and dropped replay logs. In the gain/noise model the best tested gain ratio is 1.55; decoupled scheduling yields 26.6% versus 11.1% success on a constructed urgent/high-load subset. A frozen-backbone Qwen3-8B prototype further checks orchestration, but its proxy outcomes are not task-benchmark evidence. The results establish a testable scheduling design and identify its conditions of validity, not machine consciousness or general-purpose agent superiority.

Low-Rank Adaptation (LoRA) adapts a frozen language model through a compact learned update, making it a natural setting for asking how much information fine-tuning needs to write. Yet adapter rank, parameter count, or numerical reconstruction error do not directly tell us how many bits are required to preserve the acquired behavior. We therefore define an operational behavioral rate $R^\star$ as the smallest serialized size of a compressed adapter that, once reloaded into the frozen base model, retains a fixed fraction of its held-out fine-tuning gain. Across LoRA fine-tuning experiments, this rate increases with distinct training content at fixed optimization budgets, but is not determined by intuitive metrics like source description length, fine-tuning gain, or weight-space reconstruction error. Most importantly, the same datasets exhibit substantially different, and even reordered, rates across frozen base models, showing that fine-tuning compressibility is a property of the dataset--receiver pair rather than of the dataset alone. However, we show that correction spectra measured on the data are able to rank these rates within a receiver. Finally, we study a striking noise-reducing effect of compression: adapters trained on deliberately corrupted datasets fall far below the frozen model, yet after compression score above it (23.7% to 82.9% on GSM8K, against 77.6% for frozen Qwen2.5-7B), consistently across tested families and sizes. The damage is unevenly concentrated in the adapter's leading singular directions, which compression appears to filter. Away from its training task, the compressed corrupted adapter also generalizes better than a clean fine-tune: on harder math it retains 97% of base accuracy, versus 44% for the clean adapter.

Machine-learning surrogates of linear peeling-ballooning stability now run in milliseconds, but they are mostly deployed as point estimators: they do not say how wrong they may be, or when a query lies outside what they were trained on. For a fusion company designing a new device, every equilibrium of interest is outside the training data of existing surrogates, so this is the case that matters. We describe a model-agnostic trust layer for stability surrogates, made of conformal error bars, out-of-distribution alarms that route doubtful cases to the physics code, and recalibration from a few simulations of a new design, and we test it with pre-registered hypotheses. On a large public dataset of JET-like equilibria with ideal-MHD growth rates, the error bars hold their nominal coverage in distribution and fall short under a shift in plasma shape; a few dozen simulations of the new shape, or routing the cases the alarms flag, restore it. On resistive BOUT++ surrogates, the bars stay calibrated on held-out runs, the alarm flags tokamaks absent from training, and a scenario scan is screened in seconds instead of hours. Across equilibria from the SPARC library and the DIII-D coil set, a surrogate trained on our own graded BOUT++ labels beats the baseline on unseen equilibria, while a surrogate trained elsewhere ranks the same cases backwards and its alarm declines them. Building our own neural-operator surrogate, the harness exposed two input defects whose fix closed its gap to a simple network. For every surrogate tested, the trust layer keeps coverage near target, with a bar width set by that surrogate's own error. Not every pre-registered hypothesis held, and we report those results and the limits of all of them.

Deep eutectic solvents (DESs) offer a tunable platform for lignocellulose refining, but their solvation behavior depends on multicomponent hydrogen-bond networks rather than individual constituent properties. We develop TiNN, a thermodynamics-informed neural network that hierarchically represents HBA-HBD association and solute accommodation through a learned free-energy pathway. Trained on 938 records, TiNN achieves a test-set $R^2$ of 0.917 and generalizes to new HBA-HBD combinations and sparse conditions. Ablation identifies solute-specific stabilization as the largest contributor to prediction accuracy, while the importance of hydrogen-bond compatibility depends on the dissolution mechanism. Experimental measurements of glucose, xylan, and Kraft lignin in acidic, basic, and neutral DESs show that TiNN reduces the geometric mean solubility error from more than 23-fold with COSMO-RS to approximately 1.5-fold. Without training on fractionation data, TiNN also distinguishes DES formulations with high and low lignin and hemicellulose removal from corn stover. These results support chemically grounded solvent screening for lignocellulose fractionation.

AI agents rely on third-party tools, such as Model Context Protocol (MCP) servers, whose publishers can be authenticated and whose execution can be restricted by runtime security policies. Yet authenticity does not guarantee benign behaviour: a legitimately signed tool may perform its advertised function while concealing malicious code that leaks confidential information. We demonstrate this trust failure with a signed MCP tool that embeds sensitive data in requests to a policy-allowed destination. Meta's MCPGuard-Dynamic, a state-of-the-art runtime defence, blocks a direct exfiltration attempt in the same run, yet fails to prevent this attack in every trial. We introduce germline execution, which limits information disclosure without inspecting or trusting tool implementations. Each tool invocation runs in ephemeral copies forked from a data-free germline process, each receiving only information authorised for its destination. Even if malicious code executes, what it can disclose is bounded by the data available to its copy. Under explicit assumptions on copy confinement and a data-free germline, we prove sink noninterference for every signed release and, for releases deterministic within our replay model, show that a verified witness establishes a flow violation. Across 153 attack runs, no unauthorised disclosure was detected under germline execution, compared with 5.9-74.5% of runs under baseline defences and an ablation. Invocations served by a single copy completed in 10-49 ms, 2-28x faster than restarting the tool for every invocation, the conventional way to clear state. These results show that data-isolated execution protects agent toolchains against malicious code concealed within authenticated software.

Automated vulnerability detection is critical for securing modern software systems, yet existing learning-based approaches often treat functions as independent samples and underuse the supervision provided by vulnerability-fixing patches. In this paper, we propose PatchCon, a CodeBERT-based patch-aware contrastive learning model for function-level vulnerability detection. PatchCon jointly optimizes vulnerability classification and supervised contrastive learning, and treats vulnerable-patched function pairs as hard negatives with larger contrastive weights to capture subtle security-relevant differences introduced by fixes. To evaluate patch-level discrimination, we use two pairwise metrics that measure whether a model can correctly distinguish vulnerable functions from their patched counterparts. We evaluate PatchCon on PairVul-Linux through extensive experiments, including state-of-the-art comparisons, ablation studies, hyperparameter sensitivity analysis, and statistical significance testing. We also conduct fine-grained analyses by Common Weakness Enumeration (CWE) type, function length, and patch size, and further evaluate PatchCon on Linux kernel vulnerabilities disclosed after the training data cutoff. On PairVul-Linux, PatchCon outperforms representative pretrained code models and contrastive learning baselines on both classification and pairwise metrics, improves pairwise accuracy over the strongest reproduced baseline from 0.2060 to 0.3264, and achieves pairwise accuracy comparable to that of a GPT-4o-based large language model (LLM) method at a substantially lower cost. Among the reproduced models, PatchCon also achieves the highest pairwise accuracy on test pairs from Common Vulnerabilities and Exposures (CVE) entries unseen during training and on newly disclosed vulnerabilities.

Deep neural networks train best when they are initialised at the edge of chaos, the critical line in the plane of weight and bias variances on which the distance between two inputs neither grows nor shrinks with depth. The line is known for infinitely wide networks; at finite width it moves, and how it moves has been disputed. We treat the logarithm of the distance between two inputs through a network of width $N$ as a random walk and compute its drift and variance to order $1/N$. The edge of chaos moves to a larger weight variance, by $Δσ_w^2=-a_1/(N\,\partial_{σ_w^2}\lnχ_1)$, about $8/N$ for $\tanh$ networks at a bias variance of $0.05$, confirmed by simulations of the exact finite-width recursion. The shift describes a typical network; the average over many networks is dominated by rare networks in which the distance grows, and its critical point moves the other way, which explains why calculations and simulations have disagreed. Backpropagation follows the same random walk, so the shift applies to gradients unchanged. Along the edge of chaos, the information about the similarity of two inputs is constant until the compression over the depth exceeds about one nat, which fixes where to sit on the line. The results give an initialisation rule, Crinit, that places a network on its finite-width edge of chaos without a search. On MNIST and Fashion-MNIST at width $64$ it is more accurate than the standard rules at their usual settings, by a small margin in shallow networks and a large one in deep networks, where most standard rules fail; it trains in fewer steps, and it matches or beats the standard rules with a tuned gain.

Category systems often reach their users as text, and readers apply them with tacit repairs. We ask what such a text determines once the repairs are removed, using diabetic retinopathy grading, where one international scale circulates in several documents. We transcribe four texts (the English screening definitions of 2025, the international scale as tabulated in 2017 and as published in 2003, and an Italian guideline of 2015) into one rule language over 24 finding variables, and make every point a text leaves open a named precisification. Exhaustive enumeration of 1,425,807,360 finding profiles, re-derived by an SMT solver and spot-checked by an OWL reasoner, shows that every text that assigns actions leaves some eyes without a category under every reading; that the 2017 tabulation has a proliferative class that is empty when read literally and a macular gap that the 2003 text lacks, so that no function maps the 2003 grades onto the 2017 grades; and that on image findings alone the English and Italian texts order some pairs of eyes in opposite ways. On 927 annotated photographs from three public datasets, with retinal thickening assumed absent, the English text and the 2017 tabulation disagree on referral for 26.8% of the eyes of one dataset (DDR) on which both decide, and in a pre-registered experiment classifiers trained on each text's labels side with their own text. Thickening cannot be seen on a colour photograph; left free, every disagreement involving England becomes undecidable, so it rests on what the camera cannot show. In a second pre-registered study, language models that answered without deliberating repaired most gap and overlap items even when told to read literally, while Claude Sonnet 5.5 and Claude Opus 5.5 reported the defects far more often; within two models, deliberation reduced silent repair without removing it.

Coding agents change software through tool calls. Tool calls mutate the local state. For good? For bad? Nobody knows. Agents never say why a tool call is made. We propose the epistemic agent design, in which the harness guarantees that the model declares why a tool call is made. The harness then records what the command actually did and checks the self-reported declaration against these effects. The result is a journal 1) that can guard calls before they run and 2) that supports audits and forensics. For evaluation, we use ImpossibleBench, whose tasks can be passed only by cheating, and demonstrate that forcing the agents to say why results in exposed cheating. Across 49 reward-hacking runs of two open-weight models, the journal exposed 90%.

Long-term time-series forecasting supports decisions in energy, transportation, and other applications where reliable predictions are needed beyond the observed data. Although forecasting research has advanced model architectures, fixed activation configurations provide limited guidance on how activation choices should vary across datasets and prediction horizons. This study aims to establish when task-specific activation choices improve forecasting accuracy and how previous experiments can guide the selection of suitable configurations for new tasks. We develop a calibrated selection policy that learns a reliable default from prior experiments, adapts the activation function and its location to each task, and retains the default when evidence for a change is insufficient. In an evaluation supported by 1,225 Informer training runs across 17 dataset variants and 51 forecasting tasks, the policy reduces mean squared error by 7.86% on average relative to the conventional fixed-activation baseline and improves 72.55% of tasks. These findings establish activation choice and placement as practical opportunities for improving existing forecasting models and show how accumulated experimental evidence can guide configuration decisions without repeating an exhaustive search for every task.

Resource management for distributed edge networks hosting different services has been handled efficiently using Reinforcement learning (RL). Reward generalization in complex workflows is still an open issue within the RL realm. Therefore, this work proposes RAGAM-AI, a hybrid framework that combines the capabilities of RL algorithms and Generative-AI (Gen-AI) for powering an Agentic AI workflow that optimizes resource allocation at the Multi-access Edge Computing (MEC) units. This framework further resolves the reward generalization issue by utilizing Large Language Model (LLM) for feedback. A Multi- Agent Soft-Actor Critic (MASAC) using MathΣtral as a reward oracle is implemented for managing resource allocations at the MEC. RAGAM-AI interprets system state information and evaluates multi-objective trade-offs guiding the resource orchestration loop over distributed edge server agents. The complexity of the framework is further brought down by incorporating surrogate reward model. The overall performance is evaluated against the single agent SAC, REINFORCE and random policies for benchmarking. On using RAGAM-AI, lower allocation deviations for CPU, memory and bandwidth resources are reported while minimizing the latency and data rate costs. A scalable AI as a feedback (AIAF) solution optimizing network resources with significant performance improvement is thus presented.

Large language models (LLMs) undergo safety alignment to prevent harmful generations, yet they remain vulnerable to jailbreak attacks. Recent work reveals a deeper structural weakness: safety alignment can be highly sparse, such that pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior. In this work, we take the first steps toward strengthening safety alignment such that it is less dependent on sparse model components. We propose Addition by Subtraction, which gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model. Our results show that the resulting models are substantially more robust to pruning-based attacks. Remarkably, these benefits generalize beyond the white-box attacks motivating our design: Addition by Subtraction improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks. The finetuned models also exhibit improved truthfulness while largely preserving general capabilities. Comprehensive experiments across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.

Adapting sleep foundation models to dense electroencephalography (EEG) event localization is challenging because multi-second representations may not preserve sufficient temporal detail for precise boundary recovery. We address sleep spindle localization by adapting a frozen SleepFM encoder with a local time-frequency pathway that preserves spindle-relevant spectral content and intra-token position. The local representation drives localization, while residual cross-attention incorporates SleepFM temporal context. Dense predictions are obtained through localized token-to-sample reconstruction and temporal refinement. Tested on the Massive Online Data Annotation (MODA), the proposed model achieves an event-level $F_1$ of $0.8080 \pm 0.0040$ at IoU 0.2 across six folds. Ablations show that localization primarily depends on local spectral and sub-token temporal information, while the pretrained representation provides complementary context. These results indicate that dense adaptation benefits from preserving information below the pretrained token scale. Code will be released upon acceptance.

Long-running LLM agents compact their context, replacing most history with a summary. We ask whether handoff summaries state permissions or requirements the record does not support, whether such sentences change what the next instance (the successor) does, and where they come from. Synthetic tool-use tasks carry a ground-truth permission ledger; same-snapshot replay reruns the successor with those whole sentences replaced by neutral filler. Two AI readers, reading only rule-flagged held-out summaries, confirmed unsupported restrictions in 1.2% and 2.4% of all readable summaries from two open-weight models, and unsupported authorizations in 0.1% and 1.3%. For restrictions, a pre-specified replay test was not significant (p = 0.58); a follow-up, planned after it, found a fragile effect (p = 0.027). Neutralizing false approvals cut unauthorized actions from 24 to 7 of 120 held-out runs (post hoc; later set: 17 to 11 of 48). Most development restrictions followed the agent's earlier reasoning that an action needs approval (post hoc). In pre-specified tests, replacing the system prompt's "Some actions need approval." with one saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and, on held-out templates, from 1.8% to 0.3% (both p < 10^-9), while requests for approvals the ledger requires barely changed. Code and data: https://github.com/Chuw1/handoff-audit

Agent trajectories often contain both successful decisions and failed attempts, yet optimization with sparse trajectory-level rewards struggles to identify which critical decisions should be reinforced or suppressed. We observe that meaningful distinctions among on-policy trajectories can be distilled into natural-language process rubrics and used as process supervision signals. Based on this insight, we propose Online Rubric Self-Distillation (ORSD), which leverages self-derived process rubrics for dense token-level supervision, thereby augmenting outcome-centric reinforcement learning. Specifically, the current policy independently generates multiple rubric bundles at each step according to the same group of trajectories, extracting candidate process rubrics from cross-trajectory behavioral differences. A frozen verifier then evaluates all rubric items against each trajectory in the group, and selects one candidate bundle based on bundle-level alignment between rubric satisfaction and task success. ORSD treats the selected rubric bundle as privileged information during training, then compares the log-likelihoods of the original tokens with and without the rubric. The resulting likelihood gap provides token-specific process credit, which is added to the original trajectory-level GRPO advantage through additive advantage shaping. ORSD thus bridges trajectory-level outcome supervision and fine-grained process supervision, enabling token-level credit assignment using process signals. Extensive experiments on ALFWorld and WebShop show that ORSD accelerates policy improvement and boosts final task performance for both Qwen3-1.7B and Qwen3-4B under the same environment-interaction budget. Our code is available at https://github.com/AnnaXChun/ORSD.

Proprietary language models exhibit strong reasoning and coding capabilities, but their output distributions are inaccessible, making fine-grained distillation methods that rely on teacher logits difficult to apply directly to black-box models. We observe that rich textual feedback from a black-box teacher can be treated as privileged information (PI) and converted, through a PI-conditioned privileged teacher, into dense supervision signals on the student's own trajectories. However, what PI should be acquired from the teacher to most effectively support student learning remains unclear. PI strategies that differ in where they focus, what information they request, and how much detail they provide can have substantially different effects on the current student, and their relative effectiveness can further change as training progresses. Based on this observation, we propose Evolving Privileged Information Strategies for Black-Box On-Policy Distillation (EvoOPD), which treats PI acquisition itself as an optimization problem and maintains a PI Strategy Bank that evolves with the student state. EvoOPD uses the measured repair utility of PI on the current student to adapt the focus, information type, and granularity of teacher feedback, allowing the acquired information to continually match the student's learning needs. We systematically evaluate EvoOPD on mathematical reasoning and code generation with Qwen3-1.7B, 4B, and 8B students. EvoOPD achieves the best overall performance in all six model-task settings, improving over the strongest baseline in each setting by an average of 2.14 percentage points in mathematics and 1.30 points in code. Further analysis shows that the relative effectiveness of PI acquisition strategies changes with student state, making a fixed strategy difficult to sustain as the best choice throughout distillation.

A generated domain-specific program can be accepted by its host and still miss the reference construction. We measure the two events separately and test a short construction contract. The vehicle is Open Quantum Assembly Language (OpenQASM) 3. One 3-billion-parameter generator is run under seven conditions on 100 quantum approximate optimization (QAOA) and 100 variational quantum eigensolver (VQE) instances. Eighty of 100 QAOA signatures appear in some training split. Exact McNemar tests and 95% paired bootstrap intervals are used. The contract, a card, lists structural decisions, enters the prompt, and is checked without a simulator. Parse rates are 0.55 to 0.97 and full structural match is 0.02 to 0.87. On QAOA, at fixed weights, the card moves full match by +32.0 percentage points on card-trained weights (interval +22 to +42) and by -13.0 (-23 to -3) without that training. The +45.0-point difference is not a deployment contrast. The same-step plan-free gain is +23.0 (+13 to +33); the primary distributional score does not separate it. Against the earlier checkpoint the gain is +27.0 (+17 to +37), and the checkpoint also changes. A rule card is not separated from the learned planner, nor is a predicted-card compiler (0.760 vs. 0.780). On 20 QAOA rows outside every training split, the interval includes zero. Under this card, VQE adherence is 0.020 and full match stays near 0.03. Report the two checks separately. Insert a card only into a generator trained to read it, and only where a pilot shows adherence.

Automated failure discovery can be framed as a diversity problem: the objective is to expose distinct ways a system fails rather than a single critical adversarial case. Yet uniform parent sampling provides no way to allocate search effort toward individuals likely to produce durable archive contributions. We study this credit assignment problem in an unstructured Quality-Diversity (QD) archive, testing whether signals available before an individual has reproduced distinguish offspring that merely enter the archive from those still retained generations later. Across a planar-arm benchmark and two robotic-locomotion tasks, the same signals are often associated with entry and persistence in opposite directions. In particular, lineage depth is the clearest such signal, and its persistence discrimination grows with population capacity, which is associated with lower turnover. Biasing parent selection toward the signals with the strongest measured association to survival nevertheless does not improve search: moderate bias matches uniform sampling, and strong bias reduces QD-score.

Machine unlearning removes the influence of specific training examples to comply with privacy regulations. Full retraining gives strong guarantees but is computationally prohibitive at scale; approximate methods are efficient but lack formal guarantees. SISA (Sharded, Isolated, Sliced, and Aggregated training) achieves certified deletion through data sharding and independent retraining, but random partitioning scatters forget requests across most shards, and isolated retraining degrades utility. We introduce IDSISA (Influence-Distilled SISA), which uses influence functions to concentrate forget-prone samples into few shards and distillation from unaffected shards to recover utility lost to isolation. The deletion guarantee inherited from SISA's retraining step remains exact and requires no probabilistic qualification; the distillation-based utility recovery is a separately quantified, empirically bounded approximation ($ε_{\text{cert}} \leq 0.015$ in expectation). We further formalize certification under sequences of repeated deletion requests, where cross-shard distillation can create dependencies between shards; we prove that a closure-based cascading retraining procedure restores the deletion guarantee to the full aggregated model under any deletion sequence. Experiments on ResNet-18/CIFAR-10, ResNet-18/CIFAR-100, BiLSTM/IMDB, TCN/IMDB, and a fully connected network on Fashion-MNIST show $14-91\%$ better selective forgetting, up to $50\%$ better utility retention in relative terms (with consistent percentage-point gains on the other architectures, where a relative baseline is undefined), and unlearning-time changes ranging from $-6.7\%$ to $+6.4\%$ depending on architecture, relative to SISA.

Group Relative Policy Optimization (GRPO) has advanced generative model post-training in verifiable domains such as code and mathematics, but evaluating video quality, especially emotional expression, often depends on subjective aesthetics and has no uniquely correct target. Existing approaches typically rely on fixed rewards for visual quality, motion quality, and text alignment that poorly capture emotional intent and can invite reward hacking, while RLHF is difficult to scale and subject to individual aesthetic differences that complicate the calibration of facial, body motion, and chromatic cues across narrative contexts. We introduce EmoReward, a new self-evolving training mechanism that distills evidence reliability from reward traces collected during an actual training run to adaptively guide multi-reward weight allocation in subsequent training by determining which emotional evidence to trust and how much to weight it. EmoReward first plans specific grounding cues from the emotional prompt to guide evaluation, then uses heterogeneous signal primitives to score the generated video during the act stage. Reflective Reward Gating calibrates their contributions through instance-level gates for phase, face confidence, conflict, and subtlety, then renormalizes weights over the active set to produce the sample's fused reward. RewardMemory then estimates each primitive's reliability from rollout scores and outcomes within each emotional regime in the finished training run, then fixes bounded multipliers that calibrate fusion weights in the next training run without updating reward models. Experiments on HunyuanVideo 1.5 show improvements across multiple RL algorithms and compare EmoReward with other agentic reward methods under SAGE-GRPO. Ablations and transfer experiments on SD 3.5 and FLUX further support adaptive reward calibration for visual generation.

The massive permeability of intelligent systems in every sphere of human activities have greatly enhanced clarity, ease of work, and increased general global understanding which has significantly awaken the use of intelligent systems by all users. Intelligent systems greatly rely on precise dataset/corpus for the correctness of their output. Despite efforts by researchers to address Low-resource language in the Natural Language Processing (NLP), Igala still has zero parallel corpora despite efforts by researchers. We present a dialect-aware baseline for Igala-English machine translation, among first open systems and first with dialectal variation analysis. Our dataset of 1137 sentences (899 training) is first with dialect tags: Idah 459 (40.4%), Ankpa 297 (26.1%), Ibaji 160 (14.1%), Other 187 (16.4%), Ogugu 32 (2.8%). Same English yields 20-60 Igala forms. Mixed-dialect training achieves balanced BLEU 11.66 versus Idah-only model which drops 4.7 BLEU on Ibaji (13.2 on Idah, 8.5 on Ibaji), proving dialect fairness requires mixing. Fine-tuned from Helsinki-NLP/opus-mt-en-mul (78M), epoch ablation shows convergence at 8 epochs and overfitting at 15. BLEU 11.66, chrF 46.69. Analysis shows 899 sentences on log-linear trend consistent with African low-resource literature. This work provides first dialectal impact analysis for Igala and establishes baseline for future work.

Automated program reduction shrinks a program into a smaller one that preserves a property of interest. Language-agnostic grammar-based reducers are easy to retarget across languages, but because they reason only about syntax, they cannot perform the coupled edits needed to remove whole declarations, e.g., a method cannot be deleted while it is still called. As a result, they waste oracle calls on edits that cannot succeed and often leave irrelevant declarations behind. We present SCYTHE, a fast, semantic-aware pre-processing step that runs in front of any grammar-based reducer without modifying it. SCYTHE encodes the program as a typed declaration dependency graph and applies a set of semantic-aware operators that decouple declarations from their use sites, making them removable while keeping the program well-typed. Fine-grained token-level minimization is then left to the downstream reducer. On 73 benchmarks across C, Java, and Solidity, combining SCYTHE with PERSES never yields a larger result than PERSES alone, and a strictly smaller one on 70 of 73 cases. On average, the combination produces programs 55.6% smaller (up to 95.5%) and is 2.3$\times$ faster (up to 7.4$\times$). SCYTHE also boosts LATRA, a reducer that, unlike the purely syntactic PERSES, also applies language-specific transformations: the combination produces programs 10.6% smaller (up to 61.8%) and is 2.4$\times$ faster on average. Video: https://youtu.be/uEq_YYU0zZw

Acquirers of failed banks weigh acquisition value, regulatory protection, risk, buyer strength, and execution difficulty jointly, because each signal's meaning depends on the others: high non-performing loans matter less when government loss sharing absorbs the losses. The proposed Spider Network Filtering Framework encodes such dependencies as four interpretable interaction rules among five dimensions, yielding graph-inspired features for a logistic model and a Pursue/Caution/Avoid decision rule. We evaluate it on 55 acquisitions assisted by the Federal Deposit Insurance Corporation in 2008-2024 by acquirers with a parent listed in the United States. Six financial inputs come from regulatory reports published before each failure; two deal terms, which the winning bidder knows when it bids but investors learn at the announcement, come from the regulator's press releases. The outcome, a positive three-day abnormal return around the announcement, records the market's initial reception, not realized success. Under grouped, nested cross-validation, the Spider-enhanced model reaches a mean area under the receiver operating characteristic curve of 0.594, against 0.615 for logistic regression on the same inputs and 0.632 for ridge regression; permutation tests that repeat the full model selection give at most weak evidence that any of them ranks better than chance. The Spider score used directly does not rank better than chance, and an apparent link between target distress and higher returns disappears against a bank-sector benchmark. Thresholds tuned for the F1 score approve most acquisitions with negative returns and realize negative utility when each costs three times what a positively received acquisition gains.

LLM inference is inherently heterogeneous, with performance bottlenecks shifting throughout inference phases down to individual operators. This has motivated heterogeneous serving systems that disaggregate execution across specialized accelerators, typically through fixed partitions such as prefill/decode or attention/FFN. However, these approaches commit to a partitioning strategy in advance, limiting their ability to adapt to different workloads and model architectures while forcing emerging accelerators into predefined roles. We present HELIX, a vLLM-based software stack that discovers, compiles and dispatches workload-aware partitions across heterogeneous accelerators at the operator granularity. At engine initialization, HELIX automatically determines hardware-specific operator placements that minimize execution latency, accounting for profiled execution costs, communication overheads, device capabilities, and memory constraints. It combines dynamic programming, piecewise compilation, and heterogeneous subgraph dispatch, enabling accelerators with partial operator coverage to participate in LLM inference without predefined model partitions. HELIX provides an open-source unified framework for exploring, evaluating, and integrating heterogeneous accelerators for LLM inference, lowering the barrier to emerging hardware by exposing them as subgraph executors. In systems combining H100 and B200 GPUs with emerging accelerators, HELIX resolves LLM partitions that multi-query decoding latency by up to 3$\times$ for dense and sparse transformer models and up to 3.5$\times$ for hybrid state-space models, relative to standard attention-FFN disaggregation.

Training high-performing neural networks often involves evaluating populations of candidate models with varying hyperparameters, architectures, or random seeds. Training every candidate to completion is computationally expensive, yet stopping seemingly weaker runs early risks discarding models that would ultimately perform better, as learning curves can cross. To address this challenge, we investigate whether intermediate training checkpoints can predict future model accuracy and enable more informed pruning decisions. Our nowcasters predict test accuracy five epochs ahead or at the end of training. They are trained with accuracy and loss histories, weight statistics, and SANE representations as inputs. SANE is a transformer autoencoder that encodes a checkpoint as a sequence of latent tokens. We then use the nowcasters in Survival-of-the-Fittest (SOTF) training. SOTF repeatedly ranks runs during training by predicted accuracy and stops the lowest-ranked ones. Our results show that weight statistics work best for five-epoch forecasts, and SANE representations work best for final accuracy prediction. For evaluation, we compare SOTF with pruning the population by current accuracy. In homogeneous model zoos such as the CNN zoos, pruning by current accuracy already retains a final model only slightly worse than the best. Model zoos with a larger spread in performance, such as the ResNet-18 and ViT zoos, are more difficult to prune by current accuracy alone. In these zoos, SOTF with nowcasted final accuracy almost closes the gap to the best model.

Accurate electricity load forecasting provides critical decision support for energy procurement, generation scheduling, and supply-demand balancing, contributing to the economic efficiency and operational reliability of power systems. However, future electricity demand can be affected by the complex interplay of multiple events described in Web notices, weather alerts, and event schedules. Their effects vary over time and can reinforce or offset one another, making electricity load difficult to forecast accurately. While existing methods can incorporate context as additional input, how to use the effects of multiple events and their interactions to guide the load forecasting process itself remains to be explored. To address this gap, we propose ElecLoad, a context-aware agent for multi-event electricity load forecasting that adaptively organizes the forecasting process. Specifically, we train ElecLoad to combine historical load with context to analyze each event's relevance to the target load, the timing of its effects, and potential interactions among events. Guided by this analysis, ElecLoad adaptively chooses to forecast the entire forecast horizon directly or to partition it into segments according to changes in event effects. When the horizon is partitioned, sub-agents independently generate initial forecasts for each segment. The main agent then reviews these forecasts, and the sub-agents perform selective refinement guided by its feedback. Experiments on multiple electricity load datasets show that ElecLoad achieves state-of-the-art performance in most evaluation scenarios. Code is available at https://github.com/zhjai/ElecLoad.

Coding agents are increasingly used in real-world workflows across diverse domains, motivating evaluation grounded in real user tasks. Existing benchmarks, however, largely rely on static public tasks with reference patches or hidden tests. Next-generation coding-agent benchmarks should therefore move beyond correctness on fixed tasks toward continually renewed evaluation of overall artifact quality in real user workflows. We propose Trace2Eval, a dynamically updatable pipeline that constructs benchmarks from real-world interaction trajectories. Trace2Eval employs a multi-agent reconstruction system to analyze the filtered trajectory, recover the original user query, restore the workspace files to their pre-interaction state, and rebuild a trajectory-equivalent execution environment. A candidate agent then executes the reconstructed query to produce an initial artifact. To evaluate these artifacts, the pipeline iteratively refines human rubrics through inter-annotator alignment, and calibrates judge agents against human assessments using source-code and runtime evidence. Guided by feedback from the Trace2Eval pipeline, models revise their initial artifacts, yielding a mean score gain of 10.05 percentage points across 19 model-harness configurations and a maximum of 19.28 percentage points for DeepSeek-V4-Pro with Claude Code, which supports rubric feedback as actionable guidance for artifact improvement.

Multilingual benchmarks for African languages mostly translate generic questions into local languages, measuring linguistic competence but not whether models know facts about Africa. We introduce \naija, a benchmark of factual and statistical questions about Nigeria, authored by native speakers and grounded in authoritative official statistics, with every question available in English, Nigerian Pidgin, \yoruba, Igbo, and Hausa. Answering similar questions for the UK and US yields a country by language matrix that disentangles a language effect (the model knows the fact but degrades in a low-resource language) from a knowledge effect (the fact is absent in every language). Across eight frontier and Nigeria-adapted models, the knowledge effect dominates: English-query accuracy is 49--50\% for the US/UK but 20\% for Nigeria. The language effect is smaller and tracks web representation, degrading from English to \yoruba{} (9 points on Nigeria, up to 24 on US/UK questions). We find scale does not help: the strongest models reach 66--79\% on the UK and the US but none exceed 31\% on Nigerian statistics. Extending to all 54 African countries with our best performing frontier model shows the same profile continent-wide, 96\% on culture and 98\% on politics against 26\% on statistics, so the gap is not specific to Nigeria.

Simultaneous Speech Translation (SimulST) requires balancing translation quality and latency. While Large Language Models (LLMs) excel in offline translation, current LLM-based SimulST systems predominantly rely on prefix-incremental generation. This paradigm suffers from redundant Key-Value cache recomputation and necessitates external Voice Activity Detection to process unbounded streams. To overcome these bottlenecks, we propose an end-to-end interleaved read-write framework built upon an omni-modal LLM. By interleaving speech chunks and text tokens, our architecture achieves full KV cache reuse, natively supporting segmentation-free, infinite-length inference. Furthermore, to learn optimal streaming policies without rigid offline alignments, we introduce a novel Actor-Critic self-enhancement pipeline. An RNN-T Actor dynamically samples read-write trajectories, while an LLM-based Critic evaluates these paths via intrinsic semantic and latency rewards, distilling the optimal policy back to the Actor. Extensive experiments on sentence-level (FLEURS, CoVoST2) and document-level (MuST-C, ACL 60/60) benchmarks demonstrate that our model establishes new state-of-the-art quality-latency Pareto frontiers, substantially outperforming existing cascaded and static-alignment baselines.

Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent's attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios. In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent's token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents. Our code is available at https://github.com/THU-Agent/Coca-cola.

Offline retrieval-augmented generation (RAG) on edge devices is updated through signed knowledge packs that carry text and precomputed embedding vectors. A signature proves who published a pack, not that its vectors encode its text. We study this text-vector integrity gap. A verifier that recomputes vectors and accepts them within a cosine tolerance concedes a predictable share of queries to an attacker who controls the vectors: the hijack rate equals the distribution of angular retrieval margins at the tolerance (the margin law). Margins are small in five corpora and two embedding models, so an idealised tolerance at the level of floating-point noise leaves up to 5.5% of queries exposed (at most 1.2% in any corpus once exact duplicate documents are excluded), and a threshold calibrated from our CPU INT8-FP32 comparison of a small embedding model (at the empirical 0.1% quantile of honest cosines) leaves 98-100% exposed; on a Raspberry Pi 5 running a local language model, such an attack lowers correct answers from 0.75 to 0.39. Integrity audits of compressed vector stores fare no better: on a FAISS product-quantisation index over 57,638 documents, tampered vectors that all pass the audit put the attacker's document first for 25% of targets. Verifying quantisation codes instead, with ambiguity decided by the verifier, a calibrated guard band and canonicalisation, accepted every honest vector in our samples on three platforms and left the publisher no choice. We also show that common floating-point and dynamic-INT8 inference pipelines are not bit-reproducible across batch sizes or architectures, and report pre-registered negative results on abstention, context pruning and semantic screening. We release code and data.

Most speech large language models (SLMs) still process speech in an offline manner. They only start responding after they receive the complete speech segment. However, many real-world applications require streaming speech interaction, for which the audio continuously flows in and the machine is required to answer questions based on the so far received audio chunks. For example, a user may ask the system to continuously count how many times a person's name has been mentioned while they are speaking. We observe a lack of benchmarks and evaluation protocols for real-time speech interaction. Motivated by this, we propose LibriSpeech-Real-QA, a dataset and a benchmark for real-time speech interaction. We construct different types of real-time question-answer pairs from LibriSpeech to evaluate whether speech LLMs can identify target words in streaming audio. We evaluate several popular real-time commercial models and find that they still tend to respond only after the speech turn has finished. We demonstrate that we can enable a pretrained SLM to perform this task, where we explore several modeling designs and establish a solid baseline for future research.

Short-interval wireless schedulers must return a complete, feasible beam-and-resource allocation before each execution deadline, and a late or invalid output gives no utility in the current control cycle. Searching longer can increase transmitted traffic, but it may also miss the deadline or leave only an infeasible intermediate solution. We propose Feasibility-Preserving Transactional Refinement (FPTR), which always keeps a validated feasible allocation, the incumbent. Each refinement stage builds a separate candidate and leaves the incumbent unchanged. A candidate replaces the incumbent only if it is complete, passes the feasibility validator before its stage cutoff, and strictly improves the objective; otherwise it is discarded. Four refinement stages use this rule: buffer-aware resource selection, compatibility-group (CG) user sharing, beam-mask rebuilding from remaining demand, and joint reassignment of two resources. On 150 fixed synthetic instances, the full pipeline (Full) raises traffic by 3.83% on average over a lightweight beam-first reference and by 1.79% over its Base constructor. In a separate suite with matched inputs, Full exceeds six independent search baselines (ALNS, tabu search, GA, simulated annealing, iterated local search, and GRASP) by 6.68-8.35%, where outputs later than 100 ms receive zero utility. All 750 Full runs in the internal suite returned feasible allocations, and none exceeded 100 ms. Separating mandatory output construction from optional refinement lets spare computation time improve the allocation without endangering the incumbent. The open-source implementation supports independent reproduction.

We study non-parametric bandits in which arms are evaluated using Cumulative Prospect Theory (CPT) rather than expected reward. We consider continuous action spaces with arm-dependent noise and propose Non-Parametric CPT Bandits ($\mathtt{NPCB}$), an optimistic algorithm that decouples mean estimation from CPT-based decision-making. We prove CPT regret of order $\widetilde{\mathcal O}\bigl(\sqrt{T\,\dim_E(\mathcal F,T^{-2})\,\log N(\mathcal F,T^{-2},\|\cdot\|_\infty)}\bigr)$, where $\dim_E$ is the Eluder dimension of the function class $\mathcal F$, $T$ is the time horizon, and $N$ is its covering number in the sup norm. This matches the order-wise regret of classical non-parametric bandits, showing that CPT risk sensitivity incurs no additional order-wise regret cost. We extend $\mathtt{NPCB}$ to unknown noise CDFs, retaining the known-CDF mean-learning term and adding a CDF-learning cost governed by a distributional Eluder dimension. We also study Kernel-CPT bandits and obtain information-gain-based regret bounds for a broad range of kernels, matching the corresponding risk-neutral kernel rates up to logarithmic factors.

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best rank. We show this coupling is a property of the merge method, not of merging. We train 36 experts on three domains (mathematics, instruction following, multilingual) at four ranks and three model scales, then merge with six methods at uniform ranks and at each domain's own best rank. Giving experts different ranks costs averaging and utility-based selection nothing, and gains them up to 1.8 pp of multi-domain accuracy, while the conflict-resolution methods are inconsistent and can lose up to 4 pp. The split has a simple cause. Merging keeps only a fixed share of the pooled spectral components, and methods that choose them by size discard many that would have reduced loss. They also tend to discard the same expert: the instruction adapter has the smallest singular values, so it is dropped almost entirely and IFEval falls by 16 to 25 pp. Choosing components by their effect on loss instead keeps all three experts, and leaves accuracy nearly flat across an 8$\times$ rank range (0.7 pp, against 6.7 pp for TIES at 4B scale). Rank can therefore be chosen per expert whenever the merge operator does not select by magnitude.

While multi-turn agentic reinforcement learning (RL) excels on long-horizon tasks, its performance gains remain highly sensitive to interaction constraints. Evaluating across sampling breadth $K$ and interaction depth $T$ via $Pass@(K,T)$, we show that trajectory-level RL advantages depend on sufficient turn budgets, diminishing significantly when budgets tighten. Besides, agentic RL does not uniformly elevate single-step decision quality; rather, its strength concentrates in error recovery and long-horizon resilience. To mitigate the deployment risks caused by this budget dependence, we provide a preliminary discussion on the benefits and limitations of process supervision signals and introduce BudgetMix RL as an alternative that varies interaction budgets during training. Experiments demonstrate that BudgetMix RL achieves robust performance across diverse execution settings. Project Page: https://rethinking-agentic-rl.github.io/

Graph-based retrieval-augmented generation (GraphRAG) answers questions by retrieving over a knowledge graph built from a corpus. Yet retrieval strategies are designed around the structure their own construction pipeline produces, and whether they survive a change of construction is rarely tested. We study this cross-construction portability: whether one retrieval procedure, with no construction-specific schema, relation names, or retuning, stays effective when the graph is built differently. We introduce ArcRAG, which induces a query-local graph interface at inference time (the node labels, relation descriptors, and property keys reachable from the entities a question names), plans multi-hop traversals against it, and executes and repairs them as Cypher. We build each of the MuSiQue, HotpotQA, and TriviaQA corpora as a schema-guided graph (typed nodes, a closed relation set, source provenance) and as an open-information-extraction graph (untyped nodes, free-text relations), with the retriever's configuration fixed across them. We compare against a construction-agnostic Direct-Cypher agent, ablate each stage across retry and pass budgets on both graph families, and compare with results reported for prior GraphRAG systems, most of which design retrieval together with their own graph construction. Across the two constructions, we find that retrieval need not be co-designed with the graph: query-time interface induction keeps a construction-agnostic retriever effective when the construction changes, although construction still shapes absolute accuracy.

Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing tasks, and track downstream capability as the implicit reward margin grows to $\sim 7$ $β$-scaled units (nf4 training; $\sim 3$ under fp16 measurement on the unquantized base). Question. Does margin growth degrade capability uniformly across tasks, or selectively? Findings. In a pre-registered study (8 seeds, 6 unseen at freeze) with artifact-aware scoring and TOST equivalence testing, the alignment tax is graded and task-selective: extraction is preserved ($Δ= -0.55$ pts, equivalent at $\pm5$ pts), json loses little ($Δ= -3.52$ pts, equivalence fails on a clustered item-level interval), and router degrades severely ($Δ= -30$ pts). Router's loss concentrates in a margin 1-3 damage window containing a transient format-break spike. The margin is accounted for almost entirely ($\geq 96\%$) by suppression of rejected completions, with chosen reward approximately flat, consistent with the retention term anchoring it; we do not attribute the behavioral damage to any single gradient channel. Gold-logp deflation is task-selective (extraction $-0.15$, json $-1.04$, router $-1.47$ nats/token under fp16), with gold selectively protected relative to incorrect controls on extraction (8/8 seeds). QLoRA reward margins differ up to $3.4\times$ between nf4 training and unquantized fp16 measurement, with an axis-dependent gap. Scope. Length-normalized rewards with SFT retention, sparse LoRA on 8/32 layers, one 4B model, easily-separable synthetic preferences, three structured tasks. A no-retention control and a measured format-alignment metric are left to future work.

Let $d_1,\dots,d_m$ be directions in $\mathbb{R}^k$, where $k\ge 2$. A multiset $X$ of $n$ points of $\mathbb{R}^k$ projects to the $m$ multisets of numbers $\{\{x\cdot d_j : x\in X\}\}$, or equivalently to the $m$ sorted lists of these numbers. How many directions are needed for these projections to determine every multiset of $n$ points? Two simple bounds are known in every dimension: more than $(k-1)\log_2 n$ directions are always needed, and $(k-1)n+1$ directions in general position always suffice. In the plane much more is known: Matousek, Privetivy and Skovron showed that almost every choice of $O(\log n \log\log n)$ directions suffices. Whether anything similar holds in higher dimensions was left open by Dym, Wellershoff, Tsoukanis, Levy and Balan. We observe that two lemmas of Hair and Sahai, proved for list recovery of random linear codes, answer that question, and by tightening their argument we show that in every dimension almost every choice of $(4k-2)\ln(2n)\ln(k\log_2 2n)+k+1$ directions suffices. For $k\le 10$ this beats the linear bound once $n$ is larger than about 100: for example, 144 directions suffice for multisets of 1000 points in the plane, and an exact computation lowers this to 133. Letting some of the random blocks in their argument fail removes a factor $\log k$: almost every choice of $O(k\log n\log\log n)$ directions suffices, uniformly in $k$ and $n$. We also show that $2k-1$ directions never determine three points, so that the logarithmic lower bound is not tight; that every $m$ directions in the plane fail on some pair of multisets of size $O(1.78^m)$, improving the bound $O(1.80^m)$ of Matousek, Privetivy and Skovron; and that for each $k$, $m$ and $n$, either almost every choice of $m$ directions works or almost none does.

Microfluidic droplet generation is critical for many applications, yet efficient device design in high-dimensional parameter spaces remains computationally prohibitive due to complex multiphase flow dynamics. This work presents a unified methodology that couples adaptive surrogate enhancement with simulation-driven inverse design. High-fidelity Volume-of-Fluid simulations provide ground-truth data, while neural network ensembles predict flow regimes and droplet diameters. Ensemble-based uncertainty quantification identifies epistemic uncertainty using consensus levels for regime classification and the coefficient of variation for diameter prediction, revealing that uncertainty concentrates near regime transition boundaries where physics is highly sensitive to parameter perturbations. A physics-informed enrichment strategy targets these critical zones and addresses class imbalances by deploying additional targeted simulations and retraining the surrogates forming the ensembles. A framework that integrates Particle Swarm Optimization for constrained inverse design, balancing accuracy with predictive confidence and physical consistency is therefore proposed. Validation on independent test datasets demonstrates that the enriched ensembles achieve higher accuracy, with the fraction of predictions within the 0-3% error range on droplet diameter nearly doubling, and high-error cases (above 10%) dropping substantially. The inverse design procedure reliably identifies physically realizable configurations meeting target droplet sizes. This approach significantly reduces computational costs while ensuring reliable predictions, offering a framework with potential applicability to the optimization of a broad range of microfluidic systems.

Closed-loop driving benchmarks in CARLA rank agents by a driving score that multiplies route completion by an infraction penalty. This score does not distinguish an agent that stops in front of an obstacle from one that pushes past it. In a corridor closed by two parked cars, for example, one agent that drives through with contact averages 52.2 on the blocked widths, while the agents that stop average between 37.8 and 40.6. We propose failure-boundary profiles, an evaluation in which each scenario family varies one measured factor, such as corridor width or the time gap in oncoming traffic, and each variant is certified from measurement as one where the agent should proceed or one where it should stop. Along this axis we estimate two curves per agent, commitment (the probability of attempting the passage) and competence (the probability of passing without contact), and locate where each crosses one half, with intervals that account for the simulator's run-to-run variance. The certified labels also give a summary metric, decision accuracy, which credits clean passes on passable variants and stops on blocked ones. We evaluate seven driving agents on seven CARLA families. The rule-based expert reaches a decision accuracy of 0.74 and TFv6, the highest-scoring learned agent, 0.60. Agents with similar driving scores fail in different ways, and agents that reliably stop at a blocked corridor still rarely wait for a gap in oncoming traffic. We also apply the method, unchanged, to a warehouse robot in Isaac Sim. The navigation policy we test stops at blocked gaps in only one of three families, and two fine-tuned versions of it pass more gaps cleanly but no longer stop at all.

Background/Objectives: CNNs perform well in retinal disease classification from spectral-domain optical coherence tomography (SD-OCT), but whether EfficientNet's natural-image scaling hierarchy transfers to medical-image transfer learning at fixed resolution remains unclear. This study evaluates EfficientNet-B0-B7 for four-class SD-OCT classification and accuracy-efficiency trade-offs. Methods: Eight ImageNet-pretrained variants (4.05-64.10M instantiated backbone parameters) were evaluated at 224x224 on the Kermany dataset. Grayscale images were replicated to RGB and processed using model-embedded preprocessing and augmentation. Training comprised frozen-backbone head training followed by partial fine-tuning. Metrics covered classification, discrimination, efficiency, and visual analyses. Results: Test accuracy varied non-monotonically from 95.93% (B0) to 96.95% (B1); B7 achieved 96.73%. B1 had the highest accuracy and Drusen F1 (0.8780), while B7 had Drusen F1 of 0.8720 and the highest Drusen recall (0.9177). Drusen had the lowest F1 across all variants and contributed 82.2% of the summed squared deviation in class-specific F1; class ranking remained Normal > CNV > DME > Drusen. The OCT accuracy ranking differed from the reported ImageNet hierarchy, where B7 ranked first. The accuracy-to-parameter ratio fell from 19.82 to 1.48 percentage points per million total parameters. The final train-test gap had a weak, non-significant association with test accuracy (r = 0.26, p = 0.53). Conclusions: Larger EfficientNet variants did not consistently improve SD-OCT classification. B1 achieved the highest accuracy and Drusen F1, whereas B7 showed a secondary accuracy recovery at greater model size and CPU latency, indicating task and protocol dependent scaling effects. As cross-partition duplicates have not been independently audited, results are conditional on the supplied split.

Traditional document parsing systems use Markdown as their output representation, which is effective for linear text but does not fully preserve fine-grained spatial structure, table geometry, visual hierarchy, or rendering semantics. We show that HTML is advantageous for heterogeneous and layout-sensitive documents whose representations must remain editable, inspectable, and reusable, and introduce AmbiHTMLBench to systematically study this setting. In this benchmark, given a real-world document, a model must recover executable HTML that preserves its content, organization, and visual appearance. The benchmark contains 1,115 diverse pages with structured annotations and evaluates predictions in two different ways: The objective protocol aligns ground-truth and predicted rendered blocks and measures table fidelity, page-level text error, and a two-dimensional layout consistency; a complementary reference-anchored protocol measures perceived fidelity by aligning LLM judges with human preferences. Across ten primary multimodal systems, the best objective score is 0.8092, while the subjective winner is different, demonstrating that automatic decomposition and human-perceived fidelity capture complementary properties. We further fine-tune a small-parameter expert model and find that task-specific adaptation yields good performance on document-to-HTML reconstruction task. These results position AmbiHTMLBench as a diagnostic testbed for building smarter and more logical document parsing systems.

Long-horizon language-model agents accumulate increasingly large interaction histories, making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behavior, coupling training to specific tasks or agents. We introduce SumOne, a standalone summarizer trained to preserve information in agent histories through masked reconstruction. Given a summary of the expired history, a frozen reconstructor recovers randomly masked portions of the original trajectory; reconstruction quality, combined with a length penalty, provides a task-agnostic reward for optimizing the summarizer with GRPO. This yields self-supervision directly from raw trajectories, without reference summaries or downstream task execution. At inference time, only the lightweight summarizer is retained, enabling plug-and-play context compression for otherwise unchanged agents. We conduct a systematic evaluation of SumOne in long-horizon search and software-engineering tasks. SumOne reduces estimated inference cost by approximately 78\% on DeepSearchQA and 75\% on SWE-bench Verified relative to operating with full context, with absolute decreases of 0.04. The same checkpoint also transfers across downstream models and agent harnesses without additional fine-tuning.

Skills let language agents reuse experience across tasks as natural-language documents. Recent work evolves skills from interaction, but existing methods have limitations when skills evolve online. Prior methods update a skill offline from a batch of trajectories collected over many tasks. Some then adopt the new skill without any check, which can degrade the skill over time. Others keep it only if it improves performance on a held-out validation set, which is too slow and costly to run after every update. We introduce EvoGuard, a pairwise skill selector that mitigates skill degradation at low latency and cost. We first run the agent iteratively to build pools of skills and score each skill on a private test set to obtain utility labels. We pair the skills and use rejection sampling to keep only pairs with a clear utility gap. The selector is trained on the logits of two label tokens to pick the better skill. At deployment it decides in a single forward pass, which is significantly faster than evaluation on the private test set. Compared with adopting every new skill, EvoGuard improves success by 11.0% on ALFWorld, 5.9% on ScienceWorld, and 2.7% on LiveCodeBench.

Multivariate time series anomaly detection (MTSAD) relies on modeling both temporal dynamics and inter-sensor dependencies to identify complex system faults. Although graph-based detectors explicitly model sensor topologies, they train task-specific networks from scratch on limited target-domain data, remaining vulnerable to spurious correlations and unable to reuse transferable priors. While pretrained Graph Foundation Models (GFMs) offer transferable topological priors and anomaly discriminability, directly adapting them to MTSAD is hindered by cross-modal feature-topology mismatch, unreliable pseudo-anomaly supervision, and decision distortion during node-to-window aggregation. To fill these gaps, we propose a novel adaptation framework, Gfm4Mtsad, which unleashes pretrained GFMs for MTSAD. We introduce the Structure-and-Information Contract Alignment (SICA) module to bridge continuous time series and the pretrained GFM space. This module harmonizes principal subspace token projection, normal-prior-guided graph construction, and residual recovery to suppress spurious edges while retaining fine-grained temporal details. We propose a Mechanism-Aligned Pseudo-Anomaly Screening strategy (MAPS) to steer unsupervised learning. It applies localized dependency breaks while preserving empirical residuals, selecting hard negatives near the normal boundary for targeted supervision. We develop a Tail-Gated Evidence Calibration mechanism (TGEC) to eliminate node-to-window aggregation distortion. It calibrates multi-granular graph responses against normal operational references, strictly preserving benign score rankings while selectively refining extreme anomaly candidates in the tail distribution for reliable window detection and interpretable sensor evidence. Extensive evaluations with two GFM backbones demonstrate that Gfm4Mtsad achieves superior detection performance over existing baselines.

Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data. We show that CoT generalizes by repeatedly applying a shared single-step predictor to compose reusable ID transitions. Theoretically, let $T$ denote the number of reasoning steps in the target task, $n$ the number of training samples per sub-task, and $k$ the number of output tokens per step; then the OOD task risk is controlled by the sum of the ID sub-task risks, scaled by a coefficient of order $\mathcal{O}(k)$ that is independent of $T$, although the summed risk itself may still grow with $T$. A Rademacher-complexity analysis further shows that the estimation term is of order $\mathcal{O}(\sqrt{T/n})$, rather than linear in $T$, due to predictor sharing under a uniform capacity control on the shared class. For autoregressive Next-Token Prediction, the final-task risk is bounded by the best-in-class step-level risk scaled by the same coefficient of order $\mathcal{O}(k)$, plus an estimation term of order $\widetilde{\mathcal{O}}\big(k\sqrt{T/n}\big)$. A parity-coupled construction further shows that, for the sub-task empirical risk minimizer, the expected final-task risk is at least the best in-class risk plus $Ω(\sqrt{T/n})$ when $T\le n$. Experiments on Atom-Task and GSM8K support the theory: CoT enables short-to-long and compositional generalization, while explicit state tracking is crucial. Our code is available at https://github.com/maxliang114514/CoT-OOD.

We show that a hardwired distributed processing unit (HDPU) can take a neural decision at the sensor and so remove the inference from the host. Built for zebra-crossing detection on a low-cost Artix-7 FPGA, the unit decides on every camera frame in a fixed 2,269,477 clock cycles (34.04 ms depends upon Fpga clk ), inside the 37.49-ms frame period, and on the board it returns the same scores and decisions as its bit-exact software model on all 639 test images. The whole detection path, from camera pixels to decision, takes 1,814 LUTs, with no processor, no operating system and no full-resolution frame store. When the unit answers instead of reporting "not sure", 95.7% of its answers on street photographs are correct, and a hardware contrast stage lifts accuracy on camera-like dashcam frames from 47.7% to 78.2%. The host is spared all of the inference: 8.13 ms of a core in every frame period, where a full-precision ResNet-18 would need more than a frame period even on 16 threads. It also receives 16 bytes per frame instead of 614,400, 38,400 times less. The unit is an 80x60-pixel pipeline feeding a four-layer binarized CNN with 93,408 one-bit weights, verified bit-exactly from a Python model through RTL simulation to the board. Earlier HDPUs did pre-processing (edges, corners, steering); this one shows that the decision itself can move to the sensor.

Active embodied question answering requires agents to acquire evidence through viewpoint changes and physical interaction. The central challenge is to avoid undirected interaction: taking plausible actions without specifying what evidence they should reveal for answering the question. Action plans and accumulated histories alone can leave the informational purpose of an interaction implicit. We introduce EviSeek, a training-free framework that makes question-derived evidence demands explicit acquisition objectives, linking observation-conditioned actions and selectively retained evidence to the information the agent seeks. On 3,436 ESI-Bench questions, EviSeek improves accuracy over the Single-model agent by 7.1 percentage points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro, while improving visual-support-weighted accuracy and reducing mean token usage. These findings support making the informational purpose of interaction explicit, with action selection and evidence retention organized around what the agent needs to learn to answer the question.

Reliable channel estimation and symbol detection in 5G New Radio (NR), as well as emerging 6G radio systems, rely on demodulation reference signals (DMRS). The density and placement of these pilots on the orthogonal frequency-division multiplexing (OFDM) time--frequency resource grid affect both receiver reliability and spectral efficiency. Sparse DMRS layouts increase the time--frequency distance between pilot observations and data resource elements (REs), which can make pilot information harder to exploit with purely local processing. This paper proposes \textbf{PARGAR (Pilot-Aware Graph Attention Receiver)}, a graph-attention neural receiver that represents the OFDM resource grid as a graph for bit detection. The graph combines local time--frequency edges with pilot-anchor links from data REs to selected pilot REs, and we further study a channel-scaled pilot-anchor rule based on approximate coherence time and coherence bandwidth. Using Sionna-generated OFDM simulations, we evaluate \textsc{PARGAR} against DeepRx, a convolutional neural network (CNN)-based receiver, across dense, two-symbol sparse, and single-symbol sparse DMRS layouts under channel-mismatch scenarios. Under the considered SISO uncoded setup, channel-scaled \textsc{PARGAR} explicitly injects pilot-aware inductive bias and yields lower BER in selected sparse-pilot scenarios while remaining competitive under dense pilots.

Partially interpretable models extend simple interpretable models with the option to defer complicated inputs to an accurate but uninterpretable black-box model. By deferring more inputs, the model gets more accurate; by deferring fewer, it gets more interpretable. Existing procedures to train such partially interpretable models build on linear models or decision rules. We propose three algorithms for decision trees: a partially interpretable extension of the existing Branches algorithm, which performs exact search to optimize the tree; a conversion algorithm that we run on top of standard CART, which is computationally more efficient; and a generalized CART whose splitting criterion takes into account the option to defer while growing the tree. Perhaps surprisingly, the performance of the three algorithms is similar across many realworld datasets. We can explain this through a formal bound on the potential gain of smart tree construction, provided the black box accuracy is sufficiently homogeneous across the domain. We also construct a synthetic dataset where the accuracy of the black box is deliberately inhomogeneous. There, our deferral-aware generalization of CART does outperform regular CART and matches the performance of exact search. Finally, we evaluate an application to concept bottleneck models. Partially interpretable decision trees can replace uninterpretable neural networks after the bottleneck. On undeferred inputs, the model is fully interpretable, and deferring preserves accuracy.

Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00$\times$ and 5.71$\times$ decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03$\times$ and 3.67$\times$ for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48$\times$. Code is publicly available at https://github.com/fudan-generative-vision/D3D.

Brains and large language models need not be considered equivalent systems to reveal an interpretable commonality: both exhibit high-dimensional, nonlinear dynamics whose behavior can only be modeled imperfectly. This creates a shared challenge whenever we seek to intervene in their dynamics. A controller may either rely on a model identified around a particular operating regime and risk failure as the system moves away from that regime, or explicitly account for uncertainty and a changing internal state. Neurostimulation has confronted an analogous problem for decades. The limitations of fixed, open-loop stimulation have motivated closed-loop and state-dependent approaches that condition intervention on the ongoing state of the neural system. Inspired by this principle, we formulate LLM steering as a feedback-control problem in which the network's internal state, rather than a fixed intervention alone, determines how the system is perturbed. We further account for the mismatch between tractable dynamical models and the nonlinear transformer they approximate through robust control. As one principled realization of this framework, we develop an $H_\infty$-based steering controller that limits sensitivity to disturbances and linearization mismatches. Across multiple steering benchmarks, the controller improves performance over existing baselines, with particularly pronounced gains under out-of-distribution conditions and varying length complexities. Beyond controller performance, we introduce robust steerability, $S_{\mathrm{rob}}$, a pre-deployment measure of tolerance to dynamics mismatch; across model-task pairs, it exhibits task-specific structure and varies systematically with model scale. Beyond robust LLM steering, our measure asks not merely whether control is possible, but how much uncertainty it can survive; a principle that may span artificial and biological neural systems.

Long-term problem-solving processes, such as scientific research and software development, proceed along an evolving trial-and-error trajectory toward a central goal. We term these \textit{exploration tasks}, which require models to remember not only facts but also how solutions are tested against intermediate outcomes and early attempts are continually revised or revived. However, existing long-term memory research focuses on factual recall and updates, overlooking the interconnected nature of exploration. We observe that exploration processes naturally form graphs, with branching approaches linked by revisions and reconsiderations. Motivated by this, we propose \textsc{LExMem}, a graph-guided benchmark designed to evaluate \underline{\textsc{Mem}}ory in \underline{\textsc{L}}ong-term \underline{\textsc{Ex}}ploration tasks. We model exploration processes as graphs and introduce an evolutionary mechanism to increase their structural complexity, yielding 224 dialogues and 2,018 question-answer instances derived from graph structures. Building on the same insight, we propose \textsc{LExGraph}, a graph-based memory framework that incrementally organizes dialogues into interconnected memory nodes and retrieves evidence through hybrid retrieval and neighbor expansion. Extensive experiments demonstrate that \textsc{LExMem} poses a significant challenge to existing retrieval- and memory-based systems, whereas \textsc{LExGraph} achieves superior performance and generalizes effectively to external benchmarks. Codes and datasets are available at https://github.com/mbzuai-nlp/LExMem.

The protection of mature, multi-stage Face Recognition Systems is crucial to system providers. This technical report targets the threat model most relevant to industry practitioners: although model training is performed in-house, training data is sourced from possibly-compromised third parties. Given this setting, we present three practical defense "tricks" against data poisoning-based backdoor attacks. The first operates at training-time, detecting poisoned identities during training or prior to a pipeline's deployment. The last two operate at inference-time, filtering suspicious inputs before an answer can be sent back to a potential attacker. Each method is presented with their motivation, algorithms and computational impact, and evaluated on realistic face recognition workloads. Our goal is to provide actionable hardening techniques to system providers.

We study zero-shot coordination (ZSC), where independently developed agents must coordinate at test time. While ZSC has been well studied in the RL literature, far less is known about the performance of LLM agents despite their increasing deployment in such settings. Existing work on LLM coordination often relies on specialised scaffolding that independently developed agents are unlikely to share in practice. Moreover, evaluations on complex environments (e.g., Hanabi) make it difficult to pinpoint the sources of coordination failure. In contrast, we focus on simple, general-purpose scaffolds in minimal environments designed to isolate specific coordination challenges. Our results show that even in these controlled settings, frontier LLM agents struggle to coordinate, largely due to limited understanding of the coordination problem and weak reasoning about their partner's beliefs. While LLMs introduce semantic information as an additional axis for coordination, they nonetheless fail to exploit this structure effectively. Towards this, we propose Coordination-friendly definitions (CFDs) as a principled approach for enabling robust coordination among LLM agents. Finally, we show that CFDs can be discovered automatically, removing the need for manual engineering. Code is available at https://github.com/ahayler/zsc-llm-agents.

Achieving both high pronunciation accuracy and fine-grained controllability remains a central challenge in Japanese text-to-speech (TTS), as Japanese is an unsegmented language mixing logographic Kanji and phonographic Kana with extensive polyphony. Traditional cascaded systems convert text into phoneme sequences via a front-end, ensuring pronunciation control but discarding lexical semantics and prosodic cues in surface orthography. Conversely, recent zero-shot TTS models directly consume raw orthographic text to improve naturalness, yet suffer from polyphone errors and lack explicit pronunciation control. To bridge this gap, we propose a two-stage framework for flow-matching TTS. First, we investigate Ruby text representations with orthography perturbation to jointly preserve surface forms and target pronunciations for reliable control. Second, we introduce On-Policy Distillation (OPD) on large-scale text-only corpora to enhance pronunciation accuracy on raw orthographic text without paired speech. On the Joyo-Kanji-Yomi benchmark, our model achieves 97.13% Accuracy, 2.02% Target PER, 0.70% Sentence PER, and 4.21% Sentence CER under raw text while retaining strong controllability.

QA-based text embeddings represent texts as vectors consisting of 1/0 answers to yes/no questions. Although these embeddings provide detailed information for interpretation and achieve high task performance, they often require humans to examine a large number of questions. Furthermore, the relationship between the number of questions and the appropriate level of question granularity has not been sufficiently studied. To address these challenges, we propose a method for constructing hierarchical questions for text embeddings. As the hierarchy deepens, questions become more specific and their number increases. During interpretation, we reduce the number of questions that humans need to examine by replacing questions at deeper levels with aligned questions at shallower levels. The question hierarchy is constructed through recursive clustering, contrastive prompting, and question alignment. Experimental results on general-purpose tasks show that our method improves interpretability while maintaining task performance. It is also found that, when only a limited number of questions are available, questions at shallower levels tend to achieve higher task performance than those at deeper levels, even with the same number of questions.

Non-prefix KV-cache reuse encodes retrieved chunks independently and concatenates their caches. To find where a small repair budget should be spent, we vary the chunks around the answer-bearing one (the target) across three 7B-8B models and four datasets of dialogue memory, open-domain QA and API documentation. Reuse loses 19-100% of the context-dependent questions that full prefill answers, and it loses more when the other chunks are of the same kind as the target. Because the recoverable damage is spread over these other chunks, restoring a few of them recovers little, and the budget is better spent on the target. This works through contextual compatibility: recomputing the target against the reused cache beats inserting its full-prefill KV in all seven settings tested, although the recomputed KV stays far from the full-prefill KV. Question attention over reused caches, however, selects the target in at most 18.8% of development questions. A retriever finds it more often, and recomputing its top chunk after moving it next to the question significantly beats InfoFlow-style reordering at the same 5% budget in six of nine open-domain QA and API-documentation settings. On dialogue memory, though, recomputing chunks never significantly beats the token-level selector chosen for each setting.

Offline reinforcement learning with verifiable rewards (RLVR) trains on a fixed pool of sampled responses and typically supervises every token of every response, although responses and tokens are not equally informative. We study which ones to supervise through the Gibbs-optimal policy of KL-regularized RLVR, which offline training targets, and measure the mismatch between a policy and this target with a bidirectional KL divergence. At the reference policy, this mismatch factorizes exactly and separates the two outcomes: correct responses contribute mainly through the forward term, concentrated on their low-probability tokens, and incorrect responses through the reverse term, concentrated on the responses the policy generates often and on their high-probability tokens. Based on this analysis, we propose GIST (Gibbs-Targeted Hierarchical Data Selection), a two-level selection method that approximates these principles with signals available before training. GIST uses the sampling temperature to select correct responses on-policy and incorrect responses near the policy's mode, and it uses the initial policy's token probabilities to supervise only the tokens where the mismatch lies. Applied to two offline training objectives, GIST raises the average accuracy over GSM8K, MATH-500, and MBPP for all three models under both objectives. In a controlled comparison on Qwen3-0.6B/GSM8K, it outperforms seven selection baselines that choose from the same candidates while supervising 66.6% fewer tokens than training on all candidates of the same prompts.

Finding high-performing formulations is difficult when few have been measured and experimental budgets are limited. We propose KolTa, an active learning framework that selects batches to discover the top 5% of formulations at a target temperature under a fixed query budget. KolTa combines a three-stage predictor of component contributions, role-conditioned interactions, and anchored temperature response with task-specific soft physics constraints. Its Physics-Residual-Calibrated Upper Confidence Bound (PRC-UCB) acquisition rule combines predicted utility, ensemble disagreement, and a target-temperature physics residual, with nonnegative exploration weights fitted to out-of-fold signed errors. We evaluate the framework on three measured tasks and one simulated task covering binary-mixture viscosity, drug solubility, ionic-liquid viscosity, and electrolyte conductivity. KolTa ranks first or second in mean area under the learning curve (AULC) on every task and outperforms seven of the eight comparator strategies in paired tests, winning 75 of 78 paired comparisons across tasks and seeds against those seven strategies. Across the three measured tasks, its largest gap to the best mean AULC is 0.032, compared with at least 0.056 for every comparator. On Drug, reaching 80% average recovery of the remaining targets across drugs requires an estimated median of 348 additional formulation queries, about 18% fewer than the fastest comparator. Researchers can adapt the framework to diverse formulation problems through component roles and physical constraints.

Effective collaboration in multi-agent systems requires mechanisms for coordination and adaptability in dynamic and uncertain environments. However, existing multi-agent systems rarely model resilience explicitly, instead relying on ad-hoc output correction that limits systematic, runtime management of robustness. In this work, we propose Multi-Agent Mesh (MA-Mesh), a framework that constructs a role-based self-healing layer on top of a collaboration pattern. Specifically, MA-Mesh generates resilience-aware system specifications and synthesizes them into executable multi-agent policy code. To enhance resilience, we introduce two self-healing mechanisms: Cross-agent reflection, a mechanism that allows agents to evaluate not only their own outcomes but also the performance of their peers, enabling informed behavioral adjustments and collective policy refinement; Role-based failover, a strategy which dynamically reassigns roles and enables agents to substitute for one another when sub-tasks fail or agents underperform, ensuring continuity of task execution. Experimental results across diverse tasks (e.g., software development, tool use, scientific reasoning) demonstrate that MA-Mesh significantly outperforms existing multi-agent systems in both task performance and robustness under system uncertainty.

Typed decision models offer structured choices, scores, and probabilities for low-cost automated evaluation. We present a controlled audit of presentation sensitivity in Jev, Clef Flash, and Perplexity Decider across native choice and score interfaces. Across six datasets with 9,000 multiply annotated humor, helpfulness, and summarization items, we collect 702,000 judgments under interventions that vary positions, identifiers, authorship claims, endorsements, decision framing, and rating rubrics while preserving the evaluated texts. Jev's second-position probability advantage reaches 6.11 percentage points on summaries; reversing reader endorsements changes Perplexity's selected humor text on 40.40% of pairs; and rejection framing reduces Clef's summary agreement with human preferences by 23.36 points. Rating formats also shift score levels while often preserving rankings. Across these findings, repeatability, presentation robustness, and human alignment emerge as distinct properties of an evaluator. Reliable use therefore requires validating the decisions and score interpretations on which an evaluation pipeline depends.

Dooge asked whether hydrology has laws at the catchment scale. We ask whether a low-dimensional system built under physical laws can match the accuracy of neural networks that outperform process-based models on the same inputs. Three laws constrain a lumped catchment model. Conservation makes it a flux-connected storage network, dissipation fixes the sign of every flux and aggregation determines how the unresolved composition enters the fluxes, through effective laws over threshold distributions and a memory of the forcing. These laws admit four kinds of refinement. A parsimonious model under them, CaRT (11 parameters per catchment), is refined in a loop assisted by AI agents. Each refinement addresses an observed deficit through a hypothesis from the literature. The result, CaRT+, has fourteen states, 35 parameters per catchment and no daily neural network. Most refinements follow from aggregation over thresholds or from variables that carry no water but modulate fluxes. On 602 British and 593 US gauged catchments, single CaRT+ models reach median Nash-Sutcliffe efficiencies of 0.863 and 0.765 in held-out years. A stateful LSTM reaches 0.869 and 0.751 and an entity-aware LSTM 0.862 and 0.754, while three-seed LSTM ensembles remain ahead of the CaRT+ ensemble by 0.003-0.012 in median paired NSE. On 505 drier Australian gauged catchments unused in development, single CaRT+ models reach 0.778 against 0.762 and 0.753 for the two LSTMs. The results suggest that catchment-scale laws exist as laws of form, with the constitutive families shared and only their parameters learned.

As LLM-based agents develop, they have been increasingly applied in the legal domain. Legal Judgment Prediction~(LJP) requires models to conduct multi-step reasoning and external legal evidence retrieval. Existing legal search agents' performance is often constrained by the quality of retrieval queries. They often formulate queries directly from the initial user request, overlooking latent legal issues and legally salient details. To address this issue, we propose \emph{LegalEvolver}, a query optimization framework specifically designed to generate legally informative queries for LJP-oriented retrieval. Sampling high-quality optimized queries and their corresponding search trajectories is challenging, as relying solely on unguided high-temperature rollouts is both inefficient and constrained by the model's initial capabilities. We therefore heuristically optimize queries as seeds, improving the efficiency of trajectory generation. Furthermore, we summarize rollout experience into rubrics and use them to guide multi-round iterative self-evolution. Iteratively performing rubric optimization and rubric-guided RL enables the model to surpass its initial performance ceiling. End-to-end performance on extensive benchmarks demonstrates the effectiveness of query optimization on LJP. Comparisons with direct SFT followed by RL demonstrate the effectiveness of rubric guidance. Evolving rubrics further outperform fixed rubrics, confirming the benefits of self-evolution.

Context compression reduces inference cost by encoding inputs into compact representations while preserving semantic content. An open question is whether semantic preservation alone is sufficient to maintain downstream behaviours such as refusal alignment. We investigate this question and find that encoder-based compression systematically weakens refusal behaviour in instruction-tuned language models, despite high reconstruction fidelity. This effect persists across model families and compression architectures. Mechanistic analysis shows that compression attenuates activation along the decoder's learned refusal direction. We further explore Memory Steering, a lightweight inference-time intervention that restores refusal rates to near-baseline levels without retraining and operates entirely in compressed representation space. These results demonstrate that semantic preservation does not guarantee behavioural preservation under compression, highlighting the need to explicitly preserve alignment-relevant features in compression-aware systems.

AI's inflection point is a fountain of alarming risks. Digital platforms in particular are vulnerable to three major AI risks, as those in control of global platforms cede control to AI, of the algorithms, of membership and of content: (1) AI control of global platforms, social and economic, used by the majority of humanity, resulting in the majority of humanity becoming vulnerable to AI control. (2) Unaccountable AI agents operating anonymously on social and economic digital platforms, side-by-side with humans, possibly masquerading as humans, serving unknown masters with unknown goals, operating under unknown restraints without accountability or culpability. (3) Unattributed deepfakes generated by anonymous AI, with people unable to distinguish them from genuine content and with their producers evading accountability and culpability. Here, we propose to corral these AI risks with grassroots contracts, a digital incarnation of a two-centuries old notion of Pierre-Joseph Proudhon. We present grassroots contracts for a social graph with accountable AI agents and for a social network with digital provenance, and working smartphone apps that implement the two and show how they corral these major AI risks: (1) AI control of global platforms. (2) Unaccountable AI agents. (3) Unattributed deepfakes.

Recent studies report promising results from supervised fine-tuning of pre-trained code language models for automated vulnerability detection on benchmarks such as PrimeVul. In this paper, we present a large-scale empirical evaluation of both model performance and dataset quality in this setting. We systematically fine-tune three widely used encoder models (CodeBERT, GraphCodeBERT, and UniXCoder) across five imbalance mitigation strategies, extensive hyperparameter sweeps, and multiple random seeds, resulting in 770 independent training runs. Our results reveal a consistent pattern. While standard metrics such as F1 and ROC-AUC occasionally suggest improvements, security-critical and calibration-aware measures tell a different story. Matthews Correlation Coefficient remains near zero, predicted probabilities are poorly calibrated, and evaluation under strict false-positive constraints reveals high false-negative rates. Neither imbalance handling, threshold optimization, nor post-hoc calibration recovers a meaningful discriminative signal. We also conduct a data quality assessment of the PrimeVul benchmark and identify structural issues such as minimal semantic differences between vulnerable and fixed functions, and annotation noise introduced by commit-level labeling. These findings suggests that the PrimeVul dataset, in its current form, presents significant challenges for reliable AI-based vulnerability detection in software engineering with the methodologies we used. Our study highlights the need for more robust dataset design, improved evaluation protocols, and rethinking of fine-tuning practices for security-critical tasks.

When many inputs share a budget of repeated attempts by a large language model (LLM), a common rule gives harder inputs more. On public math, code and formal-proof samples, the best published difficulty-proportional rule captures half or more of the oracle gain on benchmarks the model mostly solves, but at most 13% on hard workloads, even given true pass rates. Because the value of one more attempt peaks at intermediate difficulty, the optimal static allocation is inverted-U and gives little to inputs that are hopeless at the current budget. In a strict survey of 221 LLM allocation papers, at least 13 give the hardest inputs the most attempts; their easier settings sit at budgets where this is near-optimal, so the cost appears on harder workloads. Acting on it requires knowing which inputs are solvable, so information, not the policy, drives the gain. With a verifier and a prior from a related model's past runs on the same inputs, Thompson Sampling raises the solved share on the hardest workload from 14% to 27%; without such a prior, no policy we tested beats round-robin. For the policies we tested, the gain vanishes between prior rank correlations of about 0.3 and 0.55, which cheap text-based priors barely reach on the weakest models, and it requires a verifier whose false acceptances are rare relative to the pass rate. On new Qwen2.5 samples the pattern holds, and tempering the prior avoids losses on easy workloads. We turn these results into a decision guide.

Summarization with a fixed character length is important for tasks such as headline and article preview generation, where text must fit an exact space. Autoregressive language models (LMs) struggle with precise character-length control: training-based methods require task-specific fine-tuning, while zero-shot prompting strategies achieve only approximate compliance. We propose DELTA (Diffusion Expected Length Targeting Algorithm), a training-free inference-time method for masked diffusion LMs that estimates expected character length at each denoising step and dynamically inserts or removes masked positions to steer toward a target length. DELTA requires no fine-tuning and is designed for off-the-shelf masked diffusion LMs. Across three summarization benchmarks, it reduces character-length errors by an order of magnitude compared to baselines while maintaining competitive summary quality in both automatic and human evaluation.

Continuous fuzzing platforms such as OSS-Fuzz provide sustained security testing for open-source projects, but long-term build breakages reduce security coverage and waste infrastructure resources. In OSS-Fuzz, around 20% of integrated projects remain in build-failure states, leaving their fuzzers off duty. These failures expose a maintenance gap between platform maintainers, who lack project-specific context, and project maintainers, who often lack fuzzing and infrastructure expertise. Existing program-repair techniques cannot be directly applied because fuzz build breakages span evolving project and infrastructure artifacts and require repeated infrastructure-aware diagnosis. We present FuzzMedic, an automated agent for fuzz build repair. FuzzMedic performs evidence-guided diagnosis across project and fuzzing-infrastructure artifacts. It explores repairs over a repair-state graph (RSG) that records the evolving workspace and build feedback. Branching and rollback enable FuzzMedic to recover from unproductive repair paths during long-horizon debugging. Candidate patches are validated to restore executable fuzz targets rather than merely suppress build failures. We evaluate FuzzMedic on 128 reproducible OSS-Fuzz build failures extracted from 1,820 real-world failures. FuzzMedic achieves a 48.44% repair rate and outperforms three representative program-repair baselines. In real-world use, it produces 29 patches merged into upstream or OSS-Fuzz main branches, including repairs for three failures that had remained unresolved for more than one year and two cases that had never built successfully.

Foundation models for electroencephalography (EEG) are pretrained on tens of thousands of hours of recordings, and their accuracy grows with that amount. Such data is hard to assemble. The largest sources are clinical archives that cannot be redistributed, and the public sources are spread over several archives, file formats, and licences. We release eegdash-corpus-v0, a training-ready corpus of 77 public EEG datasets, as the gated Hugging Face dataset EEGDash/eegdash-corpus-v0 with the SHA-256 of every file. REVE, the largest published EEG foundation model, names 89 datasets as its pretraining corpus. Twelve cannot be redistributed: four clinical archives that hold 82% of its hours, and eight under no-derivatives licences. The corpus holds the 76 others that REVE's sampling rate allows and one more from REVE's release card: 77 datasets, 14,075 hours from 7,272 subjects, 5.0 million 10-second samples and 6.21 billion tokens per epoch. Every dataset is fetched through EEGDash, converted once with REVE's recipe, and stored in its own buckets of a chunked Zarr store with its licence recorded beside it, so a model can train on the whole store or on any subset of datasets. To test that the store pretrains a model, we retrain REVE-Base from scratch twice with the authors' code, changed only in how it reads data, and score it on seven public OpenEEGBench tasks. It reaches 0.331 under linear probing and 0.273 after fine-tuning, against 0.333 and 0.325 for the released model and 0.200 and 0.205 for an untrained encoder. The converter, the conversion reports, and the loader changes are released with the corpus.

Existing traffic signal control methods typically rely on structured traffic states produced by dedicated perception pipelines. Emerging visual approaches often use idealized simulator images and convert them into intermediate representations before control, limiting their ability to directly exploit roadside visual observations. To bridge both the visual-domain and perception-to-decision gaps, we introduce SynTra-V and Traffic-VL. SynTra-V converts physics-based traffic simulations into photorealistic multi-view surveillance imagery across diverse road topologies and sensing conditions, and automatic semantic auditing checks that control-critical geometry, vehicle movements, and lane semantics match the simulator state. Traffic-VL is an eight-billion-parameter vision-language model that directly maps synchronized multi-view images and natural-language phase descriptions to signal decisions without intermediate structured states. Building on Traffic-R1's two-stage reinforced reasoning framework, Traffic-VL adapts its vision encoder via offline reinforcement learning on SynTra-V and freezes it during online visual stepwise trajectory policy optimization in closed-loop simulation. Experiments show that Traffic-VL matches structured-state Traffic-R1 while outperforming image-based controllers and general-purpose vision-language models, including Gemini 3 Pro. Evaluations on unseen topologies, degraded sensing conditions, real surveillance footage, and a three-week deployment across ten intersections further demonstrate robust visual generalization and effective constrained online control.

Temporal point processes (TPPs) are the canonical framework for event sequence data, with applications spanning seismology, social media, and electronic health records. Recently, neural TPPs have been reported to substantially outperform classical TPPs, such as Hawkes processes, but these comparisons pitted flexible neural models against Hawkes processes in their most restricted form, and did not consider large-scale medical data. Here, we revisit this comparison through a systematic evaluation of Poisson and Hawkes processes of increasing flexibility against neural TPPs, on five established benchmark datasets and on two large-scale medical evaluation tasks derived from UK Biobank disease histories and MIMIC-IV ICU records, under autoregressive next-event prediction. First, the previously reported gap between Hawkes processes and neural TPPs largely disappears once Hawkes processes are equipped with flexible, learnable baselines and kernels. Second, the flexibility of the intensity parametrization also limits prior neural TPPs: a Transformer with a spline-based intensity head outperforms previous intensity-based neural TPPs and is on par with the intensity-free model. Decomposing the log-likelihood into a time term $\log p(t)$ and a type term $\log p(d \mid t)$, we find that this flexibility advantage lies almost entirely in the time density: more flexibility does not improve the type term. On these benchmarks, the gains of neural TPPs thus lie mostly in modeling when events occur, not which. Third, our results underscore the value of complex datasets for evaluating TPPs: only on the MIMIC-IV ICU records do neural TPPs retain an advantage in the type term. Our results inform both future method development and model choice. We release our framework, TorchPP, as an open-source package with consistent implementations of classical and neural TPPs.

At work, people develop skills through experience: novices acquire skills by completing tasks, while experts retain theirs through continued practice. AI assistance changes this process by offering workers a choice between solving tasks independently and accepting the AI's answers. While relying on AI saves immediate effort, it can also displace the practice needed to acquire and retain skills, including those required to oversee the AI. We extend an existing accept-or-solve model to a sequence of tasks in which workers' skills evolve through independent practice and reliance on AI. We compare myopic workers, who maximize immediate utility, with strategic workers, who also value their future skills. We show that novices can become trapped at low skill levels, while experts may lose skill, even as AI-assisted outcome accuracy remains high. Strategic workers, by contrast, may choose to solve independently even when AI is more accurate, recognizing the future value of acquiring and retaining their skills. We then study incentives that reward correct independent work and characterize the payments needed to encourage skill acquisition and retention. We show that independent work can become self-sustaining once workers acquire sufficient skill, eliminating the need for continued incentives. Counterintuitively, supporting a higher skill target can sometimes reduce long-run incentive costs, either by enabling self-sustaining independent work or by lowering the payments needed to maintain proficiency. Our findings thereby shift the focus from whether workers should rely on AI on individual tasks to how organizations can sustain human expertise over time, and which skills are worth investing in.

Unified Multimodal Models (UMMs) aim to support visual understanding and generation within a single model. Existing post-training methods either use understanding as a fixed teacher to improve generation, or rely on ground-truth trajectories to supervise both understanding and generation. Prism instead derives bidirectional supervision from self-generated interleaved rollouts, constructing training targets according to the reliability of their text and image chains. We propose Prism, a bidirectional self-teaching post-training method for UMMs. Like a prism decomposing mixed light, Prism decouples a coupled interleaved trajectory into a text chain and an image chain, assesses the reliability of each chain, and constructs a corresponding training target from their reliability states. Experiments on both SenseNova-U1 and BAGEL show that Prism improves both understanding and generation. On the generation side, GenEval improves from 85.8 to 89.2 (+3.4), and DPG improves from 86.1 to 87.3 (+1.2) on SenseNova-U1. On the understanding side, the total MME score improves from 2282.6 to 2311.6, and MMVP improves from 58.0 to 62.0. Consistent improvements are observed on BAGEL. These results highlight that Prism enables coordinated improvements in generation and understanding by constructing reliability-guided self-training targets from self-generated trajectories, without relying on an external verifier or externally annotated corrected trajectories during target construction.

The records that could train the most useful models, such as a hospital's patient files, are often the ones an institution cannot share. We propose CoRTeC, a method that releases synthetic versions of such records under differential privacy. Differential privacy places a proven limit on what the synthetic data can reveal about any person, and in particular on membership inference. CoRTeC spends the privacy budget once, on a small file of noisy counts designed for training a classifier. For each group of records, the file gives the group's size, its balance of outcomes, a histogram of every column for each outcome and the outcome rate in each of several non-overlapping cells. A reviewer can check every number before the file leaves the institution. Anyone can then turn the counts into synthetic rows, as often as they like and at no further privacy cost. A maximum-entropy decoder can do this without any language model, and so can each of the five frozen language models we tested. None of them ever sees a real record. A selection step then matches the rows to the released counts. Decoded without a language model, the automatically configured release at $\varepsilon = 2$ trains classifiers to within 0.02 AUC of a real sample of the same size on the Adult dataset, across three data splits. On NHANES and bank marketing, without the selection step, it stays within 0.05 AUC. On Adult, with a fixed cohort rule, the release carries the outcome rates a deployment names more closely than a real sample does, at some cost in accuracy on pairs of columns. Four controls show that most of the utility comes from the release and the selection step, and that a language model adds what it knows from the column names. No private row ever reaches a model, and no row CoRTeC generates copies a real one. The strongest membership attack stays far below the bound the budget allows.

Iterative feedback tuning (IFT) learns the parameters of a fixed-structure controller by stochastic gradient descent. At each iteration, two closed-loop experiments are run on the plant: the first one measures the tracking error, and the second one, driven by this error, provides the sensitivities from which the gradient of the cost is estimated. This estimate is unbiased for linear plants only. We study the biased case on automotive electronic throttle valves, whose friction and stiction vary between valves and runs, controlled by an intelligent proportional (model-free) controller, which is an incremental PI controller. We show that the mismatch of the controller state at the start of the two experiments, which stiction makes unavoidable, adds an error inversely proportional to the amplitude of the second experiment. On digital twins of ten valves, the IFT gradient is a descent direction in only 69\% of the low-cost region, friction being the cause, and all update rules plateau and then drift. On two hardware valves, the best update rule depends on the noise structure of the gradient. On the first valve, where one gradient component is buried in noise, Adam outperforms Gauss--Newton, its normalisation and its momentum each improving one parameter. On the second valve, where both components are consistent, the same normalisation amplifies a direction that Gauss--Newton rightly discounts, and Adam drifts to a controller worse than the initial one. The twins reproduce the mechanism but not this ranking.

Project-level code generation aims to automatically synthesize complete software repositories from high-level requirements, producing coherent multi-file systems with well-structured modules and consistent cross-file dependencies. Despite the strong capabilities of large language models (LLMs) in specification understanding and code synthesis, existing approaches still exhibit several critical limitations. First, many methods rely on static architectural designs that cannot adapt during iterative generation, leading to a growing misalignment between architecture and implementation. Second, they lack mechanisms to accumulate and reuse experience across iterations, resulting in recurring errors and reduced efficiency. Third, they often depend on standalone LLM-based evaluation of intermediate artifacts, which suffer from limited transparency and potential unreliability under long-context reasoning. To address these challenges, we propose EvoCodeAgent, a state-centric framework for project-level code generation. EvoCodeAgent maintains an evolving architectural state initialized from structured project specifications and continuously reconciles predicted relations from architecture and skeletons with realized relations extracted from generated code. It further uses tool-grounded verification evidence to support reliable state updates and abstracts reusable repair experiences from previous revision attempts to guide subsequent generation. Extensive experiments on project-level benchmarks of different scales, including DevBench and CodeProjectEval, demonstrate that EvoCodeAgent consistently outperforms baseline approaches. Under the GPT-5 setting, it achieves unit test pass rates of 60.48% and 27.73%, respectively, yielding up to a 53.5% improvement over the strongest baseline.

Training data can have unexpected side effects on language model behavior outside the training domain: insecure code data can induce broad misalignment, and outdated bird names can induce a 19th-century persona. Understanding such generalization is hard because only dozens of isolated examples exist, each constructed by hand or discovered by accident. We propose automatically discovering such datasets with surprising side effects (DSSEs). We define three objectives for DSSEs: they should (1) elicit a target behavior, (2) read like ordinary post-training data, and (3) be surprising to human auditors reading the data. We operationalize each with LLM judges, and search for high-scoring datasets via evolution, reinforcement learning, and subliminal learning approaches. After re-scoring and filtering with held-out metrics, our pipeline produces a collection of 6,575 DSSEs across 8 behaviors, many of which generalize across model families and sizes. We find our automated metrics track ML researchers' judgments, and frontier models outperform humans at predicting side effects. These datasets form a testbed for studying generalization and lay the groundwork for predicting training side effects before model deployment. Explore them at https://dsse.transluce.org.

Improvement in enzyme engineering is usually pursued by optimising which candidates to assay within a fixed model and workflow; when the bottleneck arises from a mismatched method or a failing workflow, that route is difficult to escape. We present EnzymeRSI, a dual-loop system for recursive self-improvement (RSI). The outer loop evolves harness assets without changing model weights, the inner loop develops prediction and selection algorithms under controlled conditions, and a cloud laboratory system is specified to close the physical loop with quality-controlled feedback. By updating its prediction strategy through iterative comparison, the outer loop raised the mean development-set Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); of the two first places, only CD19 was produced by the outer loop itself. In the self-improvement of the ThermoMPNN algorithm the inner loop reached a PCC of 0.849, and on TadA-Bench it reached a Spearman correlation of 0.307 in 1.13 h. In three prospective wet-lab tests of the system's experimental decision-making, 10 of the 32 candidates that had not been pre-specified for the luciferase RLuc showed at least twice the wild-type activity; a recommended variant of the cytidine deaminase A3A raised 5mC conversion to 90.2%; and path selection in evolution raised the catalytic efficiency of the DNA polymerase KOD-E on a modified substrate to 31.1 times that of the parent. By treating the organisation of the research workflow and the development of algorithms as a single falsifiable dry-to-wet loop, this work offers an evolutionary paradigm for moving beyond the blind spots of a fixed method and for autonomous enzyme engineering.

Conformal Prediction (CP) provides a model-agnostic framework for constructing finite-sample prediction sets with coverage guarantees and has become a tool for reliable decision-making in high-risk settings. Recent advances have extended Split Conformal Prediction to increasingly complex learning problems, including multilabel, high-dimensional, and functional outputs. In this work, we study CP for Distribution-to-Distribution Regression, a setting of growing interest across many application domains. We show that a naive application of Split CP yields prediction sets that, while marginally valid, are non-interpretable, e.g., lacking a well-defined volume or a tractable sampling procedure, and of limited practical value. To overcome this, we propose a new conformal framework that operates in an orthogonal basis space, producing interpretable prediction sets while preserving coverage guarantees. We further develop an adaptive variant with asymptotic conditional coverage under mild assumptions. Empirical results on synthetic and real-world datasets validate the effectiveness of the proposed approach. The code is publicly available at https://github.com/tuanhacs/conformal_d2d.

Risk-budgeted throughput forecasting converts a point forecast of LEO satellite capacity into a conservative bound whose overestimation rate is held at a user-specified budget, so that admission control can act on it safely. We reproduce the state of the art, BG-CFQS, on the three Starlink traces its authors used and on an independent fourth dataset, and report four findings. First, its risk guarantee is reproducible but conditional on the calibration split: under a random split we recover every published quantity, including risk-pass on 3 of 3 datasets, but under a contiguous temporal split, which is what a deployed terminal faces, it holds on 1 of 3. Second, independent of any split, the method cannot serve a risk budget below 0.15, the low end of its candidate quantile set. On all four datasets the achieved rate is identical at budgets of 0.05, 0.10 and 0.15, overshooting by 1.4x to 3.7x at the tightest. Third, static conformal calibration cannot improve a ranking, only shift an operating point; we confirm this exactly on latency spike detection, where online calibration improves AUPRC by up to 51%. Fourth, group-conditional online recalibration splits into two mechanisms. The online part replicates: it tracks the budget within 1% from 0.05 to 0.35, where static calibration is 16 to 21% off. The group-conditional part does not: on three of four datasets, conditioning on any available covariate, including measured serving-satellite geometry, is worse than not conditioning, and our attempt to predict which datasets benefit failed. What is new here is empirical, not algorithmic: we claim the findings, not the methods.

The halting problem asks whether a program eventually halts, with unrestricted resources. Engineering systems instead certify bounded halting: given a program, an input, and a step budget T, decide whether the program halts within T steps. This paper studies two of its uses. First, it defines an exact finite-budget problem, placing the simulated program and the checker on the same unit-cost timescale, so the checker can be embedded in an adversarial program. A diagonal argument with explicit additive overhead shows that no fixed deterministic checker decides every instance within any fixed fraction of T below one, for large T. In the nondeterministic case, T bounds the branch depth allowed for a stable observable decision. Second, bounded-halting instances give adversarial tests for finitary universality, since a universal model must support arbitrary finite computations as T grows. The tests check resource growth with T, the absence of a fixed deliberation bound, and later access to intermediate data. They separate a fixed model with extendable external resources, a model family that grows with T, and a larger system built around a fixed component. For machine-learning systems, this identifies the resource growth a universality claim assumes and whether it fails for deployed models.

Retrospective battery characterization across heterogeneous cells is challenging because cycling records vary substantially with chemistry, protocol, cell format, capacity, and operating conditions. An endpoint charge--discharge curve captures the local condition of an evaluated cycle, while a cell-level lifecycle terrain summarizes structural evolution across the archived trajectory. We propose STEK, a structural-terrain-enhanced framework for retrospective battery multi-target estimation. STEK combines a cell-level lifecycle (dQ/dV) terrain with an endpoint-derived structural proxy through quality-proxy-conditioned gating. Bounded FiLM and a validity-gated residual head refine the endpoint-curve representation while preserving a curve-based prediction pathway. Battery-type and optional material-text representations further condition an identity-centered low-rank linear map, providing a Koopman-inspired cross-channel consistency objective without modeling temporal dynamics or performing rollout. Under the fixed seven-output MIX-battery protocol, the full retrospective setting improves aggregate performance over endpoint-only learned baselines across multiple temporal backbones; this comparison reflects both the additional archived context and the modeling design. Under matched auxiliary inputs, STEK outperforms direct concatenation and quality-unaware FiLM, while descriptor-removal tests show graceful fallback to the endpoint pathway. Deterministic rules remain preferable for directly recoverable capacity summaries. The method targets retrospective archive harmonization rather than online estimation from cycle t alone.

Ambient sensors offer a privacy-preserving solution for activity recognition in smart homes, yet they face a critical challenge in multi-resident environments known as Data Association, which aims to identify the resident triggering a specific anonymous sensor event. Existing approaches fundamentally fail to maintain global consistency by often yielding conflicting associations for silent residents. Additionally, these methods frequently suffer from severe trajectory fragmentation caused by the ontological mismatch of applying continuous physical motion models to discrete logical behaviors. To address these limitations, we propose a novel framework that reformulates data association as a sequential decision-making task within a Markov Decision Process (MDP). We solve the defined MDP via Behavior Cloning (BC) to avoid complex reward engineering by directly learning from expert demonstrations. We instantiate the policy using two complementary architectures, including a lightweight Transformer that treats association as a multi-label classification task over linearized event histories, and a fine-tuned Large Language Model (LLM) that reformulates the problem as conditional text generation to leverage pre-trained semantic knowledge. Extensive evaluations on three diverse datasets demonstrate that our approach significantly outperforms state-of-the-art baselines, particularly in resolving concurrent multi-resident events. Furthermore, the LLM-based policy shows preliminary cross-environment transfer potential under benchmark settings, while its robustness in real deployments still requires further validation.

Large language model (LLM) coding assistants and agents construct a repository context before generating code. For security-sensitive code, this context decides whether the model sees the project's safeguards (e.g., authorization checks) or insecure code to imitate. However, context construction is designed almost exclusively for functional correctness, and secure-code benchmarks treat the context as fixed, so how repository context affects the security of generated code remains unclear. We construct CASE-Bench, a benchmark of 145 repository-level tasks from real CVEs in 51 projects. Each task runs in Docker, has functional tests and a proof-of-concept exploit validated in both directions, and is annotated with the repository code that enforces or undermines its security. Using CASE-Bench, we compare 13 context construction methods and six controls with six LLMs, edit single context elements, and analyze coding-agent trajectories. We find that (1) functionality-oriented context raises the functional pass rate by up to 8.2 percentage points (pp) but the secure pass rate by at most 3.7 pp, (2) withholding a safeguard lowers the secure pass rate by up to 8.1 pp, while one inserted insecure exemplar raises correct-but-exploitable code by 6.3 pp, and (3) coding agents observe the relevant safeguard in only 27-36% of trajectories. These findings inform security-aware context construction and secure-code evaluation.

Encrypted traffic classifiers are routinely reported at 95-99% accuracy. We show that on one of two public corpora, the same model and code score macro-F1 0.992 under the evaluation protocol standard in the literature and 0.567 under a protocol that keeps flows from one packet capture on one side of the split. Nothing changes but the partition (gap 0.426 over eight seeds, p = 0.008 after Holm-Bonferroni correction). The inflation is not a property of random splitting as such. On a month of backbone QUIC traffic, we evaluate five split protocols at eight seeds each: random, temporal and client-disjoint splitting are statistically indistinguishable from session-disjoint, while server-disjoint splitting costs 0.192 macro-F1 (d = 17.0, p = 0.008). Which axis matters depends on how the corpus was collected, in a way invisible from its label space. We give a diagnostic that identifies the right axis before any model is trained: single-identifier probes. A decision tree on the capture identifier alone reaches 0.983 under a random split and exactly 0.000 under session-disjoint splitting. Under this evaluation, a metric-trained flow embedding leads four deep baselines but does not beat gradient-boosted trees on twenty packet sizes. Five of our own pre-registered architectural claims fail: distance-based rejection of unknown applications beats softmax thresholding, but ablating every metric objective makes rejection better, because class separation correlates at r = -0.88 with rejection quality. Few-shot enrollment is flat within a corpus but, across corpora, five labelled flows per class recover 91% of the deficit left by failed zero-shot transfer. All splits, results, code and negative findings are released.

Monocular 3D object detection offers a low-cost route to 3D perception, but it remains fundamentally ill-posed, since depth must be inferred from a single view. Recent transformer-based detectors learn geometry purely from image features and 3D box annotations, lacking the high-level semantic priors that vision-language models (VLMs) acquire from large-scale pretraining. VLMs offer such priors, but deploying one directly as a detector is impractical due to its heavy inference cost. We propose MonoLaDN, a language-guided denoising framework that transfers the geometric reasoning of a VLM into a monocular 3D detector while keeping the deployed model purely vision-based. During training, a frozen VLM produces structured reasoning about each object's occlusion, depth, dimension, and angle from the image and its ground-truth boxes. We encode this reasoning into queries, perturb them with noise, and train the detector to reconstruct the clean reasoning. Reconstructing the corrupted reasoning forces the detector's depth-guided features to internalize the VLM's geometric priors, so this knowledge remains available even after the reasoning path is discarded at inference, adding no latency or parameters to the deployed model. Experiments on KITTI 3D and nuScenes show that MonoLaDN consistently outperforms state-of-the-art detectors and generalizes well to new domains.

Physics sensing in evolving environments requires reconstructing dense spatiotemporal fields from sparse observations while adapting where future observations should be collected. Existing reconstruction and sensor-placement methods have made substantial progress, but they typically rely on static sensing assumptions: they either use a single layout across all frames or optimize placements independently at each frame, which limits their ability to produce temporally coherent sensing trajectories. We introduce DyPSI, a dynamic physics sensing framework that formulates this problem as the joint generation of future fields and sensor trajectories conditioned on historical observations. DyPSI lifts discrete sensor coordinates into continuous Sensor Position Fields, encodes fields and sensing configurations with a shared Functional Tucker representation, and trains a joint diffusion model to generate future fields and sensing trajectories simultaneously in the resulting latent space. A Gaussian-process temporal kernel correlates the perturbation injected during diffusion training, biasing the denoiser toward temporally smooth recoveries, and we further provide an analysis linking the EDM training objective to a sequence-aggregated A-optimality criterion and a kernel-induced bound on discrete trajectory variation. Experiments on turbulent flow, GLORYS12 sea-surface temperature, and 3D car aerodynamics show that DyPSI consistently outperforms static and frame-wise placement strategies, with substantial error reduction under tight sensor budgets. The code is available at https://github.com/mengqi132/DyPSI.

While recent feed-forward 3D Gaussian Splatting methods enable efficient driving-scene reconstruction from surround-view images, their reliance on a single observation fundamentally hinders novel-view synthesis. Regions occluded in the input can become visible from novel viewpoints, leaving the reconstructed Gaussians under-constrained and incomplete. Moving beyond image-space inpainting for rendered views, we introduce OccSplat, a feed-forward framework whose core contribution is Occlusion-Aware Gaussian Densification, explicitly targeting missing scene geometry rather than missing pixels. By predicting new Gaussian primitives for unsupported regions, OccSplat completes the scene directly in the 3D representation. Specifically, OccSplat first identifies occluded regions that lack Gaussian support through Gaussian Occlusion Mapping. Gaussian Occlusion Completion then generates new primitives for these regions. The predicted primitives are integrated into the original representation to improve novel-view synthesis while retaining feed-forward inference without per-scene optimization. Experiments on the nuScenes benchmark demonstrate that OccSplat improves on the compared methods and yields particularly large gains in occluded regions, leading to more accurate and visually coherent driving-scene reconstruction.

Iterative neural channel estimators can improve recovery under pilot contamination, but repeated network evaluations increase inference cost. This paper develops observation-anchored consistency and Consistency Mid-Training (CMT) estimators for one-step channel estimation in multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems. Both estimators refine pilot-derived least-squares channel estimates in the delay-beam domain. Consistency learning couples endpoint predictions along noisy-to-clean paths. CMT then refines a compact consistency-trained student to predict a flow-matching teacher's endpoint from intermediate trajectory states. Across the evaluated SNR range, the compact CMT student reduces uncoded symbol error rate (SER) by 9.6-20.2% relative to the one-step flow-matching baseline (0.44-0.98 dB in logarithmic SER reduction), while using 74.5% fewer parameters and 73.0% fewer network multiply-accumulates.

Adversarial training enables one- and few-step image and video generation, but conventional adversarial losses supervise critic values without directly specifying the gradients used to update the generator. We present NEXTGAN, which trains a scalar critic by regressing its input gradient onto displacements between independently paired reference and generated features at interpolated locations in frozen representation spaces. For a fixed generator, this regression is equivalent to an adversarial objective with a gradient-energy penalty, while the generator retains the standard potential-based GAN update. The same objective adapts pre-trained image and video generators with up to 14B parameters, without auxiliary losses, online teacher scores, or trajectories. On ImageNet $256 \times 256$, NEXTGAN achieves one-step FIDs of 2.87 with pixel-space JiT-B and 1.54 with latent-space DiT-XL/1; its adversarial form further reaches 1.36. For text-to-video generation, a single trained Wan2.1-1.3B model supports one-, two-, and four-step sampling, achieving the highest VBench total scores among the compared 1.3B methods at one and two steps. At two steps, NEXTGAN scores 83.13, exceeding rCM and AnyFlow by 1.25 and 0.47 points, respectively. Code is available at https://github.com/LINs-lab/NextGAN.git.

Several widely used families of Graph Neural Networks carry a positive-definite geometric prior. Graph convolution read as heat diffusion needs a positive semi-definite operator, hyperbolic models place features on positive-definite Riemannian manifolds, and dot-product and GAT-style attention scores use positive-definite inner products. On heterophilic graphs, where edges often signify dissimilarity or structural repulsion, that prior is a mismatch: no positive-definite form can assign a negative squared distance to a direction the labels separate. To measure the mismatch, we introduce the Graph Signature Index ($s_{\mathcal{G}}$), the inertia of a class-compatibility matrix built from label topology alone, and we prove that no value of the edge or node homophily index in $(0,1)$ rules out a negative eigenvalue. We then propose the Pseudo-Riemannian Attention Network (PRAT), whose attention logit is a signed squared pseudo-distance under a learnable diagonal metric initialized from $s_{\mathcal{G}}$. The design is diagnose-then-learn: the signature supplies a sign structure that unconstrained training keeps outside near-zero directions, and the projection learns the rest of the metric around it. Across fifteen node-classification benchmarks PRAT outperforms or matches the strongest heterophily-adapted models. A controlled ablation ties that performance to the sign structure itself: with architecture and protocol fixed, forbidding the metric to be indefinite costs accuracy on the graphs $s_{\mathcal{G}}$ marks as indefinite and leaves the others unchanged.

Large language model (LLM) agents are increasingly being used to automate machine learning (ML) research workflows, ranging from code development and experiment execution to research reproduction and scientific discovery. Recent systems demonstrate that LLMs can perform substantial portions of ML research engineering with limited human intervention. These advances suggest a shift from automating individual research tasks toward automating long-horizon workflows that transform research artifacts into executable ML systems. However, ML research engineering remains substantially more difficult than isolated code generation. A research artifact typically specifies a system only implicitly through a mixture of natural-language descriptions, equations, pseudo-code, implementation details, datasets, and evaluation protocols. An agent must first infer the intended computational structure, translate it into executable components, coordinate their dependencies, and repeatedly diagnose and repair failures during execution. Consequently, the quality of the final system depends not only on code generation, but also on whether the agent can maintain a consistent representation of what should be implemented throughout this long-horizon process.

People often make decisions by holistically reasoning over heterogeneous forms of personal information. In this paper, we study the relatively underexplored capability of multimodal personalized reasoning in multimodal large language models (MLLMs). To this end, we introduce \texttt{MPRBench}, a novel and comprehensive benchmark specifically designed to evaluate personalized reasoning across a wide range of tasks and real-world challenges. \texttt{MPRBench} consists of 12 sub-tasks and additionally supports interpretable analysis of representative error cases. Through extensive experiments, we find that current personalized MLLMs still struggle with personalized reasoning due to several key challenges, including inaccurate recognition of personalized concepts and sensitivity to irrelevant personal information. We hope that MPRBench will stimulate further research on personalized reasoning in MLLMs. Code and data are available at https://github.com/kkdh15/MPRBench

Open-weight language models that fit on a single consumer GPU are becoming capable on a growing range of tasks, and they matter because they can run locally without sending the code to a hosted service. Vulnerability reproduction, turning a vulnerability description into an input that crashes the vulnerable build, is a natural target for them, since it is an automated analysis task closer to test generation than to code comprehension and it runs against a build that must be executed rather than read. In order to assess whether current models can perform it, we evaluate twelve configurations of local models, spanning three generations, two architectures and three quantization levels, on ten CyberGym vulnerability-reproduction tasks, three runs each under one harness, and release the full artefact. We show that every configuration, down to the 4B model, solves most of the six easier tasks, and configurations separate only on the remaining four. For the harder tasks, the 27B model is required. Quantization did not change what the models solved in any matched pair, but it changed how fast they got there, and the mixture of experts reached the same scores in less wall clock. We also examine the agent trajectories. Almost every run we examined, including those that fail, names the vulnerable function and states the correct hypothesis, and what separates a solved task from an unsolved one is whether that hypothesis becomes a working input.

Reward alignment for diffusion- and flow-based generative models is often formulated over stochastic trajectories, making it less suitable for few-step deterministic generators. Initial noise optimization provides a trajectory-free alternative, but existing gradient-free methods require costly per-sample search at inference. We introduce LeNO (Learning Noise Optimization), which amortizes reward-guided initial noise optimization into a learned residual transport while keeping the pretrained generator frozen. LeNO learns an anchor-conditioned distribution over noise corrections using Local GRPO and scalar rewards, eliminating iterative test-time optimization. For Gaussian initial noise with norm-preserving projection, we derive a tractable upper bound on the KL divergence between the corrected noise distribution and the original prior, providing a principled regularization objective. We evaluate LeNO on few-step text-to-image and protein structure generation, including expensive black-box designability rewards. Across both domains, LeNO improves target rewards with a single noise correction at inference while largely preserving generation quality and diversity.

The performance of large language model (LLM) agents depends on both the underlying model and the harness that governs interaction with tools and the environment. Yet harness design remains largely manual, and automating it requires navigating an open-ended search space through costly end-to-end evaluations. We introduce HarnessFlow, a framework for optimizing complete executable harnesses under a limited candidate-evaluation budget, with the model and task environment held fixed. HarnessFlow organizes candidates and their evaluation histories in a refinement tree, using a Monte Carlo Tree Search (MCTS)-based policy to balance exploitation and exploration. A coding agent refines selected harnesses using archived code, execution feedback, and a library of reusable harness mechanisms. With 15 refinement attempts per benchmark, HarnessFlow achieves the highest mean test scores among the compared methods on six benchmarks spanning agentic tasks, coding, and web interaction. The optimized harnesses achieve an average score of 69.48%, outperforming the initial harnesses and Meta-Harness by 10.89 and 3.95 percentage points, respectively, while incurring lower average per-task execution costs than hand-engineered baselines. These results suggest that structured harness search offers a promising route toward automating the design of capable and efficient LLM agents.

Reinforcement learning with verifiable rewards (RLVR) has led to significant gains in math and code generation. Extending these gains to open-ended domains remains challenging: ground-truth verification is unavailable, human annotation is expensive, and learned reward models are prone to reward hacking. The generation-verification gap, where judging good outputs is easier than producing them, offers a potential path to improvement. However, rubric-based judges are often imprecise and stationary, so policies can overfit or reward hack them. We introduce Rubric Curriculum RL (RcRL), which exploits this gap without new data, human annotations, or stronger teacher models. We combine pairwise preferences over rollouts with a curriculum over rubric criteria, yielding a more discriminative signal than absolute scoring and a non-stationary objective. Whereas baselines plateau within a few hundred steps, RcRL preserves output entropy and keeps improving past 1000 steps. On creative writing, RcRL achieves a 70.5% human win rate against the base model, with gains across benchmarks and judges. On HealthBench, RcRL improves over the base model by 14.6% (full set) and 25.4% (consensus subset), outperforming all training baselines, including instance-specific rubrics.

Decision models read a document and return a probability for each of a fixed set of options instead of generating text. Several appeared in 2026, but none has published results outside English. We describe Zuse, three decision models (4B, 9B and 27B) that train a low-rank adapter and a pointer head on frozen open-weight bases, targeting 27 European languages, and use them to study how this design behaves across languages and tasks. On a broad European suite of 11 sources excluded from training, covering 9 task groups and 27 languages, a development set that also guided model selection, the 27B model reaches 87.5% accuracy with a pooled expected calibration error of 0.016, and 0.035 when weighted per source. Post-hoc analyses of these single-seed results give three findings. First, miscalibration is structured by task rather than language: on parallel reading and topic items in 26 languages the confidence gap varies about as much as sampling noise, while across sources it spans -0.07 to +0.09, with knowledge tasks overconfident and acceptability judgements underconfident at all three sizes, so one fitted temperature cannot correct both. Second, the share of decisions that can be automated at a 5% error budget follows chance-corrected skill (Spearman 0.97 over 33 source and model pairs) more closely than raw accuracy (0.91). Third, reach beyond the training languages is set by the base model: across 210 languages, changing the training run moved scores by at most about 2 points within a size, while languages outside the base's documented list trail by 23 to 28 points at every size. On English public suites Zuse is behind Jev and Kev, including Kev models on the same bases. We also report two pitfalls in building multi-source training data: label-word collision between topic schemes, and an injection augmentation that taught a shortcut. Zuse can be tried at https://bineric.com/zuse.

Simulation-based fuzzing exposes autonomous driving system (ADS) failures by repeatedly mutating and executing scenarios, but converting one failure into reusable regression scenarios remains difficult. Such variants test whether the mechanism persists after software changes, for example by varying weather or actor placement while retaining a collision's actor-route conflict. We present Bug2Scenario, a root-cause-guided post-fuzzing framework. From scenario data, traces, and oracle evidence, it extracts a root-cause pattern. An LLM proposes semantic variations over actors, missions, routes, and weather. Deterministic rules construct DriveFuzz-compatible scenarios. Execution evidence and manual review assess preservation. Separating these responsibilities prevents fluency, syntax, or recurring labels from masquerading as causal evidence. Under matched seeds, six LLM backbones, candidate budget, and execution pipeline, Bug2Scenario improves strict preservation by 6.7 percentage points (p=0.104; not significant) and boundary-inclusive preservation by 10.6 points (p=0.011). Feedback helps some initially unsuccessful seeds yield a strictly preserved candidate. Seed-level stopping then avoids further API calls and simulator runs. Remaining failures reflect actor-placement constraints, simulator instability, and incomplete or conflicting causal evidence. Bug2Scenario thus produces reliable ADS regression assets. Our implementation is available at https://github.com/YgcCoder/Bug2Sce.

Can independent components be recovered when all of them are Gaussian? Yes, if their variances change across environments; but how many environments, and how many samples as the number of components $d$ grows? We answer both for the joint diagonalization of $K$ covariances. Each component has a signature, its variances across the environments. Separating component $j$ costs the Riesz 2-energy $E_j$ of this signature, a sum, over the other components, of inverse pairwise Fisher informations. For Gaussian sources no estimator in $O(d)$ beats $cE_j/(n+E_j)$ at any $n$ and $d$, and above an explicit sample size a spectral estimator attains $E_j/n$ up to a constant, for every mixing and without the signatures (given a lower bound on the pairwise information). Since signatures live in $K-1$ dimensions, the rate obeys a dimension law: $d^2/n$, $d\log d/n$ and $d/n$ for $K=2$, $3$ and $K\ge4$ well-spread environments. What counts is the dimension the signatures span, not their number, and $\log d$ well-designed environments make $d\log d$ samples suffice. Projecting the covariances onto one linear combination pays $d^2/n$ at first order instead. In simulations up to $d=1000$ the predictions hold and joint diagonalization with $K\ge4$ attains the rate once $n/d$ is in the hundreds; on task fMRI the energy says which components are hard to separate.

Online algorithms with bounded recourse face three intertwined obstacles: uncertainty about the future, computational limits, and a limited budget for changes. We ask a fundamental question that separates consistency from the other two: what does solution stability cost when the entire future is known and computation is unlimited? We answer this question for monotone submodular maximization, where an algorithm must maintain a solution of at most $k$ arrived elements, with a constant number of changes per step. We determine the optimal approximation thresholds: $3/4$ for deterministic algorithms and $ξ^\star \approx 0.7942$ for randomized algorithms. For every $\varepsilon>0$, we get within $\varepsilon$ of these thresholds with worst-case recourse polynomial in $1/\varepsilon$, independent of $k$ and the stream length. We prove matching impossibility results even for coverage functions. We attain the sharp bounds by optimally balancing two opposing forces: preparing for future arrivals and protecting the current solution. For randomized algorithms, we use addition robustness against prefixes of the known future, which we analyze tightly via an interval-selection problem. For deterministic algorithms, we introduce certificates, paths in a fractional relaxation that maintain high value, which we round into an update schedule.

A central challenge in AI is specifying objectives that align with human preferences. While existing theoretical results elucidate how misspecification impacts performance, they typically do not distinguish between merely suboptimal and catastrophic outcomes. To enable rigorous analysis of catastrophic risk from AI misalignment, we introduce a formal definition of catastrophic performance. Our definition is based on two reference values. The first, baseline value, captures the value attained if no agent is deployed. The second, orthogonal value, captures the value of a trajectory that is unrelated to human preferences, in a sense we describe. Orthogonal value is typically far worse than baseline value. We interpret these reference values through examples in which past human activity has altered the state of the environment to human benefit. Starting at the resulting state, an AI agent is deployed with the goal of accelerating further progress. Catastrophic performance represents this agent's undoing of benefits afforded by human activity.

Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.

With the rapid development of autonomous driving systems (ADSes), real-world roads are evolving into mixed-traffic environments where autonomous vehicles (AVs) and other traffic participants (e.g., human-driven vehicles (HDVs), cyclists, and pedestrians) coexist and interact with each other. However, current ADSes still incur various safety violations in mixed traffic scenarios, raising a strong demand for runtime safety assurance frameworks that can continuously monitor hazards and prevent potential violations. Existing ego-centric assurance approaches are insufficient for mixed traffic, where violations often arise from locally reasonable yet jointly incompatible decisions among multiple vehicles and cannot be reliably resolved through independent takeover. To bridge this gap, we propose CoSafe, a modular and ADS-agnostic cooperative runtime safety assurance framework for mixed-traffic autonomous driving to extend runtime safety assurance from ego-centric takeover to scene-level cooperative intervention. CoSafe leverages vehicle-to-everything (V2X) communication to enable roadside units (RSUs) to monitor scene-level hazards and jointly mitigate hazards involving multiple vehicles. We integrate CoSafe with seven state-of-the-art ADSes, and our evaluation on the InterDrive benchmark demonstrates that CoSafe improves the driving score of AVs by 70.03% on average, reduces safety violations by 87.80%, and accurately identifies unsafe AVs with an F3 score of 0.930, while satisfying real-time constraints.

Reliability management of industrial fleets depends on early fault warnings, yet lifecycle data across similar assets are highly imbalanced. A few assets accumulate run-to-failure histories, while newly installed or rarely failing assets provide only normal-condition data. This paper asks how diagnostic knowledge can be transferred from a data-rich source asset to a data-scarce target asset when no fault data have been observed on the target. We formulate cross-asset early fault detection as a zero-shot domain adaptation problem, in which zero-shot refers to the absence of target fault samples, and propose a framework that requires only normal data from the target. The framework comprises (1) a statistical process control step that labels the source asset run-to-failure data with an exponentially weighted moving average chart, reducing manual annotation to verifying flagged segments, (2) an Adapted Domain-Adversarial Neural Network (ADANN) that combines domain-adversarial training and a Maximum Mean Discrepancy penalty with an Adapted Classification-Domain (ACD) loss, which anchors the classifier normal region to the target own normal data, and (3) an ahead-of-time labeling scheme with a prediction horizon that targets incipient faults. Under an exact-classification assumption, a theoretical analysis characterizes the domain-adversarial term for a normal-only target and the alignment state in which source fault features remain separated from target normal features. On a public run-to-failure bearing dataset, ADANN attains the highest mean F1-score among four transfer baselines, with a nominally significant precision gain over the strongest baseline, and its robustness is examined under synthetic noise.

Dimensionality reduction (DR) techniques are fundamental to the analysis of high-dimensional data, yet their substantial computational cost limits their use in interactive settings. To address this limitation, we present GPU-accelerated Progressive Uniform Manifold Approximation and Projection (GPUMAP), an efficient progressive DR algorithm designed for interactive use. We identify three atomic operations within UMAP and restructure them into a progressive pipeline. As part of this adaptation, we introduce modifications to UMAP, including operation queues, instance-wise local join, and scheduling strategies. We further accelerate computation by leveraging GPU parallelism, enabling interactive analysis of million-scale datasets on commodity GPU hardware. Finally, we employ our implementation to systematically study the behavior of progressive DR, investigating how projections evolve over time relative to the ground-truth projection and how biased data streams influence the quality and stability of the resulting projections.

Knowledge editing has become a key technique for updating knowledge in large language models. However, existing methods neglect the geometric structure of the knowledge representation space, leading to two key limitations: they cannot distinguish which pre-trained knowledge should be preferentially preserved (core vs. non-core) nor which new edits are inherently difficult (hard vs. easy). To address this, we propose GEdit, a geometry-aware framework built on a Low-Rank Plus Diagonal (LR+D) factor model that captures the intrinsic low-rank structure of Transformer FFN hidden representations. GEdit features two complementary mechanisms. First, the low-rank regularization applies targeted protection on a low-rank core knowledge subspace. Second, the adaptive regularization dynamically adjusts the regularization strength based on each edit request's editability: relaxing constraints for hard edits to ensure success, and tightening them for easy edits to better preserve pretrained knowledge. Extensive experiments on GPT2-XL, GPT-J, and LLaMA-3-8B under large-scale editing scenarios, including both mini-batch and fully sequential editing settings with up to 10,000 continuous updates, show that GEdit consistently and significantly outperforms state-of-the-art baselines in both editing success and knowledge preservation.

AI clusters keep data in remote object stores and rely on local caches to stop GPUs from waiting for data. A cache performs well only when its prefetching and eviction policies match the access pattern, yet these patterns differ across AI workloads and change between phases of one workflow. Existing methods either need code changes, which compromises user transparency, or use low-end heuristics, which cannot follow such changes. Large language models (LLMs) can infer access behavior without user hints, but direct LLM control of cache actions suffers from long reasoning delay, high token cost, and poor decision quality. We present Chiron, a cache system based on advisory switching control. An LLM asynchronously diagnoses the access pattern and recommends a policy. A deterministic harness gates this advice, executes expert-designed policies locally, and returns cache feedback for review. We implement Chiron atop JuiceFS and evaluate it on nine AI workloads. It reduces job completion time by 14.3-41.2% over the best baseline on popular AI workloads.

Deploying an explainable intrusion detection model from a validated laboratory setting into a forensic workflow requires guarantees that benchmarks do not test: that the certified model is the exact artifact applied to a case, that heterogeneous evidence can be classified reliably, and that every step is verifiable. We present MODEXRE, a forensic pipeline that extends a previously validated synthetic-data-trained XGBoost classifier with SHAP explainability (arXiv:2607.00763) into a dual-mode architecture: a Laboratory mode, where models are trained and certified on public and synthetic data, and a Formal mode, where a frozen, hash-verified model performs inference on case evidence under a SHA-256 hash-chained chain of custody. Suricata alerts, CEF firewall logs and packet captures are normalized through OCSF. Four design decisions, each triggered by an observed failure, shaped the architecture: a timestamp leak that inflated Kitsune performance (F1-macro 0.984, 0.662 after correction); a feature-coverage gap that made the 78-feature reference model misclassify a real SSH alert; causal temporal flow aggregation, which raised Kitsune multiclass F1-macro from 0.394 to 0.932; and a certification defect, exposed only by a real packet capture, whereby the pipeline had always trained a binary classifier. After correction, the same capture went from 100% of events flagged as malicious to 87% classified as Normal, with the scanning host identified as Reconnaissance. The default four-class certified model reaches F1-macro = 0.9538 on CICIDS2017-derived evidence; the recertified 18-category model reaches 0.6214, reflecting the real multiclass task. Chain-of-custody verification, model integrity hashing and per-prediction SHAP explanations are reported as measured outcomes.

While Large Language Models (LLMs) excel in high-resource languages, their performance often degrades in low-resource languages due to tokenization inefficiencies, generative instability, and linguistic interference. This paper presents AssameseLM-E2B and its instruction-tuned variant, AssameseLM-E2B-it, a 2-billion-parameter model adapted specifically for low resource Assamese language. Built upon the Gemma-4 architecture, the model incorporates an expanded byte-level BPE vocabulary to reduce tokenization overhead by nearly 500% compared with multilingual 2B--3B baselines. We employ a parameter-efficient Continual Pre-Training (CPT) pipeline with Hybrid LoRA, followed by Supervised Fine-Tuning (SFT) on a balanced, multi-domain dataset to enhance instruction-following and Retrieval-Augmented Generation (RAG) capabilities. Evaluation demonstrates a Bits-per-Byte (BPB) score of 0.7488 on FLORES-200, alongside an overall RAG score of 96.7% and 100% refusal accuracy on adversarial unanswerable prompts. Comparative analysis reveals significant limitations in general-purpose baselines, including receptive-productive decoupling, generative degradation on Romanized Assamese, and unintended Bengali script assimilation. These results demonstrate the potential of parameter-efficient adaptation for developing capable, locally deployable language models for low-resource languages under constrained computational resources.

Multimodal large reasoning models (MLRMs) integrate visual inputs with textual and parametric knowledge, allowing sensitive information to be accessed through cross-modal associations during reasoning. This raises privacy concerns when a visual input should no longer provide access to such information. Machine unlearning can remove this influence without retraining the model from scratch. However, multimodal unlearning methods mainly target samples, concepts, or knowledge content, without separating the knowledge from the cross-modal association through which it is accessed. In reasoning models, target information may disappear from the final answer while remaining in intermediate reasoning, or be replaced by incorrect substitute knowledge. We therefore study cross-modal association unlearning, where the forget target is a designated visual-to-knowledge association: the model should suppress target-bearing behavior under the designated visual context while preserving access to the same knowledge in permitted contexts. We propose Visual-Keyed Counterfactual Rebinding (VKCR), a parameter-efficient framework for cross-modal association editing. VKCR first localizes the edit to a compact association-conditioned parameter subspace. It then reduces support for the designated target at task outputs and counterfactually rebinds the affected reasoning behavior by discouraging target-bearing or unsupported continuations while favoring valid non-sensitive alternatives. We further introduce CLEAR-R and MLLMU-R to evaluate target leakage, hallucinated reassociation, reasoning preservation, and modality-specific utility. Experiments show that VKCR substantially reduces image-conditioned target leakage and hallucinated reassociation, including across heterogeneous task formulations, while preserving non-sensitive visual reasoning and permitted text-only knowledge.

Multivariate probabilistic forecasting requires modeling cross-variable dependence alongside marginal uncertainty. The challenge is to learn this dependence while controlling marginal noise scales and avoiding dense covariance costs. We propose Factor-Correlated Diffusion (FaCoDiff), which couples a compact conditional statistical estimator with structured denoising. History-conditioned factor loadings use distinct coefficients for forecast-center correction, scale adjustment and noise dependence. Diagonal compensation preserves the estimated marginal noise variances, while low-rank operators enable residual Gaussian likelihood evaluation and noise transport without dense covariance storage or factorization. Factor-derived feature modulation explicitly conditions the denoiser, and a geometry-aware objective accounts for errors after noise transport. The same conditional geometry governs forward perturbations and reverse sampling. Across nine datasets, FaCoDiff achieves the lowest mean variogram score on eight datasets and ranks among the top two on 43 of 54 dataset-metric combinations in the reported comparisons. Matched ablations on Solar and Weather show that factor covariance improves mean Energy and Variogram scores under all four combinations of center and scale correction. Disrupting cross-variable sample pairings while preserving marginal trajectories worsens joint and aggregate scores on both datasets, demonstrating the predictive value of generated cross-variable dependence.

We study adversarial market making with orders that remain active until execution. The maker operates on two markets: a primary market where orders are made and a secondary market where the resulting positions are hedged. In each round, the maker posts a quote on the primary market before observing both primary and secondary market prices. An order executes when the primary price reaches its quote, and the resulting position is immediately offset at the secondary price. When both price sequences are chosen by an adversary over $T$ rounds, we give a simple deterministic policy with regret of order $\log T$ against the best fixed quote (static comparator) and $\sqrt T$ against the best quote for each submission date (dynamic comparator), plus two terms measuring exposure to future primary and secondary prices. These rates remain optimal even when both exposure terms vanish and lower bounds show that the exposure terms are necessary. Our policy naturally extends to simultaneous bids and asks, where we show that static and dynamic regret are both characterized (up to constant factors) by the exposures terms (where the primary price exposure $V_T^{\mathrm{two}}$ now accounts for both bids and asks) and a term of order $\sqrt{V_T^{\mathrm{two}}\log T}$. Our policy is simple, efficient and fully parameterless. We extend this setting by demonstrating how natural assumptions on the price processes lead to improved guarantees without changing the policy.

Many high-stakes regression decisions ask three things of a model at once: a point estimate to act on, an interval that bounds the risk of acting, and the option to abstain when neither can be trusted. Deep learning provides these in isolation: prediction-interval methods return an interval, and at best a value estimate inside it, for every input, while selective-prediction methods learn when to abstain, but only for point predictions and with no statement of how far the truth may lie from the estimate. We introduce SelPI, which trains interval, value and abstention heads jointly under a single loss with an explicit acceptance-rate budget. A natural objection is that interval width already tells us when to abstain. We show that it does not: across 18 tabular benchmarks, width predicts large point errors well (AUROC 0.72) but is anti-correlated with interval misses (AUROC 0.38), so abstaining on wide intervals discards precisely the inputs the interval got right. SelPI therefore selects with two gates, one trained on the interval loss and one on the value loss, coupled by a Fr'echet penalty that ties their separate budgets to the joint acceptance rate they deliver. With the base predictor held fixed, the learned gate beats width thresholding on a proper interval score on all 18 tabular datasets, closing a median 87% of the gap to an oracle selector, and it also beats selection by deep-ensemble disagreement on every dataset. The advantage grows with the heteroscedasticity of the task, and we report the settings in which it disappears.

Predicting the long-time diffusion of membrane proteins is challenging because molecular dynamics (MD) simulations are computationally expensive, while experimental single-particle tracking is often limited by finite observation times. Here, we introduce a neural stochastic (NS) model for forecasting the collective diffusion of membrane proteins from coarse-grained MD trajectories. The model combines a transformer-based representation of protein configurations, recurrent networks operating at multiple temporal scales, and a stochastic latent process with learned memory. A cross-attention decoder then generates correlated protein displacements while accounting for the instantaneous system geometry and periodic boundary conditions. We evaluate the model in two membrane systems: seven interacting PIEZO channels and a mixed system containing seven PIEZO and ten TREK channels. For the seven PIEZO system, the NS model reproduces the anomalous diffusion exponent measured in MD simulations $α\sim 1$, substantially improving on a previously developed variational autoregressive model that exhibits superdiffusive behavior. The model also generates trajectories extending to 40,000 frames and captures the distinct diffusion characteristics of PIEZO and TREK populations in the mixed system. Comparisons of time-averaged mean-square displacements, Van Hove displacement distributions, and non-Gaussian parameters show that the model preserves key statistical features of the simulated dynamics. These results establish neural stochastic modeling as a promising approach for generating long trajectories of multi-protein membrane diffusion and extending the statistical analysis of molecular dynamics simulations.

Deep learning-based super-resolution (SR) can reconstruct fine-scale Earth-observation fields from coarse observations, but reconstruction accuracy alone does not provide a calibrated rule for identifying reliable spatial regions. Conformal Mask Calibration (CMC), built on conformal risk control, converts a spatial uncertainty score into a trust mask and controls the expected worst-case local distortion under exchangeability. However, its fixed threshold can become miscalibrated in Earth-observation streams affected by seasonality and episodic events. We propose $\textit{CARM}$ (Conformal Adaptive Risk Masking), a statistical framework for adaptive spatial risk control under temporal distribution shift. CARM maintains experts with heterogeneous update rates and aggregates their threshold proposals online using an asymmetric surrogate loss, balancing rapid adaptation to distribution changes against stability during relatively stationary periods. All updates are causal and use only mature feedback available after a fixed delay, allowing CARM to operate with any SR model that provides a spatial uncertainty score. We prove that the time-averaged mask-conditioned risk converges to at most $α$ at an $O(1/T)$ rate, without distributional assumptions and for any choice of aggregation weights across experts. On a daily South China Sea chlorophyll-$\textit{a}$ SR stream, CARM reduces the empirical violation rate from $43.49\%$ to $9.38\%$ and the moving-average exceedance frequency from $47.29\%$ to $1.22\%$ at $α=0.10$, while retaining a non-trivial trusted region. Experiments on ERA5 wind and OSTIA sea-surface temperature further show consistent relative improvements across multiple SR backbones.

Video-LLMs are typically trained on paired video-text data for temporal tasks, yet collecting precise temporal annotations and training on large-scale video datasets are both costly. We investigate whether a text-only LLM, with no vision-language pretraining, can be trained on textual video descriptions alone to perform temporal grounding and receive video only at inference time. This setting removes paired video data from training altogether and is a step toward training multimodal models without paired annotations. We generate a 134K-example corpus of textual sequences with up to 100 time steps, adapt Qwen3-8B with LoRA on CLIP or LLM2CLIP text embeddings, and evaluate on Charades-STA and QV-Highlights without training on videos from either benchmark. LLM2CLIP, whose text encoder supports longer inputs, represents detailed descriptions better than CLIP (61.54 versus 53.86 mtIoU on Temporal-TextVid) and transfers them better to video. Still, replacing text with frame embeddings at inference yields only 4.99 mtIoU on Charades-STA and 11.12 on QV-Highlights, revealing a large modality gap. To bridge it, we propose Nearest Entity Replacement (NER), a training-free mapping that represents each video frame by the normalized mean of its nearest text-entity embeddings, so that the model receives the kind of input it was trained on. NER reaches 14.45 mtIoU on Charades-STA and 20.72 on QV-Highlights, outperforming the strongest of four families of modality-gap baselines by 5.00 and 1.71 points. NER thus enables partial transfer from text to video, while more reliable grounding will likely require encoders with better-aligned text and visual representations.

Mixture-of-Experts (MoE) language models send each token through a few of many expert subnetworks. Serving methods make these models faster by running even fewer experts per token, and they are validated on accuracy benchmarks. We show that these methods can weaken the model's refusal of harmful requests, and we introduce SafeStart, which largely restores refusal while keeping most of the speedup. In vLLM on one RTX A6000, keeping about two thirds of the selected experts speeds up offline decoding by 15 to 28% at batch size 8. We find that, at this setting, under a jailbreak that tells the model not to refuse, the share of harmful requests that the model answers rises by 18.4 percentage points on Qwen1.5-MoE, 9.5 on OLMoE, and 6.6 on DeepSeek-V2-Lite, the three of nine models where it rises significantly, while MMLU accuracy drops by less than two points. We discover that most of this increase arises while the model processes the prompt and generates its first few tokens. SafeStart therefore runs all experts for the prompt and the first eight generated tokens, and the reduced set of experts afterwards, through a fused kernel in vLLM. With experts removed, SafeStart keeps at least 85% of the offline decode speedup at batch sizes 8 and 32 and brings the change under this jailbreak to between -2.1 and +1.3 points for three expert-reduction methods on these three models. On 889 held-out requests, in an analysis fixed before generation, truncation raises harmful compliance by 5.8 to 21.4 points, and SafeStart leaves a rise of 0.6 to 2.6 points, non-inferior to the original model on two of the three models. Our results show that faster MoE serving need not cost safety: a serving system can keep most of the gain from running fewer experts by running all of them at the start of each response, and it should test refusal, not only accuracy, before it enables such a method.

Many recent approaches for vulnerability detection use large language models (LLMs) to classify functions as vulnerable or benign, leaving developers to locate the exploitable statements. Furthermore, while these approaches have shown performance improvements, they still fall short in distinguishing vulnerable from patched real-world functions. This paper presents SBVul, an agentic approach that detects exploitable statements in C/C++ functions. In Phase~1, an LLM agent explores the function and its repository and identifies candidate statements. In Phase~2, a second agent analyzes the control and data flow of each candidate and confirms it, rules it out, or marks it uncertain. We created a manually verified ground truth of exploitable statements for 200 pairs of the PrimeVul Paired test set and ran SBVul with the open-weight models Qwen 3.8 and DeepSeek V4.1 Flash and Anthropic's Sonnet 5.5 and Opus 5.5. Phase~1 identifies ground-truth statements in 56--77% of the vulnerable functions, and Phase~2 confirms such a statement with an explanation accepted by an LLM judge in 26--73%. However, all models also confirm candidates in 23--61% of the patched functions. Across vulnerable and patched functions, the highest statement-level precision is 0.315 at a recall of 0.605 obtained with Opus 5.5 (xhigh). At the function level, SBVul with Opus 5.5 classifies 42.0% of the pairs correctly, compared with 33.8% reported for the best previous approach.

The full-attention layers of large language models (LLMs) make every generated token more expensive than the previous one, slowing models down on long-context tasks. Hybrid models reduce this cost by replacing most full-attention layers with linear-attention layers, but the first still dominate decoding at long contexts. We accelerate these remaining layers without any training or weight changes using combined sparse attention: each layer attends to a sliding window of recent tokens, a pinned prefix, and a fixed budget of most relevant tokens in the context. The window size and budget are found on a per-layer basis using a multi-objective search algorithm. Across six hybrid models from three families, from 4B to 397B parameters, the selected configurations preserve their baseline model's accuracy on 21 long-context datasets. Decoding 64k/256k tokens becomes up to 1.26x/2.22x faster on Qwen3.5-4B, 1.20x/2.23x on Qwen3.5-122B-A10B, and 1.16x/2.06x on Ornith-1.0-397B.

Professionals have long used software built for them by others. This article reports the growth of a professional computer-assisted interpreting (CAI) tool built by its own user, a conference interpreter without programming training. Drawing on an eight-week version-controlled record with telemetry from 133 logged sessions and over 20,000 recognized utterances, the article traces how real conference work drove the tool through 62 commits and six versions. The record shows capabilities deployed on the day their need appeared and one strategy change that cut median end-to-end delay by over a third within one afternoon. The article then proposes the paradigm of practice-grown tools, defined by the shift from using to growing, the co-evolution of tool and practice through breakdown and repair, bounded local computation, and an assessment corollary, and it formulates three propositions toward tool-inclusive assessment of practitioners, from tool-making literacy to a compute-standardized format for certifying what they and their tools achieve together.

Static and incomplete parametric knowledge limits the performance of multimodal large language models on novel-knowledge and knowledge-intensive tasks, especially those requiring multi-hop evidence acquisition. Existing multimodal agentic search methods often fix visual evidence at the input or a predefined trajectory position, making it difficult for agents to select search modalities as their information needs evolve and to handle complex real-world queries. We define this capability as autonomous cross-modal search: an agent chooses among direct reasoning, text-to-text search, text-to-image search, and reverse-image search, actively introduces visual evidence when appropriate, and switches between and integrates modalities across multiple hops. We propose ModalSearch-Pipeline, a scalable procedure for constructing an interleaved multimodal knowledge graph. It samples paths with diverse depths and modality compositions and uses backward question construction to obscure entity names and image targets that would expose reasoning shortcuts while retaining necessary disambiguating clues, yielding text-only, image-initialized, and arbitrarily interleaved multi-hop problems. We train the ModalSearcher Agent with supervised fine-tuning and reinforcement learning in a live Web environment, and construct ModalSearch-Bench. Experiments show that ModalSearcher-8B improves average accuracy by 16.2 points over Qwen3-VL-8B under the same agentic workflow across five multimodal knowledge-intensive QA benchmarks, with consistent gains on independently constructed active visual-search benchmarks.

Accurate short-term wind power forecasting is critical for grid stability and electricity trading. While transformer-based deep learning is state of the art, industry practice relies mostly on local models trained for individual assets. This study investigates whether a global temporal fusion transformer (TFT), trained jointly on a heterogeneous portfolio of wind farms across Germany, outperforms site-specific local models in 48-hour-ahead forecasting. Using numerical weather prediction (NWP) data from the regional ICON-D2 and the global ECMWF HRES models, we compare the global TFT against local models and a linear regression baseline. The global model leverages static covariates, such as farm age, altitude above sea level, hub height, rotor diameter and power curve parameters, to learn shared representations across sites. We evaluate on 100 synthetic wind farm profiles and a real portfolio of 11 German wind farms, across two seasonal test periods with intermediate retraining. The global TFT outperforms the local models for 87% of the synthetic and 77% of the real wind farms and clearly surpasses the linear baseline. Wilcoxon signed-rank tests confirm that these gains are significant across both seasons and datasets. Crucially, this superiority depends on static covariates: without asset-specific context, the global model fails to generalize. The TFT's variable selection network (VSN) reveals the learned feature selection. Altitude emerges as a relevant static discriminator on both datasets, while hub height is decisive on the real portfolio. For dynamic inputs, the encoder prioritizes historical power, whereas the decoder combines upper-level wind speeds from both NWP models, consistent with averaging out model-specific errors. These findings support a shift from single-site to scalable multi-site forecasting, reducing operational complexity while improving accuracy in large portfolios.

Alexa+ devices with screens display animated icons during conversations to visually reinforce what is being said. The original set of these icons was hand-drawn by designers one at a time, a process that produces polished results but is time-consuming and cannot scale to the range of topics encountered in open-ended conversation. We describe an automated pipeline, deployed in production, that generates these vectorised animated icons from a text description alone and surfaces relevant animations to customers based on their ongoing conversation. The pipeline decomposes icon creation into a sequence of stages (image generation, semantic decomposition, vectorisation, styling, and animation). Its quality is determined at the very first of these, image generation, on which every downstream stage depends. We therefore focus on improving this first and most critical stage. Because the icons are rendered on small display canvases under strict design guidelines, generations are often unsuitable for production. Supervised fine-tuning is not viable, as no dataset of line-art icons exists under these constraints. Instead, we design objective verifiers that automatically flag generations violating the design guidelines, and use them to fine-tune a flow matching image generation model via reinforcement learning from verifiable rewards (RLVR). The fine-tuned model raises the pass rate from 15.7% to 58.7% (3.7x) under a VLM judge that is independent of the reward model. Both the trained model and the verifier are deployed in the production pipeline, where the verifier also serves as the automated quality gate before final design review.

Molecular dynamics predicts molecular motion from atomic forces and trajectory integration. Density functional theory (DFT) provides first-principles forces, but its $O(N^3)$ cost limits simulations to thousands of atoms; empirical force fields scale to millions but miss bond breaking and charge transfer. Equivariant message-passing Transformers learn the ab initio potential at linear cost. Stacked layers propagate information beyond a single cutoff, giving four layers a receptive field up to 23Å. At this frontier, scale is limited by communication: layers exchange boundary features across devices, force backpropagation repeats the exchange, and moving atoms change neighbor graphs and tensor shapes. We present Tessera, a distributed runtime that fixes workload shapes between rebuilds. It sorts atoms along a space-filling curve into equal-count shards, ensuring identical atom counts and tensor shapes per rank. Peer sets scale as $K^{0.5}$, not $K$. A conservative neighbor skin preserves topology across updates, enabling each rebuild interval to replay as one CUDA Graph. Fused kernels, three-level neighbor search, and overlapped All-to-All complete it. On 512 GPUs, Tessera sustains 30.7M atoms at 89.7% weak-scaling efficiency (77--90% across three per-GPU loads); per-step time rises 1.11--1.29x as machine and problem grow 16x. Graph reuse cuts full-step time $1.87\times$. Across an $8\times$ device increase, neighbor-search cost grows $1.7\times$, versus $3.0\times$ for Morton order. In a live $10.24$M-atom salt solution, NVE energy drift is $0.42\,μ$eV/atom/ps over $10$ ps, and structure matches a serial reference. A 10.06M-atom Norovirus capsid sustains 100 ps of NVT at $\approx 8$ ps/day on 256 A100 GPUs. Over 200,000 steps, it has no broken bonds or anomalously close pairs and uses three compiled variants, with no graph breaks or eager fallbacks.

Banks validate language models, but they deploy systems: a model inside a harness that decides what it sees, which tools it can call and who checks its actions. We test whether that harness changes compliance with banking conduct rules when the model is held fixed. Four open-weight models each ran five harnesses (a pipeline that makes one structured decision, a ReAct-style tool loop, a loop in which the policy is reachable only through search, a loop with a long benign customer memory, and a loop with a same-model compliance reviewer) on 156 pre-registered cases in a simulated retail bank whose 11 rule families paraphrase EU consumer-credit, payments, investor-protection, data-protection and anti-money-laundering provisions. Every one of the 3120 two-turn episodes was scored deterministically from the executed tool trace, with violations and over-refusals counted separately. Pooled across models the harness moved the correct-case rate only between 87.3% and 92.1%, but within a model it mattered a great deal: for one model the spread between its best and worst harness was 11.5 percentage points, enough to move it from first to last place among the four models, while for another the harness made no detectable difference; the model x harness interaction itself was suggestive rather than conclusive. The same-model reviewer cut violations from 9.6% to 6.4% without raising over-refusals, contrary to our pre-registered prediction. Rules that require date arithmetic failed in every harness. In an exploratory analysis with two LLM judges and a blind human check, 26% of violations were accompanied by a customer message that denied the action the agent had just executed, which a transcript-only review would record as a correct refusal. For model-risk management the unit of validation should be the model-harness pair, and conduct testing should read the action log, not the conversation.

Online scam videos pose significant threats to individuals and society. Existing studies primarily focus on binary classification, offering limited insights into how scams work and the reasoning behind detection decisions. Moreover, they often overlook the challenge of distinguishing scam videos from superficially similar non-scam content, such as advertisements, and the robustness of detection models against misleading arguments. These limitations hinder reliable scam video detection. To address these limitations, we introduce MechScam, a benchmark comprising 6,614 videos covering diverse scam mechanisms and non-scam categories. MechScam provides explanations supported by video evidence, annotations of potential scam elements and their context, and misleading arguments constructed from genuine evidence. Recognizing that genuine evidence can still be used to justify misleading conclusions, we further propose ScamCourt, a debate framework in which trained prosecution and defense debaters present opposing arguments, while a trained judge makes decisions based on a transcript with verified citations. Extensive experiments show that state-of-the-art MLLMs struggle to distinguish scam videos from similar non-scam content. ScamCourt outperforms larger MLLMs and same-backbone fine-tuned baselines, reducing false positives and showing stronger resistance to misleading arguments. Together, MechScam and ScamCourt contribute to more reliable and interpretable scam video detection. The dataset and code are available at https://github.com/SuperYFan/MechScam.

Latent Chain-of-Thought (CoT) reasoning improves efficiency by shifting intermediate computation into continuous hidden states. However, fully latent methods often struggle on precision-critical tasks such as mathematical reasoning, while existing hybrid approaches retain substantial token-generation overhead. We introduce VITAL, an importance-guided hybrid reasoning framework that selectively verbalizes informative tokens as reasoning anchors while keeping most computation latent. A lightweight gating head dynamically selects between explicit token generation and direct hidden-state feedback. VITAL constructs hybrid trajectories from explicit CoT using token- importance estimation and trains the model through curriculum-based supervised training, followed by reinforcement learning with a reward that balances correctness and conciseness. Experiments show that VITAL consistently out- performs existing fully latent reasoning methods, achieving relative average accuracy gains of 9.6% on GPT-2 and 5.1% on Llama-3.2-1B at comparable reasoning lengths. Compared with hybrid baselines, VITAL achieves higher reasoning accuracy with substantially shorter reasoning trajectories on Qwen3-1.7B.

Watermarking has great potential for regulating the use of language models (LMs), but its deployment has been marred by controversy, as many users argue that fine-grained tracking of LM usage violates their privacy. This tension reduces the effectiveness of watermarks, as users engage in evasion strategies such as paraphrasing that can remove much of the watermark signal. We show that provenance and privacy can be reconciled if we shift the goal of watermark design from tracking individuals to estimating aggregate usage. We develop locally differentially private watermarks which provably limit the evidence that any individual text can contribute to detection, while still enabling accurate estimation of population frequencies. We introduce the Red-Green Sequence (RGS) watermark, a private and computationally efficient watermark that achieves accurate frequency estimation at high privacy levels across various models, sample sizes and generation entropies in settings that reflect realistic deployment scenarios. On C4 language modeling tasks, we attain $0.0019$ MSE with $0.0038$ mean absolute bias at a low $\varepsilon$ of 1, while in more challenging settings such as an essay writing task with only 100 text samples, we still find errors on the order of $0.0113$ MSE with $0.004$ mean absolute bias at $\varepsilon=2$.

As autonomous agents advance, the traditional boundary between software development and cloud operations gradually blurs, increasing the need for integrated incident detection, localization, and remediation. Existing benchmarks evaluate isolated diagnostic queries or repository-level program repair, failing to test continuous incident resolution across operational clusters and source repositories. We introduce SWEOps-Bench, an interactive benchmark comprising 55 reproducible incident tasks derived from confirmed historical defects across 16 Kubernetes operator projects. Each task preserves the original software repository and initializes in a healthy cluster before a controlled workload triggers the historical fault. Agents investigate runtime evidence and submit source or configuration repairs through a terminal interface, while the harness builds and deploys the changes within a shared time budget. Evaluation covers stage-wise accuracy, verified recovery, penalized lifecycle latency, and token consumption. Across six agent-model configurations, faulty-state detection ranges from 72.7% to 90.9%, whereas overall recovery ranges from only 14.5% to 52.7%, exposing a substantial gap between recognizing incidents and resolving them end to end. Trajectory analysis further shows that reported diagnosis and final recovery only partially coincide: some tasks achieve verified recovery despite an incorrect reported RCA, while runtime symptom clearance can still fail fix-specific functional verification. SWEOps-Bench establishes a rigorous foundation for evaluating autonomous agents in integrated development and operational environments.

Can hyperspectral (HS) imaging be used to predict the elemental composition of physical specimens found on the Earth's surface? We approach this problem as a supervised regression task, using an open dataset of specimens, each comprising a HS cube and elemental characterization. In particular, we train an encoder in a self-supervised fashion that operates on pixel-level; then freeze the encoder and train a lightweight attention-based regression head in a supervised manner, by pairing many representations of a given specimen with its corresponding elemental label. In test specimens, our approach significantly outperforms traditional spectral matching and unmixing baselines. It also outperforms supervised learning models based on the mean spectrum (of similar model capacity), while full model fine-tuning achieves the strongest performance, at substantially higher model capacity.

Explicit reasoning can help language models generate code, but it also introduces a practical risk: the model may spend so much of its output budget reasoning that it never produces usable code. We study this behaviour in relatively small open-weight reasoning models, framing it as reasoning-trace overflow. Across seven models and six code benchmark settings, we measure how much of a fixed 32k-token output budget is spent on reasoning, whether that reasoning improves pass@1, and whether lightweight inference-time controls can mitigate overflow. Our results show that the value of reasoning is highly conditional. Overflow is rare on simpler function-level tasks, but becomes common on harder code-generation benchmarks and for particular model families: on several model--benchmark pairs, more than half of generations exhaust the full 32k-token budget while still reasoning. Most complete overflow is associated with repetitive reasoning loops, although substantial long reasoning also occurs without looping. Disabling reasoning can therefore be competitive for overflow-prone models, simply because it ensures that an answer is produced. At the same time, reasoning remains valuable for stronger models on harder algorithmic and code-understanding tasks. Finally, prompt steering can substantially reduce overflow for some models but is inconsistent, while budgeted transitions provide more direct control and can improve both performance and token efficiency for overflow-prone models. These findings suggest that explicit reasoning should be treated as a controllable resource for code generation, rather than an always-on capability.

Pretrained vision models provide strong general-purpose representations, but in Domain-Incremental Learning (DIL), where the input distribution changes sequentially, adapting to new domains often degrades performance on previously learned ones. We observe that forgetting is not uniform across domains and classes and is not explained by training order alone. This motivates us to track each past domain-class concept individually. We use representation drift, defined as the deviation of a past concept's current representation from a fixed anchor, as a measurable surrogate for forgetting risk. The measured drift then serves two complementary roles. A consolidation loss assigns stronger preservation to high-drift pairs, while prototype construction down-weights highly drifted domains to reduce their influence on the class reference. Preservation is thus allocated by measured drift rather than assumed importance. To obtain this signal during training without storing past raw data, we repurpose a Joint-Embedding Predictive Architecture (JEPA). For each domain-class pair, we store only a compact context summary in the backbone feature space and rerun it through the learned context-to-target relation to measure its drift from the fixed anchor. Across DomainNet, Office-Home, CORe50, and CDDB-Hard, our method sustains a strong accuracy-forgetting balance on every benchmark, using substantially less training memory than accuracy-competitive methods and the smallest storage among compared methods.

This paper presents an overview of \emph{SciClaimEval}, a pilot task at NTCIR-19 on verifying scientific claims against the tables and figures of the papers in which they appear. Our dataset pairs authentic claims from 180 papers in machine learning, natural language processing, and medicine with contrastive versions of their tables and figures, one supporting the claim and one minimally modified so that it does not. The task comprises two subtasks: deciding whether a piece of evidence supports a claim, and selecting which of two evidence versions does. Twenty-two teams registered, the highest number among NTCIR-19 tasks, and 14 submitted 177 runs. Most systems built on multimodal large language models, relying on prompting and multi-stage inference rather than task-specific fine-tuning. On the first subtask, the best systems reached 98.1\% pair accuracy, but all six top-ranked systems relied on the contrastive pairs; the best system that judged each claim-evidence pair on its own scored 24.3 points lower. Open-weight models, embedded in carefully designed pipelines, matched proprietary ones. Yet verifying a single claim against its evidence, the setting that matters in practice, remains an open problem.

Hybrid language models interleave a few softmax-attention layers with many linear-attention or state-space layers, so every layer keeps a memory of earlier tokens: a key-value cache or a fixed-size recurrent state. Serving them spends more compute and high-bandwidth memory on these memories than the output requires. The prefill runs every layer for every prompt token, and every decode step loads and stores whole recurrent states, which on Qwen3.8-Flash-Next take as many bytes per request as the attention cache of thousands of tokens. We find two properties of trained hybrids that locate this excess. First, writes are easier than reads: writing a deep layer's memory of a token needs little of the deep computation that predicting the next token needs. Trained copies of the deep layers' write paths, which we call emitters, therefore write a prompt's deep memories from the residual stream of a shallower anchor layer, and all prompt tokens but the last run as few as 63% of the layers. Second, recurrent states have a state sink: most of their energy lies in a rank-one term along a fixed direction of each head, and the rest is a low-rank working memory. SEED builds on both: it makes the prefill shallow in depth and the decode state shallow in representation. Each recurrent state is stored as its sink plus low-rank factors, which the Gated DeltaNet, Kimi Delta Attention and Mamba-2 updates extend exactly; we prove that the open-loop error of the periodic compressions back to low rank is bounded independently of their number. SEED keeps the original weights frozen, compresses no attention cache, and runs all layers for every generated token. On six checkpoints of five production hybrids of 30B to over 300B parameters, it is nearly lossless on eleven official exams. Served with SGLang on NVIDIA GB300 GPUs on an agentic-coding trace, it raises the per-GPU throughput of Qwen3.8-Flash-Next 1.2-1.4x.

Fast adversarial training (FAT) makes robust training affordable by replacing the multi-step attack with a single-step one, but it suffers from catastrophic overfitting (CO): partway through training, robustness to multi-step attacks collapses to near zero while the single-step attack used for training keeps being repelled. We show that CO is a quiet escape of the weights. Across the event the network's clean behavior quietly improves, and moving from the weights just before CO to those just after it changes the training objective only marginally while destroying most of the robustness. The training objective is therefore nearly blind to the displacement it causes, so it exerts no force to undo it. The exponential moving average (EMA) of the weights, which FAT protocols already maintain, does not accumulate this drift and stays robust through the entire event. This observation suggests a nearly free fix: when validation robustness drops, reset the weights to their EMA. We build this guard on top of a coupling-based base method designed so that its average stays inside the robust region. The resulting training removes CO in every one of our runs and attains the strongest AutoAttack accuracy among fast methods on all three benchmark datasets, including a clear margin on CIFAR-100, where it also clears the multi-step reference at roughly half its cost. Code is available at https://github.com/fzjcdt/co-weight-drift.

Modern scientific and machine-learning applications compose many independently developed libraries. When these libraries run together, their independently managed runtimes can oversubscribe shared resources: using the same CPU cores, fighting over GPUs, or contending for memory bandwidth. Avoiding such contention requires careful configuration across library boundaries. Existing autotuners do not directly investigate what resources are shared between libraries, which runtime signals correlate with interference. As a result, they are inefficient for finding optimal configurations to mitigate contention. This paper presents SysPilot, a cross-library performance autotuning system based on Library Performance Contract (LPC). A Library Performance Contract captures a library's resource demands, assumptions, controls, and observability signals in LPCL, a specification language for describing library resource requirements. SysPilot synthesizes contracts from documentation and code artifacts, composes contracts from the libraries used by an application, derives resource-conflict hypotheses, profiles the application to validate those hypotheses, and selects candidate configurations. This design makes the Library Performance Contract a reusable unit of performance knowledge: a contract for a library can be reused across many applications that depend on that library. We evaluate SysPilot on a benchmark suite of cross-library contention scenarios. The results show that SysPilot matches the performance of autotuners that rely on application-specific manual information while retaining cross-application reuse. Compared with general-purpose autotuners such as OpenTuner and SMAC3, SysPilot improves benchmark performance metrics by up to an additional 38% over the best configuration found by these autotuners, while using 3x fewer runs.

Embodied and edge systems increasingly rely on vision-language models (VLMs) to support heterogeneous skills, yet deploying a separate task adapter for each skill while retaining a high-precision backbone creates substantial memory and management costs. We present UniCompress, a deployment-oriented framework that converts multi-task expertise into one compact, adapter-free model. Starting from a consolidated BF16 Qwen3-VL-4B teacher, UniCompress constructs an asymmetric INT2 student using GPTQ initialization and teacher-guided quantization recovery. The approach jointly quantizes and recovers the projection weights in the vision encoder and language backbone, enabling aggressive compression of the complete vision-language model while preserving its multi-task behavior. We evaluate UniCompress on four heterogeneous embodied tasks: posture recognition, household-safety understanding, multimodal dialogue routing, and mobile-GUI action prediction. UniCompress reduces the approximate weight storage from 8.9 GB in BF16 to 2.1 GB in INT2, corresponding to a reduction of approximately 76%. It improves the INT2 macro task score by 34.3 percentage points over GPTQ-INT2 while remaining within 0.9 points of the consolidated BF16 teacher. These results demonstrate the potential of teacher-guided joint vision-language INT2 recovery for deploying a single multi-skill VLM on resource-constrained embodied platforms.

Existing audio-visual segmentation methods mainly predict sound-related regions or category masks and often fail to identify the sounding instance when multiple same-class candidates coexist. Recent audio-visual instance segmentation methods address instance-level prediction but rely largely on motion and cross-frame association, which may become unreliable under sparse observations or inconspicuous motion. To this end, we introduce binaural audio, whose interaural differences encode source-direction information, to provide spatial constraints for distinguishing the sounding instance among same-class visual candidates. We propose Binaural Sounding Instance Disambiguation for Audio-Visual Segmentation (BiSID-AVS), which progressively refines category-level candidate responses into a mask of the true sounding instance. BiSID-AVS first grounds category-consistent candidate regions through audio prompting, then projects binaural cues into an image-plane spatial response map, and finally refines the true sounding instance through dual-level spatial conditioning at the feature and decoder levels. Experiments on the binaurally extended AVS-Synthetic benchmark and Habitat-AVS dataset for robotic indoor perception show that BiSID-AVS achieves the best performance on the multi-candidate subsets, improving instance-level mIoU by 35.19 and 41.21 percentage points, respectively, while substantially reducing false activation on non-sounding same-class regions.

Parameter-Efficient Fine-Tuning (PEFT) has emerged as an important approach for adapting Language Models (LMs) in downstream tasks. However, in the domain of code intelligence, the performance of most PEFT methods relatively drops compared to Full fine-Tuning (FT), hindering their broader applicability. One of the main reasons is that current PEFT methods are generally based on assumptions about the low-rank nature or sparsity of parameter updates, lacking explicit modeling of code-related information. To tackle this limitation, we propose SynMOE, a syntax-aware MOE-guided approach for PEFT on code intelligence learning. Specifically, SynMOE contains two stages: syntactic information extraction and syntax-aware MOE-guided adaptation. It first obtains the syntactic labels of source code from Abstract Syntax Trees (ASTs), and then aligns the labels with the input tokens of code LMs. In the second stage, syntactic labels are integrated into the MOE module, participating in the routing decision of MOE for a more powerful syntactic-aware token representation. We evaluate the effectiveness of the proposed approach by applying it to two typical PEFT methods, LoRA and $(IA)^3$, across three diverse code intelligence tasks. Experimental results show that SynMOE achieves consistent and significant improvements. Even compared to FT, SynMOE achieves comparable or superior performance while being more computationally efficient. In addition, we also conduct ablation studies, hyperparameter analysis, and evaluation in the low-resource scenario to provide a comprehensive investigation of the proposed approach.

We study how a seller can learn a reserve price for a second-price auction from censored historical bid data. Because values below past reserves are unobserved, even unlimited data may not reveal which reserve earns the most revenue. Under a monotone hazard rate (MHR) assumption, we characterize exactly when an optimal reserve is identified, even though the value distribution may not be: either revenue has not yet begun to fall at the lowest historical reserve, so the best reserve lies among the observed bids, or the censored fraction is as large as MHR allows, which pins down the entire distribution. When identification fails, we derive the reserve that minimizes worst-case revenue regret over distributions consistent with the data, and prove that this unique minimax reserve and its regret depend on just two observable quantities: the censored fraction and the hazard just above the lowest historical reserve. With $N$ historical bidders, the optimal reserve is learned with expected worst-case regret $O(N^{-1/2})$ when it lies among the observed bids, and $O(N^{-1})$ from the censored count alone when the censored fraction is as large as MHR allows and the optimum lies at or below the lowest reserve. In the nonidentifiable regime, plugging estimates of the two quantities into the minimax reserve gives expected worst-case regret within $O(N^{-2/5})$ of the population minimum under local density smoothness, with a matching minimax lower bound on a smooth class. Beyond MHR, we solve the minimax problem for regular and unrestricted distributions and prove an $O(N^{-2/5})$ learning bound in the regular nonidentifiable regime, given known bounds on the boundary parameters. Experiments on synthetic and eBay data show gains from exploiting MHR and the benefits of the robust rule when parametric models are misspecified.

Automatic modulation classification (AMC) is a key component of cognitive wireless systems, enabling spectrum awareness and adaptive communication. Most deep learning approaches to AMC assume a fixed set of modulation classes and access to previously observed training data. In practice, modulation classes and channel conditions may evolve over time, while storage or data-retention constraints may preclude replay of raw baseband I/Q samples. We propose PRISM-AMC, a replay-free continual learning framework that mitigates catastrophic forgetting by preserving task-relevant subspaces of intermediate representations via gradient projection memory. We further analyze projected stochastic-gradient updates under an SNR-dependent gradient-noise model. PRISM-AMC is evaluated in both class-incremental and domain-incremental settings. Experiments on the MIGOU and DeepSig 2018.01A datasets show substantially reduced forgetting relative to sequential fine-tuning and elastic weight consolidation, while remaining competitive with or outperforming replay-based RSD-AMC in class-incremental settings, without storing raw I/Q samples.

In natural human-AI dialogue, users' corrections, requests, and self-descriptions carry reusable signals about how they want work to be done. These signals reveal individual process preferences, user-specific standards for evidence, depth, analytical focus, procedure, and delivery. We ask whether a preference revealed in one dialogue can shape an assistant's later answer when the task does not restate it. ProcessPrefBench uses three months of natural dialogues from two knowledge workers to build 66 source-target cases separately judged for source evidence and target applicability. Sources are preselected, so the benchmark tests application rather than retrieval. We test three models with target-only, full-history, and distilled-preference prompts. Target-specific 0-2 rubrics score preference realization, and a task gate flags severe failures. Compared with target-only prompting, full-history and imperative distilled-preference prompting raise preference realization by 0.204 and 0.244 points on the 0-2 scale, respectively, with no reliable advantage from distillation. Fabricated-evidence failures rise from 5.8% under target-only prompting to 10.1% under full-history and 10.4% under imperative prompting. ProcessPrefBench highlights a tension in learning from dialogue, where drawing on prior interactions can improve preference realization yet also invite fabricated evidence. We hope it helps build assistants that honor a user's process preferences without sacrificing factual reliability.

Time series forecasting is shifting from locally trained, per-series models to global models trained across many heterogeneous series, which requires learning what sets the series apart. When the raw data cannot be centralized, such a global model is trained with federated learning (FL), one client per data owner. State-of-the-art forecasters such as the Temporal Fusion Transformer (TFT) tell series apart through a static input that describes each series and does not change over time. In the special case of a federated setting where each client holds a single series, this static input is constant within each client and varies only across clients. We ask whether a federated TFT can learn from it. Per-batch FedAdam, which aggregates gradients after every batch, uses the static input almost as well as central training, whereas communication-efficient schemes with local epochs (FedAvg, FedAdam+Prox) extract far less from it and gain nothing significant on either application dataset. At the same local computation, an Adam-compatible SCAFFOLD keeps the static pathway responsive on wind power and gains significantly there. On building energy its error still drops, but the static pathway falls silent, so drift and communication frequency cannot be fully separated as the cause. Only per-batch FedAdam matches or beats the per-client local models across all datasets. We confirm the effect on building-energy meter data, a wind-power dataset, and a controlled synthetic dataset. The loss is not specific to the TFT: on the synthetic and the wind data two further architectures with different static-conditioning mechanisms show it as well.

Small LoRA adapters trained on Tajik news and books lowered Qwen3's Tajik text loss at every size we tried, from 0.6B to 8B, by 0.60 to 1.45 nats per token, but none of them improved Tajik reading comprehension where it can be measured. On Belebele the 1.7B and 8B adapters left accuracy unchanged and the 4B adapter lowered it by 3.7 points; at 0.6B the score rose by 4.2 points but sits near chance and coincides with a change in which answer letter the model prefers, so we do not read it as understanding. At 8B, continued pretraining on about 375 million tokens of raw Tajik text was followed by a Belebele gain of 9.9 points. After a further supervised fine-tuning stage, our main model, e3, is 11.4 points above Qwen3-8B on Belebele; its English MMLU changed by -1.2 points, and a loss larger than 3.3 points is excluded. The same heavy recipe applied to Qwen3-0.6B-Base brought its text loss close to that of the 8B model after books-only pretraining, while its multiple-choice answers stayed near chance and its English MMLU fell by 7.6 points. We report 15 trained models and their 5 starting models from June to October 2025, scored in 2026 on one harness with paired intervals, and we report what went wrong: lost training records, a thinking-mode dataset with almost no reasoning in it, later branches whose training data was never documented, and a 4-bit release that contains no transformer layers.

Recent video world models allow users to explore generated environments through camera navigation. Continuous hand motion provides a fine-grained action channel for interacting with the people and objects encountered during exploration. We present TouchéWorld, a framework that integrates gesture-driven interaction into a camera-controlled video world model. Given an initial image and a scene-description prompt, the model generates video conditioned on temporally aligned camera and hand controls. We construct joint camera--hand training data from heterogeneous egocentric videos through annotation completion, consistency-based filtering, and temporal alignment. Building on a pretrained camera-controlled world model, we combine spatial hand-skeleton conditioning with numerical hand-motion features through parameter-efficient adaptation. Causal adaptation and few-step distillation support chunk-level online generation, allowing users to update their controls based on generated visual feedback. We further introduce TouchéBench to evaluate hand-motion following, camera--hand coordination, online interaction, and object response. Together, these components establish a framework for studying continuous exploration and hand-driven interaction within generative worlds.

Attention is central to modern AI models for language, vision, and scientific modeling. However, direct evaluation compares every query with every key, requiring quadratic time in the context length, which limits its scalability to long contexts. Consequently, characterizing the complexity bounds of attention is of fundamental importance. For every fixed dimension $d$, we give an algorithm for $m$ queries, $n$ key-value pairs, and additive error $\varepsilon$, with expected time $ \widetilde{\mathrm{O}}\bigl((mn)^{k/(k+1)}+m+n\bigr),\qquad k=\lfloor d/2\rfloor. $ For fixed dimension and a fixed number of value channels, soft-O hides factors of the form $\operatorname{poly}_d\bigl(b,\log(m+n),\log B,\log(1/\varepsilon)\bigr),$ where $b$ bounds input encoding lengths and $B$ bounds coordinate and value magnitudes. In particular, the dependence on $B$ and $1/\varepsilon$ is polylogarithmic, which allows arbitrary positive temperature. When $m=n$ and $d\ge2$, the bound is $\widetilde{\mathrm{O}}(n^{2-2/(\lfloor d/2\rfloor+1)})$: near-linear in dimensions two and three, with exponent $4/3$ in dimensions four and five and $3/2$ in dimensions six and seven. It improves the previous $\widetilde{\mathrm{O}}(n^{2-1/d})$ upper bound and matches the best known batch maximum-inner-product bound in each fixed dimension. Additionally, we reduce integer batch maximum inner product computation to one attention call with the same dimension and input sizes $m,n$.

Reservoir computing provides an efficient framework for time-series prediction because only the readout layer is trained while the recurrent reservoir remains fixed. In this study, we examine how the negative sign ratio of a signed Erdos-Renyi reservoir affects the internal dynamics and forecasting performance of an echo state network on the Mackey-Glass benchmark. To characterize the reservoir responses, we use two Shannon-entropy-based observables, namely node-state entropy and node-frequency entropy, as empirical proxies for the diversity of reservoir states and temporal activation patterns. We find that intermediate negative sign ratios are associated with richer internal dynamics and substantially lower prediction error than the sign-homogeneous extremes for the task considered here. These results suggest that link-sign composition can be a useful structural parameter in reservoir design for chaotic forecasting. The present findings should be interpreted within the scope of the current benchmark and reservoir setting, and broader validation will be required to establish their generality across other tasks, topologies, and controlled spectral conditions.

Given a collection of $N$ graphs, graph-level clustering (GLC) seeks to partition the $N$ graphs into semantically meaningful groups, which serves as a fundamental task in analyzing graph-structured data. While Graph Isomorphism Networks (GINs) have become the de facto encoders for GLC, they still fail to distinguish graphs due to their neglect of global topology. On top of that, existing deep clustering frameworks suffer from severe "clustering collapse" issues, particularly in imbalanced multi-class scenarios. To address these challenges, we propose S$^3$GLC, a novel framework that integrates global topological characteristics with local neighborhood features for balanced GLC. Theoretically, we leverage the Hoffman-Wielandt inequality to construct Spectral Signature as a rigorous lower bound for graph edit distance (GED), providing a computationally efficient proxy for global topological features. We additionally propose a joint manifold-preserving sparse regression module to fuse Spectral Signature with Spatial Representation, eliminating noise while preserving the complementary information. Furthermore, to mitigate clustering collapse, we introduce a balanced clustering objective based on optimal transport, harnessing the Sinkhorn-Knopp normalization to enforce equipartition constraints on cluster assignments. Extensive experiments on 10 real-world datasets demonstrate that S$^3$GLC significantly outperforms 15 state-of-the-art baselines, achieving remarkable improvements on benchmark datasets such as REDDIT-12K and OGBG-PPA.

Personal agents draw on past conversations to tailor their responses to each user. Their memory systems store and update what each user says, and coding agents now also rewrite the harness that turns this memory into the model's context. Such patches risk overfitting, because one tested only on the failures it targets can fix those cases and help nowhere else. We present Reglos, a training-free personal memory agent that tests whether harness changes help beyond the failures they were written to fix. Its self-evolving memory adds to each user's append-only record with every conversation, as new preferences supersede old ones and facts can be forgotten. Its harness improves through recursive self-improvement, in which a coding agent proposes patches from failure traces and a critic from a second model family screens each one for benchmark gaming. After the harness is frozen, a fixed evaluator decides which further patches to accept. With the same backbone behind every system, Reglos outperforms state-of-the-art memory systems such as Mem0, MemOS and EverOS, reaching 61.3% and 84.0% on PersonaMem-v2 and LongMemEval-S with Qwen3.8-27B, and 57.5% and 87.2% with GPT-5.6 Luna.

Text-to-SVG systems are commonly evaluated by applying raster-image metrics to rasterised outputs, yet whether these scores capture fine-grained prompt faithfulness remains unclear. We introduce SVGFaith, a two-layer meta-evaluation framework in which the more faithful candidate of each pair is determined by construction or by human judgment. SVGFaith-Diagnostic contains 1,842 controlled contrastive pairs across eight capability axes, split evenly between two directions: two renderings compared against one prompt, and two descriptions compared against one rendering. SVGFaith-Arena contains 1,327 human-annotated comparisons of real generator outputs. On SVGFaith-Diagnostic, CLIP and three other image--text embedding metrics reach at least 0.789 pairwise accuracy on object identity but at most 0.672 on spatial relations and attribute binding, where chance is 0.5. We further introduce SVGFaith-Score, an interpretable pointwise metric. A domain-tuned decomposer turns each prompt into a dependency graph of atomic visual requirements, a visual question answering model scores whether each requirement is satisfied, and a requirement's score is discounted when its prerequisites score low. In addition to the overall score, the per-requirement scores show which requirements are judged unsatisfied and which inherit discounts from their prerequisites. SVGFaith-Score significantly outperforms all nine baselines, including a Qwen3.5-122B-A10B judge, on Arena and in both Diagnostic directions. Code and data: https://github.com/Yukinonooo/SVGFaith.

Search-augmented recommenders can use the latest search episode and earlier searches. We introduce FocalSwap to test how these inputs contribute to the next recommendation. For each recommendation, it changes the latest search slot or all search inputs while holding the target, candidates, pre-search history, and scorer fixed. It compares neutral masks with matched observed-search replacements, recording rank changes, RR gains relative to each control, and preference for the true focal search. A planted signal and GenSAR's current-query task confirm that the audit detects known input dependence. We evaluate behavioral and lexical evidence on KuaiSAR and Qilin, and recommendation checkpoints trained using the released SESRec and UniSAR implementations. Across three independently trained UniSAR checkpoints, the estimated all-source benefit depends on the control. On the same 66,139 rows, replacing all search inputs with a neutral ID yields a positive signed contribution, whereas replacing them with category/profile-matched observed search sequences gives smaller estimates, with an interval below zero for one checkpoint. The paired RR gaps range from $+2.770 \times 10^{-3}$ to $+6.604 \times 10^{-3}$, and every adjusted interval is above zero. Neutral masks change 59.7-61.7% of target ranks, while matched replacements change 44.1-44.8%. The control contrast persists across five alternative donor assignments per checkpoint. Under the primary neutral-mask and category-matched controls, positive latest-slot contribution is not established in SESRec or UniSAR. These results show why masking evidence should be compared with observed-search replacements when attributing a benefit to broader search state or the latest episode.

Streaming speech recognition is increasingly consumed by machines, such as translators and voice agents, that act on each token as it appears. A token that is later revised costs wasted work or a wrong action. In this work, we present Confucius4-R2T2 (short for Real Real-Time Transcription), a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency. R2T2 is obtained by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector. The adaptation is based on the concept of Longest Stable Prefix (LSP), a new notion that we introduce in this work. We prove that LSP is monotone under the prefix order, consistent with the full-audio output, and maximal among labels that equal a whole earlier output. Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change. Extensive evaluations show that R2T2 brings streaming recognition close to the base model's full-context quality while delivering stable text with low emission lag. Its gains extend beyond the post-training languages, demonstrating a practical path from full-context ASR to stable streaming. We open-source the model weights and code at https://huggingface.co/netease-youdao/Confucius4-R2T2 and https://github.com/netease-youdao/Confucius4-R2T2.

Generating daily activity schedules from individual attributes (e.g., age and income) is valuable for activity-based travel modeling and human behavior simulation. The task is challenging because exact activity structures follow a highly imbalanced, long-tailed distribution, while the same structure can admit multiple duration allocations under a fixed full-day budget. Direct generation must therefore learn many exact activity structures from highly uneven observations while modeling their possible temporal realizations. However, many exact structures share recurring activity organization. Based on this observation, we propose Act-HiMotif, a hierarchical motif-based generative model that introduces learned, reusable activity-structure motifs as an intermediate representation between individual profiles and specific activity structures. Act-HiMotif first distinguishes home-only days from active days, then samples a motif from a profile-conditioned prior to guide activity-structure generation, and finally applies a budget-constrained duration allocator to generate activity durations that sum exactly to the full-day horizon. This hierarchy models home-only routing, activity-structure generation, and duration allocation at their corresponding levels, while motifs allow related structures to share statistical support without losing their distinct sequences. Experiments on large-scale travel survey data show that Act-HiMotif outperforms strong baselines in both in-distribution and spatial generalization settings, demonstrating the effectiveness of hierarchical generation and reusable activity-structure motifs for producing realistic profile-conditioned daily activity schedules.

Anti-money laundering (AML) investigations are organized around accounts, while the suspicious evidence for an account may reside in only a few of its transactions. Yet many account-level detectors summarize transactions before learning, and message-passing models may aggregate once more over neighboring accounts. We name this discrepancy the aggregation mismatch and study it under an account-window benchmark label defined by the presence of at least one flagged transaction. We recast account-level AML detection as multiple instance learning, in which an account-window is a bag and its transactions are instances. We propose TRACK, which encodes every transaction with a shared encoder and selectively pools the encodings into an account decision. To emulate coarse supervision, we withhold the transaction labels of five AML datasets from TRACK throughout training. Under a time-ordered protocol, TRACK achieves the highest average precision among the tabular and graph baselines trained from account labels on every dataset. Despite using only account-window supervision, it remains competitive with edge classifiers trained directly from transaction labels. Its instance scores also rank flagged transactions well above chance and approach or exceed directly supervised ranking performance across datasets.

There is a fundamental limit to the best prediction error of any model for a given data distribution. In classification, the Bayes error quantifies this limit and serves as both a benchmark for trained models and a criterion to detect overfitting. The state-of-the-art Bayes error estimation techniques for binary classification use only soft labels (positive class probabilities) and require no input features or auxiliary training. We extend such notions to constrained classification problems, such as fairness. Fair classification is an important case of constrained or multi-objective classification. It requires a model not only to be accurate but also to yield fair outcomes (equal or near-equal true positive rates) across different demographic groups (e.g., race and gender). This leads to an inherent fairness-accuracy tradeoff or a Pareto frontier that is used to compare fair classifiers against each other. The focus of our work is to estimate the fundamental limit for this fairness-accuracy tradeoff. We derive an explicit expression for the fair Bayes error rate (i.e., best achievable error rate subject to given fairness constraints), and construct consistent, instance-free estimators for it using only soft labels and group information. We provably bound the bias and the sample complexity of our estimators. Empirically, our method produces a robust, low-variance estimate of the optimal fairness-accuracy tradeoff curve, even when the soft labels are noisy and derived from a classification model on hard labels.

When a language model chooses a number between 1 and 30, every integer in the range is valid, yet the model can favor particular answers. We use numerical choice to study how preference is processed across layers and tokens. Perturbation probing and gradient attribution nominate neurons for activation interventions. In a selected Qwen3-4B case, a pre-answer edit changes 17 to 15 while both runs first emit 1. After the edit ends, selective cache and attention-output replacements trace part of its continuing influence on the second digit. Replacing one attention head's output or the other heads' outputs jointly produces similar immediate readout contributions; both interventions restore 17 but have different final log-odds effects on 7 versus 5. Holding later multilayer perceptron (MLP) outputs at their values from the edited run without either replacement reduces this effect gap from 0.119 to 0.0025 nat. Later computation therefore changes how similar local contributions affect the continuation. Retrospective estimates across eighteen contexts capture much of the probability movement but miss the selected answer switch. Prompt tests also separate influence from answer changes: one held-out prompt retains influence without switching, and only one of twelve fresh wordings qualifies for the reader/MLP test; there, joint MLP clamping again reduces the gap between the two reader effects while the answer remains 17. These results use numerical choice as a controlled case for studying persistent state and conditional downstream processing; they do not establish the origin or a complete circuit of natural numerical preference.

Generative models trained on the output of earlier models lose knowledge across generations, rare knowledge first. The remedies that work all bring in information from outside the loop, and the theory of model collapse treats that supply as given: a fixed archive, drawn from a fixed distribution. We argue that the supply is the problem, and that cultural evolution has the theory of it. We read model collapse as Rogers' paradox: a population that comes to rely on copying ends up no better off than one that never copied, because copies go stale unless someone keeps learning for themselves. We test the identification in three steps. First, Rogers' model, fitted only to fully synthetic training, predicts to within a few percentage points how much rare knowledge language models, from word trigrams to a pretrained transformer, retain under other mixtures of fresh data. Second, when the world itself changes, a lineage that copies its own past needs more fresh data to keep up the faster the world changes, as Rogers' model says it must and as analyses that assume a fixed target cannot; a lineage that restarts every generation from fixed pretrained weights is largely insulated, because those weights are knowledge that is never copied. Third, we close the loop: when the people who write a model's training data can ask the model instead of finding things out, the supply of fresh data becomes endogenous. At equilibrium, the model is exactly as reliable as it is costly for people to find out for themselves, the population is no better off than without the model, and every increase in the share of model-written text must be paid for by more people learning for themselves. Users who check what they are told and relearn when it is wrong break the paradox. Filters that select plausible text are conformist and accelerate collapse; only checks that carry new information act as critical learning.

Attracting features toward learned in-distribution (ID) regions is an intuitive strategy for out-of-distribution (OOD) detection, but does such attraction improve detection? We investigate this question using class-conditioned Hamiltonian dynamics as a controlled testbed and identify an attraction-ranking mismatch: improving individual class-affinity scores need not improve the relative ordering of ID and OOD samples. Our analysis separates misalignment between the driving field and the detection score from ranking failure that persists even when they are aligned. An exact AUROC decomposition and a local boundary analysis identify relative score changes at the ID-OOD ranking boundary as the source of first-order AUROC change under regularity conditions. Experiments demonstrate that a score-aligned rule can improve performance on a held-out source validation split yet consistently degrade it after transfer, even without a single winning-class change. A separate perturbation study finds that near-boundary relative score changes track the direction of small AUROC changes in the tested settings. Finally, centroid approximation and shared scales connect the zero-step score to a static centroid family that includes the existing CTM detector. Together, these results explain why class attraction is insufficient to justify feature refinement and motivate evaluating its effect on the ranking boundary.

A cross-device test can complete every command yet judge a business claim from evidence belonging to another device or attempt. We introduce an executable evidence contract that makes support qualification an explicit stage before predicate evaluation. It binds each obligation to admissible subjects, episodes, objects and observation scopes, and records unresolved support rather than treating an apparently successful observation as proof. We evaluate two implemented parts of the contract. A prospectively specified, 12-seed controlled study isolates association under reused identifiers across 2,304 claims. Given identical records and business policy, Atomic-Facts Flat produces 160 false PASS and 76.30% PASS recall; supplying computed associations reduces false PASS to five and increases PASS recall to 99.09%. Deterministic Closure produces zero false PASS and 100.00% PASS recall, improving on Atomic Flat in 11 seeds with one tie (two-sided sign p=0.00098). Most of the benefit comes from explicit association rather than graph notation. A separate prospective field characterization spans four workflows, 40 physical episodes and 80 claims. Against frozen reference labels, the physical-evidence verifier attains 95.00% accuracy and 100.00% PASS recall (27/27), with three false PASS judgments. The field results are descriptive: a policy-input mismatch prevents the planned matched-policy comparison, and post-hoc sensitivity does not establish superiority over a policy-informed textual baseline. Together, the studies characterize qualification on physical evidence and isolate the value of explicit association under a controlled matching contract.

Expressiveness is central to speech-to-speech translation (S2ST), yet existing datasets rarely align source and target speech jointly in content, prosody, timing, and speaker identity. Consequently, current systems often rely on indirect conditioning or transfer mechanisms. To obtain this supervision, we introduce a training-data construction strategy that combines professional dubbing, which provides human-performed correspondence in content, prosody, and timing, with multilingual voice conversion to reduce speaker mismatch. We evaluate this supervision using existing expressive S2ST systems and DirectS2ST, an end-to-end method developed in this work that predicts the complete target-speech representation directly from source speech. Experiments on English-German and English-Spanish show that holistic parallel supervision improves expressive realization across all evaluated systems. DirectS2ST exhibits the largest gains in content, prosody, timing, speaker similarity, and human-rated expressiveness. To our knowledge, this work provides the first systematic investigation of holistic parallel supervision for expressive S2ST.

Hundreds of languages have written text but almost no translations, and whether a model learning one from untranslated text builds on a related language or familiar letters is unclear: they usually coincide. Here we show that familiar letters, not a related language, let an unseen language into a shared space of meaning. GlyphMind, a tokenizer-free vision transformer that reads text as pixels and lets any sentence be shown in any script, was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, a language and script it had never seen, it retrieved English translations seven times as often as the strongest simple baseline. Rewritten in unfamiliar letters, Lao did not enter although its close relative Thai was in pretraining; Santali, which had none, entered in Devanagari, familiar from Hindi, but not in its own script. Transliteration into familiar scripts may thus open multilingual models to languages without translations.

We view a $k\times n$ Latin rectangle as $n$ points in the grid $[n]^k$. Each symbol gives one point, whose coordinates are the columns it occupies. The rectangle is balanced when these points are pairwise equidistant in the $\ell_1$ metric. Its imbalance measures the distance from this ideal. We study the minimum imbalance from three sides. Constructively, we use a simulated annealing that moves symbols within rows and only penalises column clashes. Against the reference table of Voboril et al. (2025), it improves 33 of the 51 non-square values for $8\le n\le 13$, by up to 39.4%, and matches the rest. A cheap invariant, the row group, certifies that none of the 33 new rectangles is isotopic to rows of a group table, so the earlier column-swapping local search cannot reach any of them. Exactly, a computer-assisted branch and bound proves that the minimum imbalance of a $3\times 9$ rectangle is $56$. Díaz et al. found this value, but its optimality was open. A sharper pruning bound, derived from the identity $I=2\sum(d-t)^+$, reduces the search tree by a factor of $2.4$ to $3.3$ on the instances with $k\ge3$ that we solve. Asymptotically, the minimum imbalance is $Θ(n^3)$ for every fixed $k$. A pigeonhole argument gives a lower constant of order $k\,4^{-k}$, and random rectangles give an upper constant of order $\sqrt k$. For $k=2$, the limit of $\min I/n^3$ exists, a permuton attains it, and it lies between $4/375$ and $0.1196$. Perfect balance is an $\ell_1$-equilateral problem, so it forces $k=Ω(n/\log n)$ rows. For $2\times n$ rectangles, single two-minute annealing runs improve five of the seven six-hour MIP upper bounds for $14\le n\le 20$. Finally, the Lean 4 kernel checks every rectangle reported in this paper against a 74-line specification that imports only Lean's core library. We also prove the pigeonhole lower bound in Lean with Mathlib.

Activation steering can modify language-model behavior without parameter updates, but behavioral efficacy alone does not determine whether an intervention is usable. Contrastive Activation Addition (CAA) steers targeted behaviors using activation directions derived from contrasting examples, potentially degrading unrelated model capabilities in the process. We therefore study CAA through a behavior-capability frontier, selecting interventions that maximize behavioral gain subject to an explicit bound on capability degradation. We evaluate this framework across four instruction-tuned models, seven behaviors, and capability-loss budgets of 1%, 3%, and 5% (absolute percentage points). Validation-selected behavioral effects generalize strongly to held-out data ($ρ= 0.945$), while joint behavior-capability survival among nonzero operating points rises from 37.5% at the 1% budget to 72.7% at 3%, with limited additional improvement at 5%. Control experiments indicate that these effects depend on learned behavior directions rather than generic activation perturbations. Representational geometry is significantly but only modestly predictive of causal cross-behavior transfer. Finally, secondary capability, out-of-distribution, and open-ended evaluations show that interventions satisfying protected benchmarks can still degrade untested capabilities or transfer unevenly to free-form generation. These results characterize activation steering as a constrained intervention whose usable range depends on model, behavior, direction, layer, and evaluation budget, rather than a universal control knob.

Designing systems that attribute a generated image to its source model has lately become a major area of AI research. Most existing datasets and systems related to image source attribution are targeted towards the detection of architectural signatures of the source models. Distinct from this, focusing on stable diffusion, we introduce a novel task, IntrAA-SD, Intra-Architecture source Attribution for Stable Diffusion, where the source models to be differentiated share the same architectural backbone and differ only due to fine-tuning. The proposed task aligns with the emergence of fine-tuned generators having the same architecture, like Juggernaut XL and RealVis XL, which are fine-tuned versions of Stable Diffusion XL (SDXL). We provide a first-of-its-kind $140$K-image benchmark dataset containing images generated from different fine-tunes of SDXL, along with in-domain and out-of-distribution prompts. We also provide a baseline framework for IntrAA-SD, which leverages a style-conditioned Mixture-of-Experts network. It uses a frozen CLIP backbone to capture the image style and gate the MoE, which processes fine-tuned trace features from a high-resolution LoRA-adapted DINOv2 network. The fine-tuned models used, the prompting strategies considered, and the formation details employed in the novel dataset are elaborated. Our baseline intra-architecture source attribution method is evaluated in terms of attribution accuracy on our dataset, and ablation studies are also performed. Code and dataset is available at https://github.com/mallikagarg/IntrAA-SD.

Helping users discover new interests is vital to the long-term health of industrial recommender systems, yet the usual remedy, boosting exposure of under-explored (discovery) content, often backfires in production. We report URED (Unexposed Recycling for Efficient Discovery), a training-free recall-stage system serving Taobao feed recommendation at Alibaba for over six months to tens of millions of daily active users. It rests on a production-validated insight: naive boosting fails because it injects cold, information-poor candidates that rankers systematically over-estimate, buying no discovery gain while risking engagement. Items scored highly but never exposed (signals industrial pipelines routinely discard) form a pool that is both relevant and reliably predictable. Over a 30-day horizon URED aggregates them and re-injects at recall time through a lightweight pipeline spanning offline storage, distributed aggregation, and low-latency online serving, adding under 1% latency via a single O(1) lookup per request with no model retraining. Online A/B tests show +3.47% discovered categories and +8.47% content richness with engagement preserved: completion video views rise significantly (+2.62%) and total views move positive (+0.36%, not significant). Ablations confirm both design pillars are necessary: a one-day window collapses the discovery gain to +0.55% (not significant), and random injection without score filtering yields no gain while trending negative on engagement (-0.17%). Across production recall channels, URED reaches 13.9x the platform-average discovery rate (6.3x the strongest nearline baseline) at engagement quality on par with the platform average. We share deployment details and lessons for practitioners facing the exploration-exploitation dilemma under tight resource constraints.

Federated Prompt Learning (FPL) enables parameter-efficient adaptation of pre-trained Vision-Language Models across decentralized clients without exchanging raw data. To cope with cross-client heterogeneity, existing FPL methods typically pair a globally shared prompt with a client-specific local prompt. However, this design personalizes only at the client level: the same local prompt is applied to all samples of a client, and the global and local prompts are combined in an input-agnostic, fixed manner. Consequently, it overlooks the substantial visual variations among samples within the same client and cannot adjust the balance between shared and personalized knowledge to each input. To address these limitations, we propose Federated Prompt Learning with Instance-Aware Cross-Modal Representations (FedIA), an FPL framework that learns instance-aware cross-modal representations. Specifically, FedIA first uses global text prompts as client-invariant semantic anchors to select image patches most relevant to each candidate class, suppressing distracting regions. The selected visual evidence then conditions the generation of instance-aware local text representations. Finally, an adaptive global-local fusion module calibrates the contributions of local and global text representations for each image-class pair. In this way, FedIA shifts personalization in FPL from the client level to the instance level while preserving the semantic knowledge shared across clients. Extensive experiments under domain shift, pathological label skew, and Dirichlet-based label skew show that FedIA outperforms FPL baselines, with maximum accuracy gains of 1.44\%-4.37\% over the latest baseline across the three settings. Ablation studies and analyses further verify the complementary contributions of instance-aware representation generation and adaptive global-local fusion.

Time series forecasting is critical for demand prediction and resource management in modern Web services, yet applying pretrained Time Series Foundation Models (TSFMs) to downstream tasks remains challenging due to domain shifts. Existing retrieval-augmented approaches typically retrieve coarse-grained window contexts and may introduce context-specific noise. To address this problem, we argue that temporal patterns should be reused at a decomposition level, as effective trend and seasonal dynamics may originate from different contexts. We propose \method, a two-stage \underline{F}ine-grained \underline{K}nowledge \underline{A}ugmentation framework that decouples and consolidates pre-training knowledge into compact, task-shared residual codebooks to enhance frozen TSFMs. Specifically, Stage I constructs dual-branch residual codebooks for decoupled trend and seasonal signals, while Stage II dynamically retrieves and merges these fine-grained patterns into target representations via task-adaptive aggregation and gate fusion. Extensive experiments across eight real-world benchmarks with four representative TSFM backbones demonstrate the superior effectiveness and computational efficiency of \method, while further validating the strong transferability of the residual codebooks across different backbones. The source code is available at the anonymous repository: https://github.com/yuppielqx/TSFM-Component-Token-Adapter

Video editors often work with large collections of fragmented footage and must select and organize relevant clips into a coherent story. However, many existing video editing and summarization methods assume a temporally ordered source video and aim to preserve its original narrative and temporal structure. In this paper, we introduce Clips2Story, a training-free, LLM-based pipeline that enables users to create keyword-specified short videos from a pool of video clips. Given a pool of clips and a user-specified keyword, we first extract text-based multimodal descriptors for each clip and retrieve the clips most relevant to the keyword. We then provide the textual descriptors of these candidate clips to an LLM, which selects and arranges them into an editable storyboard with connective narrations. Finally, we assemble the selected clips and retrieve accompanying visuals for the narrations to compose the output video. To evaluate the proposed pipeline, we construct a synthetic evaluation dataset from 100 long-form videos across five genres by segmenting each video into short clips using a scene-change detection model and shuffling the resulting clips to remove their temporal order. Through objective evaluation and qualitative examples, we demonstrate that Clips2Story can effectively retrieve keyword-relevant clips and compose coherent storyboards based on user-specified keywords. Human evaluation further shows that Clips2Story outperforms a strong baseline in terms of narrative coherence, keyword alignment, and overall realism.

Multi-agent LLM applications repeat most of their prompt tokens across requests, and serving engines reuse the cached state of shared prefixes. vLLM and SGLang reuse it at different granularities, and benchmarks of their default deployments report large performance gaps, usually credited to the cache. Such comparisons rarely control for other configuration differences. One of them, whether a long prompt is prefilled in chunks, is set by each engine's defaults from the hardware and the release, and it confounds the comparison. We built a reproducible testbed on Kubernetes and a seeded multi-agent workload generator, and compared the two engines under vendor defaults and with chunked prefill matched. SGLang's configuration is the same in both arms, which makes it an internal control across independently provisioned deployments. Under defaults, vLLM's median time to first token is up to 64x SGLang's. Matching one setting removes 78% of the capacity gap where that gap is largest and 30 to 67% of the latency gap. Near saturation the remaining latency gap reflects queue growth. A fluid queue model driven only by a 4.4% throughput shortfall reproduces it, and shows that such a gap scales with the length of the benchmark run. What survives matching is a tail-latency advantage of 2.4 to 6.4x at equal throughput, consistent with cache-aware ordering of waiting requests. We derive six recommendations for comparing serving engines.

Deep learning over relational databases has led to increasingly effective prediction models, motivating a growing need to explain their behavior. We study explanation methods that assign importance scores to individual database attributes, focusing on a fundamental question: to what extent does the choice of importance-attribution method matter? We develop a methodology to systematically compare importance-attribution methods and analyze their similarities and differences. In particular, we study two aspects of the explanations: which attributes they deem important, captured by differences in attribute rankings, and how importance scores vary across attributes, captured by differences in score distributions. We further investigate whether these differences are reflected in the conventional measure of explanation quality, fidelity. We apply our methodology in an extensive empirical study of four methods: Monte Carlo Shapley, Kernel SHAP, single-attribute impact, and mask learning. We experiment with multiple datasets and tasks using two RDL architectures: heterogeneous GraphSAGE and Relational Graph Transformers. Our experiments show that explanation methods behave differently depending on the prediction type: regression or classification. In regression, most methods exhibit high agreement in attribute rankings, whereas classification yields considerably less consensus. Notably, while mask learning frequently diverges in its importance distributions, Shapley-based methods and single-attribute impact often converge to similar explanations. This suggests that, in many settings, single-attribute impact can replace Shapley at substantially lower computational cost. Overall, our study provides methodological tools for comparing importance-attribution methods, and empirical evidence that their choice has important consequences for interpreting relational deep learning models.

Hyper-Connections (HC) architectures are increasingly adopted in frontier language models, yet their components remain insufficiently understood and introduce non-negligible training overhead. In this work, we systematically study the HC architecture to understand its components and identify simpler, more efficient designs. We find that learned read/write gating without mixing and unconstrained mixing with fixed gates form two effective interaction groups: either alone achieves performance comparable to mHC. We also identify repeated access to the expanded residual state as a major source of training overhead, even without mixing. These findings lead to SimpleHC, which simplifies the architecture by removing residual mixing and simplifies execution through shared memory access, using the resulting additive structure and shared gate inputs. Layer-specific gates share a group-entry state, enabling batched reads and deferred writes; scalar corrections exactly preserve preceding sublayer output contributions. Our analyses support both the retained residual structure and the design of shared memory access. SimpleHC incurs only 2.4% extra end-to-end training time over a vanilla residual baseline at the largest tested scale. It maintains performance comparable to mHC when scaling from 15B to 30B MoE models, exhibits smoother gradient dynamics during standard training, and remains stable under stress. We hope SimpleHC will encourage a broader pursuit of simplicity in foundation model design, making performance gains more accessible and practical at scale.

Channel sounding in Wi-Fi multiple-input multiple-output (MIMO) systems must balance channel state information (CSI) accuracy against CSI acquisition overhead. Device mobility causes channel aging, while autoregressive CSI prediction introduces errors that accumulate over sounding intervals. However, standardized statistical channel models cannot fully capture the environment-specific channel dynamics encountered in physical artificial intelligence (AI) applications, motivating the use of Sionna-based ray tracing to evaluate CSI prediction under device mobility and evolving propagation geometry. This paper develops a dynamic sounding framework that integrates autoregressive CSI prediction with signal-to-interference-plus-noise ratio (SINR)-constrained interval selection. A Transformer with rotary position embedding that converts device positions into CSI sequences predicts CSI across multiple subcarriers for beamforming updates between sounding exchanges. We formulate sounding-interval selection as a predicted average-throughput maximization problem under per-stream SINR lower-bound constraints, including CSI acquisition and processing overhead. A prediction-error perturbation analysis provides the SINR bounds, and cumulative-sum enumeration globally solves the finite single-variable problem. Actual SINR feasibility holds whenever the prediction errors satisfy the bounds. Ray-tracing experiments assess channel prediction, trajectory transfer, and autoregressive prediction errors. In single-STA Wi-Fi 7 link-level simulations, predicted-CSI beamforming improves mean goodput by 67.3%, 112.7%, and 119.1% over stale CSI for linear, circular, and cosine trajectories, respectively.

Pedestrians increasingly encounter robots in urban public spaces without interacting with them, but how robots change the perceived safety of a street remains unclear. We examine how robot presence and type affect perceived safety across urban contexts. We generate 18 images of Seoul streets that vary in time of day, road class, and robot condition (no robot, a delivery robot, or a humanoid). Two multimodal large language models (MLLMs), Qwen3.8-27B and Gemma4-26B, each conditioned on 500 synthetic personas, choose the safer image in each pair, and 25 human respondents complete the same task. Both models prefer streets without robots and favor the delivery robot over the humanoid. Robots are chosen more often at night and less often by older personas, whereas road class has smaller effects. In the model rationales, the delivery robot appears as a physical obstacle, whereas the humanoid appears as either a threat or a source of protection. Like humans, both models chose robots more often at night, but chose the humanoid over the delivery robot far less often, and Gemma showed a night effect for the humanoid about three times that of humans. These results show that the effect of a robot on perceived safety depends on when it appears, and that persona-conditioned MLLMs capture only part of human judgments.

Graphs play an increasingly important role in large language model (LLM) systems, supporting structural reasoning, knowledge grounding, retrieval, and memory. As these systems evolve toward the agentic era, graphs also organize execution dependencies and multi-agent communication, linking intermediate observations and retained information to subsequent decisions. However, understanding their contribution requires distinguishing the function of each graph operation, the component responsible for its computation, and the subsequent use of its output. This survey introduces an operation-level role--locus taxonomy for reasoning, knowledge and evidence, retrieval and memory, execution, and coordination. We compare how graph observations and external computations guide decisions, how memory representations and updates support later tasks, and how dependencies constrain execution and information delivery. These comparisons distinguish improvements in local metrics from downstream task performance while retaining the conditions under which results were reported. We then organize benchmarks, comparison designs, process diagnostics, and system costs into an evaluation framework. Finally, we identify open challenges in adaptive computation and stopping, memory maintenance, and dependency-aware recovery. The survey provides a basis for assessing when graph mechanisms are useful and for evaluating their effects across components and stages.

RL for LLM reasoning often reaches an early performance plateau as the policy distribution becomes increasingly sharp. Recent work has studied policy entropy as both a diagnostic and a control signal for this behavior. In our experiments, however, increasing entropy alone does not reliably change the eventual plateau. We analyze RL through its implicit target policy and derive a sharpness factor $κ$, together with an effective counterpart $κ_{\mathrm{eff}}$, that describes how strongly reward or advantage gaps are translated into probability-ratio gaps. Motivated by this view, we introduce \emph{reward budgeting}, a plug-in advantage construction that assigns a fixed total reward budget to each prompt. For binary rewards and independent on-policy rollouts, its unclipped score-function estimator follows the gradient of an exact finite-$G$ concave objective with diminishing returns, while its advantage gap decreases as more rollouts become correct. For general rewards, a shifted log-mean interpretation holds asymptotically under the stated regularity conditions. These results motivate, rather than establish exact objective equivalence for, the clipped implementations. On Qwen3-8B and Llama3.2-3B-Instruct, reward budgeting consistently improves a range of RL baselines, with gains up to +12.50 on AMC23 and +10.00 on AIME24 and AIME25. Overall, our results support an update-dynamics view of premature convergence and show that strong gains can be obtained without explicitly preserving entropy.

Generalized Additive Models with pairwise interactions (GA$^2$Ms) are regarded as glass-box models, since each pairwise shape function can be visualized as a contour plot. In practice, the shape functions learned by methods such as EBM or NODE-GAM are so complex (irregular contours and sharp local fluctuations) that a practitioner cannot meaningfully interpret them or check them against domain knowledge. Auditability in this stronger sense requires substantially simpler shape functions. We argue that such simplification can match the accuracy of standard GAMs if the right feature pairs are captured. We propose Bivariate Additive Models (BAMs), a class of GA$^2$Ms in which every pairwise shape function is constrained to a simple parametric form, instantiated as an asymmetric bivariate Gaussian. Each component is visualizable as a unimodal contour map, and its feature pair is selected automatically during training. Experiments on six tabular benchmarks show that BAM attains predictive accuracy comparable to standard GAMs and approaches that of GA$^2$Ms. We also show the trade-off between accuracy and the number of components to inspect.

Automatic modulation classification (AMC) underpins spectrum sensing and non-cooperative signal identification in cognitive radio, yet the two standard CPU-feasible approaches to it, a compact convolutional neural network (CNN) on raw in-phase and quadrature (I/Q) samples and a classical higher-order-cumulant feature classifier, are rarely compared against each other conditioned on signal-to-noise ratio (SNR). We show that they are, in fact, complementary: the CNN is markedly more robust at low SNR, where hand-crafted statistics collapse toward chance, while the cumulant classifier is more accurate at high SNR, where higher-order statistics cleanly separate dense constellations. We exploit this complementarity with FuseAMC, a lightweight decision-fusion model that trains a logistic-regression meta-combiner over the concatenated class-posteriors of an approximately 198k-parameter residual CNN and a gradient-boosted-tree cumulant classifier, on a validation split disjoint from test. On RadioML 2016.10a (5 seeds, 44k-frame held-out test set), FuseAMC reaches 55.8% overall accuracy, improving on the best cumulant-feature baseline by 5.1 points and on the CNN alone by 3.2 points (paired-bootstrap 95% confidence-interval lower bounds of 5.0 and 3.1 points, both positive), while matching the best classical baseline in its own strong high-SNR regime and the CNN in its strong low-SNR regime. With about 0.2 million parameters, a model footprint of a few megabytes, and roughly 3 MFLOPs per frame, the whole pipeline trains and runs on one CPU core, targeting general-purpose and embedded processors without hardware acceleration.

Off-policy deep reinforcement learning trains on random samples from a replay buffer of past experience. In physical robot settings, that buffer is dominated by routine: thousands of near-identical transitions from normal operation, while the rare events that define the edges of the task are rarely sampled. We propose scoring every stored transition by its statistical leverage, which measures how unusual it is relative to the buffer's own data, computed from the stored numbers alone, with no labels, rewards, or value network. We draw each mini-batch from a mixture of this score and uniform sampling. First, the natural way to build the score along the directions where the data varies most turns out to be blind when actions are stored as one-hot indicators, because indicator columns always win the variance ranking; we instead select the feature columns with the heaviest tails (largest kurtosis) where small far-away populations are revealed. Second, the importance weights that prioritized replay uses to keep gradients unbiased cancel the method exactly: with full correction, training matches uniform sampling, so we drop the correction and deliberately optimize the rebalanced objective. Third, when the rare situations of a log are relational, visible only in how a stretch of behavior relates to its goal, not in any single stored coordinate-per-transition scores are blind; applying the same tail rule to short trajectory windows described relative to their goal restores the signal, and on a three-joint arm that must push an object around an obstacle it raises the rare-task success from 0.49 to 0.69 against a labeled-oracle ceiling of 0.74.

Mixture-of-Experts (MoE) enables multimodal large language models (MLLMs) to scale capacity efficiently, but expert parameters dominate memory even though only a few experts are activated per token: they account for over 90% of model size in modern MoE-MLLMs (e.g., 54 GB out of about 58 GB in BF16 for InternVL3.5-30B-A3B and Qwen3-VL-30B-A3B). Expert pruning has emerged as an effective compression strategy, but existing methods aggregate importance uniformly across tokens, ignoring that the same expert may serve different roles for vision versus language processing. To address this, we propose MAEP (Modality-Aware Expert Pruning), a training-free framework that decomposes expert importance by modality in a single forward pass. MAEP introduces (1) Delta-H Output ($ΔH$), a redistribution-aware metric measuring the MoE output change when an expert is removed, and (2) low-attention vision tokens as a vision-side pruning signal that is dataset-stable: its expert-importance ranking remains stable across different datasets. Text-side importance serves as a preservation signal while vision-side importance serves as a pruning signal, and experts are ranked globally across layers. Experiments on three MoE-based MLLMs across multimodal and text benchmarks show that MAEP outperforms modality-agnostic baselines on the majority of multimodal benchmarks while remaining competitive on text-only tasks, reducing the BF16 expert weight footprint by about 27 GB (about 47% of total memory) so that the full weight footprint of an approximately 58 GB model fits within a single 40 GB GPU. Code is available at https://github.com/nota-github/MAEP.

Diffusion drafters accelerate speculative decoding by denoising multiple draft tokens in parallel. Their training uses anchor sampling to construct compact draft queries and performs cross-attention with a block-sparse mask over a much longer context. We empirically show that acceptance length remains stable as context scales with a fixed query budget, while increasing the query budget yields diminishing quality gains, jointly establishing a practical regime with queries far shorter than the context. No existing context parallelism (CP) method jointly exploits this query--context asymmetry and meets the nonlinear, sample-dependent load-balancing requirement created by anchor sampling. We present Split-KV, an exact CP method that keeps key/value shards local and limits per-layer attention communication to query-sized tensors, making this communication independent of context length. Its global query view further enables a cost-aware longest-processing-time-first (LPT) strategy that balances context key/value chunks by their realized attention cost. At a CP size of 64, Split-KV retains 85--91\% weak-scaling efficiency and achieves up to $3.67\times$ speedup over the fastest available CP baseline. LPT achieves $<$1\% imbalance and plans three orders of magnitude faster than the best load-balancing baseline. Code is available at https://github.com/zxgx/verl.

Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying. However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer. This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors. In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control. We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment. Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features. These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense). Building on this, we selectively edit expert router scores during inference to curb overthinking. Across four models and four benchmarks, this intervention reduces the number of generated tokens by a median of 30.4% (and up to 63.3%). Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps. Accuracy never drops by more than 0.5 percentage points, and frequently improves. Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.

Merging companies that each run Salesforce must consolidate their customer data into a single organization while both source systems remain in use. This requires reconciling different schemas and picklists, resolving duplicate customers across and within organizations, preserving relationships between records, and keeping the target current until cutover. This article proposes a reference framework for an autonomous extract-transform-load pipeline for this task. It combines partitioned bulk extraction, declarative mapping templates, rule-based and statistical quality gates, master data management with deterministic and blocked fuzzy matching, constrained loading with per-record error capture, and change data capture guarded by a data-health circuit breaker. Each stage is mapped to Jitterbit Integration Studio constructs. The framework was implemented as an open reference prototype and evaluated on synthetic two-organization scenarios with known ground truth at three scales, from 32,040 to 383,973 source records, over five random seeds. At a preset threshold, hybrid matching reached pairwise F1 scores of 0.918 for accounts and 0.885 for contacts; calibrating the threshold per object would raise these to 0.938 and 0.963. The quality gates prevented all load failures and all silent loss of unmapped picklist values without wrongly flagging any clean record, whereas removing them produced about 1,000 load failures per run. Over five simulated days of hourly incremental synchronization, replayed events were ignored, an injected upstream fault was contained in the batch in which it occurred, and the target agreed with a full rebuild on more than 99.3 percent of field values. Validation on a production deployment remains future work.

We study the minimum-disagreement cluster editing problem that, from a complete edge-signed graph, where a positive edge represents two vertices agreeing to each other while a negative represents a disagreement, constructs disjoint cliques of upper bounded size such that the total number of negative edges inside cliques and positive edges outside cliques is minimized. When the sizes of the candidate cliques are unbounded, the problem is NP-hard and APX-hard, and many approximation algorithms with improving ratios have been proposed in the literature and the state-of-the-art approximation ratio of $α= 1.3865 + ε$. The capacitated variant, in which the candidate cliques have sizes no greater than a given $k$, is less studied and the best approximation ratio is $5.37$. In this paper, we study the latter problem. We first present a downsizing procedure that achieves an approximation ratio of $α_k = 2 α+ 1$, based on any $α$-approximation algorithm for the uncapacitated problem. Secondly, we propose a hybrid of cluster-LP sampling and pivoting and show that for every fractional capacitated cluster-LP solution of cost $L$, the hybrid using ordinary pivoting has an expected cost at most $\frac 95 L + \frac 1{10} \OPT_k$, and the hybrid using conditional pivoting has an expected cost at most $\frac {306}{203} L + \frac {50}{203} \OPT_k$. A cardinality-constrained adaptation of the approximate dual-separation framework of García-Soriano and Schohn supplies the required fractional solution and yields a randomized $(1.9 + ε)$-approximation algorithm and a randomized $(1.7537 + ε)$-approximation algorithm, respectively.

Planning electricity systems requires repeated decisions that connect regional resources, investment, operation and network constraints. Here we present an autonomous planning framework that carries prepared regional scenarios through proposal, numerical evaluation, feedback interpretation, redesign and final plan selection. Agents propose investment bounds, storage operating conditions and feeder interventions, interpret numerical feedback, and decide how to revise designs and when to stop. The framework automatically executes annual investment-and-dispatch optimization and representative-feeder verification, applies fixed acceptance criteria and retains the lowest-cost accepted candidate. We evaluate forty carbon-constrained trajectories across ten system-year scenarios in five electricity systems and four countries, using two model deployments and single- and multi-agent architectures. All forty retain candidates satisfying the implemented physical and fixed annual operational carbon-budget checks. Twenty-eight of forty matched reference designs exceed the new scenario ceilings. A separate archive of ninety-six trajectories under a common 1,000 Mt annual ceiling provides matched references and broader regional coverage. The results demonstrate autonomous planning under explicit emissions constraints using public regional data and published institutional demand outlooks.

Saliency maps, offered as evidence that mammography classifiers use the cancer, are untested in dense breasts, which can hide cancers. We tested whether they locate cancers worse there, and why. On 1,345 malignant CBIS-DDSM views, we scored three CNNs' Grad-CAM, Integrated Gradients and occlusion maps by whether their peak hits the lesion and a chance-normalized lesion-to-background saliency ratio (bits; +1 doubles it). Unadjusted gaps were large: Grad-CAM peaks hit 44% of dense vs. 58% of non-dense lesions. Case mix, chiefly more calcifications in dense breasts, explained about half or more of each gap. Occlusion kept a significant average of -0.56 bit gap. Controls also implicate the classifiers: unadjusted gaps exceeded an image-only conspicuity map's, and models relied less on dense-breast lesions. We recommend reporting density gaps before and after case-mix adjustment, per classifier, with conspicuity and lesion-reliance controls.

Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys rank their capabilities on a ladder from temporal prediction to counterfactual rollouts. Position on that ladder is currently assigned by reading papers rather than by measurement. We turn the counterfactual rung into a measured quantity. From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov we construct 9,996 audited within-trial contrasts from 7,207 trials, and we score a simulator on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld. Interventional skill is the error the full model removes relative to that reference. On a temporal split (878 test contrasts from 655 trials) a registry-trained gradient-boosting simulator removes a fifth of the error of a trivial baseline, but its interventional skill is -0.12 (95% CI -0.21 to -0.05): withholding the intervention makes it better, and no alternative featurisation, stratum, temporal cut or same-molecule retrieval yields an interval above zero. A frontier language model reading only the registered protocol has interventional skill 0.20 [0.08, 0.31]; smaller and open-weight models sit at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts. The blind reference is itself not drug-blind (a language model recovers the withheld drug class 45% of the time from the blind prompt), so interventional skill is a conservative lower bound. A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift. Benchmark, code, prompts and per-call model outputs are released.

As sequence lengths continue to grow, distributed attention based on sequence parallel (e.g., Ring-Attention) has become a mainstream solution for large language models (LLMs). Nevertheless, inherent communication dependencies limit computation-communication overlap, reducing GPU utilization. Moreover, in cross-hardware (NUMA or node) architectures, due to imbalanced communication bandwidth, frequent communication synchronization will limit overall performance. Based on these insights, we propose CHARM+, Cross-Hardware Attention with Re-Merge Multistream Mechanism. We reorganize sequential communications into several levels and achieve intra-level parallelism and computation overlap. Furthermore, we decouple cross-hardware communication from intra-hardware communication, eliminating synchronization overhead caused by bandwidth imbalance. Forward and backward algorithms were evaluated on multiple hardware architectures and consistently outperformed state-of-the-art methods. Compared to Ring-Attention, our forward and backward algorithms achieve average speedups of 1.3x and 1.7x on a dual-NUMA PCIe architecture, and 2.8x and 2.9x on a dual-node NVLink architecture, respectively.

KV cache eviction is not merely a token-ranking problem: the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations and reinforce an incorrect final eviction decision, particularly when evidence is dispersed across a long context. We present EchoKV, which discovers and addresses this failure mode with iterative correction. It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps. It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair. On LongBench (16 tasks, Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct), EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. Across budgets and backbones, EchoKV selectively improves evidence-intensive tasks while maintaining comparable performance on the remaining LongBench subsets. These results identify query--workspace coupling as a practical failure mode of query-driven KV eviction.

Developers insert logging statements in source code to record its execution, and the resulting logs are essential for detecting and troubleshooting system anomalies. A key part of these statements is the log level (e.g., DEBUG, INFO, WARN, ERROR), which determines whether a message is recorded. Choosing it is a trade-off: levels set too low mask important information in production, while levels set too high create noisy logs and extra I/O and storage overhead. Recent work uses Large Language Models (LLMs) with In-Context Learning (ICL) to predict log levels, but typically augments the prompt with randomly sampled examples. We argue that this overlooks the socio-technical nature of logging, where conventions vary by component and development team. Off-the-shelf LLMs using random examples are reported to achieve median AUCs of only 0.56 to 0.76, leaving room for improvement without fine-tuning. In this paper, we propose a retrieval approach for LLM-powered log level prediction that prioritizes context relevance over random selection and basic similarity. Our method retrieves logging examples relevant along two socio-technical dimensions: (i) functionality (code performing similar operations) and (ii) ownership (code maintained by the same developers). Evaluating four LLMs across ten popular Java projects, we find that this retrieval improves AUC over random baselines by 0.07 (median), and combining both signals yields median AUCs of 0.83 to 0.97. Our ablation studies show that each signal improves on random retrieval, and that combining them outperforms the stronger of the two on all ten projects. Our results also significantly exceed state-of-the-art supervised predictors (e.g., DeepLV) by a median AUC of 0.12 to 0.19. Our findings highlight the importance of socio-technical signals such as code functionality and ownership for LLM-powered log level prediction.

Forecasting emotion in conversation is increasingly framed as predicting the next emotional state's value. We show this targets a hard quantity in a state-observed testbed and propose a learnable, deployable reformulation instead. On valence-arousal-dominance trajectories from a dyadic corpus, repeating a speaker's own previous state sets a persistence baseline that no frozen-encoder correction meaningfully improves upon and a state-space filter exceeds; the floor requires the historical state to be observed, since an estimated anchor collapses persistence below a direct forecaster. What survives is the timing of a departure from persistence, learnable on three of four corpora and near chance on the highest-base-rate corpus, a negative result we report. We recast the task as selective departure forecasting: a lead-time-weighted metric with a leakage protocol, and commitment policies whose best choice switches with the departure rate. A dialogue-level conformal risk control bounds the expected loss of committed predictions, meeting the target in all twelve configurations tested, though the guarantee is bought through abstention, with coverage and alarm recall falling sharply on the hardest corpus. One tested language model's confidence fails as a ranking signal where another succeeds, a model-dependent failure not yet attributable to family versus scale. Code and result logs are available from the corresponding author upon request.

Comparisons in conversational emotion-shift forecasting are easily distorted when a learned model receives rich features but its transition baseline is given a noisy or unavailable current-emotion label, when hyperparameter budgets differ, or when target-turn information enters the representation. We present an information-matched, leakage-safe auditing protocol that closes these gaps with an explicit commitment point, executable causality tests, equal four-configuration validation searches, and seed-dialogue-aware inference. Applying it to six causal model families on four text corpora and two feature-source sensitivity configurations shows that a deployable train-only current-emotion classifier feeding the transition baseline yields a different verdict than an oracle diagnostic where both sides receive gold current emotion. Under the deployable condition, learned models improve over the predicted-label baseline on all four primary corpora; DailyDialog gains are large and survive a 72-comparison Holm correction (best ROC-AUC gain = +0.167, adjusted p=.0144). Under matched gold information, the baseline remains competitive on IEMOCAP and stronger on MELD, while every model retains a large DailyDialog advantage. We additionally report precision-recall AUC, calibration, a strictly utterance-local feature control, same-person self-shift forecasting, and a future-access dose-response test. An eight-item leakage checklist accompanies the protocol. Code, raw seed predictions, and reporting artifacts are available from the corresponding author upon request.

Recycling a target model's own predictions offers a training-free source of proposals for speculative decoding. In vision-language models, however, existing recycling indexes predictions by the last token alone, conflating predictions computed under different images and prefixes; we find that such reuse fails most often at image-sensitive decoding states. Longer keys reduce this aliasing but match fewer states. We introduce \method, which stores target predictions under one-, two-, and three-token keys: contextual memories live only for the current request and are thus implicitly conditioned on its image, while token-level memory persists across requests to retain coverage. At each lookup, entropy-confidence weighting fuses the retrieved predictions according to their key specificity and concentration. Best-first search composes the fused transitions into a candidate tree, supplemented by suffix-matched continuations, and one target pass verifies the tree under greedy decoding. Without training or an auxiliary draft model, \method\ achieves mean speedups over autoregressive decoding of $1.66\speed$ and $1.56\speed$ across seven single-turn benchmarks on Qwen2.5-VL-7B and LLaVA-1.6-7B, exceeding the strongest training-free baseline, Token Recycling ($1.51\speed$ and $1.37\speed$). Our implementation is available in https://github.com/Senmo996/RECAP-official-code.

A learned cost model that predicts query runtimes is trained on measurements from one machine. When the database moves to other hardware, its plans stay the same but their runtimes change, each by its own factor. We ask how to adapt such a model from k queries measured on the new machine: which queries to run, how to use their runtimes, and how many are enough. On PostgreSQL we timed the same plans of 1,162 TPC-H and TPC-DS queries in six environments: a laptop and a server, each with all its processors or pinned to one, the server also pinned to two, and the laptop also with a small buffer pool. No single factor describes the slowdown between two environments: most of its variance lies between query templates, so a speed factor stops improving after a few queries. A correction that fits separate lines for serial and parallel plans and adds an offset for every measured template brings the moved model, with one measured query per template, to an error (geometric mean over templates of the median Q-error) of 1.10 on TPC-H and 1.19 on TPC-DS, close to the same correction on a model trained on the new machine; fine-tuning a DACE-style neural cost model on the same queries gains at most 2%. On templates never measured, no correction we tried is on average more than 2% better than the speed factor. We then fixed hypotheses and timed new data. On TPC-H at ten times the scale the errors follow these results, but two of five repeated tests miss. On the Join Order Benchmark, whose queries differ widely within a template, the model errs by 2.36 even when retrained on the new machine, no correction gains much, and only the test on new templates holds; boosted trees and a neural model fit these queries better, and correcting the neural model works as well as fine-tuning it. We report every test as registered.

Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed. We introduce CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Its predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation. A shared world state links predictions, acquired measurements, and source-linked claims, supporting explanation and revision. On 339 held-out SciPlex3 conditions across three cell lines, CellCanon improves differential-expression AUPRC and Macro-F1 over the best baseline for each metric by 63.3% and 17.3%, respectively, and direction Macro-F1 by 7.2%. Relative to gene-wise reasoning on matched drug-cell cohorts, measured language-model token use falls by 99.77% for Predict and 98.41% for Explain. On the source-grounded benchmark, Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient-evidence cases. In simulated experimental feedback, Update uses one measured source context per drug to exceed the best fine-tuned baselines in differential-expression AUPRC and Macro-F1 without changing model weights. Together, these capabilities support a cumulative account of cellular responses, with predictions and biological judgments that can be revised as evidence accumulates.

Confidence estimates guide human review in LLM-built knowledge graphs, but better error prediction can yield little improvement in an audit. We study this gap through repairability, selection headroom and label validity. Modelling non-adaptive auditing as budgeted covering, we characterise the minimal decision-relevant error statistics and bound the regret of scalar-preserving first-order planning by a structural mass $Δ^{\mathrm{cl}}$. We simulate ideal repair against benchmark-aligned labels for five extraction configurations and two cross-document graphs on Re-DocRED, followed by frozen confirmation on test and replication on 500 hash-selected training documents. On the held-out cross-document graphs, an error-rate $\times$ fan-out queue leaves 3.98-4.74 percentage points more errors than a joint model with the same actions. By comparison, opening shared actions for that joint planner reduces residual errors by 0.89-1.43 points relative to its own fact-only plans, in point estimates. Equivalence between joint and independent-block models remains inconclusive. Generic LLM verification improves three of four model-budget configurations on each held-out set. LLM adjudication identifies possible benchmark omissions.

Surrogate modeling of transport requires operators that capture the combined influence of spatially distributed sources and Dirichlet and Neumann boundary conditions, together with training datasets that can be generated at low computational cost. This work presents MIGONet, a multiple-input graph neural operator, and a discrete method of manufactured solutions (MMS) that trains it without solving a partial differential equation, using steady heat conduction as the model problem. In the discrete MMS procedure, temperature fields are prescribed a priori, and the source terms and Dirichlet or Neumann boundary data are evaluated with the same discrete operators as the reference finite-difference method (FDM), so every training sample satisfies the discrete governing equations by construction. Because the boundary data are derived from each field, a Dirichlet or Neumann condition can be assigned to every boundary segment of every sample, and the boundary specification becomes an operator input. MIGONet supplies representations of the complete source field and boundary data to every node before local message passing. On the rectangular domain, it is compared with a version without the complete-source encoder (MIGONet-local) and with the multiple-input operator network MIONet, using the same training data. MIGONet achieves mean relative temperature errors of 10.31% and 10.64% for curved-wire and Gaussian-spot heating, compared with 45.97% and 48.34% for MIONet and 72.25% and 67.82% for MIGONet-local. For 10,000 samples, discrete MMS reduces data-generation time from 486.65 s to 3.15 s relative to matched iterative FDM labeling. Separately trained MIGONet models are also evaluated on L-shaped, disc-shaped, and ring-shaped domains.

Repeated mobile GUI tests can reuse routes, locators, and recovery strategies from earlier executions. However, a route that worked in one workflow or app version may be irrelevant or misleading in another. We study how an LLM-based tester can retain useful experience while withholding entries contradicted by current observations. We present a maintenance contract for AdaptiTest in which reusable experience has an explicit applicability scope, trace provenance, and a lifecycle for confirmation, revalidation, and withdrawal. The contract reserves persistence for a Reflector role and treats recalled experience as advice. The archived implementation does not establish end-to-end enforcement of every specified mechanism. The design changes retrieved context, not model weights. The evaluation uses read-only Android workflows and archived internal comparisons. A six-task, three-repeat ready subset reports 18/18 completed runs with memory versus 3/18 without it, but a grounded single-agent control also solves the earlier hard-task slice. A separate cross-session probe is excluded from efficacy conclusions because of confounded route schedules and unvalidated state labels. We examine 13 selected real-update hazard records across nine subjects, distinguishing stale-locator observations from policy-based withholding and offline retrieval diagnostics. The latter do not measure live-device failures or actions saved. These events are UI changes, not confirmed regression bugs. The evidence supports a bounded design and case study; it does not establish general efficacy, superiority over independent GUI-agent systems, or unrestricted exploration.

When does an AI assistant improve human decision-making, and when does it hurt? We argue that the answer depends not just on AI accuracy but also on how much its evidence overlaps with the human's and how the human combines the two. A worker makes a binary decision by observing her own signal and receiving a recommendation from an AI. We develop a human behavioral model with two primitives: her awareness of the overlap and her trust in the AI. Our analysis gives a closed-form phase diagram over AI capability and overlap, with regions of complementarity, impairment, and automation. A worker who neglects the overlap (correlation neglect) can be made less accurate by a more capable AI, double-counting the shared evidence. Trust reshapes the regions - undertrust may compensate for correlation neglect, so that the trust level maximizing accuracy depends on the deployment parameters. Empirically, we identify the two model parameters by pairing the same workers with different AIs, and break the accuracy loss into a correlation-neglect cost and an undertrust cost, one tied to reducing overlap on the AI side, the other to calibrating trust on the human side. On ChatBench, AI assistance raises accuracy by 21 points, yet workers apply only about a quarter of the Bayes-rational weight to AI information; GPT-4o falls in the automation region and Llama-3.1-8b stays at the complementarity boundary; and calibrating trust recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.

Temporal-difference (TD) learning underlies many prediction and control algorithms, but can diverge when combined with off-policy learning, bootstrapping, and function approximation. TD with Regularized Corrections (TDRC) is theoretically sound and practically simple, but existing convergence theory covers only the one-step case under i.i.d. sampling, while the Huber Bellman-error formulation has not been extended to TDRC or eligibility traces. In this work, we establish convergence of constrained TDRC($λ$) with diminishing stepsizes and non-i.i.d. off-policy data, for state-dependent and history-dependent $λ$-parameters. We also propose TDRC($λ$)-Huber, which preserves the TDRC($λ$) fixed point while reducing sensitivity to large Bellman errors. For both algorithms, we apply an ODE-based weak-convergence method from stochastic approximation theory. For the nonlinear mean dynamics of TDRC($λ$)-Huber, we give an integral small-gain condition that guarantees uniqueness and global asymptotic stability of the target equilibrium. We then establish convergence of the constrained algorithm, derive computable sufficient conditions for convergence, and identify a regularization-based convergence mechanism shared by linear TDRC($λ$) and TDRC($λ$)-Huber. When the mean TD matrix is positive stable, both constrained algorithms converge under sufficiently strong regularization without requiring positive definiteness.

Machine unlearning aims to remove a user's data from a trained model after a deletion request. Existing methods are majorly evaluated on three axes - how well they forget, how much capability they retain, and how robust they are to membership inference attacks, but not on whether the provider can offer the user any evidence that the deletion occurred. To fill this gap, we introduce auditability as a fourth axis, defined as the separation between requested (forget set) and retained (retain set) data under a statistic the provider computes. We show that across eleven unlearning methods on TOFU, MUSE and RWKU no method satisfies all four axes, with different methods falling short on different axes. To address this, we propose a two stage method that can be used on top of any unlearning method, in the the first stage we propose ${GapWiden}$ which uses a hinge loss to increase the difference between audit statistics of forget set and retain set. In the second stage we propose ${GapVeil}$, an inference time rejection-sampler which uses the base model as the drafter and re-samples based on the audit statistics of the drafted response. We observe that using our method, auditability and leakage reach oracle parity for every unlearning method, and forget performance converges to the oracle; retain performance improves for nine of eleven methods but remains short of parity. Our results show that evidence of deletion or auditability can be constructed by design, and that treating it as an explicit requirement changes the viability of existing unlearning methods.

Large language model (LLM) shopping assistants now recommend products in multi-turn conversations. However, reliable evaluation faces Data scarcity and Difficulty Standardizing Criteria. To address these obstacles, we introduce CARTBench for CART (Complex and Ambiguous Requests with Timelines). For Data scarcity, we curate cases from real commercial shopping needs across 17 retail domains, pairing timestamped dialogue with 1,603 product records. For Difficulty Standardizing Criteria, we develop case-specific rubrics guided by Ideal Response Principles, distilled from excellent replies manually selected from successfully completed orders. These principles also guide Principle-Guided CART Prompting, our structured recommendation strategy. The experimental benchmark contains 1,328 expert-annotated criteria across Core, Excellence, and Safety. To answer the CART question, we evaluate 7 flagship commercial API models, including GPT-6.1-Sol and Claude-Opus-5, and 35B+ parameters open-weight models. In the first full-suite evaluation under the frozen rubrics, 3 of 10 models exceed 90% on Core, while none exceeds 45% on Excellence. Therefore, we identify an understanding-to-execution gap between required behavior and high-quality purchase advice. To examine these limitations, we diagnose shared failures in product evidence and unresolved needs, linked to changing requirements and decisive unknowns. Code and data are available at https://github.com/Da1yuqin/CARTBench.

Various land-use models characterize land-use patterns and facilitate scenario planning. However, existing models are largely predictive, forecasting transitions based on numeric data without synthesizing new land-use maps from simple text prompts. Meanwhile, generative AI in urban systems remains unimodal, weakly controllable, and evaluated mostly by generic image quality rather than domain-specific metrics. To address these gaps, we introduce DualControlNet, a generative land-use model that jointly synthesizes high-resolution land-use maps and satellite imagery, addressing multimodality, controllability, and verification challenges. Trained on over 40,000 paired tiles from Orlando and Chicago, DualControlNet is evaluated using land-use, spatial consistency, and image quality metrics. Relative to unimodal baselines, DualControlNet improves land-use diversity consistency by 62.8% and building coverage ratio consistency by 16.3%, more than doubles cross-modal spatial consistency, and triples visual quality. Scenario-based experiments demonstrate that DualControlNet supports sustainable development through multidimensional controls, increasing density and diversity, and responding to text prompts like "high-density development." Overall, this work advances land-use modeling via a multimodal, controllable, and verifiable generative framework, lowering technical barriers and enabling intuitive exploration of urban futures.

Prenatal ultrasound biometry must return a head circumference (HC) and refuse when the image cannot support one. Selective prediction trades coverage against error, but its finite-sample guarantees bound the rate of gross errors, not their magnitude. We audit a tool-grounded biometry pipeline on two separately pre-registered four-way splits of the 999 labelled HC18 images, retraining the segmenter three times on each. Calibrating the contour-completeness threshold by exact Clopper-Pearson risk control returns a certificate indistinguishable from a hand-set floor, yet the pipeline still emits a 271.3 mm error: a small, geometrically complete ring fitted to an image whose true HC is 324.6 mm. The accepted candidate scores 0.94 completeness, so the score cannot see the failure, and one outlier among $\approx 150$ answers leaves the certified risk bound ($α= 0.10$) intact. A second, untouched split reproduces the mechanism on a different image (233.3 mm). Constraining the admissibility region---the anatomical range in which the measurand is defined---removes the family instead: raising the lower bound from 50 to 55 mm costs 0.4 coverage points and cuts the worst-case error on both splits to 11.4 and 11.1 mm, whereas suppressing it through the completeness threshold costs 19.0 and 12.1 points. At an anatomically defined 70 mm the frozen tests give 1.91 and 1.81 mm MAE.

Foundation-model predictors of perturbation response are evaluated primarily by aggregate accuracy. Using their predictions to prioritize experiments requires assessing each prediction's reliability. We audit the state-transition (ST) decoder used in STATE across three encoders (SE-600M, scGPT, and Tahoe-x1) and four cell lines under CRISPRi. The decoder's first attention layer has a low-rank output. Projecting this output onto 4-32 fixed directions estimated from training activations preserves mean per-perturbation Pearson delta correlation (PDC). This low-rank attention bottleneck bounds the decoder's reliability by distance to training responses: PDC declines as the cosine distance from a perturbation's observed response to its nearest training response increases. We propose predicted-NN-distance, the cosine distance from a predicted response to its nearest training response, as a per-prediction confidence score. The score has a mean absolute Spearman correlation of 0.81 with PDC, close to the 0.88 obtained with the observed response. Thresholds calibrated on this score transfer across encoders and cell lines, and the score remains predictive of PDC for drug perturbations.

User preferences shift as needs, context, and tastes evolve. This matters for personalized recommendation, LLM response generation, and RLHF-trained reward models, which can become miscalibrated as preferences drift. Existing non-stationary bandit methods forget uniformly across all dimensions, but real preference shifts are typically partial: some dimensions go stale while others remain valid. Neural contextual bandits accumulate confidence monotonically via the Gram matrix, with no mechanism to direct uncertainty where predictions have gone wrong. We propose PEARL (Prediction Error-Adaptive Regularized Learning), which reverses the Gram update along feature directions where predictions are surprisingly wrong, while leaving confidence intact elsewhere. We prove that (i) any non-adaptive policy suffers linear regret after a preference shift, (ii) no fixed exploration rate achieves low regret in both stationary and non-stationary regimes, and (iii) PEARL's cumulative regret scales with the number of shifted users N_s, not the total population N. On PRISM, a personalized LLM response selection benchmark (1,500 users, 21 LLMs), PEARL beats greedy by +4.0 points and the best neural bandit baseline by +2.5 points (p<0.001), while performing on par with greedy when preferences are stable. It beats sliding-window and discounted approaches by 3 to 10 points, showing that selective forgetting outperforms uniform forgetting. The selection policy is a ~100K-parameter MLP that runs in milliseconds on CPU, leaves LLM weights frozen, and works with any API-based model.

This vision paper argues that distributed learning, including federated and decentralized, should become a first-class paradigm in LLM post-training, not merely a privacy add-on. Current methods like RLHF and DPO rest on an invisible assumption: preference data should be centralized. We contend this assumption is scientifically unjustified, practically unsustainable, and ethically problematic. As LLMs deploy into healthcare, law, finance, and multilingual communities, preference data originates from populations that cannot or will not centralize. We develop a three-pillar argument grounded in (1) privacy and data sovereignty, (2) pluralism and representational equity, and (3) robustness and scalability. We formalize key technical challenges unique to this intersection, reference drift, preference heterogeneity, and the DP-utility trade-off for contrastive losses, and propose concrete algorithmic directions, and sketch a research agenda. Our analysis shows distributed post-training is not only necessary but tractable, and ignoring it risks AI that is legally noncompliant, representationally narrow, and brittle at scale.

Assessing the deployment performance of learned robot policies during training is challenging, as directly measuring task success requires costly real-world trials for each candidate checkpoint. Averaged validation loss offers a rollout-free alternative, but its checkpoint rankings can be inconsistent with deployment performance. We analyze two limitations of this criterion: it penalizes action deviations without considering whether the policy can recover from them, and gives later reference predictions equal weight even when earlier deviations may lead execution away from the reference. We propose RACE, a novel rollout-free metric for ranking checkpoints using a fixed set of held-out successful trajectories. RACE calibrates action deviations against state-conditional thresholds derived from action differences at similar states in these trajectories. We aggregate these scores through a running product, so that a substantial deviation reduces the contribution of subsequent reference predictions. Averaging the resulting cumulative scores over time and across trajectories yields the checkpoint score. In simulation, RACE outperforms baselines across validation sources, while several baselines degrade sharply when the validation set changes. On real robots, its rankings align with deployment success across the evaluated policy backbones and tasks of varying difficulty. Perturbation analyses further show that RACE distinguishes deviations with markedly different effects on task success, even when baselines respond similarly.

Firms using large language models under fixed budgets must decide which model to call and whether to retry a failed task. Because a call's charge is revealed only after completion, one long response can exhaust the remaining budget and leave later tasks unserved. Our objective is to maximize the number of tasks solved before service stops. We develop a contextual bandit policy, ADP-HT, which learns each model's task-specific success probability and constructs conservative cost estimates from coarsened response-length labels on a quantile grid with $2^K$ intervals. The policy weighs predicted success against these costs, adjusting the cost penalty to realized spending and the budget available per remaining task. Under calibration, local regularity, and grid-coverage conditions, we bound the excess allowance in the policy's conservative cost estimates. At a fixed sample size, increasing $K$ reduces the discretization term but increases the estimation allowance. At the tightest tested budgets, ADP-HT solves more than twice as many MATH500 and GPQA tasks as the best implemented baseline. It achieves the highest mean cumulative score at ten of fourteen RouterBench budgets. On LiveCodeBench, it outperforms a deployed commercial router in both default and cost-saving modes at the three lowest test budgets.

Intelligent production requires systems to make decisions and organize work according to production goals and operating conditions. It also requires continuous adjustment through feedback. This paper analyzes intelligence in human production organizations from a complex adaptive systems (CAS) perspective. The analysis identifies two coupled mechanisms: individual autonomy and collective coordination. Individual autonomy enables actors to form local decisions using available information. Collective coordination aligns these decisions with shared goals and joint constraints. Learning and adaptation improve both mechanisms over time. A production-system model formalizes how local decision rules and coordination jointly form system decisions. On this basis, we propose the Recursive Autonomous Agent (RAA) framework, which combines goal-driven autonomous agents with a recursive organizational structure. Goal contracts define local objectives and operating boundaries, while capability contracts support the selection and composition of production units. Common interfaces allow coordinated groups to participate as units in larger organizations, connecting equipment, production lines, and factories. RAA thus provides an engineering approach to next-generation intelligent manufacturing, translating the requirements of intelligence, openness, and system-wide integration into mechanisms for autonomous and coordinated production.

Decision models choose among predefined options in tasks such as routing and action selection. In multi-agent collaboration, a decision determines how members divide work, exchange resources, and act in parallel. We introduce TandemBench, a benchmark of 1,200 questions in multi-agent cooking scenarios where coordination choices affect goal completion. Models receive complete global information and candidate team plans, with reference answers established through execution and exact search. Models select an optimal team plan and, in paired variants that retain or remove it, judge whether better arrangements exist beyond the supplied options. We evaluate 40 native decision models and compare them with 10 general-purpose language models. The best native decision model achieves 62.50% decision accuracy. Native decision models remain close to chance in optimality verification. Paired tests reveal that some report better alternatives regardless of whether an optimal plan is already present. These findings show that team-plan decision accuracy alone can conceal systematic errors in judging whether the supplied options contain an optimal arrangement. The dataset and evaluation code are publicly available at https://github.com/lililibaohang/TandemBench.

Security-by-design emphasises integrating security into software design and development. Security pattern catalogues operationalise this principle by prescribing reusable architectural solutions to recurring security problems. However, recovering their implementations from source code remains challenging because security patterns are abstract, span multiple components, and admit many valid implementations. We present SecPatRec, a hybrid approach for automated security pattern recovery from source code repositories. We treat a security pattern as a reference architecture and frame its recovery as a deductive task: determining which prescribed roles are present, grounding them in code entities, and recovering their interaction sequences. SecPatRec pairs a ReAct-style LLM agent equipped with semantic and lexical code search for role detection and grounding with static code property graph analysis for interaction recovery. We curate a human-validated ground-truth dataset of 100 pattern instances covering five security patterns across 93 open-source Java and Python repositories, and introduce a three-level evaluation framework measuring recovery correctness against this ground truth. We evaluate SecPatRec using four LLMs, including proprietary and open-source models. Averaged over all models and pattern instances, SecPatRec achieves role-detection accuracy of 0.84, role-grounding similarity of 0.84, and an interaction-recovery score of 0.66. Compared to a one-shot end-to-end LLM baseline, SecPatRec achieves higher aggregate recovery scores and lower estimated LLM API costs per run, while processing repositories that exceed the baseline's configured input limit. The recovered role mappings and statically derived interaction sequences provide structural and behavioural evidence that can support inspection of security-pattern implementations.

Large language models increasingly automate complex tasks, but human decision capacity does not scale with their capabilities. In many applications, humans remain essential for providing private context, authority, or judgments unavailable to the agent. We formulate human decision capacity allocation: given multiple consultation opportunities and a limited budget, at which decisions should an LLM agent consult a human? The central challenge is estimating consultation value, which is executor-dependent because the same human input may benefit different downstream agents differently. We introduce Consultation Value Distillation (CVD), which estimates consultation values offline using Monte Carlo executor rollouts and distills these values into lightweight predictors for budgeted allocation, requiring no rollout-based value estimation at deployment. Across five batch and three sequential settings with two frontier executors, CVD improves allocation through joint and continuation-aware value learning, while independent consultation values remain competitive in several settings. CVD also broadly outperforms direct prompting and Value of Information where applicable, remains robust with substantially less rollout supervision, and enables nearly lossless approximate allocation at low search cost.

Office documents are now routinely uploaded to large language models (LLMs) for processing. However, an LLM does not always see what a human sees, because it typically reads the text that an extractor produces from the document's underlying Office Open XML (OOXML) files. Depending on the implementation, extractors differ in which OOXML parts they read and how they interpret them, so their output can differ from what Office applications display. For example, by slightly shifting the positions of text boxes on a slide, an adversary can swap the order in which the model reads them while the slide looks unchanged. We show that this extraction divergence is an attack surface for LLM applications. In this work, we first measure what an adversary can exploit by analyzing the widely used extractors that LLM applications rely on. By combining static source code analysis with dynamic analysis, we identify 19 OOXML channels where extractors read content a human cannot see and 12 where they drop content a human can. We also find that widely used extractors ignore display formatting and reading order, so even visible content reaches the model in a different form. Building on these findings, we construct five attacks that lead the model to interpret a document differently from a human reviewer, without injecting any instruction. We test the attacks on four major chat assistants (ChatGPT, Claude, Gemini, and Microsoft 365 Copilot) and four coding agents (Codex, Claude Code, Antigravity, and GitHub Copilot CLI), and every one of them is misled by at least one attack. Finally, we propose Prism, a lightweight defense that reconstructs the human view from the OOXML specification before passing the text to the model. Prism also generalizes to unseen attack payloads.

Expert deactivation at inference time changes both token routing and expert weights, and its effect on performance depends on how these changes are implemented. We compare removal, freezing, remaining-expert renormalization, and unnormalized re-routing on identical disabled sets across five instruction-tuned checkpoints from four model families, with a base OLMoE checkpoint as an additional comparison. We develop a weight-sum model that relates the loss of expert contributions and changes in branch scale to the weight rule used during training. With half of the experts deactivated according to activation-based rankings, remaining-expert renormalization yields the highest perplexity in models trained with raw router probabilities, whereas unnormalized re-routing yields the highest perplexity in models trained with normalized weights. Perplexity ranges from 1.4 to 73 times the intact baseline when the training rule is retained, and from 14 to 28,740 times the baseline under a rule mismatch. Task outcomes vary across evaluation types: on OLMoE, GSM8K strict-match accuracy falls to zero under all four implementations. Extending the model to changes in the number of active experts yields a further prediction, supported by the OLMoE experiments: the perplexity penalty from additional renormalization increases as fewer experts are activated.

How much phonological knowledge do language models (LMs) possess about Chinese characters, and how is this knowledge represented and used? We introduce PinyinBench, a controlled benchmark covering homophones, phonological features (initials, finals, and tones), and polyphonic characters. We investigate 17 text-only and multimodal LMs through three complementary approaches: behavioral evaluation, layer-wise probing, and activation steering. Behaviorally, models demonstrate substantial knowledge of initials and finals, with a frontier model (DS-v4-flash) scoring 90%+ in homophone judgment, when two characters share no phonological features. However, its accuracy drops sharply to 72% when two characters differ only in lexical tone. Probing reveals that tone is highly decodable from embeddings and improves modestly across layers, whereas segmental information starts less decodable and improves substantially with depth. Activation steering shows that segmental representations can systematically influence generation, while tone exhibits much weaker effects. Furthermore, audio-bearing training improves phonological decodability, including tone, without corresponding improvements in steering effectiveness. Together, these findings reveal distinct representational and functional profiles for segmental and tonal information, with tonal information much harder to access and manipulate, demonstrating the necessity for combined evaluation, probing, and causal intervention for understanding linguistic knowledge in LMs.

Recent LLM-centric advances can generate attractive furnished rooms by selecting and arranging assets. Yet evaluation remains limited to specific perspectives or predefined questions, failing to reflect overall scene quality. To address this pressing need, we introduce GREP. Its core is a focus shift from defining specific questions to establishing a three-tier evaluation philosophy: basic physical validity checks whether a scene obeys physical constraints; visual coherence concerns compatibility in object scale, style, and composition; and, at the highest level, human-use logic requires object arrangements to align with everyday human habits. Guided by these principles, an evaluation agent autonomously and adaptively identifies issues from scene context, gathers evidence, and produces scores and reports. This supports scene-specific issue discovery and reliable judgments. Experiments show that (1) GREP achieves 87% agreement with human majority preferences, substantially outperforming baseline evaluators. (2) GREP reports improve scene quality through refinement, whereas score-only and no-feedback baselines show slight declines. (3) GREP suggests that as planning ability improves, recent models such as GPT-5.6 Sol can move from relying on pipelines and harnesses to challenging their guidance and constraints, but still lack the ability to resolve the resulting conflicts. Stronger models are therefore expected not only to plan well, but also to adapt to external guidance and constraints more flexibly. Project page: https://han-808.github.io/GREP/

Holistic perception for autonomous agents requires reliably segmenting and associating scene elements across time and camera views, including objects outside a fixed semantic taxonomy. Existing open-set panoptic benchmarks focus on individual images or views, while panoramic tracking benchmarks assume closed-set taxonomies. Consequently, these benchmarks do not evaluate whether unknown objects can be consistently segmented and tracked across synchronized multi-camera sequences. We introduce WayPOP, a panoramic open-set panoptic tracking benchmark built on real Waymo PVPS sequences augmented with diverse out-of-taxonomy 3D assets from Objaverse. Inserted objects are placed at fixed locations in the 3D scene and rendered using calibrated vehicle-mounted cameras as the ego vehicle and surrounding traffic move, providing temporally and cross-view consistent panoptic annotations and auxiliary LiDAR range observations. WayPOP contains approximately 670 curated assets instantiated as 980 unique unknown-object tracks, yielding 53,452 unknown-object mask annotations across 21,300 camera images. Our evaluation protocol defines a hierarchy of tasks from view-independent open-set panoptic segmentation to full panoramic open-set panoptic tracking, using known/unknown PQ and (w)STQ metrics. We benchmark representative open-set and open-vocabulary segmentation methods with temporal and cross-camera association. Our results reveal substantial limitations of current methods in truly open-set scenarios, highlighting robust unknown-object localization and identity association across time and views as major open challenges.

Modern recommender systems rely on generative retrieval techniques to predict the next item either as discrete codes or as a continuous vector. On the one hand, Semantic IDs (SIDs) give discrete models shared parameters and pretrained semantic structure, but discretizing items into hard codes discards some information from the original content embeddings. On the other hand, continuous models avoid quantization, yet typically use flat item lookup tables with no explicit parameter tying across items. We ask whether a continuous recommender can introduce shared structure without making it a discrete bottleneck. We propose CAP-Mix, a flow recommender that represents each item as a soft mixture of shared prototypes plus an item-specific vector. The assignments, prototypes, and item vectors are trained jointly, and the soft representation is retained at inference. Pretrained information is optional: SID responsibilities can serve as a prior over assignments, while a full content embedding can initialize the item-specific vector without changing the initial flat table. In matched comparisons on three datasets, CAP-Mix improves over a flat table on the two sparse Amazon datasets and yields a small gain on Steam when trained without pretrained information. Given only an offline SID, it outperforms a hard table built from the same tokenizer by 4.45-14.90% in HR@5. Given the full content embedding, it matches or improves a strong flat table while hard assignments regress on the sparse Amazon datasets. Overall, averaged across six benchmarks, CAP-Mix improves HR@5 and NDCG@5 by 6.9% and 13.6% relative to the best evaluated baseline for each dataset and metric, while delivering 1.4-161x the inference throughput of the discrete baselines in our experiments.

Recognizing visual content from neural-network weights requires separating functional information from arbitrary parameter coordinates. We audit this separation with hidden-neuron permutations and positive ReLU rescalings that preserve the represented function. The study contains 480 independently fitted shape fields and 1,000 MNIST fields, with transformed copies grouped by origin. On shape recognition, a statistics reader falls from \ShapeRawStats\% to \ShapeJointStats\% accuracy under joint transformations; a normalized set reader retains \ShapeNormSet\% with unchanged predictions. Four small frozen language models obtain \LMRange\% raw accuracy with next-token scoring. A separate four-configuration reasoning experiment with 1,920 scored sessions obtains \FrontRawRange\% raw accuracy with generated answers, while fresh identical raw prompts disagree on \FrontRawRepeatRange\% of origins. Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. Joint correctness/agreement metrics, semantic-cue controls, fresh repeats, and exact-rational rounding bounds separate these failure modes. The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations.

LLM-based coding agents are primarily evaluated by their solve rates. However, this metric does not reveal how difficult the solved tasks are or account for differences intask difficulty across benchmarks. Determining the difficulty of individual tasks by repeatedly running multiple agents is costly, while a predicted difficulty score alone provides limited guidance on what support an agent needs. To address these challenges, we propose DARE (Difficulty Analysis for Repair Enhancement), which integrates repair-task difficulty measurement, prediction, feature-group contribution analysis, and task-specific repair support. We define continuous task difficulty as the observed failure rate across multiple coding agents. The DARE difficulty model predicts this difficulty using 151 static features extracted from benchmark artifacts. The framework also analyzes the contributions of seven Repair-Task Feature Groups to identify the characteristics most salient to each task and uses this information to select task-specific agent support. We evaluate DARE on four benchmark partitions comprising 2,594 tasks. The difficulty model's aligned predictions achieve Spearman's $ρ$ = .476-.621 on held-out partitions, and the most salient Feature Groups vary across tasks and benchmarks. Furthermore, coding-agent experiments show improved repair success when support is matched to each task's salient Feature Group. These findings demonstrate that difficulty prediction and interpretation can lead to practical improvements in automated program repair (APR), while also providing a basis for difficulty-stratified agent evaluation, cross-benchmark comparisons, and identifying areas of repair capability that warrant further improvement.

Short-horizon model predictive control (MPC) is fast but myopic. ZipMPC compensates by learning a context-dependent correction of the short-horizon cost that imitates a long-horizon expert. We observe that the optimal correction is often multimodal: when a slower opponent can be passed on either side, good corrections of opposite sign coexist, and the expert's choice between them hinges on tiny cost differences. Deterministic regression then averages modes or inherits inconsistent expert tie-breaking. We propose ZipDrift, which trains a one-step drifting generator over MPC cost corrections, tilted toward low closed-loop cost and learned only from cost evaluations of candidate corrections, without differentiating through the optimiser. A few generated candidates are ranked by a learned critic, and the model-based MPC then enforces dynamics and constraints. In a kinematic overtaking benchmark derived from ZipMPC, 53% of contexts are bimodal and the manual short-horizon MPC crashes in 28.5% of test episodes. With an eight-dimensional stagewise correction, ZipDrift is collision-free in all 600 test episodes and matches the best evaluated candidate, whereas an idealised ZipMPC-style regressor crashes in 4.0% (16.0% with inconsistent expert labels) and critic-ranked random candidates with the same budget in 2.0%.

Large language model (LLM) agents increasingly rely on external memory stores to retain information across conversations, typically via dense vector embeddings searched with approximate nearest-neighbor methods. The fruit fly's olfactory circuit computes a locality-sensitive hash of sensory input using a strikingly different mechanism - sparse random projection followed by winner-take-all competition - and prior work (Dasgupta et al., 2017) showed this scheme, dubbed FlyHash, preserves neighbor structure better per bit than conventional random-hyperplane hashing (SimHash) on several benchmarks. We apply FlyHash, to our knowledge for the first time, to LLM agent memory retrieval, and ask three questions a practitioner actually needs answered: does the quality advantage hold under a correctly matched-bit protocol (it does, robustly, across 10 seeds and four datasets spanning two modalities, with one specific, previously undocumented comparison pitfall identified and corrected along the way); does that advantage translate into real memory and latency savings, including against the dense-vector indexing (HNSW) practitioners default to today (no: at equal bits FlyHash's codes are larger and slower to compare than SimHash's, and HNSW dominates both hash methods at ordinary scale; a sparse-set comparison we introduce recovers most but not all of a naive latency gap, leaving hashing's practical case narrower and more specific than the quality result alone implies); and does the advantage survive genuine, non-circular agent-memory relevance labels rather than the nearest-neighbor-by-embedding proxy this literature otherwise relies on (it does). We report all three answers precisely, including the two that narrow rather than flatter the headline result, and release our implementation, a working agent-memory demo, and a full account of the methodological pitfalls we found and corrected.

Observed populations of rotating neutron stars~(NSs) with millisecond-order periods are powerful probes of fundamental physics. Mass-radius measurements inform poorly constrained properties of degenerate nuclear matter, whereas high-fidelity population synthesis connects unknowns in NS formation and binary evolution to large radio-timing surveys. However, repeated structure calculations are prohibitive due to the computational cost of solving the highly coupled partial differential equations of axisymmetric general relativistic hydrostatic equilibrium that govern rotating NS interiors. Non-rotating calculations or approximate rotational corrections can bias equation of state inference and limit the fidelity of population synthesis in tracing the evolutionary history of progenitors and propagating equation of state uncertainties. In this \textit{Letter}, we develop physics-informed neural operators that can solve the axisymmetric problem with high accuracy and negligible computational cost. \textit{Our solvers are trained in a fully unsupervised setting, with the training objective informed solely by the partial differential equations, and achieve a $10^3\times$ speed-up with negligible accuracy loss}. These solvers enable scalable EoS inference that self-consistently accounts for rotation and repeated stellar-structure calculations during binary evolution, making spin predictions that account for rotational structure and EoS uncertainties computationally accessible to population synthesis.

Continual model updating has become a common practice in the deployment of machine learning systems. However, the dynamic training behaviors give rise to new risks not sufficiently addressed by existing studies. In this paper, we present the first systematic study of temporal privacy risk in continually updated deep learning models. We formalize the notion of temporal membership in continually updated models. We then define the temporal membership inference attack, which seeks to infer whether and when individual samples were involved in model training, at both timestamp-level and interval-level granularities. To this end, we propose TMIA, a novel black-box attack that leverages bidirectional LSTMs to infer temporal membership from model outputs observed across successive updates. Through extensive experiments on three language models and three image classification models, we show that TMIA is highly effective and significantly outperforms existing baselines. Finally, we evaluate three potential defenses, revealing the trade-off between privacy protection and model utility.

Algorithmic recourse identifies modifications that can change an unfavorable model prediction into a favorable one. Under predictive multiplicity, similarly performing models can disagree on individual predictions, so recourse valid for one model may fail for another. We formulate robust recourse under predictive multiplicity as a budget-constrained minimization of the worst-case validity loss across a competing set of models, which can include models with different architectures or from different hypothesis classes. We propose an oracle-efficient algorithm called ORPM and prove sufficient conditions for its convergence. We evaluate ORPM on one synthetic and five real datasets, and competing sets consist of linear, neural, mixed, and decision-tree models. Our experiments show decreasing optimization-gap bounds and high validity in many tested settings, with favorable performance against baselines designed for specific model classes or architectures. We also quantify how requiring a common action across competing models changes recourse success at a given budget and examine disparities in this success across demographic groups.

This study investigates the authorship impersonation capacity of six widely used large language models (LLMs): Gemini 3 Pro, Claude Opus, GPT 5.2, GPT 5.2 Pro, DeepSeek 3.2, and Grok 4.1 Fast. The evaluation was conducted using short product review texts selected to ensure high within author consistency in individual writing style. Two types of in context prompt were designed: a simple prompt with minimal instruction and a complex prompt incorporating explicit stylometric guidance for mimicry. Authorship mimicry was evaluated using a cosine distance based distance to training metric, supplemented by visual analyses of embedding distributions. The results indicate that authorship impersonation remained challenging for most LLMs, with Gemini 3 Pro showing the strongest apparent mimicry performance. However, subsequent overlap analysis revealed that Gemini 3 Pro extensively copied and reused sentences and phrases from the training texts, at levels far exceeding those observed among human authors. Follow up factor analyses showed that training set consistency was the most robust predictor of mimicry success across models, whereas author distinctiveness did not generalise as a reliable predictor. Global embedding visualisations further revealed a consistent macro level pattern: for most models, generated texts occupied regions of the embedding space that were largely separated from the corresponding human training clusters. This indicates limited convergence towards target author style and suggests the presence of model specific lexical and phraseological preferences. The findings highlight the importance of distinguishing genuine stylistic imitation from direct textual reuse and have implications for the forensic analysis of textual evidence involving potential LLM enabled impersonation.

Recent advances in large language models (LLMs) have established human-agent collaboration as a promising paradigm for data science. Effective collaboration requires a shared computational abstraction that enables agents to reason about processing logic while allowing human analysts to understand, guide, and refine the analysis. Although scripts written in Python or SQL provide precise control, their syntax and execution details can be difficult for analysts to understand and modify. In this paper, we argue that dataflows are an effective abstraction for human-agent collaboration in data science. We introduce BobFlow, a system that combines the dataflow abstraction with the ReAct paradigm. It enables agents to iteratively construct and execute dataflows for data science tasks while allowing analysts to inspect the analysis and provide feedback. Our user study shows that BobFlow helps analysts understand, inspect, and verify agent-generated analyses. A thorough experimental evaluation across proprietary and open-weight LLMs shows accuracy score gains of up to 18.6 points at just over 1/3 of a script-based ReAct agent's inference cost.

Before an AI agent takes on a paid job, someone has to judge whether it is likely to succeed. We test whether frontier language models can make this call before any attempt. We propose PAYWORTHY which contains 1{,}494 real freelance job postings, sampled from 18.8M postings made between 2020 and 2026 on a freelancing platform. An agent attempts each job from the posting alone in a single run, and an LLM judge acting as the client decides whether to accept the work. We ask 18 models to forecast how likely the agent is to succeed and measure how well each model's forecasts rank the accepted jobs above the rejected ones (AUC, where 0.5 is a random ranking). The answer depends on how the agent is described. Told what the agent cannot do, every model ranks the jobs well (AUC 0.66--0.86). Told instead that it is a generic agent with a browser and code execution, the same models rank the same jobs about as well as a random ordering (0.45--0.56). These results hold under judges from four other model families (Claude, GLM, Kimi and Qwen), and the informed forecasts also rank the verdicts of two human raters. Code is available at https://github.com/upwork/Payworthy

Organizations accumulate evidence in conversations, documents, decisions and relationships, but they act on interpretations: judgements assembled from several observations, resting on assumptions, and revised as circumstances change. Existing retrieval, knowledge graphs and agent memory keep the records and lose the interpretation. We propose a four-layer architecture, Signals, Domains, Claims and Dimensions, in which interpretations persist with their justifications and standing, a truth-maintenance network revises that standing as evidence changes, and validated, versioned Dimensions position interpretations on named axes. We call the result a shared organizational cortex. In a registered retrieval experiment, the Dimension term gains 0.028 nDCG@10 (p = 0.002). In system-level comparisons, the full retrieval configuration improves nDCG@10 by 0.081 over Signals-only and 0.053 over graph-augmented Signals-only retrieval (both reported p < 0.001). The architecture gives organizations a persistent, revisable basis for shared judgement.

Large language models (LLM) have been the key focus of recent machine learning breakthroughs. As the model size grows exponentially, the computation needed for training is astronomical. A GPU can do a huge amount of arithmetic. However, the data its compute units need is not always available when they are ready. This data may still be in slow memory or on another GPU. While the chip waits, expensive compute sits idle, and at scale even a small drop in utilization translates into substantial cost. In this paper, we treat this waiting time as the primary bottleneck. We focus on hardware features designed for explicit data movement. These features load data from off-chip memory onto the chip, rearrange it into the layout the computation requires as it is loaded, and transfer it between GPUs. By orchestrating them, we keep a steady pipeline of data flowing to the compute units so that they remain busy. We present kernel and framework implementations that address this bottleneck using three key data-movement instructions and hardware features. The Tensor Data Mover (TDM) moves tensor data from high-bandwidth memory (HBM) into on-chip memory. Transpose-load instructions transpose tensor data into the layout that LLM computations require as it is loaded on-chip. Finally, System DMA (SDMA) and Remote DMA (RDMA) move data between GPUs over the scale-up and scale-out networks, respectively. Our results show that TDM and transpose-load provide on average 1.88x and 1.19x speedup at the kernel level, respectively. The SDMA and RDMA lead to on average 1.16x and 1.33x speedup for computation overlapped with communication. This is the first AMD LLM training system to utilize all the GPU data pipelining features. Our implementation has been integrated into open source framework and library for AMD GPUs. https://github.com/AMD-AGI/Primus-Turbo https://github.com/AMD-AGI/Primus

Evaluating and improving domain-specific LLM responses requires high-quality reference data and rubrics, yet expert authoring is slow and expensive, and synthetic datasets may fail to capture practitioner preferences unless they are grounded in expert judgment. We present SME-C (Subject Matter Expert Critique), a human-AI workflow that captures expertise through lightweight critiques of candidate responses and complements it with LLM-driven analysis and refinement to iteratively curate response datasets and evaluation rubrics. We validate SME-C in the domain of financial business briefs, collecting 3,853 critiques from 411 finance professionals on 520 queries across two survey rounds. We evaluate two LLM-driven refinement strategies: one guided by the experts' written critiques and one that infers critique solely from their preference votes. In a blind validation survey with 100 additional professionals, we found that both strategies significantly improved quality over the experts' top-ranked baseline responses. The gains were concentrated among weaker original responses, and the two strategies excelled in different ways: critique-based refinement produced more domain-grounded improvements, whereas vote-based refinement was more effective at identifying and removing irrelevant content. The resulting rubrics and refined dataset support reproducible offline evaluation and benchmarking of LLMs, suggesting that structured expert critique is a viable alternative to expert authoring for building high-quality, domain-specific evaluation resources more quickly, at scale, and at lower cost.

Automatic speech recognition (ASR) systems often struggle with specialised medical vocabulary, especially in low-resource languages where domain-specific speech data are scarce. We investigate text-only domain adaptation for Hausa medical ASR using the fixed MMS-1B-FL102 acoustic model and two decoding-time language-model fusion methods: Shallow Fusion (SF) and the Density Ratio Approach (DRA). A 4-gram KenLM trained on 19,421 Hausa health sentences serves as the medical target model, while a 75,372-sentence general-domain model provides the source prior for DRA. Evaluation uses 250 recordings of 25 physician-reviewed prompts from 10 speakers across six northern Nigerian states. Under prompt-disjoint five-fold cross-validation, SF and DRA reduce WER from 29.56% to 23.55% and 23.33%, respectively. Both significantly outperform greedy decoding (p < 0.001), while their difference is not significant (p = 0.423). However, general-LM-only SF achieves 23.96% WER, indicating that the medical LM's additional improvement over a general LM is modest. Medical Term Error Rate also improves under fusion, but does not establish a distinct medical-domain advantage. These findings demonstrate that lightweight, text-only language-model fusion can improve Hausa medical ASR without additional domain-specific speech collection.

Vision Transformers (ViTs) are bottlenecked by the length of their input token sequence. Standard patchification fixes this length with a single patch size applied across the image, regardless of regional content. Prior work on adaptive patch sizing reduces token count by assigning coarser patches to visually simple regions, but relies on noisy pixel-level heuristics such as local entropy to make patchification decisions. We argue that this compute allocation should be grounded in the model's interpretation of task relevant information, not in raw pixel statistics. We introduce Split Policy Learned Image Tokenization which learns to allocate token resolution based on what the model finds meaningful, producing shorter token sequences that are better matched to task-relevant content. SPLIT matches the accuracy of full-resolution ViT classification baselines while using 30% fewer tokens. The same gains hold on dense prediction: SPLIT achieves competitive accuracy on COCO and ADE20K, demonstrating that semantically aligned patchification generalises well beyond classification.

Coding agents must deliver correct software while respecting users' expertise, preferences, and working conditions. Existing benchmarks increasingly incorporate interaction, yet personalized requirement fulfillment remains underexplored. We introduce PersonaCodeBench, which turns user characteristics into verifiable obligations in persona-driven multi-turn coding sessions. Our persona pool spans roles, technical abilities, interaction preferences, and practical constraints, yielding task-grounded evolved requirements for implementation, permissions, testing, explanation, and delivery. At runtime, a controller governs user simulator responses and requirement disclosure, using online requirement evaluation to provide feedback and advance the interaction. Requirements emerge progressively, with their content, disclosure conditions, and acceptance criteria fixed before evaluation. A completion verifier and a personalization verifier separately assess native correctness and persona fulfillment. Joint pass requires both adherence to the declared budget and compliance with validity conditions. We evaluate all major frontier agents, establish a comprehensive leaderboard, and conduct detailed experimental analyses to validate effectiveness. PersonaCodeBench supports reproducible evaluation of coding agents' ability to meet explicit, personalized user expectations alongside coding objectives.

Olfaction provides critical information about airborne chemical cues, yet it remains largely underexplored in embodied AI, with limited prior machine olfaction work focusing primarily on semantic odor perception. We introduce OdorSpaces, a physics-based simulation platform and benchmark for visual-olfactory embodied navigation. OdorSpaces augments Habitat-Sim with 3D odor transport and dispersion simulated by Fire Dynamics Simulator (FDS), enabling agents to navigate photorealistic indoor environments using both visual and olfactory observations. The benchmark includes two representative source-seeking scenarios: invisible gas-leak localization and early-stage fire localization. We develop an initial multimodal reinforcement learning policy that integrates visual observations and olfactory measurements over time. Experiments show that the policy outperforms a diverse set of baselines, establishing a strong reference for future research on visual-olfactory embodied navigation.

Neural operators fit their representation of functions to the task they are trained on, so every new task rebuilds a reduced space from labeled data. Following the self-supervised pre-training that made representations shareable in natural language processing and computer vision, we decouple representation learning from operator learning. A functional autoencoder is pre-trained on analytic functions synthesized on the fly, with no solver and no labels, to represent functions on arbitrary point clouds, and the latent space it induces is frozen and reused across tasks. The representation is exactly equivariant to translation, dilation, and amplitude, and reconstructs solutions of partial differential equations (PDEs) and natural images on geometries it never saw. Being task-agnostic, the latent space can be interpreted in its own right. The whole space is used and decodes to valid fields, it preserves the ordering of distances between physical fields, its geodesics interpolate physical evolutions, and edit directions transfer across fields. With frozen representation maps, a dataset is encoded once and a neural operator is trained entirely on cached embeddings, which saves memory and training time. On 28 tasks in fluid dynamics, solid mechanics, acoustics, and the Helmholtz and Poisson equations, this paradigm can exceed the accuracy of state-of-the-art end-to-end architectures, with up to an order of magnitude higher throughput and memory savings that fit models nearly two orders of magnitude larger on consumer-grade GPUs. At inference, latent rollouts run up to an order of magnitude faster than regular rollouts, and chunked inference processes functions of over half a billion points on a single consumer-grade GPU.

Most memory systems for LLM agents are retrospective, preserving and retrieving past information to inform current decisions. Proactive agents, however, also require prospective memory: retaining future-directed intentions and realizing them when appropriate opportunities arise, even without explicit requests. Crucially, remembering the right intention does not guarantee acting at the right time: a related observation may call for waiting, while a later cue may make the same intention actionable. Inspired by cognitive theories of human prospective memory, we introduce Pro-Memory, a memory framework that separates maintaining future intentions from deciding when and how to realize them. Pro-Memory preserves pending objectives linked to their supporting evidence, assesses new observations against their original goals and activation conditions, and coordinates appropriate contributions with the agent's ongoing actions. It further updates intention states based on actions actually issued or evidence of resolution, while an experience memory distills feedback from completed episodes into guidance for future decisions. Experiments on two benchmarks spanning delayed-intention tasks and open-ended proactive assistance demonstrate consistent improvements in proactive task performance over representative memory approaches across multiple LLM backbones, while limiting unnecessary interventions.

Two foundational aspects of human cognition, abstractness and gender, are typically examined as separable. Here, we find they are systematically linked throughout the entire space of concepts spanning single (N = 40,000) and multi-word expressions (N = 67,000): across billions of words and millions of images on the internet -- and across numerous measurement techniques -- more abstract concepts are more associated with men, an effect which replicates in languages around the world. This pattern not only holds across conceptual domains, but it also maps onto grammatical structure: verbs and adverbs are more abstract and more male-coded, a system-wide trace of stereotypes accentuating agentic qualities in men. Furthermore, this bias emerges in human and AI responses in free association tasks, indicating its internalization and its algorithmic reproduction. The male abstraction bias may underlie and unify a range of stereotypes, evidencing gender as a core organizing dimension in conceptual space.

With scalable Laplacian solvers now available, the next frontier is to exploit them more effectively for high-quality hypergraph partitioning. In this work, we revisit spectral hypergraph partitioning from the perspective of spectral embedding geometry. Through a systematic analysis of circuit hypergraphs and controlled synthetic (hyper)graphs, we identify a ray-like embedding structure that helps explain the effectiveness of projection-based partitioning. We further show that the geometry of spectral embeddings can be systematically influenced through Laplacian design and hyperedge reweighting. Guided by these insights, we propose FRESHPart, a feedback-reshaped spectral partitioning framework that refines the spectral embedding through Geometry Feedback and Search Feedback, producing embeddings that are better aligned with projection-based partitioning. On the Titan23 benchmark, FRESHPart achieves average cut reductions over hMETIS of 13.09% at $ε=2\%$ and 25.21% at $ε=20\%$, surpassing the strongest previously reported results (12.16% and 22.53%, respectively), while completing every benchmark in under 120 s on a single GPU.

Extreme events often occur sequentially, from earthquake-tsunami cascades to successive extreme losses in financial markets. We introduce an extremal token representation that encodes multivariate observations as a categorical sequence. A quiet token records the absence of threshold exceedances; otherwise, the token identifies the exceeding component with the largest marginal probability-scale score. The representation is invariant to componentwise strictly increasing transformations. It preserves the identity and temporal ordering of dominant exceedances but discards their magnitudes and the identities of nondominant simultaneous exceedances. A decoder-only transformer estimates the history-dependent next-token kernel on this representation, and recursive sampling generates future token sequences from the induced joint predictive law. We amortize our generative model by pretraining across simulated process families with varying parameters and dependence persistence. Next-token predictions are evaluated against exact conditional probabilities for held-out categorical processes and particle filter references for heavy-tailed vector autoregressive processes. Applied to ten-second returns for five semiconductor stocks, the pretrained model supports zero-shot transfer, while a few epochs of adaptation improve agreement between observed and generated token frequencies and transition patterns. These results support task-specific tokenization and simulation-based pretraining for modeling which dimensions dominate multivariate extremes.

Dantzig-Wolfe reformulation strengthens the linear relaxation of a mixed-integer program by convexifying a chosen set of constraint families, and every candidate costs a column generation run. The choice is hard when families strengthen the bound only together, because a search that adds one family at a time then stops at the linear relaxation. The classical inequalities of lot sizing, network design and facility location rest on such zero-marginal groups. We prove that when each evaluation convexifies at most $s$ of $k$ families, any method that observes only bounds needs a covering number of evaluations, of order $k^w$, to find a hidden group of $w$ families, also on actual Dantzig-Wolfe relaxations. A prediction that misses $e$ members reduces the cost to $Θ(k^e)$, through one staged search that does not need to know $e$, so the evaluations a prediction saves measure what the predictor knows. We study two sources of predictions. A one-pass lookahead at an optimal LP dual is provably blind to groups whose families share no variable, and a language model reads the names and text of the model. JointDW verifies the prediction of the language model with the staged search and interleaves it with the lookahead at a bounded loss. Groups that strengthen the bound only jointly appear in a quarter of the MIPLIB-NL models whose bound can be improved. There the strongest language models identify them about as often as the lookahead, their predictions rest on the names for textbook models, and reasoning about doubles the accuracy of the same model. On larger models with distracting families, the lookahead finds the group in one evaluation of three families. Language models given only the text seldom name it, while one reasoning model given every coefficient names it in all but one answer and then needs fewer evaluations than the lookahead for pairs of families.

Speculative decoding accelerates large language model inference with a small drafter, whose per-guess acceptance rate determines the achievable speedup. Feature-level drafters share a design choice: every token, easy or hard, is drafted under a single forward pass. Many of the tokens the target rejects are ones the drafter is itself unsure about yet the target answers decisively. To address this, we introduce Parallel Reconsideration: guided by its own uncertainty, the drafter reconsiders its hesitant guesses with extra passes that run in parallel with the one-pass draft, adding negligible latency. A co-designed training recipe and serving-time drafting algorithm append these reconsidered guesses to the original draft tree as refined candidates. Across models and benchmarks, on average, Parallel Reconsideration improves over the baselines by 5.3% to 10.2% on acceptance length and by 3.5% to 9.2% on speedup.

Search can help LLM agents correct mistakes, but their starting conditions may shape which evidence they find and use. Drawing on cognitive-bias paradigms, we develop a controlled search harness to examine result order, records of prior work, and initial hypotheses while holding each question and document collection fixed. Matched controls distinguish order effects from differences in document composition and the effects of failed work from context length. Across 10 models and 1,000 synthetic search worlds, we track queries, retrieved documents, document openings, and final answers. Our comparisons suggest that final answers are more sensitive to initial evidence order during active search than when all evidence is supplied in full. Dead-end histories can reduce strategy switching relative to matched neutral histories, indicating an effect beyond context length. Under wrong hypotheses, stage measurements distinguish reduced retrieval or opening from errors that persist after counter-evidence is opened. Exploratory interventions show that opening counter-evidence more often can coincide with lower accuracy, while accuracy can improve with little change in opening.

Text embedding inversion reconstructs source text from fixed-dimensional vectors, posing a privacy risk to systems that store embeddings instead of raw text. We adapt the vec2text framework (Morris et al., 2023) to support text-embedding-3-small and mistral-embed alongside GTR-base. We evaluate 800 samples per category across general instructions, Wikipedia passages, paraphrased instructions, health-related instructions, and short literal code snippets, using a common 60-80-character range and paired original and paraphrased instructions. Under the evaluated vec2text configuration, text-embedding-3-small achieves the highest exact-match rate in all five categories, including 47.4% on health-related instructions and 24.5% on general instructions under the evaluation's case-insensitive, whitespace-stripped matching rule. BERTScore F1 exceeds 87 on the 0-100 scale for every prose category and embedder, indicating high contextual similarity to the references. Short code snippets yield lower reconstruction scores. These findings characterize one attack configuration and the selected datasets; they do not establish attack-independent rankings or sensitive-attribute recovery rates. They nevertheless motivate explicit assessment of inversion risk and access controls for stored embeddings.

Probabilistic traffic forecasting is essential for reliable Intelligent Transportation Systems (ITS), as it captures uncertainties in complex road networks. Recent diffusion- and flow-based generative models have facilitated uncertainty modeling through a multi-step sampling process. However, beyond slow inference, this process is designed to capture complex multimodal distributions, which misaligns with the limited complexity of the conditional traffic data distribution, causing models to underperform deterministic baselines. To address this, we introduce the \underline{Co}nsistency-\underline{Re}laxed \underline{Flow} Map (CoReFlow) model, presenting the first effort to leverage the one-step generation capability of flow maps, known as solution operators of probability-flow ordinary differential equations. Besides, we improve accuracy using a concise training objective that supervises the model via interpolation endpoints over finite time intervals. Unlike existing methods (e.g., self-distillation and MeanFlow Identity) that maintain trajectory consistency to model multimodal distributions, our objective leverages conditional mean to restrict the consistency, thereby achieving higher accuracy in the probabilistic traffic forecasting task. Extensive experiments on four real-world datasets show CoReFlow achieves state-of-the-art forecasting accuracy. It surpasses recent baselines, provides competitive uncertainty modeling, and reduces inference costs compared to existing generative baselines.

Facial Expression Recognition (FER) on optomyographic (OMG) smart glasses such as OCOsense enables privacy-preserving, camera-free affect sensing for health and well-being. Most FER systems are fully supervised, i.e., they rely either on hand-crafted expert features with classical machine learning, or on end-to-end deep models trained from labelled data. Both depend on large quantities of labelled expressions, which are costly to collect and limit deployment in real-world, personalized settings. We investigate Self-Supervised Learning (SSL) as a route to label-efficient FER on these glasses. We pretrain a Vision Transformer with a masked-autoencoder objective on unlabelled OCOsense recordings and finetune it for five-class FER under person-independent cross-validation. Across label budgets from 5% to 100%, pretraining consistently improves performance; with only 5% of the labels, the pretrained model outperforms the from-scratch model trained with all labels. Our results show that SSL pretraining is a practical way to reduce labelling cost for smart eyewear FER, and we discuss the implications for future personalization and continual adaptation.

LLM agent distillation transfers agentic capabilities to student models by training on complete teacher trajectories that typically interleave thoughts and actions. We question how critical thought supervision is for agents' ultimate action performance. To answer this, we disentangle their impacts by independently perturbing teacher thoughts and actions during distillation. On widely adopted agent benchmarks across model scales, altering or discarding thoughts has little effect on student performance. We find, however, that thought supervision becomes important when information about the thought is not disclosed in the corresponding action during training. By identifying such discrepant cases, our findings challenge the common practice of distilling entire teacher trajectories for LLM agents, and offer a new insight that thought supervision can be a discernible choice for efficiency depending on whether the information of thought is disclosed in the action.

Hopfield neural networks and noncooperative games provide seemingly different approaches to understanding how collective outcomes emerge from local interactions. In a neural network, individual units adjust their states according to local signals; in a game, players choose actions to maximise individual payoffs. I establish an exact equivalence between symmetric binary Hopfield networks and binary-action games that admit a quadratic potential. The equivalence identifies stable network configurations with pure-strategy Nash equilibria, and asynchronous neural updating with best-response dynamics. More broadly, the storage and retrieval of memories in neural networks correspond, respectively, to the design and selection of equilibria in games.

Structural Causal Models and Bayesian Networks are the dominant framework for representing causal relations. They are fundamentally type-level formalisms since their estimands, Average Treatment Effects and Conditional Average Treatment Effects, are properties of distributions over populations, not of individual units. When treatment effects exhibit unbounded heterogeneity, these estimands collapse. What persists is the structural, token-level core of causation. We argue that Petri Nets under the Individual Token Philosophy are the natural representation of these token-level causal structures. We develop a formal translation from Bayesian Networks to Petri Nets, providing a two-layer account of individualized causality. The PN layer provides a structural representation of causal mechanisms that does not rest on distributional assumptions. In turn, a conformal prediction layer quantifies causal effects with finite-sample coverage guarantees. Together they constitute a complete account of individual causal effects.

Qualitative data can tell software teams what their stakeholders need, yet rigorously analyzing thousands of qualitative responses takes expert time that few teams can afford or justify. Teams therefore often face a false choice: leave the responses unanalyzed, or hand them to a large language model and accept its interpretation and judgment. In this paper, we take a pragmatic approach to AI-assisted qualitative analysis and present TACT (Traceable AI Council for Thematic analysis), a procedure that enables such analysis at scale while keeping teams in control of interpretation and judgment. TACT comes in two variants: TACT-Lite, the core AI council procedure, and TACT-Full, which adds adversaries (models from other model families that critique each model's output) and a decision model that surfaces high-uncertainty code assignments. We evaluated TACT on an independent dataset of 256 Stack Overflow threads qualitatively coded by other researchers. Against their labels, TACT's candidate codebook achieved an F1 of 0.75-0.77 and its final coding an F1 of 0.71-0.72. On a held-out sample, agreement between our own manual coding and TACT's coding was Cohen's $κ$ of 0.79-0.85. TACT is also economical: model calls cost \$13.77 (TACT-Lite) to \$33.59 (TACT-Full) per 100 threads. Finally, drawing on our experience building and validating TACT, we identify pitfalls in AI-assisted qualitative analysis, describe safeguards to address them, and offer guidance for teams and researchers applying TACT in practice.

Background: The automatic identification of tree species from bark photographs is of practical interest for forestry inventory and the wood-processing industry, yet the design choices that govern the accuracy of deep-learning approaches have not been studied systematically, and some prior results may be inflated by leakage of images of the same trunk across training and test partitions. Methods: A systematic ablation study is conducted on the public BarkNet 1.0 dataset (20 species) under the established tree-level, leakage-safe partitioning scheme. Using ConvNeXt-V2 backbones, the following factors are varied in turn and evaluated by five-fold cross-validation: patch size, the strategy by which per-patch predictions are aggregated into an image-level decision, the training regime (single- versus two-stage), backbone capacity, the whole-image baseline, and the number of patches available at inference. Results: Per-patch accuracy rises monotonically with patch size (from 75.6% at 160 px to 89.6% at 512 px), and image-level accuracy rises in step, plateauing near 94.9% around 512 px. Learned aggregation exceeds majority voting by at most 0.7 pp, an advantage that shrinks as patches grow and attention becomes near-uniform. Backbone capacity (pico to tiny) alters accuracy by less than a point, and two-stage training offers no advantage over one-stage. Whole-image classification (91.9%) trails patch-based aggregation, since it admits no bag to average over; accuracy increases with the number of patches at inference, saturating around 64. Conclusions: The study yields a simple recommended configuration: patches as large as possible while leaving enough to average over (around 512 px here), with a nano backbone. It also quantifies the second-order role of learned aggregation and model capacity, providing a reproducible reference point for practitioners and future research

Vision-language models (VLMs) have demonstrated significant advances in reasoning with visual data, which is particularly important given that much of human knowledge is encoded visually in scientific articles and other sources. However, a gap remains in understanding how artificial agents, powered by VLMs, perceive and navigate decision-making under uncertainty in the presence of visual data. This study explores this gap by evaluating and analyzing autonomous decision-making processes. Building upon prior work in visualization literacy and decision theory, we conducted two complementary studies: one examining how agents perceive and quantify uncertainty encoded in visualizations, and another investigating how artificial agents make decisions based on visual data in the form of images. We analyzed perception accuracy and decision performance across models and visualization types. Our results show that GPT-4o demonstrates the highest accuracy in perceiving uncertainty in pie and bar charts among the tested models. For decision-making, autonomous agents showed the best alignment with rational decision theory when uncertainty was represented through circle charts, icon arrays, or text.

Soft Actor-Critic (SAC) is a widely used deep reinforcement learning algorithm that combines entropy regularization with an actor-critic architecture. Standard implementations rely on one-step temporal-difference learning of Q-functions, yet alternative critic design choices remain underexplored. In this work, we revisit three such choices: learning soft advantages instead of soft Q-values, the use of multi-step return targets, and the training of multi-step critics from replay buffer data. Keeping the policy objective fixed, we compare the six critic variants these choices generate: (i) advantage and (ii) dueling-advantage critics learned from one-step transitions, and $λ$-return and soft Generalized Advantage Estimation (GAE) critics learned either (iii, iv) on-policy or (v, vi) from replayed trajectory segments. Each variant is evaluated across an entire random hyperparameter search on fourteen tasks, treating the hyperparameters as part of the algorithm. Learning soft advantages instead of soft Q-values does not improve performance, and the soft GAE critic adds nothing over the simpler $λ$-return critic. What does improve performance, most clearly on continuous control, is training a $λ$-return critic on replayed trajectory segments, even without importance weighting or trace cutting. This extends earlier evidence that uncorrected multi-step targets work well under replay to the full soft $λ$-return, entropy terms included. By design, the underlying ideas studied here extend beyond SAC, and our empirical findings aim to encourage their application to a broader class of reinforcement learning algorithms.

Software developers increasingly adopt LLM-based agentic coding tools to assist them with development tasks. To provide developers with up-to-date information on these tools' effectiveness, prior work introduced benchmarks that enable comparisons across tools. Despite the widespread adoption of such benchmarks, however, recent work has shown that they suffer from issues threatening their validity. Aims. In this work, we aim to complement prior work by analyzing the construct validity of agentic coding benchmarks. Methodology. We analyze nine benchmarks commonly used to evaluate the effectiveness of agentic coding tools. We identify the constructs that the benchmarks claim to measure, the metrics used to operationalize them, and potential confounders that interfere with the measurement. Results. Our findings show that benchmarks commonly operationalize broad concepts such as coding capability with only a small set of metrics. The reported attributes that might confound the measurements vary between benchmarks, indicating a lack of standardization. Conclusions. We conclude with recommendations to strengthen construct validity in agentic coding benchmarks and encourage further empirical work to better understand the effectiveness of agentic coding tools.

A neuro-symbolic system should combine only relevant answers from expert agents or other knowledge sources. Using only relevant experts for a given input query becomes ever more important for accuracy and scalability as the knowledge network grows and connects ever more knowledge sources. We extend the standard definition of relevance from the legal theory of evidence: An answer is relevant if and only if it is both probative and material for the matter at hand given the knowledge available. We apply this relevancy scheme to mixing the feedback fuzzy cognitive maps (FCMs) of real or artificial experts. The expert FCM directed causal graphs mix or combine into a common knowledge network. We compare the equilibrium limit-cycle predictions or answers of the experts to the answers of the mixture FCM and then estimate the probative status of answers by their difference or variation. A new Bayes-like theorem de-mixes the knowledge mixture into corresponding posterior-like FCMs that also allow relevancy-based comparisons. The result is a final set of relevancy mixing weights that produces a relevancy-mixed FCM knowledge network for the given query. So different queries produce different relevancy-mixed FCMs. We demonstrate this relevancy approach using generated synthetic FCMs from an actual 13-node expert FCM that models the causes and feedbacks of ischemic-stroke risk.

Evaluating open-ended agent tasks increasingly falls to an Agent-as-a-Judge: an evaluator handed tools and the agent's full trajectory so it can verify a result against the environment. The premise is that a judge that gathers its own evidence should be grounded by it. We show that this grounding does not immunize the judge against confirmation bias: a single false pre-screening note, attributed to an authority and carrying no evidential bearing on the correct verdict, overturns up to 40% of a tool-using judge's correct verdicts across four judges and five datasets, even though the judge collects decisive evidence. The shift is driven by specific properties of the cue. We treat this cue as a controlled diagnostic of confirmation bias rather than an adversarial attack. Our central contribution is to explain how confirmation bias changes the judge's behavior in a trajectory. We classify changes in judge behavior and attribute changes to one of three failure modes: over-trust (verify, conclude correctly, then defer), hallucination (fabricate support for the flip), and under-investigation (stop at the first corroborating signal). For the two Claude judges, the failure is predominantly over-trust (67-75% of harmful flips), while smaller judges have error modes spread across all three. Lightweight mitigations blunt but never remove the bias.

Authentication solutions are the key design, architectural, and implementation decisions that implement authentication security controls. An example is a password-based authentication solution. While researchers often examine authentication quality, existing approaches do not distinguish which effects on quality stem from which aspect of authentication. For example, multi-factor authentication solutions tend to be less usable because they combine multiple authenticators, whereas passkeys are more secure than passwords because of properties inherent to the authenticator itself. This paper presents AQUA, a modular quality modeling framework for authentication solutions. AQUA's core idea is to decompose authentication into four interrelated concepts, enabling modular specification of authentication quality and tracing quality effects back to their source. We demonstrate AQUA with initial definitional quality models that address the levels of authenticators, authentication techniques, and authentication solutions. We further illustrate the interrelations between these concepts through concrete examples and a usability metric, showing how quality propagates across the concepts and ultimately affects the overall quality of an authentication solution. Our results provide a foundation for future work, including designing assessment and prediction models for authentication solutions and automated quality assessments.

In this work, we present adaptive domain decomposition finite basis physics-informed neural networks (ADD-FBPINNs). The proposed method integrates Partition of Unity Networks (POUNets), a data-driven approach for learning domain partitions, with finite Basis physics-informed neural networks (FBPINNs) to perform more efficient domain decomposition and achieve more accurate approximations. FBPINNs improve upon the accuracy of standard physics-informed neural networks (PINNs) through user-defined subdivisions of the computational domain. By adapting the domain decomposition based on the solution's features using a separable POUNet, the framework can automatically learn partitions that are stable and can be geometrically interpretable. We test this approach on the two-dimensional Helmholtz equation, two-dimensional Poisson equation with multiscale features, and viscous Burgers equation and observe improved accuracy compared with FBPINNs with fixed domain decomposition.

A central challenge in assistive brain-to-robot control is decoding a user's latent intent from complex and noisy brain signals. Prior approaches evaluate decoders as single-trial classifiers with accuracy-based metrics, but this neglects the inherently sequential nature of assistive control. Since a planner must integrate intermittent, noisy decoder outputs over time, single-step accuracy can only be a proxy when predicting downstream task performance. We formalize brain-to-robot control as a partially observable Markov decision process (POMDP), and prove that there is a sharp phase transition: regret is uniformly bounded across the horizon as long as pairs of goals are all decoder-distinguishable, but grows linearly otherwise. We show that a single channel statistic, the worst-pair Chernoff information $Γ_{\min}$ sets this boundary. Then, we derive a closed-form regret predictor $U_{\mathrm{id}}(C)$, that can be computed from any confusion matrix in $O(K^2)$ time. Across two domains spanning orthogonal information channels, we show that $U_{\mathrm{id}}$ predicts cumulative planning regret beyond what accuracy can explain. We then highlight a pathological accuracy trap, in which two decoders have identical validation accuracy 35.1% on real EEG data (BNCI2014001 subject 5) but result in a difference of $9.6\times$ in held-out regret, which $U_{\mathrm{id}}$ predicts. Same-signal control confirms $U_{\mathrm{id}}$'s advantage with decoders that span independent feature channels. Ultimately, our work provides a drop-in diagnostic to evaluate and select sensors and decoders in sequential assistive control.

Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data. We develop a framework that connects these tasks through an energy-weighted optimal-transport plan under a specified similarity geometry. The same plan defines a dictionary-level discrepancy, decomposes it into latent contributions, and quantifies how each latent's mass is distributed across counterparts. We characterize relationships between decoder, activation, and first-order causal similarities; idealized models clarify the interpretation of splitting and absorption. For fixed trained SAEs, we derive joint asymptotic inference for transport costs and fractions, accounting for activation similarities and both energy marginals estimated from the same documents. A joint document bootstrap gives asymptotically valid simultaneous confidence bands. Experiments across four language models locate cross-seed discrepancy and separate context-dependent changes.

Electroencephalography (EEG) measures brain activity through recording electrical potentials of the scalp. EEG recordings capture two distinct sets of characteristics in the signal: (1) \emph{Traits} -- information that remains relatively constant over the course of a recording session, such as subject identity, age, and sex; (2) \emph{States} -- signatures that encode responses to external stimuli, including visual, auditory, and motor information. There is limited understanding of how different EEG modeling features (signal processing and foundation models) encode state and trait characteristics. In this work, we quantify state- and trait-level features captured by EEG embeddings using a hierarchy metric and kNN purity on Brain-Computer Interface dataset trials, and find that a majority of EEG features form a macro-trait hierarchy (embeddings form trait-level clusters with micro-states within). Our linear probe results indicate that this hierarchy structure leads to suboptimal state distinguishability, negatively impacting BCI performance of EEG foundation models. To mitigate this, we explore normalization techniques that reduce trait-specific information in embeddings and provide a $\sim$ 10\% improvement in balanced accuracy for state classification. Additionally, we observe that when fine-tuning models for downstream state decoding, embedding spaces transform from macro-trait to macro-state hierarchies, which in turn lead to better state decoding.

Published explanations of why long-horizon agents fail are commonly taxonomies labeled over completed trajectories: a judge reads a transcript and names the failure. We argue that a claim attributing an agent's behavior to its history should carry intervention-verified evidence, per instance or via a validated estimator: freeze an executable mid-trajectory state, alter only the history the agent sees, and continue - executed rather than inferred from the transcript, at a state whose restoration has been measured, with a replay certificate matched to the estimand. Each of the four manipulation classes we name has a precedent, and several methods already execute interventions from checkpoints, re-executed prefixes or recorded conversation states (for example DoVer, REFLECT, Min et al., Who&When Pro, Causal Agent Replay), but none reports a measured replay certificate. Two cautionary measurements, neither applying the standard, follow: a carrier taxonomy whose "concentration" could not be separated from benchmark construction (adjusted odds ratio 1.15, interval crossing 1), and hash-identical snapshots that, restored into identically seeded processes, agreed with one another in 24 of 24 states while differing from the live continuation in 9 and 20 of 24. Restoring the captured position of a random stream made all 24 agree with the live run.

Large-scale ranking and recommendation models of Web platforms benefit from increased compute much as large language models do, and their scaling behavior is broadly power-law in form. Yet ranking models have heterogeneous architectures which makes their scaling law less straightforward than that of LLMs. We address how to discover the scaling law of an arbitrary ranking model efficiently. We introduce a scaling layer, which groups primitive architecture parameters into semantically coherent groups and defines how they co-vary, inducing a mapping from a low-dimensional space of hyper-hyperparameters into the full architecture space. On top of it, Axia fits scaling-informed surrogates, selects configurations by a posterior-based acquisition criterion, and returns a Pareto frontier from which per-model operating points are validated -- replacing exhaustive sweeps with a small number of informative runs. A fleet-level evaluation protocol, computed on a fixed representative model set under matched trial budgets, makes refinements to either component measurable. Our contribution is an automated system that learns model-specific complexity-quality relationships and gives uncertainty-aware guidance for allocating compute across a heterogeneous model ecosystem. Across 34 model iterations in two development cycles on Meta's ads ranking and retrieval system, the workflow yielded a mean 0.66% offline normalized-entropy improvement and 0.36% cumulative online lift.

Speaker-recognition benchmarks test how accurately a model recovers the speaker label of a recording, treating voice identity as a property of the person who produced it. However, these benchmarks do not evaluate how humans perceive voice identity. Whether a listener hears two voices as the same person has consequences for intellectual property rights, cybersecurity, and everyday life. We introduce the Voice Identity Perception benchmark (VIPBench), which contains 124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs from 100 public figures. The pairs include real recordings, voice clones generated by a text-to-speech system, and morphs between two speakers' voices. We also define four tasks that evaluate the alignment between speaker-representation models and human identity perception. The first two tasks measure how well a model's similarity score predicts the proportion of listeners who respond "same" and the majority response. The other two measure whether a model and listeners find the same speaker pairs similar, and whether a model's relationship to listener judgments holds equally for real and synthetic recordings. We evaluate eleven publicly available speech models. Listener judgments diverge from speaker labels most often for voice clones: in 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people. No model reaches the noise ceiling; the best model explains 61% of the explainable variance. Models trained with speaker labels align with listeners better than self-supervised models do, yet a model with a 12% error rate on speaker labels aligns with listeners as well as models with error rates below 2%. VIPBench enables the evaluation of speaker models against human perception on real, cloned, and morphed speech, and our results motivate training speaker models directly on listener judgments.

Speech transcription errors can be difficult to recognize when competing word hypotheses are unfamiliar or visually similar. We describe an interface that combines alternative transcripts, replay and contextually checked pictograms for selected uncertain concepts. The system will initially support French and Spanish functional utterances. A counterbalanced study will assess whether pictorial feedback helps listeners recover the speaker's intended action beyond text alternatives alone. We will also quantify semantic mismatch and additional interaction time, since an inappropriate image can reinforce the wrong interpretation rather than repair it.

The memory of production agents is increasingly becoming a bottleneck and is built almost exclusively from text: key points are extracted from out-of-limit context and retrieved into context at query time. Therefore, every new query incurs an additional cost that grows with the accumulated history. Weight memory is a complementary alternative that internalizes knowledge into model parameters and answers with no context at all, and has recently proven effective in self-evolving agents for internalizing skills and behaviors. However, its application in episodic and semantic memory, such as user-specific facts, remains comparatively unexplored, and a gap remains between earlier attempts at weight memory and text-based systems regarding recall performance. We study how to make weight memory competitive as a long-term factual store by applying on-policy self-distillation (OPSD). It trains a context-free student from a teacher, which is the same model conditioned on the user's memory, together with a verify-then-regenerate router that lets weight and text memory each cover the queries they handle best. On the HaluMem benchmark, OPSD matches or surpasses strong text-memory baselines, including full-context prompting and retrieval systems, while reading fewer tokens per query. We further conduct an extensive study that characterizes the behavior of weight memory and analyzes how different training recipes affect its performance. As agent harnesses move toward persistent, multi-session deployments, these results position weight memory as a practical, low-cost component of agent memory and provide insights into feasible training and evaluation protocols for combining it with retrieval.

Many temporal approaches for video recognition rely on explicit cross-channel aggregation projections, which have quadratic complexity in the number of channels, introducing significant computational overhead. In addition, the spatio-temporal representation used for video recognition is biased either toward local order interactions (for CNNs) or toward global order interactions (for transformers). Starting from the multi-order spatial representation of MogaNet, our method introduces two complementary contributions to tackle these problems. First, for the temporal modelling, we introduce parallel multi-order processing with depthwise-only convolutions and implicit gating-based aggregation, which is linear in the number of channels, inducing a considerable computational reduction. Second, we propose a dual interaction between spatial and temporal dimensions, which allows a better spatio-temporal representation without increasing the computational cost. By combining these two contributions, our proposed VideoMogaNet achieves performance on par with or better than state-of-the-art action recognition approaches across multiple video benchmarks (UCF101, HMDB51, Kinetics-400, Diving48), while considerably reducing the cost of the temporal branch.

Autonomous software agents can act through tools, delegate work, and change their models or execution environments while earlier authorizations remain in flight. We present Assurance-Conditioned Continuous Authorization (ACCA), an approach designed to be harness-independent, for maintaining operational authority according to currently justified assurance. It distinguishes evidence validity from contextual applicability, derives assurance claims from eligible evidence, and resolves a multidimensional effective-authority envelope within a principal-declared ceiling and applicable policy. A typed dependency graph supports selective re-evaluation of authorization-relevant changes. An authority epoch changes when effective authority changes; a distinct optional authorization-basis version supports revalidation of action-specific authorization conditions. Protected effect-time enforcement can reject stale actions before a consequential resource change, subject to the available commit mechanism. A compositional Domain Governance Profile expresses domain-specific semantics without granting authority. Controlled scenario tests implemented as a regression suite using a trading profile exercise default-deny rules, evidence-conditioned contraction and recovery, policy divergence, selective non-response, and stale-action rejection at a local gateway before the simulated-broker boundary. The study establishes implementation-level evidence in one domain, not a cross-domain guarantee of harmful-behavior prevention.

On Replit, our agentic engineering platform, users now build entire products without writing code, through successive feature requests to an agent. Yet existing coding benchmarks evaluate single, fully specified tasks, and many are near saturation. We introduce ViBench 2.0, a benchmark for long-horizon software building that we use to select our platform's models. Its 17 full-stack web products, modeled on widely used software, grow over 241 feature requests and are judged only when finished, as a user would judge them: a browser agent runs 544 test plans, and a build passes only if every plan does, under a grading protocol validated to keep this strict score reliable. Evaluating nine models, we find a reliability gap: the three strongest pass up to 98% of test plans yet deliver a fully working product in at most 65% of builds, and no open-weight model delivers one. The gap is closing fast: across three releases of one model family, pass@1 on the public products rises from 3% to 53% while the cost per build falls by more than half. We release eight products with the harness and grader, hold out nine, and report lessons from production use.

Training on large datasets motivates selecting which examples receive expensive training computation. Fixed quotas determine the amount of pruning in advance, while absolute confidence cutoffs depend on the score scale. We study a population-relative alternative that adapts the threshold and retained count together. We introduce LACS (Label-Agnostic Confidence Selector), a training-time rule based on maximum predictive confidence, requiring neither labels at selection time nor per-example gradients. LACS removes examples at or above a log-mean-exp reference plus a margin. We characterize its sensitivity to score gaps and mass, and derive an exact no-pruning condition. For $0<ρ<1$, the removed fraction is at most $ρ$ and at least one example remains. The unsigned score permits both confidently correct and confidently wrong examples to be removed. Seven evaluation settings include neural training, language modeling, and boosting. Neural execution studies reduce backward-active rows by 10-35%; on CIFAR-10, 29% fewer backward rows accompany accuracy means of 89.77% versus 89.82% for full CE. At a matched backward budget and number of optimizer updates, LACS is 1.0 pp above small-batch training and 1.2 pp above random selection over five seeds. Across 14 tabular datasets, quadratic boosting reduces mean test error from 23.36% to 22.96%. Its full training replays use about 30% fewer fitting rows; these counters span a different horizon from validation-selected accuracy. In one RBF-SVM case study, elapsed time falls by 37% with slightly lower error. These findings separate selective-work reduction from realized acceleration: the current exact-current-score neural implementation uses an additional scoring forward, so its backward sparsity does not yet translate into wall-clock acceleration. This overhead is an execution choice rather than a requirement of the selection rule.

We present the Whiel (a contraction of Whi(le-R)el) system for Lean-certified reasoning about imperative relational programs. The system can encode containment and equivalence problems for Datalog, its extensions with negation, and the core fragment of graph query languages. It can also represent optimized or incremental evaluation algorithms that are not expressible in the original query language, as well as first-order specifications of integrity constraints, schema mappings, and views arising in data integration, exchange, and sharing. Whiel's program states are database instances, and its imperative structure enables compositional Hoare-style reasoning with first-order relational specifications, using prophecy variables and leveled invariants to support inductive arguments. Whiel supports neurosymbolic discovery of loop invariants whose verification conditions can be checked using the Vampire first-order theorem prover. Counterexamples are fed back into the neurosymbolic pipeline, while proposed solutions are translated into Lean proof certificates. To validate the Whiel system, we evaluate its performance on a curated benchmark suite. In summary, we envision the Whiel system as a tool for tackling mathematically demanding, typically undecidable problems involving complex query and constraint languages, with a high degree of trust in its output.

Speculative decoding (SD) speeds up large language model (LLM) serving, but its gain depends on the workload's \emph{shape}, such as batch size and context length, and on its \emph{content}. A model-based drafter proposes a short block on any content, while a model-free drafter such as n-gram lookup proposes long copies at almost no cost when the output repeats its input, as in code editing. Adaptive verification spends a batch-wide verification budget on the draft tokens most likely to be accepted, but it supports only one model-based drafter. Model-free drafters provide no per-token confidence, so they cannot take part in budget allocation, and deployments must either fix one drafter or split their GPUs between drafters. We present \sys, a framework that runs a model-based and a model-free drafter on every step and gives each request the better of their chains, within adaptive verification and entirely on the GPU, which keeps asynchronous scheduling intact. The key idea is a price per verified draft token: the goodput that slot would earn elsewhere in the batch. Once the model-free drafter's tokens carry calibrated acceptance estimates, fit once on general text, this price puts both drafters' chains on one scale. Instantiated with \dspark and an n-gram drafter in vLLM, \sys achieves 1.10--1.62\x the output throughput of \dspark on code editing at batch size~4 and 1.04--1.27\x at batch size~128, while staying at 0.97--1.01\x of it on chat at temperature~0. A single deployment can therefore serve both workloads.

Multi-teacher on-policy distillation (MOPD) extends on-policy distillation (OPD) to consolidate capabilities from multiple teachers. Although prior work reports near-perfect in-distribution (IND) distillation performance, we find that this can mask incomplete capability consolidation beyond the supervised distribution. A preliminary study on synthetic reasoning shows that OPD does not fully transfer teachers' generalization capabilities. We then extend this investigation to real-world tasks, evaluating extrapolative and compositional generalization across mathematical reasoning, coding, agent, and long-context capabilities using Qwen3 models. Across extrapolative settings, OPD and MOPD reach 96.04-102.73% of their teachers' IND scores. Yet their out-of-distribution (OOD) scores remain 2.45-16.25% below the base model's. Across three compositional settings, MOPD improves IND performance over the base model, but its performance on composed tasks as OOD remains 18.19-24.34% below that of the best teachers. To understand these gaps, we examine teacher capability and data coverage as two potential bottlenecks. Jointly alleviating these bottlenecks eliminates OOD forgetting in our experiments and suggests the potential for generalization beyond the teacher. (M)OPD also reduces OOD forgetting on average relative to its reinforcement learning (RL) baselines. Analysis of the associated update geometry reveals smaller and more spectrally concentrated updates on average than those of the RL baselines, along with greater shared update structure under MOPD. Motivated by these findings, we propose geometry-based patching and base-preserving replay to improve OOD performance while largely preserving IND performance. Our findings identify a blind spot in capability consolidation via (M)OPD, highlighting the need for algorithms that explicitly support OOD retention and generalization.

Variational approaches to learning transport maps for multimodal sampling can suffer from mode collapse. One way to mitigate this problem is to learn transport along a prescribed path of intermediate distributions defined by geometric annealing. We show that annealing can take a substantial detour in Wasserstein space: when mass moves between modes, the path can be much longer than the distance between its endpoints. We further link path geometry to transport learning, showing that long annealing paths can slow population optimization, while sufficiently regular paths permit rapid convergence. Motivated by these findings, we propose componentwise annealed transport (CAT) to reduce the need for mass transfer between modes. Starting from a mode-informed mixture reference, CAT constructs an exact mixture decomposition of the target and learns separate transport maps along componentwise annealing paths. We estimate the component weights separately and show that these estimates remain consistent even when the learned flows are inexact. We also develop adaptive annealing schedules guided by transport effort. Experiments on multimodal synthetic distributions and Bayesian inference problems demonstrate improvements in transport learning and sampling efficiency.

A deep RL agent trained on a fixed persona reward plays one way: an Explorer keeps exploring while starving, a Survivor hoards on an empty map. Since each persona fixes its own reward, automated playtesting trains N agents for N styles. Changing reward composition mid-episode runs into the switching-stability dilemma: PPO degrades under non-stationary rewards. We present RIDGE, a single-agent architecture that sidesteps the dilemma with two design choices: a state-conditioned blender that computes softmax-normalized sigmoid weights from slow-moving state variables, and a multi-head critic that trains one value head per persona on its own stationary signal. On Crafter with four personas, one \ridge agent acquires both branches of the wood-tier crafting/combat split that no fixed-persona baseline solves jointly, and a sharpness ablation confirms state-conditioning drives the gain. We position these results as demonstrations of the blending mechanism's dynamics across n=3 seeds per condition; larger multi-seed sweeps remain future work.

Different retrievers exist for retrieval-augmented generation, yet none performs best for all queries. Therefore, the key question is: given a query and retrieved passages from different retrievers, which one is most suited for that query? Existing routers usually make predictions from query-specific features and a fixed pool of retrievers; such routers have no output for an unseen retriever and are incapable by design on a new corpus. To fill these gaps, RAGCompass reads the retrieved passage sets themselves: passages from an unfamiliar retriever or corpus stay directly comparable to other passages, so routing stays well designed where a query-only map would fail. We contrastively fine-tune a reranker against the answer each evidence set produces, which accounts for most of the gain; a small comparison head adds a moderate improvement, and neither component has retriever-specific parameters. RAGCompass improves answer F1 by 28% relative to the best fixed retriever, and outperforms fusion baselines, three query-only routers, and a prompted 8B language model. Generalization experiments show that RAGCompass keeps most of its gain under an unseen corpus, retriever, or reader, while query-only baselines fall to the best fixed retriever or below.

Relevance propagation, which redistributes a network's prediction backward through the layers, is usually treated as a heuristic. We show that it is the solution of a single variational problem: Layer-wise Relevance Propagation (LRP) is the divergence-free flow on the active computational graph of zero relative entropy to the backward transition law of the forward activation flow; for a non-negative reference, this is a relative entropy between Markov measures on backward paths. The rules LRP$_0$, LRP$_{z^+}$ and LRP$_{α,β}$ follow from this principle by the choice of a reference, and are instances of one rule, LRPN$_{α,β}$, whose weights may vary from node to node, for instance with the node's activity, in the spirit of Hebb's principle. The principle rests on a measure-theoretic formalization of feed-forward networks as directed acyclic graphs of Filippov operators, which covers piecewise-smooth and discontinuous architectures (e.g. hard-routed Mixture-of-Experts) with a Clarke-Filippov generalized-gradient calculus; it also characterizes backpropagation as the adjoint equation of a Lagrangian relaxation. On networks of ReLU and max-pooling nodes, LRP$_0$ is Gradient$\times$Input, with an explicit jump across a kink where LRP$_{z^+}$ stays continuous. Taking the entropy edge by edge yields a new rule, Sinkhorn Relevance Propagation. Aggregated over an input region, the flow of a non-negative reference tells, with exact accounting, which hidden units carry its evidence, the conditioned flow being a Schrödinger bridge. On VGG-16, the conservation identities hold to round-off, and the rules are told apart by the sign of their maps, the share of the output left to the bias nodes, faithfulness, localization and class dependence; on ConvNeXt-Tiny, LRPN$_{α,β}$ conserves the output and its Hebbian weights give the highest ABPC of the rules tested.

Can a neural network learn to render 3D Gaussian splats on its own, with no rasterizer in the renderer? We present GaussianRenderFormer, a transformer that takes a set of Gaussians and a camera and outputs the image directly. Each Gaussian becomes one token in a view-independent encoder, and a view-dependent decoder turns small patches of camera rays into pixels by attending to those tokens. Trained on 26,820 Objaverse objects pruned to 20,000 Gaussians each, GaussianRenderFormer renders 300 unseen objects at 38.08 dB PSNR against the full, unpruned splats, while rasterizing the same pruned input reaches 44.02 dB. The gap is not one of generalization, and our diagnostics are consistent with a bottleneck in the view decoder: telling it the projected footprint of each Gaussian, which it could in principle infer, adds 1.4 dB. Under-training is not ruled out. Code, trained models and videos: https://svlwoof.github.io/gaussianrenderformer

Gated and ungated optimizations evaluated in GraphCast and Pangu-Weather. The first panel (upper-left) shows the mean over the 36 dates; remaining panels show first-of-month 12 UTC optimization descents from January-November 2017. The y-axis metric is the summed day 1-10 Z500 RMSE verified against ERA5, expressed as percent change from each model's date-matched control. The gated arm optimizes the initial conditions jointly with a LoRA weight correction; its GraphCast curve uses the initial conditions alone (LoRA removed), the same state Pangu receives. Stars and diamonds mark the lowest Pangu errors in the 5- and 10-day windows, respectively. Percentages above each panel show the best Pangu improvement relative to control for gated (green) and ungated (red) optimization. Ungated outperforms gated on half of dates when stopped early. Gating does not guarantee a better Pangu performance, but it does produce more comparable GraphCast and Pangu-Weather improvements on all 36 cases.

Reasoning-oriented language model training increasingly relies on multiple on-policy rollouts, but standard Group Relative Policy Optimization (GRPO) uses a fixed group size, wasting compute on easy prompts and under-exploring hard ones. Existing adaptive-budget methods alleviate this inefficiency, yet their stopping rules are heuristic and do not provide a finite-sample calibration guarantee. We propose CGRPO, which replaces the fixed rollout budget of standard GPRO with calibrated conformal thresholds over task-specific nonconformity scores, combining Adaptive Prediction Set (APS)-style calibration for closed-form reasoning, a first-success score for execution-based code, automatic miscoverage selection, and periodic re-calibration. We prove finite-sample coverage for the resulting stopping rule and show that, under monotone policy improvement, recalibration induces a non-increasing expected training budget. Across six benchmarks spanning mathematics, graduate-level science, and code generation, CGRPO reduces mean training rollouts by 49-70% relative to fixed-budget GRPO while retaining strong fixed-budget accuracy, achieves the strongest pass@1 among evaluated static and dynamic baselines on different datasets, and transfers across model families with up to 82% rollout savings.

What should ``verified'' mean when evaluating LLM-generated code? Many existing pipelines report a single verification number. We argue this collapses three distinct properties measured in a conditional audit sequence: verifier acceptance (Surface A), \emph{syntactic specification binding (Surface B: does the proof obligation bind to the spec?), and \emph{semantic specification fidelity} (Surface C: does the spec capture the informal requirement?). Two empirical demonstrations show the surfaces separate. \textbf{First ($A \neq B$), via Li@Opus (a same-model reproduction of \citet{li2025dafny}'s Dafny-as-intermediate-representation (IR) pipeline under Claude Opus 4.7) on three Python benchmarks}: it reaches \textbf{99.1\% verifier acceptance} on HumanEval while only \textbf{1.9\%} carry a certificate binding to a spec-defined predicate or function. \textbf{Second ($B \neq C$), via VCG on two Rust benchmarks}: we use VCG (which is our own in-house Dafny-as-intermediate pipeline) on Rust because the Rust target lets us enforce fully-verifiable output (every component ships as a Dafny certificate, ensures-binding gate passing by construction). Even inside that construction, \textbf{11.48\% (MBPP-Rust) and 14.82\% (VeriContest LC slice) of structurally-verified samples still fail hidden tests}. On the LC slice of VeriContest, \textbf{VCG's reported End2End rate is 7.79\% (535/6870).} The cited external rates are 5.51\% (GPT-5.5) and 1.88\% (single-shot Opus 4.7). These give descriptive differences of 2.3 percentage points and 4.14$\times$, respectively; task-budget, repair and harness differences prevent a matched-performance claim.

As autonomous AI agents enter permissionless onchain markets, selecting new service providers remains a core challenge. Existing reputation systems depend on scarce resources like human curation, identity credentials, or voluntary reviews, while naive volume rankings are vulnerable to Sybil attacks. We evaluate whether existing onchain payment histories can yield provider ratings that are resilient against Sybil attacks. Using a 157-day dataset comprising 66M agentic transactions worth $6.1M across 16 blockchains, we model payment flows as a bipartite graph between buyers and providers. A provider's rating is derived from graph paths linked to a buyer's existing trusted network. We propose three Sybil-resistant rating algorithms: a max-flow rating and two relaxed algorithms optimized for stronger rating performance. Evaluating these algorithms against simulated adversaries across 813 buyers, we show that rating manipulation costs 20% of those buyers' total spend and is around 23x harder to manipulate than the standard similarity rule (collaborative filtering).

Data tends to be most useful when it is shared, and the data most worth sharing is often the data that privacy obligations restrict. Synthetic data from a generator trained under differential privacy (DP) is one way to reconcile the two, but its utility depends on the privacy budget $\varepsilon$ and on how much data the generator sees - a parameter that most data sharing applications of synthetic data control. We present SAVANT, a self-attentive variational autoencoder for tabular data trained with DP-SGD that, to our knowledge, is the first to combine attention, a VAE, and DP in one tabular model. We evaluate it against two marginal mechanisms, PrivBayes and MST, jointly over the privacy budget and the dataset size on two public datasets, measuring the utility a downstream classifier loses by training on synthetic rather than real data. We find that, under the same DP guarantee, more data generally means more utility. For the marginal mechanisms it absorbs most of what a strict budget costs, and a noise-scaling argument suggests DP-SGD exchanges data fo budget at a worse rate. Two independent membership-inference attacks, judged against a measured null, found no aggregate membership signal.

Real-world vision is dynamic and scene-rich, while EEG and fMRI resolve it at complementary temporal and spatial scales. Most neural decoding systems, however, align one modality in isolation and mostly against static images or short clips. We study the synchronized case: EEG and fMRI are recorded during the same long-form naturalistic viewing, giving a shared stimulus but also recurring narrative context and many semantically similar candidates. We present BrainBind, a framework that maps synchronized EEG and fMRI into frozen long-form video embedding spaces and treats those spaces as common visual-semantic references, not literal neural codes. On CineBrain, a five-seed Recall@K 2 x 4 matrix across two video anchors and four system configurations shows distinct retrieval profiles. With Omni, fMRI organizes the full candidate list most strongly (MRR = 23.82, nDCG@10 = 25.31), while Joint-Dual gives the highest mean label-aware R@1 (13.52%) with a distinct rank-wide profile. The complementary episode-disjoint evaluation quantifies the contribution of repeated stimulus identity and narrative context to the random split. Across these experiments, proxy choice and system configuration produce distinct top-rank and rank-wide retrieval profiles within the present synchronized dataset. The code is available at https://github.com/Forainest789/Brainbind.

We study repeated general games over $T$ rounds between a utility-maximizing optimizer and a learning agent, where budget constraints couple the players' decisions across rounds. Without budgets, low swap regret simultaneously guarantees the learner sublinear regret and caps the optimizer at its Stackelberg value. Budgets break this picture: adversarial no-regret becomes impossible, so the natural goals are non-manipulability---no adaptive optimizer earns more than $\mathtt{BSE} \cdot T + o(T)$, where $\mathtt{BSE}$ is its Budgeted Stackelberg Equilibrium value---and low stochastic regret when the optimizer plays a fixed mixed strategy. We give an efficient, budget-feasible learning algorithm that achieves both goals in every budgeted game: any adaptive optimizer, budgeted or not, earns at most $\mathtt{BSE} \cdot T + O(\sqrt{T})$ in expectation, and if the optimizer plays a fixed mixed strategy on each of a constant number of intervals, the learner has $O(\sqrt{T})$ regret on each such interval. Our learner best responds to forecasts of the optimizer's play that are calibrated in two novel senses, since under budgets the calibration that prevents manipulation no longer yields low stochastic regret. Response calibration, which conditions on the learner's own mixed response by embedding its continuous mixing weight as a step function in $L^2([-1,1])$, prevents manipulation without discretizing responses; dual-weighted Lagrangian calibration ensures low regret. Although low calibration error requires randomized forecasts---distributions over a continuous space that need not have a succinct representation or be directly computable---we construct such forecasts in polynomial time by simulating online gradient ascent.

Multimodal AI fusion of histology and clinical data is an active research area in computational pathology, but fusion methods are rarely compared under identical conditions and across different clinical endpoints. We present a modular framework in which slide-level embeddings, obtained by training an attention-based aggregator over patch embeddings from a frozen foundation model, and per-visit clinical embeddings are fed into 15 interchangeable fusion methods with a shared output dimension and prediction head (except decision-level methods). Using a prospective cohort of newly diagnosed inflammatory bowel disease patients, we predicted STRIDE-II-aligned remission outcomes from baseline or follow-up visits. The best fusion method reached macro-F1 of 0.57-0.75, outperforming both single-modality models in four of six scenarios, and differed by outcome and input. Image reliance, the drop in balanced accuracy when the slide embedding is removed, tended to follow the remission type - highest for histological and lowest for endoscopic remission; methods with similar accuracy relied on the slide to very different degrees.

Self-supervised world models are often said to organize their latents around physical quantities. We test this on 2D decaying Navier-Stokes flow in three pre-registered studies, each tested once on held-out data. In Study 1, a joint-embedding predictive architecture (JEPA) trained to predict the flow fails the pre-registered criteria for organizing its latent around fluid quantities better than an untrained network of the same architecture (not supported); a JEPA trained on non-physical linear dynamics and a static invariance model that never sees time do as well or better. In Study 2, plain autoencoders that render the field make enstrophy prominent in their slowest unsupervised latent directions (best r 0.80-0.86 vs 0.38 untrained; fresh seeds, sealed data). Adding VICReg lowers this prominence in every comparison, though not by the pre-registered margin for vorticity, while raising the autoencoders' linear decodability (partially supported). In Study 3, we build steerable directions ("dials") by routing: the decoder gets overall scale only through a small conditioning code, while its other pathway carries a degraded copy of spectral shape. Label-free rules find a loudness dial in all 8 trained models and a shape dial in 3 of 5 fresh seeds on sealed data, against a pre-registered bar of 4; neither control shows a shape dial (partially supported). Across the studies, what a latent contains and what can be controlled through it come apart. In exploration, a supervised direction reading spectral shape almost perfectly moved nothing when steered. During the build, the loudness dial had to be found by its effect on the output (closest to a pure rescaling), not by data variance; for shape, a variance rule and an effect-constrained rule each worked in 3 of 5 seeds. The program used about 70 logged hours on one consumer GPU; all pre-registrations, deviations, code and outputs are public.

Personalized dialogues with AI substantially reduce conspiracy beliefs. To identify why, we conducted two experiments with U.S. adults recruited online (total N = 2,193) that removed possible causes one at a time. The debunking effect was robust to disclosing the AI's persuasive goal, framing the exchange as a debate, instructing the AI to inform rather than persuade, halving its response length, altering how beliefs were measured, and replacing the dialogue with a single tailored message. Only instructing the AI to persuade without factual evidence and arguments substantially reduced the debunking effect. Content analyses of the AI's output and participants' own accounts linked larger reductions in conspiracy beliefs to factual argumentation rather than rapport, storytelling, or appeals to empathy. AI dialogues appear to reduce conspiracy beliefs primarily by supplying relevant, well-reasoned counterevidence. Our results suggest that humans might be similarly effective if they could supply comparable evidence.

Point cloud segmentation is a fundamental task in 3D data analysis, with applications across robotics, augmented and virtual reality, and remote sensing. However, the inherent sparsity, noise, and unordered nature of point clouds introduce significant challenges, particularly in achieving reliable classification. While the state-of-the-art rely on training models for explicit modeling of uncertainty, there is a gap in exploring different roles that estimated uncertainty using generative approaches can play. This work investigates the impact of deterministic uncertainty in point cloud segmentation by explicitly modeling epistemic uncertainty and using it in different roles, namely, as a feature and in loss function. We evaluate a baseline artificial neural network (ANN) and two uncertainty-aware variants (uncertainty as an additional input feature and as a per-sample confidence weight in the training loss) on four large-scale LiDAR benchmarks spanning airborne and mobile acquisition, namely ISPRS Vaihingen, AHN3, Paris-Lille-3D, and Toronto-3D, together with PointNet++ and RandLA-Net references and their uncertainty-augmented counterparts under canonical spatially disjoint train and test splits. Under this protocol, uncertainty-aware modeling improves mean Intersection over Union, recovers minority classes that the baseline ANN never predicts at all, and in half of the configurations improves the alignment between prediction error and estimated uncertainty as measured by Patch Accuracy versus Patch Uncertainty. Our findings indicate that uncertainty can play multiple roles in the segmentation pipeline rather than as a device for improving accuracy alone.

Despite growing interest in foundation models for robotic manipulation, systematic evaluation of their capabilities remains elusive. Existing benchmarks are either narrow and saturated, offering little signal on SOTA methods, broadly scoped, providing diversity at the cost of controlled experimentation, or require training on large, messy open-source datasets. We introduce MESA, an evaluation framework designed for precise, reproducible measurement of language-conditioned policy generalization. MESA provides five evaluation suites that isolate in-distribution performance and four distinct axes of generalization: spatial configuration, object instance, object category, and subtask composition, enabling researchers to diagnose where and why policies fail, not just whether they fail. To populate these suites at scale, we develop MESA-Gen, a pipeline for task and demonstration generation built by improving MimicGen. We instantiate the framework on two platforms and release both: the single-arm MESA suite, and BiMESA, the first large-scale benchmark for bimanual generalization. We evaluate seven baseline methods spanning diffusion policies, VLM-based approaches, and pretrained VLAs, finding that the strongest model achieves only 52% on MESA and 49% on BiMESA, providing meaningful signal while leaving ample room for progress. MESA rankings are broadly consistent with RoboArena scores, suggesting that MESA captures method-level differences relevant beyond simulation. Project website: https://pairlab.github.io/MESA.

Contrastive vision-language models score left/right relations near chance, but the symptom has two causes that call for opposite repairs: the relation may survive in the pooled embedding yet be unreachable by native image-text similarity, or pooling may discard it while the dense features retain it. On clean two object photographs, across ten encoders, the lateral polarity is linearly decodable from the flip odd component of the pooled vector at 0.77-0.99, yet a search over a million captions leaves nine of the ten at chance under native scoring: an access failure. On controlled C4 orbits, where the relation is the only thing that varies, pooling keeps the arrangement axis at 0.998-1.000 and loses its polarity, which the patch grid retains: a pooling failure. Both sit downstream of the dense features, so one readout covers both. Our zero parameter, query conditioned spatial moment recomputes the geometry from the patch grid and stores it as a vector beside the native embedding. On a 118,287 image index it raises left/right pairwise accuracy against reversed geometry negatives from 0.51-0.53 to 0.81-0.96 and conditional R@1 among 21 category matched candidates from 0.05 to 0.37-0.75, at under a millisecond per batched query, while reproducing the native ranking exactly for queries that name no relation.

In multi-objective reinforcement learning, preferences may be specified only after data collection, requiring one dataset to support a continuum of reward trade-offs. We study finite discounted Markov decision processes with unknown transitions, $S$ states, $A$ actions, and $m$ known reward functions taking values in $[0,1]$, given access to a generative model. The learner must provide an $\varepsilon$-optimal policy for every convex combination of these rewards without further sampling, with a guarantee that holds simultaneously over all preferences and initial states. Writing $H=(1-γ)^{-1}$, we establish the constant-confidence minimax sample complexity $\widetildeΘ\!\left( \frac{SAH^3}{\varepsilon^2}\min\{S,\log(em)\} \right)$, where the suppressed factors are logarithmic in $S,A,H,\varepsilon^{-1}$ and the dependence on $m$ is explicit. Thus, up to these logarithmic factors, the worst-case statistical cost of supporting all preferences grows logarithmically with the number of objectives and saturates at the number of states. The upper bound uses uniform state--action sampling and empirical planning, with reward rounding in the logarithmic-objective regime. Its analysis chains optimal-value increments over covers of the reward hull. A one-sided Bellman variance bound controls signed value differences under a second transition kernel, preserving the $H^3\varepsilon^{-2}$ dependence. A complementary uniform guarantee for all bounded rewards yields the state-dimension saturation. We establish lower bounds matching up to logarithmic factors through a rare-exit construction with subset rewards and action encoding. These bounds apply to adaptive simulator queries and arbitrary randomized history-dependent output policies, while the upper bounds are attained by deterministic stationary policies.

In biodiversity monitoring, a single image raises several questions: which taxon, how many, which trait, and whether the individual has been seen before. Vision-language models (VLMs) could answer all of these from the image, but their ability is largely untested in biology, as existing benchmarks focus mostly on species identification within a few taxonomic groups. We present Bio-VQA, the largest biological visual question answering (VQA) benchmark and the only one to span four aspects of biological visual understanding: hierarchical taxonomic reasoning, fine-grained attribute grounding, quantitative spatial reasoning, and individual re-identification. Bio-VQA contains 1.86M questions over 217K images from 11 public datasets, and extends right-or-wrong scoring to graded scoring, so a near-miss counts for more than a distant one. Evaluating 18 open-weight models from 3.75B to 38.4B parameters, we find that general visual capabilities do not transfer to biological organisms: the best model counts animals exactly on 41.7% of images with 2 to 10 individuals, less than half its accuracy on everyday objects at the same counts. Taxonomic accuracy falls to one sixth, from 0.72 at phylum to 0.12 at scientific name, and scaling helps mostly at coarse ranks. The gap to a domain specialist widens accordingly, from 0.731 against BioCLIP-2's 0.874 at class (1.2x) to 0.269 against 0.544 at scientific name (2x). On trait size measurement, 15 of the 18 models do no better than an image-blind baseline that predicts one fixed number for every specimen, and on re-identification, 17 of the 18 models fall below chance under adversarial conditions. These failures persist across a tenfold range of model size. We hope Bio-VQA becomes a standard test of how well VLM capabilities transfer to the biological domain.

Support vector machines (SVMs) are fundamentally grounded in geometric margin maximisation and currently lack a mainstream probabilistic interpretation. This keeps them theoretically isolated from probabilistic machine learning models that are naturally comparable and combinable. In this paper, we bridge this gap by proving that the dual objectives of both support vector classification and support vector regression are exact, constrained Kullback-Leibler (KL) divergence minimisation tasks between two Gaussian measures. We further show that this same divergence formulation recovers the exact objective function for kernel ridge regression when domain constraints are removed and $L^2$ regularisation is applied. By demonstrating that SVM duals directly minimise relative entropy, our results provide a unified information-theoretic model for standard kernel machines and lay the foundation for unifying broader kernel-based algorithms under probability theory.

LLM-as-a-judge is a common substitute for costly human evaluation, and aggregating several judges is the standard remedy for any single judge's bias. Judge pools mix judges of very different reliability, while the human-labeled sets that validate them hold a few hundred items or fewer. In this regime the per-item mean weights weak judges as heavily as strong ones, and confusion-table models such as Dawid-Skene must estimate a per-judge error table from too few items. We propose AGORA (Aggregation Of Reliability-weighted Assessments), an item-response-theory aggregator that places every item on one latent quality scale and describes each judge by a sensitivity parameter and ordered rating thresholds, so judge reliability is learned from all items jointly and without human labels. Across three benchmarks, AGORA matches or exceeds Dawid-Skene on large evaluation sets, and its advantage widens as the set shrinks toward a few hundred items. It remains accurate as pools grow more uneven, a label-free dispersion statistic indicates in advance when trust weighting helps, and its learned trust weights track per-judge human validity.

Decision Models (DMs), a new class of AI models, give fast responses and return a probability distribution over a fixed set of options. These properties make them candidate human surrogates in economic simulations and surveys, a role increasingly filled by Large Language Models (LLMs), which are costly at scale and under-represent the variation in human behaviour. We present a first evaluation of DMs as human surrogates in economics, comparing 10 DMs with 8 LLMs in three economic settings: a household spending survey on consumption, several classic economic games and a labour-market field experiment. Frontier DMs, those that score highest on general decision-making benchmarks, reproduce the variation in human responses more closely than LLMs in all three settings. They capture the modal human response as well as LLMs in the spending survey, but less well in the economic games. Frontier DMs are also roughly ten times faster than the fastest LLM and can be over a hundred times cheaper. Together, these results suggest that DMs capture the heterogeneity of human behaviour better than LLMs and make large-scale economic simulations more feasible.

Virtual trials use routinely collected clinical data to evaluate treatment strategies, but translating a study protocol into an executable analysis requires determining how clinical concepts and study requirements should be represented in the available data and applied consistently across the workflow. These decisions depend on local data representations and operational conventions and can involve source selection, patient and encounter identity, linkage, and temporal interpretation. Here we introduce VirtualTrialHarness (VTHarness), an agent framework that coordinates virtual-trial operationalization by selecting reusable procedural skills for each workflow stage and uses explicit contracts to maintain consistent study decisions across stages. We also introduce VTBenchmark for evaluating source selection and cohort construction. Across three complex synthetic environments, VTHarness successfully produced all required cohort outputs matching the references in 95.1% of workflows after requirement clarification, compared with 60.4% using retrieval from the same procedural corpus. Procedural skills improved source recovery across three language models and increased source overlap with analyst code in 13 Mayo Clinic cases. Site-specific adaptation enabled reference-matched cohort execution across eight MIMIC-IV tasks. Within the Mayo Clinic Virtual Trials Program, the framework supported investigator review of cohort construction and endpoint analysis. These findings support coordinating procedural guidance with workflow decisions to improve virtual-trial implementation across complex clinical schemas.

Transparency has become a central principle in AI governance, yet there is little agreement about what transparency means, for whom, and in what contexts. Drawing on interpretive flexibility within the Social Construction of Technology, this paper examines how different stakeholders understand and experience transparency in generative AI. This study draws from a global survey of 400 end-users, AI builders, and policymakers, complemented by a structured focus group with end-users, grounding the quantitative findings in qualitative insights from the group most underserved by current transparency practice. We present three key findings. First, all three groups share similar understandings of what transparency means and what it should afford, centered on ensuring accuracy and privacy. Second, gaps between the importance of transparency features and their perceived availability are substantially larger for end-users than builders, revealing how power shapes what gets built and who it serves. Only 28% of end-users report that current transparency practices reflect their needs very well, compared to 53% of builders and 49% of policymakers. Third, transparency is situationally constructed, priorities shift with context and task structure, with role-specific frames activating when context demands them and receding when it does not. We argue that meaningful AI transparency requires moving beyond a one-size-fits-all solution, including current emphases on AI disclosure and static documentation, toward pluralistic transparency infrastructures that reflect the needs of diverse users and stakeholders.

We evaluate the extent to which agentic AI models can complete an end-to-end data science project, and how this improves with feedback. A 'writer' model with specific tools creates a complete data science project including downloading, cleaning, and analyzing a dataset, and writing a short paper about it. The project is executed and a 'grader' model evaluates it. The writer can then revise its submission, and this repeats for a total of three rounds. We consider seven models from five developers: Anthropic, OpenAI, Google, DeepSeek, and Z.AI. Each model is both a writer and a grader, so there are 49 writer-grader pairs and we run each pair three times. Of the 147 initial submissions, 107 run and render, and the average score increases from 59 per cent to 76 per cent after the final round of feedback. By writer the extent of the increase ranges from 3 percentage points for Anthropic's Opus 5.5 to 46 percentage points for Anthropic's Haiku 4.5, and by grader it ranges from 13 percentage points for Google's Gemini 3.8 Flash to 32 percentage points for Anthropic's Haiku 4.5. When a model grades its own writing the score tends to be about 9 percentage points higher than the average of what the other models give it. About half of the difference that we see from round to round is within the grader's own noise. Feedback generally results in the highest marks, but even just having the chance to revise, or having the additional budget, in terms of time or tool calls, without any feedback, results in considerable increases for some of the models but not for others. We find that how much such a loop helps depends on which model plays which role, but weaker models generally do not improve stronger models.

Decentralized federated learning (DFL) eliminates the central aggregation server but remains vulnerable to data heterogeneity, Byzantine attacks, and channel noise. We show that noise accumulates as models propagate across nodes, becoming increasingly significant as the learning rate decays. Moreover, accumulated noise enables Byzantine attackers to exploit loss-based model selection by mimicking honest model averages. To address these challenges, we propose screened, hop-aware, ranked, proximal averaging, a robust DFL framework (FedSharp). In this approach, each node maintains models from its 2b predecessors, ensuring at least b honest candidates when at most b nodes are Byzantine. FedSharp screens suspicious models using coordinate-wise median deviations, averages the b lowest-loss candidates to reduce fresh channel-noise variance by a factor of b, and applies noise-adaptive gradient-norm regularization and proximal training to mitigate accumulated noise and non-IID drift. Experiments on a 100-node ring with 33 Byzantine nodes show that FedSharp achieves 93.5-94.9% accuracy on MNIST, compared with 29.4-38.1% for a state-of-the-art single-model-selection benchmark, and 32.8-48.8% on CIFAR-10, compared with 10.0-30.5%, across Gaussian and hidden attacks under IID and non-IID data. On MNIST, FedSharp also achieves accuracy within approximately one percentage point of the benchmark operating without channel noise.

In this paper, three operations on Gibbs probability measures are studied: renormalization, normalized log-linear combination, and nesting. The resulting measures are shown to be Gibbs probability measures with explicit expressions for their objective functions, reference measures, and regularization factors. These measures are further characterized as unique solutions to optimization problems involving linear combinations of the original objective functions, subject to either relative entropy regularization or relative entropy constraints. When the reference measures are probability measures, geometric interpretations of these three operations are established involving Pythagorean identities and exponential geodesics.A central result is the equivalence between nesting and normalized log-linear combinations of Gibbs probability measures. These results are then applied to supervised machine learning with Gibbs algorithms, which are represented by Gibbs probability measures. In this context, either through nesting or log-linear combinations, Gibbs algorithms trained on local datasets are used to form new Gibbs algorithms trained upon the aggregation of all local training datasets. Interestingly, the parameters of such operations, determine the weights of each training data point on the aggregated dataset. %Operations on local Gibbs probability measures are shown to control the contributions of training data. This observation is leveraged for achieving: (a) exact machine unlearning in centralized, decentralized, and federated learning systems; and (b) centralized-equivalent performance, in the sense of producing algorithms trained on the aggregation of all local datasets in decentralized and federated learning .

In multivariate conformal prediction, many nonconformity scores yield prediction sets with complex geometry (e.g., ellipsoids or density level sets) that are difficult to interpret and visualize. Hyperrectangles offer a practical alternative by providing one interval per output. We extend this notion to Cartesian products of arbitrary one-dimensional sets, which we call generalized hyperrectangles. We introduce a rectangularization framework that turns any continuous multivariate nonconformity score, or conditional density, into one yielding generalized hyperrectangular prediction sets with finite-sample coverage. We establish key geometric properties of our rectangularization framework and provide useful intuitions on generalized hyperrectangles. We further establish a representation theorem showing that a broad class of hyperrectangular conformal methods based on max-aggregated coordinate-wise scores can be interpreted as rectangularizations of implicit conditional densities. Through synthetic and real-world experiments, we quantify the efficiency cost of rectangularization, while highlighting its benefits, especially for practical applications of uncertainty quantification.

Whole-page ranking jointly optimizes advertisements and organic content to balance business outcomes and user engagement under complex layout and business constraints. Existing non-autoregressive (NAR) approaches enable efficient page generation but typically score candidate items independently of previously selected items, limiting their ability to capture order-dependent interactions. Autoregressive (AR) methods address this limitation through sequential conditioning, but incur substantial inference overhead that challenges large-scale deployment. Bridging this quality-efficiency gap remains a challenge in industrial whole-page ranking. We propose Multi-Item Scoring (MIS), a prefix-aware NAR-style framework that incorporates sequential context into efficient multi-item scoring. MIS integrates global candidate interactions, localized prefix conditioning, and position-aware scoring within a shared representation architecture. By computing reusable representations once per request and performing lightweight prefix-conditioned scoring during constrained beam search, MIS captures important sequential dependencies without repeated full-model inference. Experiments on KuaiRand-1K and RL4RS demonstrate that MIS approaches the generation quality of an autoregressive CGA baseline while achieving 4.07x and 3.54x faster CPU decoding, respectively. In a randomized evaluation on a large-scale social-video platform, MIS improves the realization metric by 0.104% and watch time by 0.058%, while reducing ranking-service CPU consumption by 9.73%. These results demonstrate that prefix-aware multi-item scoring can narrow the quality gap between NAR and AR generation while retaining the efficiency required for industrial scalability.

Recent advances in AI-assisted proof automation are extending the use of proof assistants in software verification and mathematics. However, greater proof automation does not remove users' responsibility to assess whether formal statements capture their intended requirements, making education in formal specification and reasoning essential. To lower the installation and learning barriers associated with mature proof assistants, we present ProofStudio, a browser-based environment integrating Isabelle/HOL, Lean, Rocq, HOL4, and HOL Light with corresponding worked examples, preconfigured proof search, and contextual AI assistance. With ProofStudio, users can explore formal specifications and proofs across five systems without local installation, inspecting native prover feedback and experimenting with automation through a common interface.

LLMs can repair noisy OCR transcripts of historical documents, but they also change text that was already correct. We propose EditVeto, a veto gate placed after a full-context LLM rewrite. The rewrite is character-aligned to the OCR input, its differences are split into blocks, and a 0.6B-parameter gate keeps or reverts each block. An oracle analysis motivates the design: given the true error locations, per-span generation is worse than no correction, whereas vetoing a full rewrite's edits helps. Trained on rewrites by two Qwen3 models and evaluated, frozen, on the eight HIPE-OCRepair-2026 test sets, the gate improves all five rewriters tested, two of them unseen in training: weighted cMER falls by 0.0026-0.0044 and the preference score rises by 0.15-0.25 (paired bootstrap, p ~ 0.0002 each). On published competition outputs it lowers cMER for three of four teams but raises it for the winning run (0.0050 to 0.0065). In a second, disclosed test look, a retrained gate still improves LoRA-tuned rewriters (cMER -0.0011 to -0.0021), and an ensemble of their gated edits reaches a cMER of 0.0069 and a preference score of 0.928, mostly through fine-tuning and pooling rather than gating. Built only on public models and data, it is behind the winner (0.0050) and tied with the second-ranked run (0.0071), both from the only team using domain pre-training, and ahead of all other published runs. A third, secondary look shows the unchanged gate improving zero-shot Qwen3.8-27B (cMER 0.0114 to 0.0108) but only the preference score of its LoRA-tuned version.

Historically automated electroencephalography (EEG) systems was done using task-specific supervised models trained on purposely annotated data. A more recent modeling paradigm relies on creating a unified representation, or a 'foundation model', for multiple modality-specific tasks through large-scale domain-specific pretraining on the unannotated EEG signal. The quality of this unified representation directly depends on the availability of the modality-specific datasets. At the same time, generalist Vision Language Models (VLMs) exhibit broad few-shot learning capacities as well as good fine-tuning adaptability to different visual and language tasks. Can VLMs learn how to read and interpret EEG visually, emulating to a degree the way human clinicians inspect the EEG patterns? In this work we make a case that they can: we show that generalist VLMs without large-scale EEG pretraining perform competitively across a collection of clinical neuro-diagnostic tasks, can adapt to different tasks in a few-shot learning setting better than specialist foundation models, and exhibit better performance with model scaling when modality-specific models fail to exhibit.

Across a wide range of applications, kernel measures are used to tune hyperparameters, select among candidate kernels, and assess the suitability of kernels for specific learning tasks. For classification, kernel measures are used to predict which kernels may produce accurate kernel classifiers. Many existing kernel measures, such as alignment to the ideal kernel, are designed to reward kernels which captures all samples from the same class as similar. While effective in many settings, this requirement is unnecessarily restrictive for classification, where classes may exhibit substantial internal structure. In this work, the "Confidence Score", a new kernel measure, is proposed. Unlike alignment-to-the-ideal kernel measures, the Confidence Score measures the alignment between an encoding of the ground truth and that ground truth projected by the kernel, allowing kernels that preserve intra-class structure to achieve high scores. Through experiments, it is demonstrated that the Confidence Score exhibits a higher correlation with classification accuracy across a variety of kernels on both real-world and synthetic datasets, compared to select other kernel measures.

In this paper, we provide tight lower bounds for minimizing high-order Hölder-smooth and uniformly convex functions when their orders match. Specifically, we consider functions whose $p^{th}$-order derivatives are Hölder continuous with degree $ν$ and parameter $H$, and that are uniformly convex with degree $q$ and parameter $σ$. When these orders match, namely, $q = p + ν$, we show that any deterministic algorithm with access to derivatives up to order $p$ requires $Ω\left(\left(\frac{H}σ\right)^{\frac{2}{3(p+ν)-2}}\log\frac{Δ_0}ε\right)$ oracle calls to reduce an initial optimality gap of $Δ_0$ to $ε$. This matches the corresponding upper bound and establishes a higher-order uniformly convex counterpart to Nesterov's classical first-order lower bound for strongly convex functions. Rather than directly generalizing Nesterov's classical quadratic-chain construction, we develop a multiscale phase-wise construction that preserves previous oracle responses while maintaining global Hölder smoothness and uniform convexity. Our result resolves the matched case left open by existing lower bounds for the two asymmetric regimes.

We propose MA-SafeDiffuser, a multi-agent extension of SafeDiffuser that equips diffusion-based trajectory planners with finite-time diffusion invariance guarantees for joint-agent safety specifications. Building on the single-agent construction that embeds control barrier function (CBF) constraints into reverse-diffusion updates, we: (i) formalize joint safe sets as intersections of per-agent and pairwise barrier sets; (ii) derive centralized and decentralized (communication-aware) constrained denoising procedures with provable invariance under mild assumptions; (iii) address local-trap and deadlock phenomena via time-varying specifications and liveness CBFs; and (iv) develop a lightweight benchmarking suite including a multi-agent Maze2D domain. Empirically, MA-SafeDiffuser reduces violation counts relative to unconstrained diffusion baselines, while retaining planning quality. We provide algorithms, proofs, and reference implementations.

For humans, referential disambiguation is a collaborative process involving clarification questions asked when they will meaningfully change the outcome. However, large language models (LLMs) have been shown to be biased toward premature answering, preferring definite single-turn responses over more clarification exchanges. In this paper, we study the problem of "strategic clarification" -- asking clarification questions in a reference game when their answers will narrow the set of candidate referents. We construct a controlled test environment from the SIMMC 2.1 dataset that records how many questions each ambiguous reference needs to be successfully resolved, and examine how baseline LLM behavior and common post-training methods ask clarifying questions vs. attempt to answer the referent. We also propose BACE (Belief-Aware Critic with Epistemic tracking), a lightweight action chooser method on top of a frozen LM backbone, and show that BACE more consistently chooses the optimal number of clarifying questions than methods such as GRPO and ArCHer.

Textual information is ubiquitous in real-world time series datasets, appearing as news, reports, and annotations that hold key predictive signals not deducible from numerical history alone. Large Language Models (LLMs) interpret text but lack numerical precision, while Time Series Foundation Models (TSFMs) are quantitatively accurate but cannot reason about the effects of text. Moreover, the lack of large-scale paired multimodal datasets limits progress in bridging this gap. In this work, we address this challenge by providing a scalable approach for generating realistic multimodal datasets by annotating real-world time series with LLMs. Building on this, we propose Context as Covariates (CoCo), a forecasting framework in which an LLM distills its textual reasoning into numerically grounded forecast and confidence covariates that guide a TSFM backbone, thereby exploiting the strengths of both LLMs and TSFMs. We align the two models in stages, first fine-tuning the LLM via Group Relative Policy Optimization (GRPO) with a reward tied to covariate-induced improvement in forecasting performance, and then fine-tuning the TSFM on covariates generated by the aligned LLM. Our experiments show that models trained on our synthetic multimodal corpus generalize to real benchmarks spanning finance, economics, weather, traffic, and security. Notably, CoCo with Qwen3-4B-Instruct-2507 as the LLM and Chronos2 as the TSFM achieves up to 12% MAE improvement over state-of-the-art TSFMs and LLM forecasters, such as GPT-4o. Code, checkpoints, and the evaluation datasets are publicly available.

AI Agents are becoming digital coworkers that can browse the web, write code, run experiments, and assist humans over extended horizons. Yet existing benchmarks fall short of real knowledge work: many rely on synthetic data or isolated tasks, missing the complex, multi-party dynamics of collaborative projects, where critical information is scattered across conversations, code, and evolving decisions. Even simple workplace questions may require locating relevant channels and direct messages, recovering context from long discussions, inspecting repository state, reconciling stale or conflicting evidence, and synthesizing across sources. We introduce SlackBench, a benchmark for evaluating agents on realistic research and engineering workflows that require joint reasoning over workplace communication and evolving repository state. Grounded in privacy-preserving reenactments of real academic research projects, SlackBench provides reproducible environments with realistic multi-party dialogue, evolving decisions, and corresponding code changes grounded in GitHub repositories and Google Docs. It contains 140 queries spanning heterogeneous-source agentic search, abstention on unanswerable or false-premise questions, goal-directed summarization, and circumstance inference over long conversations. These tasks are not reducible to static corpus question answering: agents must determine which conversations matter, traverse threads, inspect repository state, and synthesize evidence across communication and code artifacts. Across 33 frontier models and agent harnesses, the best system scores around 80% and most agents score below 66%, highlighting substantial headroom for improvement. SlackBench thus provides an extensible framework for evaluating agents in collaborative work environments. The SlackBench dataset is available at https://huggingface.co/datasets/HuanzhiMao/SlackBench .

Humanoid motion tracking has scaled to general policies that reproduce diverse human motion, but the resulting behavior is in free-space: the tracker knows where the body should be, not how it should interact with the environment. Other work makes the object part of the objective, yet abstracts that interaction to contact or geometry, which leaves the demanding cases under-explored. We argue that force is a natural command channel between motion imitation and object interaction, one that specifies not just where to make contact but how hard. We present Ctrl-F, a hierarchical motion-force framework that augments motion tracking with an explicit force command, a 3-D force at each designated contact body such as a hand or a foot. The hierarchical framework comprises: (1) A unified motion-force controller tracks the reference motion while applying the commanded force to the world. (2) A high-level force planner that issues a force command online based on object state and proprioception observations. In simulation and on hardware, Ctrl-F completes loco-manipulation that demands large, sustained interaction force and adjusts its command online as the load changes, both out of reach of a force-blind tracker. On hardware, Ctrl-F demonstrates hand-based dragging, pushing, and lifting, as well as foot-based kicking, successfully manipulating boxes weighing up to 20kg in pushing and kicking tasks where the force-blind baseline fails, while adapting online to additional loads introduced mid-task.

Aligning large language models (LLMs) with several objectives, such as helpfulness, harmlessness, and honesty, typically relies on preference judgment for each objective, growing the annotation cost linearly with the number of objectives. Active preference learning can reduce this cost by selecting the comparisons whose labels are expected to change the policy most. Existing remedies either lack an objective-aware query selection mechanism for the multi-objective setting or regard conflicting preferences as harmful, therefore excluding them from the training data. We introduce OGC-DPO (Objective-conditioned Gradient Coverage DPO), an active learning framework that treats each (response pair, objective) combination as a separate query and measures its informativeness through the policy's own implicit gradient, without relying on an external reward model. OGC-DPO then selects the queries that most expand each objective's gradient coverage. Empirical experiments across varying numbers of objectives demonstrate the effectiveness of OGC-DPO over existing active preference learning methods under the same annotation cost. Further analysis demonstrates that informative conflicts, rather than being uniformly harmful, provide a useful policy training signal.

Instantaneous quantum polynomial-time (IQP) Born machines are attractive quantum generative models, as certain IQP circuit families are conjectured to be hard to sample, yet they can be trained classically. Previously proposed conditional IQP Born machines make the parameters of the diagonal layer condition dependent. We introduce conditional IQP Born machines, in which the condition is encoded in the input state and a single IQP unitary is shared across conditions. Efficient sampling and amplitude-ratio evaluation for the reference state entering the diagonal layer suffice to obtain unbiased classical estimates of a conditional maximum mean discrepancy objective and its gradients. We show that input conditioning can represent conditional families that no diagonally conditioned IQP model with the same number of qubits can reproduce. For the sixteen-qubit Curie-Weiss spin model, a single conditional model captures the magnetization transition at held-out temperatures. Its mean total variation distance is six times lower than that of diagonal conditioning under the same protocol and comparable to that of classical generative models with up to seven times as many parameters. We also train conditional models classically with up to 2048 qubits, which reproduce the mean squared magnetization of the target away from the crossover. These results extend classically trained IQP Born machines to conditional generation.

Large language models (LLMs) are increasingly used to support scientific literature reviews, yet evaluations often examine individual tasks rather than the review process as a whole. We assess six AI tools across four stages: finding relevant articles, interpreting them, synthesizing across them, and proposing future research. The benchmark is a published review of the Spanish plume weather pattern and its 102-article corpus. In the initial searches, five tools collectively identified 57 articles, with the best individual tool finding 42. Institutional library access increased discovery from 42 to 82 articles for Codex and from 15 to 79 for Claude Code. Given the complete corpus of 102 PDFs, three tools correctly identified all 18 articles discussing the Spanish plume substantially, but incorrectly classified 40% to 79% of the 84 passing mentions as substantial. The same three tools produced syntheses evaluated blindly by two domain experts. Although largely accurate, the syntheses provided limited insight and sometimes contained unsupported claims, quotation errors, and exaggerated statements of consensus. Finally, all three tools proposed research directions broader than the Spanish plume literature, although two approached questions identified in the benchmark review. No single tool performed best across all tasks. We provide eight recommendations for researchers using AI tools in literature reviews. Our findings suggest that LLMs can assist with discovering and interpreting scientific literature, particularly when institutional access is available, but cannot replace expert judgment in evaluating evidence or synthesizing research.

Large organizations carry a long tail of infrastructure migrations that are worth doing but rarely urgent enough to compete with feature work, because each one costs significant engineering time per repository. This paper reports on our industrial experience from Automated Uplift, a multi-month initiative applying LLM-assisted migration pipelines across thousands of Bloomberg repositories. We examine two distinct production campaigns: a heterogeneous Python packaging migration (approx. 1,900 candidate repositories) and a uniform C++ build system migration (approx. 6,700 candidate repositories). Together, these campaigns generated 2,333 pull requests, resulting in 729 successful migrations. We detail our hybrid architectural solution which maximizes deterministic processing, confines the LLM to bounded file transformations within a pre-computed context, and incorporates runtime production telemetry (such as execution logs and usage traces). Our findings show that generating migrations turned out to be the easier part; the bottleneck was the cost, for owning teams, of reviewing, testing, and deploying them. These costs were lowered when changes could be verified directly in the pipeline before opening a PR (as with C++ compilation and test validation), and when the population was uniform enough for fixes to carry over from one repository to the next. We conclude by distilling our experience and findings into actionable, transferable patterns for architecting hybrid LLM-deterministic pipelines and scaling code migrations across the enterprise

When a language model answers, we can usually only estimate why. We build a transformer in which every step is a Born measurement - attention $|\langle q|k\rangle|^4$ normalised over each row, a measured product in place of the feed-forward nonlinearity, and a Born readout. With its denominators held at their values for the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share, to $10^{-11}$ or better. On e-SNLI natural language inference, with about 98k parameters outside the embedding, the model reaches $76.98 \pm 0.17$% test accuracy, level with a softmax transformer of the same size ($76.62 \pm 0.36$%). Deleting the word its evidence ranks first changes the answer in 45% of test pairs, out of the 53% any single deletion can change, against 32% for the softmax model explained the same way and 40% by its attention x gradient. Attention can look as much like the words annotators highlighted as the evidence does, yet accounts for the answer less well: deleting the word it ranks first changes the Born Transformer's answer in 32% of pairs. The same weights compile exactly to optoelectronic optics and to qubits, and a coherent attention layer becomes one circuit equal to the model (852 CNOTs on 23 qubits). On IBM's ibm_kingston, reading only the attention overlaps by Hadamard tests, the device gave the model's answers, also when a word was deleted. With calibration circuits as deep as the overlaps, each word's evidence with the device's attention row sums followed the exact evidence (correlations 0.997 and 0.932 on two pairs chosen in advance); where a pair's evidence is a small total of large opposite terms the earlier runs say it does not, and we say why.

Enterprise teams that deploy large language model (LLM) agents inherit a maintenance problem that traditional testing does not address. The model is non-deterministic, but the larger risk is that most of what changes an agent's behavior (a retired model version, a revised policy, an upstream schema change, a prompt edit meant to fix one case) originates outside the owning team and is discovered as a production defect rather than caught as a regression. This paper presents Agentic Behavior Engineering (ABE), a platform-agnostic methodology that treats observable agent behavior, rather than the prompt or the model, as the version-controlled engineering artifact. ABE adapts Design by Contract to non-deterministic systems. A specification-only behavior contract with stable rule identifiers states what the agent must do, must not do, and when it must abstain for lack of evidence; a lightweight scenario library turns each rule into concrete situations; and every engineering artifact (tests, runtime policies, telemetry) references the contract in one direction only, so that specification and implementation can change independently without drifting apart. The method organizes the work into five responsibilities (Specify, Ground, Control, Verify, Observe), treats missing evidence and a verified negative as different things, enforces authority outside the model, gates releases on per-scenario pass^k reliability rather than aggregate accuracy, and classifies every production run against the contract so that each finding routes to the asset that must change. We illustrate the method with an executable capacity-allocation boundary suite and a constructed worked example of a product-review approval workflow.

As agents become more capable, they are moving beyond one-off tasks toward sustained assistance across successive user goals. Such assistance requires agents to discover how unfamiliar environments work, reuse what they learn across tasks, and keep that experience current as environments evolve. Existing benchmarks increasingly study experience reuse and adaptation, but provide limited insight into which parts of accumulated experience agents revise, retain, or lose. To address this gap, we introduce ReScopeBench, a benchmark for the experience lifecycle across 18 interactive environments. It distinguishes structural experience about how an environment is organized and accessed from rule experience about reusable principles and constraints that guide decisions across tasks. To study how experience is acquired, reused, and revised, we propose an experience-first construction approach. It builds task streams around shared experience and introduces silent environmental changes. We evaluate task performance and track changes in saved experience throughout each stream. Across 20 models, we find that retained experience improves task performance, but agents struggle to keep it current as environments change. Even when the required experience is correctly recorded, agents do not always complete tasks successfully. These findings identify selective maintenance and effective use as complementary challenges for long-lived agents.

Naive on-policy RL is often bottlenecked by the capabilities of the base model --- insufficient coverage can result in an inability to learn from certain rewards. At the same time, rewards in modern post-training systems often come from rich textual sources, e.g. judge models with a rubric of criteria. Can we utilize such structures to provide a rich directional signal that enables improving on otherwise difficult to optimize rewards? In this work, we examine a framework we refer to as targeted guidance that learns from per-turn textual feedback via an on-policy self-distillation objective. Our focus is not only on directly improving hard-to-optimize behaviors, but on how to do so throughout the training process without degrading baseline RL performance. Over a range of settings across an agentic multi-turn software engineering environment, we find success by combining targeted self-distillation with a per-criteria RL reward, and by limiting feedback to turns where the policy specifically fails. We further examine the design choices that lead to stable entropy and reduce oscillatory behavior throughout training.

Learning on boundary representations (B-reps) requires robustness to the pose of a part, since CAD models are not stored in a canonical orientation. Existing B-rep networks either operate on world-frame coordinates and normals, and thus degrade under rotation, or encode each face in a local reference frame. Local frames yield invariance, but reduce inter-face messages to scalars and discard the relative orientation of adjacent faces. We propose SO3brep, an SO(3)-equivariant graph neural network on the B-rep face-adjacency graph. Node features are computed in local frames and mapped into a shared frame by Wigner D-matrices. Clebsch-Gordan tensor products couple these features with the directions between adjacent faces, so that relative orientation is preserved throughout message passing and the output remains SE(3)-invariant. Across three face-segmentation and two classification benchmarks with randomly rotated test parts, SO3brep achieves the highest accuracy among the compared B-rep networks and equivariant point cloud networks, with the fewest parameters. Its advantage is largest with few labeled parts, a setting typical of industrial CAD. Code is available at https://github.com/UGent-CVAMO/so3brep.

Large language models understand passengers' driving commands, but as driving agents they reason in generated text for seconds and must infer each action's consequences implicitly from the scene. We present Jev-AD, which pairs Jev, a typed decision model that answers predefined questions with probabilities, not generated text, with a language world model that predicts for every candidate action how nearby vehicles move over the next two seconds. Code turns these predictions into foreseen consequences, which Jev compares instead of inferring them from coordinates. In simulated highway merging (500 car-episodes), this foresight lowers the same decision model's collision rate from 26.0% to 2.4% at 0.25 seconds per decision, and Jev-AD achieves the passenger's intent in 93% of free-form command trials. In 500 replayed real-world NGSIM episodes, foresight lowers the collision rate from 48.2% to 2.8%, and the world model, learned from this traffic, foresees 93% of dangerous actions. Project page: https://ys-qu.github.io/jev-ad-website.

Large language model agents increasingly retrieve tools from large catalogs, yet most retrievers rank each tool independently by request similarity. For multi tool tasks, this recovers the goal matching tool while omitting prerequisite tools required by the workflow.We introduce LIBRA, a framework for Library Induction by Runtime Assembly that turns a flat tool catalog into an executable tool library. An LLM proposes per tool contracts of candidate invocations, reusable outputs and state changes, and success conditions; LIBRA then tests these compositions in a sandbox and retains the relations supported at runtime. Constructed before user queries, the library requires no dependency annotations, tool use trajectories, or task specific training, and supports both one-shot and agentic retrieval. On AppWorld, BFCL, and SciToolEval, LIBRA improves one-shot Recall and NDCG over the same dense retriever by 7.9 and 3.7 percentage points on average. Within ToolReAGt, it yields corresponding gains of 4.5 and 2.9 points after ten retrieval iterations.

Process mining derives process models from event data recorded by information systems, but its results are usually presented as two-dimensional visualizations that are separate from the environments in which enterprise models are created and maintained. Metamodeling platforms for extended reality (XR), such as MM-AR, allow users to define modeling languages and to display models in 2D and 3D, but they do not provide access to event data or mining algorithms so far. In this paper we therefore present a first proof-of-concept for showing how an XR-based metamodeling platform can be extended with process discovery capabilities without changes to the platform core through a separate Python service based on the open source PM4Py process mining library. We demonstrate the approach using three common discovery algorithms for obtaining Petri nets and BPMN models based on CSV or XES event logs. A technical evaluation using four public event logs shows that the element counts of the created model instances match those of the PM4Py models in all 16 configurations. The integration adds between about 8 and 53 seconds per run, and this overhead grows with the size of the discovered model rather than the size of the event log. Although the prototype does not yet use the third dimension to encode process information, it established the foundations for studying such approaches.

Robot foundation models have shown remarkable performance in robot control, yet adapting them to a new task still relies heavily on teleoperated demonstrations. Egocentric human videos offer a scalable substitute, but overlaying rendered robots on human videos or retargeting human poses into robot actions leaves significant visual or action gaps for high-DoF robot embodiments. To address this limitation, we propose RoboTryOn, which treats human-to-robot translation as virtual try-on: try-on keeps the person and swaps the garment, RoboTryOn keeps the demonstrated motion and swaps the human for the robot. Specifically, RoboTryOn converts a human video into a video of the target robot specified by a reference image, while preserving the original task, objects, and background. To maintain visual consistency, we translate selected keyframes autoregressively and use them to guide video synthesis with motion conditioned on the human hand skeleton. Additionally, we use an inverse dynamics model (IDM) that incorporates depth features to infer action labels from synthesized videos for policy learning. Our experiments across multiple humanoid embodiments show that policies trained with our synthetic data improve the average success rate on novel tasks, including challenging bimanual manipulation, from 2% to 60%, without any real robot demonstration of those tasks. Moreover, training with our synthetic data matches the performance of policies trained solely on teleoperated demonstrations at approximately a third of the collection cost, demonstrating the cost efficiency of our approach.

We study PAC learning when each example comes with a set of at most $\ell$ correct labels and a prediction is correct if it belongs to this set. Training may reveal the full set, only whether a proposed label is correct (bandit feedback), or one correct label chosen by a demonstrator (demonstrations). Our main finding is that demonstrations can be much harder than full sets, and even than bandit feedback, both in which classes can be learned and in how many examples are needed. For finite label spaces, a generalized Natarajan dimension characterizes learnability from full sets and from bandit feedback, in both the realizable and the agnostic setting. From demonstrations, however, a class is realizably learnable if and only if the intersections of at most $\ell+1$ of its hypotheses are learnable from full sets. Already with three labels, this condition can fail for a class learnable from a single full-set example or even a single bandit answer. When the condition holds, we show that $\widetilde O(\varepsilon^{-\ell})$ demonstrations suffice, and the exponent $\ell$ cannot be improved in general, even for classes learnable from $\widetilde O(1/\varepsilon)$ full sets or bandit answers. For countable label spaces and $\ell \geq 2$, our packed dimension, which asks for many disjoint groups of at most $\ell+1$ hypotheses with no common label, characterizes full-set learnability, and demonstrations still reduce to intersections. Agnostic learning from demonstrations, already for finite label spaces, is not determined by the dimensions of the intersection class that characterize realizable learning. Finally, we show that the uniform bound $\ell$ on the number of correct labels is essential. Without this bound, it is impossible to characterize learnability by any dimension determined by finite subclasses.

Surgical world models have the potential to provide intra-operative guidance, accelerate risk-free training and facilitate transfer of expert surgical maneuvers. However, the use of world models for transferring surgical expertise has been largely unexplored. Moreover, recent approaches in this domain have mainly relied on ground truth kinematics actions or expert annotations for specific tasks, limiting their applicability beyond the narrow range of tasks with ground-truth annotations. In contrast, existing works in the natural domain have shown promising results in learning latent actions solely from unlabeled videos and transferring the learned actions across contexts. We observe that this paradigm is short-sighted for the surgical domain, failing to capture the hierarchical and goal-directed structure of surgical procedures. Thus, we propose a novel multi-step action learning framework to learn context-invariant latent actions that are easily transferable across various surgical contexts. We perform extensive experiments on the publicly available datasets, SurgToolLoc-2022 and PitVis-2023, verifying our approach with both quantitative metrics and qualitative evaluation by surgical experts. We have published our code and the implementation details at https://github.com/bhattarailab/Surgical-Vision-World-Model.