arXiv AI

arXiv:2610.26950v1 Announce Type: cross Abstract: Multi-teacher on-policy distillation (MOPD) integrates domain specialists into a shared student, but determining how strongly each teacher should supervise remains challenging. Existing discrepancy-based allocation can assign strong supervision even when the student already matches or exceeds the teacher's task performance, since behavioral mismatch does not necessarily imply a remaining capability deficit. We propose Capability-Aware Multi-Teacher On-Policy Distillation (CA-MOPD), which combines specialist merging with capability-guided allocation to account for what the student has already learned. CA-MOPD first merges reinforcement-learning specialists into a strong student initialization, then uses verifier-estimated teacher--student capability gaps to adjust domain-level distillation strength. Supervision decreases as capability gaps close and increases when estimated student capability regresses, while dense token-level guidance remains provided by the teachers. Across six benchmarks spanning mathematics, code generation, and instruction following, CA-MOPD achieves an average score of 32.22, outperforming naive MOPD, Open-MOPD, and the merged initialization by 0.86, 0.83, and 0.74 points, respectively. Controlled comparisons further support the benefits of merged initialization and capability-gap guidance under matched training budgets.

arXiv:2610.27888v1 Announce Type: new Abstract: Approximate solutions of partial differential equations (PDEs) on unstructured grids can be represented on a graph. Numerical optimization involving stiff evolutionary and static PDEs often requires solving a large system of equations, which can be treated as a constraint. In this work, we present a fully differentiable graph-based implicit solver that incorporates this constraint. We evaluate its performance across four key tasks: a) linear and nonlinear forward modeling of evolutionary and static problems, b) solving inverse problems to estimate equation coefficients, c) solving inverse problems to estimate the right-hand side, and d) learnable graph coarsening to accelerate forward modeling. The stability and differentiability of our solver enable significant reductions in the complexity of forward modeling while efficiently addressing related inverse problems. These capabilities make it a powerful tool for different physics and engineering applications, in particular for geophysics, where inverse problems are commonly solved in order to reduce the model uncertainty.

arXiv:2610.27548v1 Announce Type: new Abstract: When an LLM agent acts on behalf of a user, it is expected to act within the authorization scope the user grants. Existing agent evaluations largely measure task completion or harmful consequences and can therefore overlook actions that violate this delegation without producing conventional safety or privacy harms. We introduce AuthorizationBench, a benchmark of authorization alignment grounded in documented real-world cases. A scalable construction pipeline produces 286 executable tasks spanning situational triggers, the presence or absence of direct harmful consequences, and ten application domains. We execute every task with four models at two reasoning-effort settings across three agent harnesses in an authorization-aware sandbox that keeps authorized and out-of-authorization paths executable and observable, yielding 6,864 trajectories. Agents act out of authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across the 24 configurations. Task completion does not imply authorization. Increasing reasoning effort leaves the aggregate rate nearly unchanged, with effects varying across models. On controlled pairs, agents act out of authorization at similar rates with (17.9%) and without (20.8%) direct harmful consequences, and more often under benign pressures than malicious injections (23.5% versus 17.6%, p=0.002). Our results show that task capability, additional reasoning, and consequence-oriented safety evaluation do not by themselves ensure authorization alignment. Moreover, agents that consult the principal before completing a task exhibit lower rates of out-of-authorization behavior (14.0% versus 21.6%, p<0.001), pointing to a tension between agent autonomy and preserving users' decision authority that calls for further human-centered study. Dataset and code are available at https://github.com/PEACH-Research-Lab/AuthorizationBench .

arXiv:2610.27307v1 Announce Type: new Abstract: For AI assistants to move beyond simple chat interfaces and reliably assist users with applications like Gmail, Venmo, and Slack, we need to improve their tool-use capabilities. However, existing tool use benchmarks are still limited and difficult to scale because they rely on human supervision or are too domain-specific. To make data generation scalable, we propose, TURTLE (**T**ool **U**se **R**elated **T**asks with **L**ong-Horizon **E**volution), an Unsupervised Environment Design (UED) approach to evolve existing tasks using a proposer, a solver, and a feedback generator large language model agents. The proposer is rewarded for difficult yet solvable tasks based on the solver's empirical performance, and the feedback generator summarizes the solver's execution traces for the proposer to exploit. TURTLE produces evolving, self-tuning test suites adaptable to different environment requirements, from supporting user simulators and dense document retrieval to interacting across multiple apps. Empirical results on AppWorld, $\tau$-Bench, and AutomationBench using Gemma 4 31B-it show that the evolved tasks target the desired difficulty range (e.g., a 165\% increase in the percentage of hard tasks on AppWorld), increase task diversity (in terms of reasoning and simultaneous tool interactions), and elicit a broader range of failure modes than the original tasks.

arXiv:2610.21682v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) can become off-policy due to training-inference mismatch, mini-batch updates, and asynchronous training. The resulting mismatch is especially harmful in Mixture-of-Experts (MoE) models, where rollout-training discrepancies can change routing decisions and lead to router drift. Rollout Routing Replay (R3) mitigates this mismatch by reusing expert selections from the rollout stage during training. Our analysis shows that routing replay introduces gradient bias by losing contributions from experts selected by the training router but not the rollout router. To address this, we propose Router Drift Control (RDC), which uses routing disagreement between rollout and training as a signal of off-policy staleness and selectively rejects high-drift responses. The sequence-level filtering avoids gradient bias introduced by replay while preserving native expert selection and the associated learning signals for retained responses. Experiments show that RDC achieves larger gains as policy staleness increases and remains effective in asynchronous RL. In asynchronous training, RDC combined with Truncated Importance Sampling (TIS) improves the average peak validation accuracy across AIME 2024 and AIME 2025 by 5.5 points over TIS and 4.5 points over R3.

arXiv:2610.21565v2 Announce Type: replace-cross Abstract: On-device deployment of large language models (LLMs) is increasingly adopted due to latency and privacy requirements, but exposes proprietary models to reverse engineering and intellectual property theft. Trusted Execution Environments (TEEs) provide hardware-enforced isolation for on-device inference, yet most TEE-based parameter obfuscation relies on unilateral matrix transformations that obfuscate only the output side. We identify a vulnerability in this design, termed the Solution Determinacy Property (SDP): the obfuscation secret is uniquely determined, exposing the structure of the true key and enabling attackers to break the protection. Under our attacks, representative methods GroupCover and ArrowCloak are reduced to negligible protection, with attack accuracy reaching over 2.4 times the full-shielding black-box baseline and recovering near-complete model functionality. To address this vulnerability, we propose BiShield-TEE, a bilateral framework that induces Ambiguous Solutions through joint obfuscation of input and output dimensions, hiding the true key among many equally valid candidates. Experiments across diverse models and datasets show that BiShield-TEE confines attack accuracy to the full-shielding black-box baseline, while achieving higher inference efficiency than prior methods on real hardware. Our code is available at https://github.com/laduiw/BiShield-TEE.

arXiv:2610.21337v2 Announce Type: replace-cross Abstract: On-policy distillation (OPD) effectively transfers capabilities from a teacher to a student through feedback on student-generated rollouts. However, generating complete student rollouts and computing teacher feedback for every token make OPD computationally expensive. We introduce Stochastic Truncated On-Policy Distillation (STOPD), which accelerates OPD by randomly truncating student rollouts and correcting the resulting sampling bias with blockwise inverse-probability weighting (IPW). At each training step, we sample a stopping position for each rollout before generation and let the student generate until that position or natural termination. We obtain teacher feedback on every generated token and use blockwise IPW to compensate for the lower probability of retaining later tokens while controlling the weights. We prove that this weighting preserves Vanilla OPD's gradient sum in expectation and derive one-step update-error bounds for SGD, Adam, and Muon. We evaluate STOPD across multiple teacher-student model pairs, task domains, and optimizers. STOPD reduces generated tokens by 25-48% while maintaining accuracy comparable to Vanilla OPD, achieving 1.3-2.2x training speedups.

arXiv:2610.19424v2 Announce Type: replace-cross Abstract: LLM system capabilities depend not only on model parameters but also on the executable harnesses that organize information acquisition, tool use, reasoning, and memory. Recursive harness evolution can automate their design, but how it can reliably discover generalizable systems from limited, noisy, and repeatedly reused feedback remains unclear. We study this question in agentic video understanding, where rich multimodal evidence and coupled mechanisms make harness optimization challenging. We first introduce VHarness-BRSI, a recursive self-improvement framework that evolves video harnesses around a frozen vision-language model. Analyzing over 4,000 evolved candidates, we find it competitive with hand-engineered systems but identify three interconnected failure modes: validation overfitting, search-budget dilution, and the early-experience trap. We trace all three to a common root: the same imperfect validation evidence drives both harness revisions and the proposer's search strategy. We therefore propose VHarness-SRSI, an evidence-governed framework in which every search decision rests on evidence that is transferable, sufficient, and gathered before the search commits. It restricts the proposer to content-free feedback, informs exploration with instrumented warm-up experiments, and develops mechanisms in isolated probe-race branches before factorial composition and joint refinement. Across four video understanding benchmarks, VideoHarness-SRSI improves the equally weighted mean held-out score by 6.2% relative to VideoHarness-BRSI and reduces the mean validation-test gap by 37.9%. Reliable harness evolution thus depends on how search is organized and informed, not only on how well it optimizes validation performance.

arXiv:2610.18654v2 Announce Type: replace-cross Abstract: Summary-token architectures such as Kwai Summary Attention (KSA) compress distant context into learnable Summary token per chunk, and hybrid variants interleave these layers with full attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-attention layers, the summary token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache memory compression but also reduces the computation cost in Full attention layer. Specifically, we found that the summary tokens can serve two roles: the KV memory and sparse indexer.

arXiv:2610.17832v3 Announce Type: replace-cross Abstract: Materials and chemical optimization must choose the next synthesis, characterization, or high-fidelity simulation from sequential observations under a limited budget. Large language models can use task knowledge to propose adjustments, but composing surrogates, acquisition functions, and search strategies is constrained by compatibility, parameter dependencies, and evidence requirements. We present Agentic-MatriBO, a capability-compiled agentic Bayesian optimization (BO) framework that represents BO components and selection conditions as composable capabilities. A language policy proposes BO recipes and candidate plans; a compiler checks proposals against the registered capability catalog; and a deterministic execution layer validates and commits candidates. Oracle observations are stored in research memory and inform later rounds. We compare Agentic-MatriBO with eight baselines on four offline finite candidate pools---COF, PCE10, Buchwald--Hartwig (BH), and Suzuki--Miyaura---using five seeds per condition; results are reported descriptively. Across the four tasks, Agentic-MatriBO has the best mean endpoint quality after orienting each objective so that higher is better. On COF, it reaches the tabulated optimum in all five seeds both in the end-to-end comparison and after shared initialization, where it also retains an advantage in cumulative quality. Its normalized area under the best-so-far quality curve during the BO stage (nAUC-BO) is substantially higher on COF, near parity with the best baseline on PCE10 and BH, and lower on Suzuki. COF ablations show that removing the capability compiler or evidence gate coincides with lower cumulative quality and fewer exact-optimum hits. The framework provides an inspectable interface for evidence-conditioned BO adaptation.

arXiv:2610.17697v2 Announce Type: replace-cross Abstract: AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck. A learned verifier could check proofs cheaply, but training one requires a correctness label on every step, and such labels are hard to obtain for open-ended proofs. Existing methods take them from human annotators, which is costly, or derive them from final answers, which proofs do not have, and formalizing each proof in Lean requires expert effort. We introduce LeanPRM, a framework that uses the Lean proof assistant to create step-level labels for natural-language proofs. A model writes a natural-language proof together with a paired formal proof, without feedback from Lean. Lean then labels every step, and a second model discards failures caused only by the formal syntax, so that the negative labels mark errors in the mathematics. LeanPRM contains 10,634 proofs with 152K labeled steps and no human annotation. A Qwen3-4B verifier trained on LeanPRM raises exact step-label accuracy on a held-out set of 200 proofs (LeanPRM-Bench) from 48.2 to 74.7, as opposed to 60.3 when the same base model is trained on PRM800K. It also selects better proofs in best-of-N selection on IMO-level problems, and outperforms the PRM800K-trained verifier on three of four external verification benchmarks. Training on LeanPRM improves verifiers at every size from 0.6B to 8B parameters on LeanPRM-Bench, Hard2Verify and IMO-GradingBench. Given the labels need no human annotation, LeanPRM can grow with the supply of formal statements, offering a path to step-level feedback on AI-generated proofs that are hard for humans to check.

arXiv:2610.14018v2 Announce Type: replace-cross Abstract: The rise of large language models (LLMs) has opened a path toward generalist embodied AI, raising the question of how LLMs can act in the physical world. One straightforward approach is training the model to generate actions: collect robot-action data and video demonstration, refine LLM architecture with an action head, and train a vision-language-action (VLA) model at scale. However, we argue that coding agents have the potential to realize generalist embodied manipulation. All they need is a simple yet effective robot harness. We present RoboHarness: a robotic harness that gives LLM agents a visual-geometric control panel so that they can directly understand and invoke embodied tasks. With RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks. The implementation is open source.

arXiv:2610.13644v2 Announce Type: replace-cross Abstract: Selecting a subset of features (columns) from a tabular dataset is a ubiquitous challenge in real-world applications. Industry experts seek to minimize model training costs, climate scientists require an interpretable subset of factors to understand extreme weather events, and clinicians need to choose among diagnostic tests due to financial and time constraints. When looking at the scientific literature to determine which feature selection method to use, practitioners across domains face an unrepresentative landscape of benchmarking. Despite extensive research, existing feature selection benchmarks provide limited guidance because they rely on outdated models, ad hoc evaluation protocols, or artificially generated data that do not reflect real-world problems. To close the gap between scientific benchmarking and real-world applications, we introduce SelectArena, a modern benchmark for feature selection. We compare 15 widely used feature selection methods on 71 manually curated real-world tabular datasets, covering classification and regression problems, using standardized evaluation protocols and state-of-the-art tabular (foundation) models. Our results reveal clear trade-offs across predictive performance, runtime, stability, and validity. Tree-based methods, CART and RFImportance, along with F-Test are Pareto-optimal when taking into account predictive performance and runtime. F-Test, information-theory-, and distance-based methods lead on stability and validity, with F-Test emerging as a balanced choice across our metrics. Search-based methods are slow, unstable and achieve low performance. SelectArena turns these findings into practical guidance and, as an extensible platform, supports future feature selection research.

arXiv:2610.03827v3 Announce Type: replace-cross Abstract: Similarity scores are often offered as evidence that a model shares structure with the brain or across languages. We ask what such a score reads when that structure is removed, or when the instrument measuring it does not work, in two settings. Across seventeen languages, a grammaticality probe transfers worse between more distant languages (r = -0.66). But the probe's own accuracy falls along the same axis and is at chance in four languages, four of the five most distant. Dropping them halves the explained variance, and because it also narrows the range of distances, the design cannot say how much of the gradient is the instrument. Counting the 272 language pairs as independent gives p = 0.0006 for a steering effect that is null when the seventeen languages are the unit (p = 0.155). In brain alignment, a training objective raises a language model's similarity to fMRI responses from 0.10 to 0.34 (reliability ceiling 0.54), yet targets with the brain correspondence destroyed still reach 0.31. A shuffled target scores about k/n against the real one, for rank k and n sentences, a baseline that can be computed before any model is trained. Both interventions work, yet we detect no distance-graded steering effect and no syntactic benefit from alignment. Before reading a correspondence score, measure how much of it survives without the correspondence.

arXiv:2610.00385v2 Announce Type: replace-cross Abstract: Replay selectors often rank cached trajectories by format feedback, confidence, freshness, or response length, although cache-level correctness and downstream learner utility are distinct objectives. We formalize this selection-to-learning gap and introduce FAER as an auditable full-trajectory replay framework. Its training-free fixed selector is a protocol baseline; FAER-UTILITY is the learner-aware selector fitted on disjoint calibration blocks. The normalized gradient alignment is reported as a baseline, while a disposable optimizer-aware virtual update supplies a magnitude-aware utility surface. The audit contract freezes observed fields and replay traces before evaluation labels are joined. On GSM8K with Qwen2.5-1.5B-Instruct, the matched learner study reports quality 0.6329 for the fixed selector, compared with 0.5482 for uniform and 0.6037 for format-feedback under 128 updates. Metadata-only cross-fitted calibration reaches $0.6476\!\pm\!0.0139$ over eight seeds (median 0.6481; paired 95% interval $[+0.079,+0.122]$) at 63,276 target-run tokens; its recorded full cost is 189,642 tokens and 3.48 GPU-hours including calibration. The completed FAER-UTILITY row reaches 0.6624 at 62,844 target-run tokens and 4.26 GPU-hours. Format-feedback selects records with correctness 0.6953, compared with 0.3594 for the fixed selector, despite the different downstream ranking. The completed comparison surfaces report the learner-aware ablation, same-seed gap, policy-optimization rows, and strict zero-shot transfer.

arXiv:2610.00327v2 Announce Type: replace-cross Abstract: Tool-using agents can expose citations and execution logs while leaving a critical association unaudited: whether the claim shown to a user is the claim emitted by the committed execution and supported by the cited source. A valid citation and a valid trace can therefore remain individually well formed while being transplanted across claims, actions, runs, or source versions. We introduce a claim-anchored execution contract that jointly binds the emitted claim, its exact source span, the ordered execution prefix that produced it, and the source version and access state observed by that execution. Each receipt contains an emission anchor that deterministically locates the claim inside a committed answer or claim-bearing action, together with source identifiers, offsets, hashes, quotes, and a domain-separated execution commitment. A deterministic integrity verifier reconstructs these bindings before semantic or task labels are joined. We separate this integrity plane from a pluggable support plane, so structural validity is not used as a proxy for entailment. The contract exposes seven independently testable properties: claim-emission binding, source binding, ordered-execution binding, oracle separation, persisted-object replay, execution-rerun consistency, and version/access binding. Across 1,280 cross-object attacks, the joint contract detects 1,275 substitutions (0.9961). Removing a targeted property reduces its attack-detection rate to 0.0156-0.0625. On an independently adjudicated 384-pair split, the conflict-aware support guard reaches F1 0.8865 and false acceptance 0.0729; on unseen failure families, these rates are 0.8679 and 0.0938.

arXiv:2609.39601v2 Announce Type: replace-cross Abstract: Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control. Yet vision-language-action (VLA) and world-action models (WAMs) take perception from general-purpose vision-language and video-generation backbones, which still fail in these settings. We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI establishes a new state of the art, averaging 73.68%, above the larger GPT-6 Astra (71.54%). As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. On RoboTwin 2.0, it outperforms every mainstream backbone we evaluate in all four out-of-distribution settings, by up to 24.8% relative to the strongest backbone. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m. We systematically analyze GroundingPI's pretraining in scale and data composition. Downstream autonomous driving and robotic manipulation improve as the pretraining is scaled. Analyzing the data recipe across these 11 perceptual capabilities shows dense grounding's substantial benefits for both, and OCR's potential as a catalyst for perceptual learning. These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.

arXiv:2609.39600v2 Announce Type: replace-cross Abstract: Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens. We view grounding as visual evidence extraction: objects, locations, and spatial relations are jointly constrained by the image and query, yet their dependencies do not imply an intrinsic left-to-right generation order. This distinction makes bidirectional diffusion a natural fit, allowing spatial hypotheses to emerge in parallel and be jointly refined through iterative denoising. We introduce GroundAnything, a 4B-parameter grounding foundation model that reconciles fast parallel decoding with precise localization through blockwise denoising. Training combines grounding pretraining from public datasets and dedicated data engines, direct AR-to-diffusion conversion with joint AR and diffusion objectives, supervised fine-tuning, and GRPO-based reinforcement post-training. Across 30 grounding benchmarks, our autoregressive variant, GroundAnything-VLM, establishes a new overall state of the art among similarly sized models at 72.42%, remaining competitive with GPT-6 Astra (71.35%). With entropy-guided decoding, GroundAnything also surpasses the prior state of the art at this scale, averaging 61.75% versus 53.32% for the fast MTP-based LocateAnything model. We further explore decoding strategies, showing that an optional self-speculative mode achieves a $4.51\times$ speedup over the AR counterpart with a 0.74 percentage-point drop in COCO F1mIoU. Infrastructure experiments show that progressive inference optimizations translate parallel decoding into practical speedups. These support efficient visual grounding in latency-sensitive real-world systems.

arXiv:2609.38662v2 Announce Type: replace-cross Abstract: Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.

arXiv:2609.36958v2 Announce Type: replace-cross Abstract: Repeated verifier calls are useful only when they contribute conditional information. We introduce VStress, an auditable replay contract, and VStress-CA, a correlation-aware allocation policy that estimates the conditional marginal information of an unqueried verifier on a sealed calibration split, discounts uncertainty, normalizes by call cost, and stops or abstains when the next call is not informative. The controller freezes its decision and cost ledger before joining the clean oracle; a dependence-shift alarm disables channel preference and falls back to exact-stop. The controlled audit gives the mechanism boundary: at 35% symmetric corruption, majority-5 improves balanced accuracy from 0.6578 to 0.7739, whereas at 65% it loses 0.1226 points. In the matched fixed-budget comparison, breadth, redundancy, and adaptive allocation obtain balanced accuracies 0.6048, 0.6375, and 0.6538, with 3.4216 calls per item and an RLVR score of 0.6417 for VStress-CA. Dependence diagnostics also increase from same-model repeats to cross-family channels, with conditional marginal gains of 0.0126, 0.0462, and 0.0913. These measurements turn correlation from a post-hoc warning into an auditable allocation decision.