PHYSICAL AI · 2026-11-02

Physical AI Brief

Daily cross-source signals for the Physical AI supply chain — silicon photonics, CPO, VLA models, humanoid hardware, embodied AI. Three streams, one page, zero filler.

355 items today · 288 arxiv · 5 SEC 8-K · 62 humanoid · 0 CN photonics

01 ARXIV · PHYSICAL AI PAPERS

288 items
  1. arxiv:2610.42149 · cs.LG
    Latent Reasoning in Continuous Space for Unified Multimodal Models
    Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho +1

    Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space. To bridge this gap, we introduce a Latent Reasoning framework in Continuous space (LARC), which alleviates the token-level bottleneck by reasoning over continuous hidden states. Specifically, the training framework consists of two stages: (i) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity. Notably, LARC achieves state-of-the-art performance on GenEval with a 7%p gain over BAGEL and demonstrates clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies.

    benchmark
  2. arxiv:2610.42148 · cs.CL
    Flide: Accelerating Decision Inference with Exact State Forking
    Bobo Li, Jiajia Song, Hao Fei, Wynne Hsu +1

    Decision models such as Jev have emerged as a highly promising paradigm for LLM agent systems, providing lightweight and reliable categorical reasoning for rapid System One execution. However, existing generation-oriented inference engines fail to support these workloads effectively, suffering from severe prefix recomputation overheads or distorted output probabilities. To address this challenge, we propose Flide, an efficient and exact inference engine that decouples the logical tree structure of a decision request from its physical execution within a single flattened token stream. Through ForkAttention, Flide navigates request dependencies using GPU-resident indices across both standard attention and recurrent Gated DeltaNet layers, guaranteeing in exact arithmetic that every query branch evaluates identically to an isolated forward pass. By expressing branch topologies directly in GPU-resident metadata, Flide accommodates diverse request mixtures with a compact set of pre-recorded execution graphs while enabling expanding agent histories to resume from checkpoints saved in earlier passes. Extensive evaluations on public hybrid and decision models demonstrate that Flide answers queries up to 9.5 times faster than vLLM and SGLang while staying closest to a float32 reference in total variation distance. By unifying physical execution throughput with mathematical exactness, Flide establishes a principled foundation for scalable, latency-critical decision intelligence in agent systems.

    agentllm agentagent system
  3. arxiv:2610.42147 · cs.AI
    Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention
    Ales Kubicek, Andrea Caciolai, Liangzu Peng, Sukrit Sumant +6

    Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context windows. However, at ultra-long context (above 1M tokens), full attention is prohibitively expensive. A promising approach is sparse attention, in which a dedicated indexer selects a subset of keys for the attention. Yet current methods of indexing either still have quadratic cost in context size or lose accuracy at ultra-long context scale. We address this with Latent Sparse Attention (LSA), an efficient attention method that maintains accuracy at ultra-long context. In LSA, keys are hierarchically clustered in a learned latent space, and queries score the hierarchy of clusters instead of scoring every key, making prefill, decode and training faster at long context. We propose two variants, Distilled-LSA (D-LSA), which can be applied to pre-trained models via distillation, and Native-LSA (N-LSA), which is trained natively on the next-token loss alone. On GLM-5.3-FLASH (QWEN-3.8-27B) at 10M tokens, D-LSA speeds up one attention layer by 22.7x (72.4x) in prefill and 4.6x (42.5x) in decode over its original sparse attention (full attention), while keeping similar accuracy across a variety of tasks. N-LSA matches the accuracy of full attention with a training speedup of 1.6x per layer at 128k tokens. LSA offers practical and efficient attention for models to benefit from ultra-long context.

    long contextagentic
  4. arxiv:2610.42146 · cs.RO
    Hierarchical World Models with Implicit Dynamics
    Gaoyue Zhou, Yvonne Wu, Zichen Jeff Cui, Nicolas Ballas +3

    Planning with world models enables test-time adaptation for embodied agents, offering superior generalization over direct policy learning. However, long-horizon tasks remain hindered by cumulative errors and search complexity. Existing hierarchical methods often use rigid macro-actions or latent skills, leading to brittle representations and limited fine-grained control. We introduce ID-WM (Implicit Dynamics World Models), a hierarchical world model that decouples state-space reachability from local control. Our framework utilizes a high-level model to sample feasible transitions across the state manifold and a low-level world model as an inverse dynamics solver to ground these transitions into precise actions. This allows the planner to identify reachable subgoals without being constrained by low-level execution details. By prioritizing local physical feasibility over global behavioral patterns, ID-WM synthesizes novel long-horizon plans from short segments, transcending the specific behavioral modes seen during training. We demonstrate its effectiveness across three navigation and manipulation suites, achieving a 38.5% improvement over state-of-the-art policy and diffusion-based planners. Crucially, by leveraging this hierarchical structure to mitigate compounding errors, our approach yields a 135% performance gain over single-level world model MPC, particularly in complex, unseen configurations for long-horizon tasks. Videos can be viewed at https://id-wm.github.io

    embodiedmanipulationworld modelembodied agent
  5. arxiv:2610.42145 · cs.CV
    SemEllipsoid: Semantic-aware Shape Abstraction with Ellipsoid
    Kuankuan Cheng, Haocheng Yuan, Changjian Li

    Representing 3D shapes using a set of primitives - known as shape abstraction - is an actively studied problem in computer graphics with wide applications in shape understanding, analysis, and simulation due to the compactness and computational efficiency of the primitive-based representation. Prior work has primarily focused on geometry-aware shape abstraction, where the objective is to approximate the input shape with a minimal set of primitives while preserving geometric fidelity. However, these methods often overlook semantic consistency, inevitably leading to primitives that span across multiple semantic parts. Such inconsistency can cause undesirable artifacts in downstream tasks such as part-level shape editing or manipulation. In this paper, we introduce a new task - semantic-aware shape abstraction, where the goal is to produce abstractions that are both geometrically faithful and semantically consistent, ensuring that each primitive adheres strictly to a single semantic part without crossing semantic boundaries. We propose SemEllipsoid, a novel method to address this problem using ellipsoidal primitives. The core idea is to inject semantic constraints into the primitive optimization process leveraging the availability of semantic cues from large-scale foundation models. We design a fully differentiable projection-blending-evaluation scheme to measure semantic divergence against input semantic masks, and formulate an alternating optimization strategy to iteratively solve for optimal semantic assignment and primitive configurations. Our method is non-learning-based, requiring no data collection or network training, and generalizes well to shapes from arbitrary categories. It also supports different levels of abstraction according to user demand. Extensive experiments and evaluations validate the effectiveness of our approach across several 3D shape collections.

    manipulation
  6. arxiv:2610.42141 · cs.RO
    RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory
    Shengguang Wu, Suning Huang, Cherie Ho, Ria Doshi +2

    Recent work shows that a frozen robot policy performs far better inside an agent harness, where control code decides when to query the policy, consult vision tools, or command the arm directly. Yet today these harnesses are built by hand and redesigned for each new robot or benchmark. We show that harness design can be automated, offering a strong alternative to finetuning on a novel setup. In RoboArchitect, a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the harness around the frozen policy. The harness evolves on a set of episodes for its optimization and is evaluated on held-out scenes and tasks. Without updating any model weights, RoboArchitect raises average success from 9.7% to 43.3% on unseen scenes (4.5x) and from 12.7% to 40.2% on unseen tasks (3.2x), across multiple benchmarks and policies. The gains are largest where the benchmark is far from the policy's training data: wrapping a general-purpose policy, the evolved harness yields a specialist that often outperforms even the finetuned version. Intuitive control strategies are automatically designed, such as lifting a jammed arm clear or reaching in horizontally when an obstacle blocks the way from above. The harness also evolves its own persistent memory, carrying learned experience, such as effective grasp depths, into later decisions. These results suggest that substantial headroom in robotic manipulation lies in the agent harness, and its design can be effectively automated and optimized without weight updates.

    manipulationrobot policygraspmemorypersistent memoryagent
  7. arxiv:2610.42139 · cs.LG
    Counting on Dummies: Privacy Amplification with Observable Counts and Truncated Batches
    Vitaly Feldman, Audra McMillan, Kunal Talwar

    In many practical deployments of Differentially Private Stochastic Gradient Descent (DP-SGD), the server processes data in batches. Randomly sampling these batches allows for better privacy bounds via privacy amplification results. Most sampling schemes with provable privacy amplification, including Poisson sampling and random allocation, result in variable batch sizes. The privacy guarantee of these schemes requires that the batch sizes are not revealed. In distributed settings, the server performing the aggregation or any party observing the communication traffic may learn these batch sizes. Furthermore, in both centralized and distributed settings, memory and computation constraints may require each batch to be of some maximum size $s$, necessitating truncation of the batch when it exceeds this limit. It is known that this truncation can result in a significant increase in privacy loss. In this work, we address both of these challenges via a simple mechanism that uses a random number of \emph{dummy points} to privatize the batch size and maintain the privacy guarantee even under truncation. We demonstrate that our mechanism is easy to implement in a distributed setting, and provide a general analysis of this mechanism for both Poisson sampling and random allocation. Our analysis is designed to support efficient and relatively tight privacy accounting. It also effectively amplifies the privacy of the count releases by subsampling, making the expected number of dummy points asymptotically smaller than what is needed to directly privatize the (sensitivity-one) batch sizes with non-negative noise. For Poisson sampling with truncated batches, we demonstrate empirically that our mechanism achieves better privacy--efficiency tradeoffs than the direct accounting for truncation recently given by (Ganesh, 2025).

    memory
  8. arxiv:2610.42128 · cs.CL
    DecisionRank: On-Policy Distillation for Decision-Model Rerankers
    Yanran Wu, Dian Yu, Chongyang Bai, Jieyi Zhang +9

    Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. However, transferring a generative LLM's ranking knowledge to a decision model remains challenging: the teacher expresses preferences through a ranked list of document ID sequences, while the student determines the ranking using scores from the first-token position. To bridge this gap, we introduce DecisionRank, a decision model trained through item-space on-policy distillation (OPD). Item-space OPD constructs the student's rollout by sorting its document scores into a ranked sequence. A generative teacher takes the resulting ranking list as inputs and provides conditional feedback to directly supervise those scores. Combined with offline ranking supervision, DecisionRank achieves state-of-the-art performance on both general and reasoning-intensive retrieval benchmarks such as TREC DL, BEIR, and BRIGHT. Across four benchmarks, item-space OPD consistently outperforms alternative training objectives. Even at 1.7B parameters, DecisionRank surpasses its 27B generative teacher on TREC DL19--23. At 8B, it outperforms its generative listwise counterpart while running $24.5\times$ faster on TREC DL and $16.7\times$ faster on BRIGHT.

    benchmark
  9. arxiv:2610.42116 · cs.AI
    ACCT: Asymmetric Cross-LLM Context Transfer
    Yuzhe Liu, Haolin Tian, Tonghan Wang

    In real-world deployments, large language model (LLM) systems are increasingly evolving from single-model invocation toward heterogeneous deployments in which different models are dynamically invoked for inference. When the same context is reused across models, however, each newly selected target model typically has to process it again, repeatedly incurring substantial computational costs. Therefore, we propose ACCT (Asymmetric Cross-LLM Context Transfer), which targets KV cache transfer between models of different scales within the same model family. ACCT transforms the KV cache produced by a source model into sparse KV states that can be directly consumed by a target model, thereby enabling cross-model reuse of context computation and decoupling context computation from downstream inference. ACCT consists of three lightweight components: the Cartographer identifies query-relevant context tokens for each target layer and KV head; the Allocator determines how many context tokens each target layer and KV head retains under a shared total KV budget; and the Compiler transforms selected source KV into sparse target KV. The target model can therefore perform inference using the transferred KV without reprocessing the original context. We evaluate ACCT in two complementary settings comprising five general-capability benchmarks and six highly context-dependent tasks. Experiments show that ACCT matches or exceeds the target model's original inference performance even when reducing the KV cache size by 75%. At a context length of approximately 19K tokens, ACCT's KV conversion achieves a 6.5-21.2 times speedup over target context prefill. These results demonstrate the feasibility of transferring context computation across models under a limited KV budget.

    benchmark
  10. arxiv:2610.42112 · cs.AI
    Subquadratic Attention with a Hierarchical Tree-Structured Indexer
    Leon Lufkin, Cutter Dawes, Alexander Speicher, Kamesh Krishnamurthy +2

    The quadratic scaling of attention is a fundamental bottleneck limiting the long-context capabilities of large language models (LLMs). A variety of approaches have attempted to circumvent this, but they either fail to match attention's expressiveness (e.g., sliding-window or linear attention variants with fixed-size state) or only achieve constant-factor speedups (e.g., sparse attention methods whose indexers themselves scale quadratically). In this paper, we derive an asymptotically subquadratic sparse attention method based on a hierarchical, tree-structured indexer, which we call Zyphra Hierarchical Attention (ZHA). ZHA constructs a tree of recursively compressed key summaries over the KV cache, which future queries traverse from the top down to retrieve a sparse subset of keys. Crucially, by iteratively growing the tree's depth, the total cost including routing and final attention grows sublinearly in sequence length, resulting in overall subquadratic (but superlinear) scaling for attention. We demonstrate that ZHA can be cheaply applied to existing pretrained models with only a short KL distillation training stage over a few billion tokens, and maintains performance on long-context retrieval and challenging reasoning, mathematical, and agentic coding evaluation tasks. Moreover, we show that our method is highly robust to tree geometry including very deep trees of up to 6 levels. Finally, we develop a detailed performance model and show that our approach scales subquadratically, dramatically outperforms competing methods beyond 1 million context, and ultimately has the potential to extend effective context windows to the hundreds of millions or billions of tokens.

    long-contextagentic
  11. arxiv:2610.42105 · cs.MA
    Strategic Tanking in Stable Matching Markets
    James P. Bailey

    Motivated by a broad literature in psychology, organizational behavior, and matching markets, we introduce strategic tanking: an agent deliberately makes themselves less attractive to selected potential partners in order to improve their eventual assignment. Unlike traditional manipulation in stable matching, a tanker does not misreport whom they prefer; instead, they influence how others rank them. We show that tanking has a striking connection to classical preference manipulation. Cutoff tanking, in which an agent tanks with every potential partner below a chosen threshold, has exactly the same effect on the stable set as truncating one's own preference list at that threshold. Thus an agent can reproduce the power of a classical truncation while leaving their own preferences unchanged and every match acceptable. Arbitrary tanking, however, can go substantially further: under proposer-deferred acceptance, tanking equilibria can lie outside the sincere stable set. More broadly, whenever the sincere stable set contains multiple matchings, profitable strategic behavior remains available through either traditional preference manipulation or tanking, even when one side is restricted to honestly reporting their own preferences.

    manipulationagent
  12. arxiv:2610.42104 · cs.LG
    Hyperbolic Inverse Reinforcement Learning
    Iulia Rusu, Anoop Praturu, Sergei Gepshtein, Tatyana Sharpee

    Inverse reinforcement learning (IRL) seeks to infer the reward function underlying observed behavior, but its predictive accuracy and interpretability depend on how that reward is computed from the underlying features. While prior IRL reward functions have been implemented through a variety of techniques including linear weights, neural networks, or other function approximators, none make assumptions about the geometry of the underlying feature space. Inspired by geometric organization observed in neural circuits, we introduce Hyperbolic inverse reinforcement learning (Hyperbolic IRL), which parameterizes reward as a function of Hyperbolic distance between the feature vector and a learned parameter vector, as well as its Euclidean distance variant. We tested our methods against a range of linear and nonlinear IRL baselines under both deterministic and stochastic policies. Our Hyperbolic approach achieved comparable performance to state-of-the-art models on established benchmarks and excelled in the low data limit and across dense environments. Applied to human navigation among multiple exhibits in a museum room, Hyperbolic IRL reveals reward structure that depends not only on individual exhibit-distance features, but also on non-separable conjunctions between them. Together, these results demonstrate that a Hyperbolic reward parameterization can improve both the predictive accuracy and interpretability of IRL, providing a framework for studying how non-separable, conjunctive structure can begin to explain behavior.

    benchmark
  13. arxiv:2610.42097 · cs.RO
    ROOF: Learning Humanoid Fall Recovery under Overhead Constraints
    Yidan Lu, Jiayuan Yu, Zongrui Cai, Jingwen Sun +4

    A fallen humanoid beneath an overhead obstacle faces a coupled recovery problem. It must create clearance while organizing support for a stable rise, since attempting to stand before making room can further increase obstruction. We introduce a reinforcement learning method with contact-supported geometric escape guidance, which rewards incremental reduction of robot-obstacle overlap without prescribing an escape path. To achieve this, a learned motion prior, together with posture-dependent support and joint-stall costs, is proposed to shape a single policy that learns when to clear the obstacle, how to establish support, and when to complete the rise, all using proprioception alone. In simulation, it recovers in 88.5% of trials beneath plates that can only be lifted, compared with 18.9% without geometric guidance, and in 98.6% of trials beneath free plates with randomized size, mass and placement. On a Unitree G1, the deployed policy recovered in 95 of 150 attempts across seven real-world obstacle scenes, including furniture and multi-object arrangements absent from training. Project page: https://lu-yidan.github.io/ROOF/

    humanoid
  14. arxiv:2610.42096 · cs.RO
    CausalWAM: Learning the Causes of Actions in World Action Models
    Juntao Cheng, Jinzhou Tang, Fan Feng, Henrik I. Christensen +1

    World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes actions. WAMs fitted to demonstrations do not, in general, guarantee that they will generate actions solely from the latent causes of the demonstrated actions (the action-relevant latent variables), as the model could just as well learn to imitate in this way whilst entangling the action-relevant latent variables with the nuisance variables. We therefore separate out what causes action from what prediction depends on. We prove that under a sparsity constraint the latent variables used to generate actions are an invertible function of the action-relevant latent variables alone and the true structure is identified up to a permutation of the latent variables. This requires neither sufficient temporal dependence nor any assumptions on how the unobserved regimes evolve and holds even when the action-relevant latent variables are statistically dependent on the rest. We therefore propose CWAM (Causal WAM), which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged. Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.

    action headworld model
  15. arxiv:2610.42085 · cs.LG
    Cost-Aware Stopping for Freeze-Thaw Bayesian Optimization: When Does One-Step Lookahead Suffice?
    Long Vu, Jongseong Brad Choi, Binoy Ravindran, Zhoulai Fu

    Freeze-thaw Bayesian optimization (BO) evaluates candidates in resumable increments but usually stops at a budget or convergence test. We maximize a budget-free utility, the best observed score minus a cost per increment, with a one-step lookahead: advance the candidate with the largest expected improvement (EI) under a fixed surrogate, and stop once no EI exceeds the cost. This rule, EI stopping, never falls below stopping immediately in expected utility, yet it can stop too early. We construct a certificate, a lower bound on the value of two-step lookahead, that proves such a stop premature, and repeating this correction never lowers expected utility under the surrogate. Under the surrogate, EI stops are premature on average at low costs and lose little at high costs. On three hyperparameter optimization benchmarks, a pre-registered test finds EI stopping noninferior to the correction in most high-cost settings, while the correction gains at low costs on two of the three; more information sharing in the surrogate raises EI stopping's utility. On the surrogates of two published methods, EI stopping trades lower scores for fewer training increments. EI stopping is a simple cost-aware reference whose utility depends on the surrogate.

    benchmark
  16. arxiv:2610.42082 · cs.CV
    Think in 360°: Benchmarking and Bootstrapping Panoramic Video Reasoning
    Lidong Lu, Lidong Zou, Xinyi Mao, Jiaxin Lv +4

    A panoramic video captures all directions around the camera, but relevant evidence can appear in different directions and at different times. Understanding it requires models to connect observations across the viewing sphere and time. We introduce PanoVideoBench, the first manually annotated benchmark for comprehensive panoramic video understanding. Its 1,000 questions span five categories and 15 question types, with diagnostic annotations for target search and panoramic capabilities. Evaluation of 11 proprietary, open-source, and panoramic-specialized models identifies cross-view tracking and sphere-time evidence integration as their weakest capabilities. To strengthen these capabilities, we construct PanoVideoCoT, the first reasoning dataset grounded in real panoramic video, through sphere-time-grounded synthesis that converts verified panoramic observations and successful agentic explorations into 13,465 QA-CoT pairs. Starting from a general video MLLM, PanoVideoReasoner learns from this corpus through Panoramic CoT SFT and Evidence-Privileged On-Policy Self-Distillation, a panoramic instantiation of Vision-OPD's OPSD in which the teacher uses verified perspective views to guide student-generated trajectories. It reaches 55.7% on PanoVideoBench, improving the backbone by 12.5 points and the previous SOTA by 2.6. The gains concentrate on the same spatial, cross-view, and temporal capabilities exposed by the benchmark, showing that grounded sphere-time supervision can equip a general video MLLM with panoramic reasoning.

    agenticbenchmark
  17. arxiv:2610.42081 · cs.LG
    Differentially Private Synthetic Data from Sliced Chebyshev Moments
    Lucas Rosenblatt, R. Teal Witter, Christopher Musco

    Research on practical differentially private (DP) synthetic data has largely focused on methods that approximately preserve low-dimensional marginals of a target distribution, and can thus only hope to preserve higher-dimensional, global structure as a lucky by-product. At the same time, existing methods that target global accuracy guarantees pay a price exponential in the data dimension, $d$. In this work, we address this ``curse of dimensionality'' by presenting a new DP synthetic data algorithm, SPLICE, that comes with a compelling global distributional guarantee, even in high-dimensions, and moreover, is practically state-of-the-art. Our algorithm is based on measuring noisy sliced Chebyshev polynomial moments of the target distribution and fitting a distribution to match, a twist on recent polynomial methods for 1-D synthetic data. We prove that, given a dataset of $n$ examples, SPLICE returns a private distribution with Sliced Wasserstein-1 distance $\tilde{O}(1/\sqrt{n})$ from the target, for constant privacy parameters. Notably, this bound is independent of the data dimension, $d$. Along the way, we prove a moment matching result of independent interest: for a distribution on the unit ball, matching the first $O(1/α)$ Chebyshev moments along $\tilde{O}(d/α^2)$ random directions guarantees $α$ Sliced Wasserstein error. Finally, we provide a practical implementation of the algorithm that borrows tools from the tensor decomposition literature and achieves state-of-the-art performance compared to prior DP synthetic data methods on an extensive benchmark.

    benchmark
  18. arxiv:2610.42068 · cs.LG
    RAMA: Retrieval-Augmented Multi-Agent Routing with Episodic Memory for LEO Satellite Networks
    Xin Ming Li, Peng Hu

    Low-Earth orbit (LEO) satellite networks present a challenging routing environment due to rapidly changing topology, traffic, and link conditions. Existing deep reinforcement learning (DRL) routing methods typically rely on current observations and information encoded in network weights, limiting their ability to exploit successful past experiences. We propose RAMA, a retrieval-augmented multi-agent DRL framework with Approximate Information State (AIS), in which each satellite combines its local observation, recent routing history, neighboring state information, and retrieved entries from a per-satellite episodic memory. Retrieved experiences are integrated through a Transformer-based decision network trained with AIS-based auxiliary objectives. With weights frozen in a 140-satellite Kepler constellation, RAMA reduces mean end-to-end latency by 13.4% and 99th-percentile latency by 66.7% relative to the state-of-theart double deep Q-network (DDQN) routing method, mainly by avoiding long detours. Updated online, RAMA remains stable while DDQN degrades. On the 1,584-satellite Starlink shell 1, the reduction in mean latency grows to 93.2%, and updating RAMA online lowers its mean latency by a further 25.5%. These results demonstrate the effectiveness of retrieval-augmented memory for scalable decentralized routing in LEO satellite networks.

    memoryepisodic memoryretrieval-augmentedmulti-agent
  19. arxiv:2610.42066 · cs.AI
    The Knowledge Layer Model (KLM): Evidence-Visible AI Output, Routing Diagnostics, and Negative Results
    Omer Asim Ozturk, Cihan Aydiner

    AI answers require more than apparent correctness: reviewers need to inspect their recorded basis. We introduce the Knowledge Layer Model (KLM), an open, implementation-neutral specification for inference records with versioned epistemic assessments, evidence relations, and explicit unknowns. SEDIM, the knowledge and evidence architecture of Nage AI OS, is a reference implementation. We reference KLM v0.3 (repository tag v0.3.0). Empirical material consists of separate internal experiments, not end-to-end validation. A historical platform-domain routing benchmark reports 137/150 classifications across 15 classes (91.3%; nominal 95% Wilson interval 85.7-94.9%); the numerator is recovered from a rounded summary, and historical example membership is unverified. A native-model weight-path configuration yielded 0/15 successful answers in a corrected two-source test. A capacity-growth comparison could not establish a reliable effect because of a defective legacy scorer and missing original outputs. Lexical and scoped diagnostics each yielded AUC 0.008 for silent-versus-contradicted examples, below the registered minimum of 0.65. Legacy packing lost gold-containing text in 118/221 selected cases; repaired packing retained it in 221/221. These measure text presence, not answer quality. Python and TypeScript verifiers reportedly agreed on six checks per synthetic fixture; a separate production disclosure was not accepted. Restoration, re-derivation, human-rubric support, boundary isolation, and bounded-response evaluations remain proposed. Findings are internal and author-reported; historical observations are distinguished from later reproductions. Evidence visibility does not establish correctness, causal reliance, security, or regulatory compliance. Independent replication and a complete public audit package remain future work.

    benchmark
  20. arxiv:2610.42054 · cs.AI
    CRoS: Causality Reward on Scrambled CoTs Improves Accuracy and Reasoning Faithfulness
    Xiao Yu, Justin Chih-Yao Chen, Zaid Khan, Hyunji Lee +1

    Reinforcement learning with verifiable rewards (RLVR) has become the standard method for training language models to reason. However, RLVR typically relies solely on outcome correctness, which does not guarantee that the CoT causally contributes to the final answer. Without this connection, outcome rewards may fail to reinforce high-quality reasoning. Existing causality metrics assess CoT causality by perturbing the CoT, such as deleting reasoning steps or injecting errors, and measuring changes in the final answer. However, these perturbations inevitably alter information in each step. Consequently, high causality scores may reflect sensitivity to information changes introduced by perturbation rather than true causality. Such confounding effects further induce undesirable reasoning behaviors when these metrics are used as training rewards, including generating less informative CoTs and reducing exploration and self-correction. To address this problem, we propose CRoS (Causality Reward on Scrambled CoTs), a novel CoT causality metric that perturbs the order of reasoning steps without altering content, thereby preserving information within each step. On Reasoning Gym with Gemma-3-4B, CRoS outperforms existing causality and outcome rewards in out-of-domain accuracy, CoT quality, and CoT faithfulness. Specifically, CRoS achieves 46.8% accuracy on five unseen logic and games datasets, outperforming the strongest baseline (43.8%). It also achieves better CoT quality, consistently outperforming baselines in CoT-Pass@K from K=1 to 128, averaging 43.4% vs. 41.4%. For CoT faithfulness, CRoS achieves a hint acknowledgment score of 34.0% vs. 20.2%.

    self-correction
  21. arxiv:2610.42053 · cs.RO
    REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
    Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang, Haoyuan Chen +4

    Foundation models are becoming remarkably reliable at high-level reasoning, yet they remain too slow and costly to serve as fine-grained controllers. Conversely, fast policies like VLAs and whole-body controllers offer more efficient reflexes, but struggle to generalize. While pairing a deliberative brain (foundation models) with a reactive cerebellum (fast policies) seems natural for embodied systems, two obstacles remain: methods are built and scored on incompatible stacks, and their asynchronous coordination is difficult. For example, an agent that calls the VLA in the agentic loop can stall the fast execution, while a VLA that calls the agent on errors tends to re-trigger fallbacks repeatedly or completely miss silent failures. To address this, we propose REA (Routed Embodied Agent), a unified runtime where a deliberative System 2, an embodied router System 1 and reactive System 0 policies cooperate under a single golden rule - no running policy waits on the brain. Instead of querying the brain at each failure, REA employs a route memory that stores each policy's measured success ahead of time, a boundary-triggered judge gate that consults the brain on policy execution, and a shared-budget escalation ladder for bounded recovery. Evaluated on 3,000 test episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success and beats the best single-agent baseline in its own policy pool (94.7%) while reducing control steps by 25% and control time by 17%. In a separate 100-episode RoboTwin 2.0 study, REA achieves 84% success with VLA policies and 86% with VLA+code, compared with 80% for the strongest single VLA baseline MEDex. These results demonstrate learned and programmatic policies operating within the same runtime. REA also transfers to BEHAVIOR-1K and a real Unitree G1 humanoid, achieving the highest observed success among the compared methods.

    vlaembodiedhumanoidwhole-body controlliberorobotwin
  22. arxiv:2610.42042 · cs.LG
    A richness-sensitive Acoustic Biodiversity Index from species-level detections: classifier benchmarking, false-positive control and validation against standard soundscape indices in an African savanna and an urban zoo
    Muhammad Saeed, Dmitry Mikhaylov, Mohamed H. Alhosani, Yasser F. Al Wahedi +3

    Acoustic indices are widely used as rapid proxies for biodiversity, but they summarise the distribution of sound energy rather than the species producing it and often track richness poorly. We revise and test an Acoustic Biodiversity Index (ABI) built from species-level detections. Its primary form, ABI_E, is Shannon diversity of detections normalised by the regional species pool expected at the site and season (H / ln S_ref); temporal stability T is reported as a separate component. We first benchmarked four classifiers on a strongly labelled tropical dataset (1,216 calls, 24 classes) with recording-grouped cross-validation. BirdNET embeddings + logistic regression performed best (macro-F1 0.750 +/- 0.029), and random frame-level splits inflated accuracy from 0.41 to 0.70. We then applied BirdNET v2.4 with a location and season filter to recordings from a riverine savanna in the Zambezi Valley, Zimbabwe, and from three habitats at Emirates Park Zoo and Resort, Abu Dhabi, United Arab Emirates. A stricter post-hoc location filter removed 517 high-confidence detections of 9 implausible species, all at the zoo. After diel-matched rarefaction to equal effort (6 h), the savanna site held 21 species (Hill number 1D = 9.83; ABI_E = 0.38) against 1.3-13 species (ABI_E = 0.03-0.21) at the zoo sites. Across 149 recorded hours, ABI_E correlated strongly with richness (Spearman rho = 0.93), whereas six standard indices did not (best: ACI, rho = 0.48). The stability term was lower at the savanna site (0.39) than at the zoo (0.65-0.67), showing that it penalises the natural diel cycle of bird activity. Species-level indices with explicit false-positive control, a declared regional reference pool and effort standardisation provide a more interpretable biodiversity indicator than energy-based acoustic indices.

    benchmark
  23. arxiv:2610.42031 · cs.CV
    Any Forcing: Any-Step Generality and One-Step Stability for Autoregressive Video Generation
    Jiahua Xu, Yijie Fang, Nannan Wang, Xunhao Lin +5

    Autoregressive video generation enables efficient inference through streaming prediction. However, existing methods are typically restricted to fixed sampling steps, limiting the flexibility to balance generation quality and efficiency, while one-step generation suffers from degraded visual quality and error accumulation over long sequences, leading to content drift and generation collapse. To address these issues, we propose Any Forcing, a unified framework for achieving Any-Step Generality and One-Step Stability. Specifically, we introduce a unified adversarial discriminator that jointly performs real-distribution alignment, inter-step consistency alignment, and first-frame quality enhancement, enabling flexible sampling while preserving one-step generation quality. Furthermore, we propose a training-free adaptive long-term memory selection mechanism that selects historical states based on global semantic relevance and local value contribution, while remapping their temporal positions to preserve temporal relationships within the memory cache, thereby enabling stable one-step long-video generation. Experiments demonstrate that Any Forcing achieves high-quality and stable one-step autoregressive video generation, achieving 20.9 FPS and 24.7 FPS for short and long video generation on a single H100 GPU, respectively.

    memory
  24. arxiv:2610.42030 · cs.LG
    There Is No Free Loop: When Does Looping a Trained Model Pay for Its Inference Compute?
    Rhea Sanjay Sukthanker, Pashmina Cameron

    Looping re-applies a block of layers so that each token passes through the block more than once. Adding a loop to a trained model is attractive because it leaves the parameter count, and hence the disk size, the memory footprint and the training cost unchanged. Unfortunately, every generated token pays for the loop. Re-applying m of N layers makes each token about (N+m)/N times as expensive. Loops that a model learns during large-scale pretraining justify this price, but whether loops added after pretraining justify this cost remains unclear. We compare a looped model with two simple baselines. We first compare it with the same model trained identically without the loop (equal parameters). Secondly, we compare it with the best unlooped use of the same inference floating-point operations (equal FLOPs). Across Qwen3 and Qwen3.5 models from 0.8B to 9B parameters, only the looped 0.8B model beats both baselines. At 4B and 9B, the looped model does not beat the first baseline, i.e. the same model trained without the loop. This is not because the benchmarks are too easy, since on the same benchmarks the unlooped 9B model still scores up to 13 points higher than the 4B model, showing that there is ample room for improvement. Introducing loops in an already pretrained model therefore only pays for small models with a fixed parameter budget.

    memorybenchmark
  25. arxiv:2610.42029 · cs.LG
    Empirical Evaluation of Quantum Kernel Methods for Classification: A Comparative Study with Classical Kernels
    Ronak Bhalgami

    Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical $γ$, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs $5\times10^{7}$ circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.

    benchmark
  26. arxiv:2610.42028 · cs.LG
    CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
    Yanming Zhang, Eric Papenhausen, Klaus Mueller

    Reinforcement learning is a powerful paradigm for training autonomous agents and has achieved impressive performance in complex environments. However, this success often comes at the cost of interpretability, diminishing trust and complicating efforts to debug and improve agent behavior. To address these challenges, we introduce CausalXRL, a novel framework for explainable reinforcement learning (XRL). A key feature of CausalXRL is its use of causal graph reasoning, which provides transparent, structured, multi-level explanations of agent decision-making. We validate CausalXRL through comprehensive case studies and a two-part evaluation: (1) a quantitative analysis of explanation fidelity and causal-structure learning efficiency in benchmark RL environments, and (2) a qualitative expert study assessing explainability in the real-time strategy (RTS) benchmark MicroRTS. The quantitative results show that CausalXRL can provide faithful explanations while efficiently learning causal structures, and the qualitative expert study suggests that participants found CausalXRL useful for inspecting high-level RTS strategies.

    agentautonomous agentbenchmark
  27. arxiv:2610.42026 · cs.LG
    GhostPrefix: Synchronization-Free Universal Prefix Attacks for ASR Manipulation
    Shengding Liu, Ashwin Innuganti, Qiben Yan

    We present GhostPrefix, the first synchronization-free universal adversarial prefix attack against automatic speech recognition (ASR) systems. Unlike prior audio attacks, which are typically crafted per utterance, require synchronization with the victim waveform, or target narrow objectives, GhostPrefix learns a short, reusable audio segment prepended to arbitrary live speech to exert real-time control over ASR outputs without knowledge of the future utterance or temporal alignment. A key challenge is maintaining stable adversarial effects across unknown, evolving speech contexts. GhostPrefix combines token-level losses with a novel representation-level objective for stable, input-agnostic effects. GhostPrefix supports three attack modes: (i) prompted prefix attacks for semantic steering, (ii) targeted prefix attacks for transcript hijacking, and (iii) untargeted prefix attacks for transcription degradation. To remain perceptually inconspicuous, we regularize prefixes toward benign environmental-sound distributions and improve robustness against over-the-air distortions. We evaluate GhostPrefix across 13 open-source ASR models from four families. In digital settings, it achieves average attack success rates of 97.6% for prompted steering and 93.4% for targeted hijacking, while driving word error rate to 1.18 under untargeted attacks. Over the air, it achieves success rates of 91.5% and 89.5% for prompted and targeted modes, respectively. These manipulations propagate into downstream AI-assisted pipelines, including voice assistants and AI note-taking tools, causing systematic semantic misinterpretations. GhostPrefix exposes synchronization-free universal prefixes as a practical, under-explored threat to real-world ASR and downstream applications. Artifacts and demos are available at https://ghostprefix.github.io/.

    manipulation
  28. arxiv:2610.42021 · cs.RO
    Nexus-VLA: Bridging Human and Robot Experiences via a Shared Hierarchical Semantic Latent Space
    Biaoliang Guan, Tingyu Yuan, Wen Ye, Ziming Ji +4

    Large-scale egocentric human videos provide rich and scalable manipulation experience for robot policy learning, yet substantial embodiment gaps in visual appearance, kinematic structure, and action spaces hinder effective knowledge transfer. Existing methods often decouple cross-embodiment knowledge alignment from robot policy learning, making it difficult for human manipulation knowledge to directly and fully contribute to robot action learning. To address this issue, we propose Nexus-VLA, a hybrid policy learning framework that enables cross-embodiment knowledge transfer through end-to-end human-robot joint training, allowing human manipulation experience to directly contribute to robot policy learning. We first introduce a hierarchical cross-embodiment semantic alignment paradigm that progressively establishes shared manipulation semantics from task intent to fine-grained interactions, enabling more complete use of the manipulation knowledge contained in human videos. Building on this hierarchy, we design a shared manipulation latent representation that jointly encodes multi-level shared knowledge into an internal representation directly used for continuous action generation, improving policy learning efficiency while effectively leveraging human knowledge. Across RoboCasa-GR1, OOD settings, and real-world robot experiments, Nexus-VLA achieves an 80.5% task success rate on RoboCasa-GR1 and demonstrates strong advantages in human-data utilization, learning efficiency, and cross-scenario generalization.

    manipulationrobot policy
  29. arxiv:2610.42010 · cs.CL
    TurnTurk-GPT: Text-Based Turn-Taking Prediction for Turkish
    Ahmet Tuğrul Bayrak, Fatma Nur Korkmaz, Nur Yağmur Tuncer, Mustafa Sertaç Türkel

    This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish conversation against rule baselines, a zero-shot Turkish language model and a widely used multilingual end-of-turn detector. Evaluation uses word-level decision points on unpunctuated, lower-cased text approximating the output of streaming speech recognisers, recording-disjoint and speaker-disjoint folds, nested threshold selection and paired cluster bootstrap tests. TurnTurk-GPT outperforms the multilingual detector under both recording-disjoint and speaker-disjoint evaluation. It outperforms the zero-shot language model only when the test speakers occur in training; on unseen speakers the two are statistically indistinguishable. Removing the suffixes of the turn-final word reduces AP by up to 44%, against 2% for a sentence-internal word, locating the Turkish end-of-turn cue in verb-final morphology. Synthetic dialogue alone yields the weakest trained model, and when a speaker pauses the zero-shot and multilingual baselines remain close to chance. The model and the evaluation protocol are publicly released.

    evaluation protocol
  30. arxiv:2610.42004 · cs.AI
    TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents
    Xiangyu Kong, Ruirui Yang, Xia Hou

    Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes-without updating base LLM weights-by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (*100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC's final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.

    memoryagentllm agentbenchmark
  31. arxiv:2610.42002 · cs.CV
    GeoFlip: Benchmarking Geometry-Sensitive Physics in Video Generation
    Yuqi Li, Junming Liu, Bowen Hu, Ruize Niu +3

    Video generation models synthesize realistic motion, but can they respond correctly to changes in initial geometry? We introduce GeoFlip, a video generation benchmark centered on physical outcome flips induced by geometry. With shared physical settings and textual prompts held fixed, GeoFlip applies controlled interventions to support extent, object placement, and mechanism configuration, constructing paired scenes with contrasting reference events, such as remaining stable versus falling or tipping. These pairs test whether models translate geometric differences in the first frame into correct physical evolution. The benchmark comprises 482 scene pairs and 964 reference videos across six physical categories, with reference outcomes established through procedural scene construction, physics simulation, and numerical validation and verification of physical events. We evaluate 16 models for image to video generation along three dimensions: event correctness, structural integrity, and similarity to reference videos. We examine evaluation consistency with independent vision language judges. On the 442 conditions shared by all models, the highest event correctness averaged across tasks is 27.98%, revealing substantial limitations in physical generation conditioned on geometry. Model strengths vary across physical mechanisms, and the three evaluation dimensions capture distinct capabilities. Typical failures include ignoring geometric changes, generating incorrect object motion, bending rigid objects, and introducing additional components. Rank correlations between judges reach 0.906 for event correctness and 0.929 for structural integrity. By combining controlled geometric interventions, verifiable physical references, and multidimensional evaluation, GeoFlip assesses progress from visual realism toward physical consistency conditioned on geometry.

    benchmark
  32. arxiv:2610.42001 · cs.AI
    AutoAnalysis: An automated, self-updating meta-analysis of human-AI collaboration
    Michelle Vaccaro, Abdullah Almaatouq, Thomas W. Malone

    As generative artificial intelligence (AI) becomes more capable and widely used, understanding when and how humans should work with these systems is becoming increasingly important. However, the empirical evidence in this space offers conflicting recommendations, which leaves individuals and organizations without clear guidance on when to combine human and AI capabilities. Existing meta-analyses can help reconcile these findings, but they require substantial manual effort and cannot readily incorporate new evidence as it emerges. To address this challenge, we introduce AutoAnalysis: an automated, self-updating meta-analysis of when humans and AI work better together, and when they work better alone. AutoAnalysis' most recent run synthesizes 1,955 effect sizes from 420 papers. Across this body of work, we found that, on average, human-AI combinations performed significantly better than humans alone, but they performed significantly worse than the better of humans or AI alone. These findings highlight the important distinction between improving human performance and achieving human-AI synergy and underscore the need to evaluate collaboration against both human and AI performance alone. As new studies emerge, AutoAnalysis will integrate their findings into the meta-analysis, providing cumulative evidence about when humans and AI perform better together and how human-AI performance varies across AI agents, tasks, and processes.

    ai agent
  33. arxiv:2610.41998 · cs.AI
    SecPO: Principled Adversarial Training for Prompt Injection Security
    Sizhe Chen, Yibo Peng, Jaewon Chang, Chawin Sitawarin +1

    Recognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's operation. The existing best-defended LLM, Meta-SecAlign, is trained with the DPO loss but still suffers from attack success rates (ASRs) approaching 100% under adaptive prompt injections. One reason is that DPO anchors the model around the undefended reference model, which is insecure. We show an idealized reference model---an injection-invariant, utility-preserving oracle---can nevertheless be approximated, as its output distribution in training is interestingly accessible. This new observation motivates a theoretically principled SecPO loss. Another Meta-SecAlign shortcoming is that it trains on weak simulated attacks. We implement an efficient adversarial training (AT) pipeline that synthesizes strong injected inputs (best@240), and train the model on them. Our AT reduces ASRs by 1.5-14.5$\times$ across three SoTA adaptive attacks, compared with Meta-SecAlign; the SecPO loss reduces them by a further 2.4-4.6$\times$ compared with using DPO loss. In agentic workflows, our SecPO + AT Qwen3.6-27B is much more secure than Meta-SecAlign: 8.8% vs. 78.8% AgentDojo ASR@800 under the AttackerMovesSecond attack. Moreover, our defended model shows no noticeable reduction in benign utility in nine benchmarks covering agentic tool calling, instruction following, science, and math. We release our $\href{//github.com/Sizhe-Chen/SecPO}{\text{code}}$, $\href{//huggingface.co/datasets/Sizhe-Chen/Qwen3.6-27B-Instruct-SecPO-trainset}{\text{training data}}$ (with optimized injections from AT), and $\href{//huggingface.co/Sizhe-Chen/Qwen3.6-27B-Instruct-SecPO}{\text{model}}$.

    ai agentagentictool callingbenchmark
  34. arxiv:2610.41994 · cs.AI
    DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
    Zhangchi Feng, Harry Jiang, Yangyifei Luo, Ray Zhang +1

    Linear attention scales linearly with sequence length, but its finite-state memory can limit long-context retrieval. In gated linear recurrences, memory retention depends on state decay, which is shaped jointly by gate parameters A_log and dt_bias and by input activations. In controlled Gated DeltaNet (GDN) experiments, both parameters remain highly correlated with their initial values after training, highlighting the importance of initialization. Therefore, we introduce Dual-Axis Memory Initialization (DAMI), which jointly configures these parameters before training. DAMI assigns geometric timescales across heads for varied retrieval distances and a low-high-low input-sensitivity profile across layers. A small activation sample calibrates the GDN-family implementation to match both behavioral targets without adding trainable parameters. Across GDN, KDA, GDN2, Mamba-2, and RWKV-7, DAMI improves average RULER-style retrieval accuracy by 5.29 to 16.71 percentage points over the corresponding random-initialization baselines. Performance on language modeling, commonsense reasoning, and 2K real-world retrieval remains broadly comparable to those baselines. Our code is available at https://speedai.com/research/dami.

    memorylong-context
  35. arxiv:2610.41992 · cs.CV
    MoETok: Sparse Expert Routing for 1D Visual Tokenizers
    Jiahao Guo, Bao Tang, Haoxiang Cao, Jingfeng Yao +5

    Modern visual tokenizers aim to preserve fine-grained visual details while learning high-level semantic representations for downstream generation and understanding. Query-based 1D visual tokenizers encode images into flexible latent sequences, yet how individual learnable queries gather and encode image information remains poorly understood. We reveal that these queries spontaneously develop complementary roles under semantic supervision: a small set of Content queries captures high-level semantic information, while a larger set of Detail queries preserves fine-grained visual details for reconstruction. To adapt token processing to these heterogeneous roles, we introduce MoETok, a 1D visual tokenizer that routes tokens through sparsely activated experts. Jointly optimized with reconstruction and semantic objectives, the router learns role-specialized processing paths adapted to each token's representation. Across model scales and token budgets, MoETok consistently improves reconstruction, semantic representation, and autoregressive generation over dense tokenizers under matched supervision and active computation. On ImageNet, it combines strong reconstruction and semantic representation with state-of-the-art class-conditional generation within the 1D tokenizer family. We further validate MoETok at scale in a unified autoregressive model jointly trained for understanding and generation with a next-token cross-entropy objective. The resulting model combines competitive visual understanding with state-of-the-art generation among unified discrete models, matching or surpassing diffusion-based models such as BAGEL and Qwen-Image on benchmarks including DPG-Bench (87.61), GenEval2 (40.04 GM), and WISE (0.49).

    benchmark
  36. arxiv:2610.41987 · cs.CV
    FoAViS: Analyzing Learned Fixation Strategies in a Foveated Active Vision System
    Valentin Hassler, Jonathan Henrich, Yohaï-Eliel Berreby, Haoze Zhu +1

    Human vision is foveated and active: resolution is highest at the fovea and falls off towards the periphery, and fixations are actively directed to locations of interest. Building a machine system that processes information in the same way lets us study gaze behavior in a controlled setting where the system's components and objectives are known. The emergence of gaze characteristics can therefore be examined systematically, which may in turn inform our understanding of biological vision. This requires a model that can sample like the fovea, handle various tasks from classification to dense prediction, and be trained concurrently with a fixation policy. In our study, we present such a system and study its capabilities and the fixation policies that emerge from it. Our system, FoAViS, samples visual input by adopting the established log-polar foveation and accumulates information in a 2D spatial memory. Due to its 2D structure, the same memory can be read out globally for classification or densely for semantic segmentation, without any architectural change. Fixations are drawn from a 2D map by a policy trained concurrently with the task. We find that learned fixation policies select more informative, task-specific locations than non-learned heuristics. These policies also shape the task representations in a beneficial way: for both classification and segmentation, models trained concurrently with a policy outperform models trained with random locations, even when the latter are provided with learned scanpaths. Moreover, these scanpaths exhibit similarities to human gaze behavior by capturing aspects of empirical scanpath characteristics. Our work marks an important step away from standard passive, single-pass architectures, towards biologically grounded models of vision. It opens a path for studying goal-directed gaze behavior and information representation in the visual system.

    memory
  37. arxiv:2610.41983 · cs.CV
    TwinPast: Benchmarking Historical Evidence Use in Streaming Video Understanding with Evidence Contrastive Video Pairs
    Haotian Yang, Yong Chen, Cheng Peng, Haobo Wang +3

    High accuracy on streaming video benchmarks does not establish whether a model uses visual history. Our review of OVO-Bench and StreamingBench finds that recent frames suffice for 32% of sampled historical questions, while models also answer many questions correctly from the question and options alone. We introduce TwinPast to test whether answers correctly follow changes in history, using 350 human-reviewed video pairs spanning five historical tasks. Each pair holds the recent video context, question, and ordered options fixed while changing critical historical evidence and the correct answer. Each video is evaluated independently, and PairAcc requires correct answers for both histories. We evaluate 34 model and input configurations. For Qwen3-VL-8B and DeepSeek-Flash, historical inputs improve performance on TwinPast, reversing the recent-frame advantage on existing benchmarks. The strongest evaluated streaming system reaches 49.14% PairAcc, compared with 66.57-91.43% for the evaluated general-purpose models. This gap persists across all five tasks. Case studies further show that models select states from the wrong time or contradict recorded event order despite retaining the relevant evidence. These findings highlight the need to connect historical memory and retrieval with reasoning that uses evidence according to the question's temporal references and event relations.

    memorybenchmark
  38. arxiv:2610.41982 · cs.RO
    Suspended-Fabric Robotic Manipulation Surfaces: Modelling, Design Space, and Multimodular Extensio
    Pratik Ingle, Jørn Lambertsen, Kasper Støy, Andres Faiña

    Robotic Manipulation Surfaces (RMS) reposition objects by deforming the surface on which they rest. A recent class of RMS suspends a continuous fabric sheet from a sparse grid of vertical linear actuators lowering the actuator density of such surfaces but how to design such surfaces so that objects can be controlled everywhere on it, and how to extend it beyond a single module so that objects can pass between modules and achieve parallel, remain open. This paper presents a separable catenary model of the suspended sheet that is fast to compute and provides both the surface shape for given actuator heights and the actuator heights for a desired local slope. The closed form inverse lets a simple PID controller position objects on the sheet, and the model relates fabric length, module span and actuator stroke to the reachable workspace, yielding a minimum fabric length bound and design charts for these surfaces. The approach is then extended to a multimodular platform in which neighbouring modules share actuators. On a $2 \times 2$ prototype, we demonstrate that centimetre-scale objects are transferred between modules and positioned within 0.03 m over a $1 m^2$ workspace with only nine actuators, enabling multi-object manipulation and opening the way to parallel manipulation.

    manipulation
  39. arxiv:2610.41968 · cs.AI
    Prune Gradually, Adapt Sparsely: Efficient Pruning of Large Language Models at Extreme Sparsity
    Jean Lucien Randrianantenaina, Hung Nguyen, Nasib Ullah, Jinbin Zhang +3

    Weight pruning can reduce the storage and inference cost of large language models (LLMs), but preserving quality at high sparsity levels remains challenging. Existing methods face a quality-cost trade-off in this regime. One-shot pruning is cheap but degrades sharply as sparsity increases, even when followed by post-pruning recovery. A growing line of work instead optimizes the mask directly from the dense model under the end-to-end objective; among these methods, [Lee et al.,2026], the recent state of the art at extreme sparsity, reaches high quality but at the cost of model-sized optimizer state and auxiliary variables. To address this gap, we propose ETAPS, which prunes gradually and adapts sparsely: it tightens a global sparsity constraint in progressive stages while training a sparse delta with a small, fixed budget. Our key observation is that gradual weight removal keeps each pruning step small enough for sparse adaptation to compensate, allowing the adapted model to provide end-to-end gradients that guide the next pruning decision without model-sized optimizer state. Across six models from 125M to 13B parameters, ETAPS achieves lower perplexity than ELSA at every evaluated sparsity from 50% to 90%, with perplexity reductions of up to 42%. ETAPS is also far cheaper: at 90% sparsity, it prunes LLaMA2-7B on a single H200, where ELSA uses four, with $5.0\times$ smaller total persistent state and 49% fewer GPU-hours. Together, our findings point to a promising direction for achieving high quality in demanding regimes such as extreme sparsity at substantially lower cost.

    persistent state
  40. arxiv:2610.41962 · cs.LG
    When Is a Game a Markov Decision Process? Correspondences Between Game Theory and Reinforcement Learning
    Eduardo C. Garrido-Merchán

    Reinforcement learning and game theory describe the same object from two sides. Game theory says what a rational outcome of a strategic interaction is; reinforcement learning says how an agent could reach one through experience. The two literatures nevertheless keep separate vocabularies for identical structures, and the treatments that bridge them present the correspondence as analogy rather than theorem. This monograph supplies the theorems. A game is a Markov decision process in three distinct senses, and every table entry is labelled with the one that applies: a definitional identity, an exact reduction carried by a value- and solution-preserving map, or a structural analogy that fails under identifiable conditions. We take every standard class of game, from normal-form through repeated, stochastic, partially observable, extensive-form with imperfect information, Bayesian, graphical, Stackelberg, potential, coalitional, mean-field and evolutionary games, and identify for each the decision-theoretic model that reinforcement learning already studies under another name. The same treatment is applied to solution concepts: Nash equilibrium and the fixed point of the Bellman optimality operator, quantal response equilibrium and the soft Bellman operator, the Shapley value and credit assignment, Blackwell approachability and no-regret learning. We map each model class onto the training procedures sound for it and record the complexity of the associated solution problems. A final part collects the correspondences that do not hold: the non-stationarity that invalidates the independent-learner reduction, the impossibility results that forbid uncoupled convergence to Nash equilibrium, and the gap between the coarse correlated equilibria that no-regret self-play reaches and the correlated equilibria it is credited with.

    agentself-play
  41. arxiv:2610.41958 · cs.LG
    WASABI: Wasserstein Sampling via Bridge Interpolants
    Sebastian Salazar, Dilum Fernando, Michal Kucer, Emily Casleton +2

    Sampling from Boltzmann distributions is a fundamental problem in physics, chemistry, and biology. We introduce WAsserstein SAmpling via Bridge Interpolants (WASABI), a method that trains a diffusion model to generate samples from a target Boltzmann distribution given access only to the energy function. WASABI minimizes the KL divergence between the distribution generated by its diffusion model and the Boltzmann target using Wasserstein gradient flows. This approach builds on a principle underlying classical sampling methods such as Langevin dynamics. However, applying Wasserstein gradient flows to diffusion models is challenging because the updates require the gradient of the model's own log density, which is generally intractable. We derive a practical training algorithm that avoids computing this gradient. Once trained, WASABI can generate new samples without further evaluations of the energy or its gradient. On alanine dipeptide, WASABI outperforms the best reported baselines using an estimated 3.3 times fewer energy evaluations. It also achieves competitive performance on interacting particle system benchmarks.

    benchmark
  42. arxiv:2610.41956 · cs.LG
    Chronos-2: A foundation model for multivariate time series forecasting
    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer +21

    Accurate forecasting with uncertainty quantification is critical in a variety of scientific applications ranging from epidemiology and economics to energy systems and earth sciences. Yet most forecasting models still need to be developed and trained separately for each domain-specific dataset. Time series foundation models (TSFMs) aim to remove this limitation and enable the application of the same model across domains through large-scale pretraining. However, most existing TSFMs treat each time series independently and therefore cannot model interactions among time series and external drivers. Here, we introduce Chronos-2, a foundation model for zero-shot forecasting that models the interactions between multivariate time series and heterogeneous covariates, which commonly arise in scientific applications. The model infers relationships among time series directly from historical context and generates probabilistic forecasts without task-specific treatment. Across various benchmarks, Chronos-2 achieves state-of-the-art accuracy on average by improving skill scores by 5.2 percentage points relative to the strongest baselines and increasing win rates by 10 percentage points. Chronos-2 is also computationally efficient, generating forecasts for 100 time series in under one second on a single mid-range GPU, 50x and 1000x faster end-to-end than the best task-specific point and probabilistic forecasters, respectively. In addition, we demonstrate its effectiveness in diverse scientific settings -- including infectious-disease forecasting, macroeconomic modeling, carbon-flux prediction, and runoff simulation -- where the model matches or surpasses specialized methods developed for each domain. These results show that a single pretrained model can capture structure shared across temporal systems.

    benchmark
  43. arxiv:2610.41951 · cs.AI
    Aligning MLLMs with the Latent Structure of Human Cognition via Behavior-Derived Semantic Dimensions
    Ning E, Changde Du, Yizhuo Lu, Huiguang He

    Multimodal Large Language Models (MLLMs) have achieved impressive performance across a wide range of reasoning tasks, yet it remains unclear whether their internal representations and decision criteria align with the latent structure of human cognition. In this work, we propose a novel framework to align MLLMs with the multidimensional mental representations underlying human similarity judgements. We build on a previously established behavior-derived semantic embedding space, in which objects are represented along 66 interpretable dimensions derived from 4.70 million human Odd-One-Out (O1O) judgements. For each triplet, we infer the most salient decision dimension and convert it into a dimension-grounded linguistic rationale. This enables MLLMs to learn both the final choice and a justification consistent with the inferred behavioral criterion. Experimental results demonstrate that our approach improves model consistency with human similarity judgements while retaining competitive performance on standard multimodal benchmarks. Furthermore, using searchlight Representational Similarity Analysis (RSA) on an independent fMRI Natural Scenes Dataset (NSD), we observe increased representational alignment between model and human. Ultimately, our findings point to a data-driven route for incorporating human cognitive structure into MLLMs alignment. The code is available \href{https://github.com/ening-open/cogalign}{here}.

    benchmark
  44. arxiv:2610.41946 · cs.RO
    Sensorium: An Open-Source Multimodal Data Collection and Curation Library for Embodied Surgical AI
    Lorenzo Mazza, Ariel Rodriguez, Stefanie Speidel

    Embodied surgical AI requires reliable multimodal data, yet surgical data collection remains costly, difficult to repeat, and strongly dependent on the hardware setup. We present Sensorium, an open-source Python library for recording, processing, annotating, and curating multimodal surgical data from configurable robot and sensor streams using the Robot Operating System (ROS) and ROS2 middleware. The library combines operator-facing data acquisition, configurable processing, semantic annotation, and episodic dataset generation. We evaluate Sensorium across 1516 recordings of phantom, ex vivo, and in vivo porcine experiments involving multiple surgical tasks and sensing configurations, and report preliminary usability feedback.

    embodied
  45. arxiv:2610.41936 · cs.RO
    Adaptive Coverage Algorithms for Mobile Ad Hoc Networks in Dynamic Worlds
    Hugo Gandhi, Edwin Meriaux, Antonio Loría

    This paper extends an existing framework for the coordination of mobile agents in unknown worlds, by considering the possibility of dynamic instability. In the original problem, known as the Partially Observable Cooperative Guard Art Gallery Problem (POCGAGP) (arXiv:2512.09495), the multi-agent system must establish a communication network which completely covers an initially unknown constrained space. However, the POCGAGP fails to account for the unpredictability often seen in real-world deployment. To address this, the dynamic case (DPOCGAGP) introduces two settings which can influence the mobile network during deployment: Agent Loss and World Change, which represent dynamic changes to the world. Within this study, we present advancements to two state-of-the-art algorithms, the Centralized Agent-Directed Exploration for Network Coverage (CADENCE), and its decentralized analogue, DADENCE, to provide solutions in these cases. We then compare their performance against another state-of-the-art benchmark algorithm, and show that the adapted algorithms outperform across our benchmark test set, which comprises over 4500 test cases of varied size and structure.

    agentmulti-agentagent systembenchmark
  46. arxiv:2610.41933 · cs.LG
    Binary Sparse Codes: Where are we on Energy Efficient Representations?
    Jose Miguel Ramos, Luis SaCouto, Andreas Wichert

    Biological neural systems achieve impressive computational capabilities under tight metabolic constraints, relying on sparse cortical representations that minimise total neural activity. This combination of efficiency and representational power has motivated growing interest in sparse coding as a computational strategy for artificial systems. Artificial models designed to produce binary sparse codes span three broad paradigms, theoretical methods grounded in efficient coding, deep learning methods trained via backpropagation, and biologically plausible methods that learn through local Hebbian rules, each yielding codes whose quality depends not only on achieving low activation density but also on preserving the informational structure of the input. The latter property, however, lacks an agreed-upon quantitative definition, leaving the field without a principled basis for comparing methods. We propose a set of metrics that formalise information preservation as similarity preservation, measuring how faithfully a binary sparse code reproduces the relational geometry of the input space, and apply this framework to compare four representative methods under matched conditions. The results reveal a consistent trade-off in which models that achieve high information preservation do so at the cost of sparsity, while models approaching the uniformity of an optimal random sparse code tend to sacrifice it. Competitive Sparse Codes, a biologically plausible approach, is the only architecture to occupy a region of near-maximal sparsity while retaining meaningful information preservation. These findings frame the simultaneous satisfaction of both factors as an open challenge for the field and provide a reusable evaluation framework for measuring progress toward it.

    evaluation framework
  47. arxiv:2610.41910 · eess.SY
    Safety Verification for Voltage Control in Inverter Dominated Radial Distribution Networks under Actuator and Sensor Adversarial Attacks
    Georgios Pantazis, Revi Boelen, Michelle S. Chong

    This paper develops a safety verification framework for inverter-based radial distribution networks under sensor and actuator attacks. Each customer is equipped with a droop-based controller to regulate its voltage within a desirable bound of a reference, i.e., to achieve safety. In our setting, attacks and disturbances enter the droop-based controllers, while actuator attacks corrupt the reactive power injected into the network. Consequently, invariance of the controller state alone does not guarantee compliance with inverter capacities or customer voltage limits. To address this challenge, we derive conditions to guarantee that the reactive power state remains inside a predefined ellipsoid, while at the same time ensuring that the attacked reactive power injections and customer voltages respect predetermined operating limits that hold uniformly for the entire predefined ellipsoid. Together, these conditions guarantee safety under all admissible disturbances and state-dependent sensor and actuator attacks. Finally, we express the resulting containment condition as a linear matrix inequality (LMI) feasibility problem. A five-customer radial network based on the CIGRE benchmark corroborates our theoretical developments in simulations.

    benchmark
  48. arxiv:2610.41909 · cs.AI
    Structure-Aware Skill Retrieval and Arbitration for Optimization Modeling
    Anran Liu, Benzhao Hou, Zongsheng Cao, Daichao Wang +2

    Large Language Models (LLMs) can turn natural-language optimization tasks into solver-checked answers, but current agents remain brittle under narrative rewrites and often store reusable experience at the instance level. We introduce \textsc{ArchOR}, an archetype-centric skill agent that builds on \textsc{OptSkills} by adding two retrieval-focused modules. \emph{Centroid-Aware Representation Learning} (CARL) refines the ordering of cached skill representations through a low-rank prototype residual. \emph{Signature-Guided Retrieval} (S\textsuperscript{3}R) attaches open-dimensional structural signatures to skills, prioritizes candidates by weighted Jaccard overlap, and bounds semantic re-ranking to a top-$K$ set. A baseline-preserving arbitration rule determines when the structural route should replace the original selector. With the same DeepSeek-V4.1-Flash backbone and released skill library, \textsc{ArchOR} reaches 65.00\% Micro-Avg.\ Pass@1 across five benchmarks, improving over \textsc{OptSkills} by 1.64~pp Micro-Avg.\ and 2.68~pp Macro-Avg.

    agentbenchmark
  49. arxiv:2610.41905 · cs.CL
    Completeness and Cost of Retrieval Structures for Multi-Tool Requests in LLM Agents
    Shuyang Zhang

    An agent that must call several tools to fulfill a request fails when the tool retriever omits any one of them, yet tool retrievers are mostly built and evaluated on queries requiring a single tool. We compare representatives of six retrieval structures (flat lexical and dense retrieval, LLM query decomposition, LLM-navigated hierarchies, server-then-tool matching, dependency graphs, and trained set-level retrievers) across two multi-tool testbeds: 345 AppWorld queries with expert dependency labels and 103 MCP-Bench tasks over 257 MCP tools. We measure completeness@k, the share of queries whose entire required tool set appears in the top k, alongside the LLM calls, tokens, and latency that each structure incurs at retrieval time. A motivating analysis on all 7,961 ToolRet queries shows that aggregate completeness mixes two regimes: completeness at 10 is .58 to .67 on single-target queries and .08 to .17 on queries with four or more targets, and the ordering of retrievers on single-target queries reverses on multi-target ones. Across the two testbeds, none of the compared structures that is untrained on the testbed and makes at most one LLM call per query completes more than 10% of AppWorld queries at k = 16. The two structures that do better incur different costs: a set-level retriever trained on 75 to 90 in-domain queries reaches .33 on AppWorld and .58 on MCP-Bench with no LLM calls, while an LLM-navigated hierarchy reaches .30 and .40 with a median of 31 to 38 LLM calls and about five minutes per query. Their failure modes also differ: the trained retriever mostly ranks the missing tools just below the cutoff, whereas the hierarchy finds the target API and stops before collecting its prerequisites. In a downstream check with the official AppWorld agent, tasks with a complete retrieved set succeeded 21 of 21 times, whereas incomplete ones succeeded 2 of 39 times.

    agentllm agent
  50. arxiv:2610.41903 · cs.LG
    Statistical Foundations of Reinforcement Learning: A Non-Asymptotic Perspective
    Yuxin Chen, Yuejie Chi, Jianqing Fan, Gen Li +2

    Reinforcement learning (RL) embodies a foundational paradigm for sequential decision-making in the face of uncertainty. The rapid growth of model complexity in emerging applications has significantly amplified the challenges of achieving data-efficient RL, particularly in sample-starved scenarios where data collection may be costly, time-consuming, or even high-stakes. Consequently, understanding and improving the data efficiency of RL algorithms has become a central problem of both fundamental theoretical interest and practical importance. Despite decades of research, the statistical underpinnings of RL remained remarkably underdeveloped until relatively recently, particularly in the finite-sample regimes most relevant to modern practice. Over the past few years, there has been a surge of activity advancing the frontiers of non-asymptotic statistical RL theory, drawing on ideas from high-dimensional statistics and probability, information theory, and statistical learning theory. In this monograph, we present a non-asymptotic treatment for RL based on tabular Markov decision processes, focusing on three canonical sampling mechanisms --- RL with a generative model, online RL, and offline RL --- together with two algorithmic paradigms: model-based and model-free approaches. Centered around non-asymptotic sample complexity analysis, our treatment emphasizes minimax-optimal algorithm design, particularly methods that attain statistical optimality without significant burn-in costs, as well as sharp statistical characterizations of several widely used RL algorithms. Whenever possible, we provide largely self-contained analyses together with intuitive explanations. We also briefly discuss several extensions, such as robust RL, competitive multi-agent RL, federated RL, uncertainty quantification, and RL from human feedback.

    multi-agent
  51. arxiv:2610.41892 · cs.RO
    Online Learning of Heterogeneous Locomotion Dynamics under Reinforcement Learning Policies for Precise Loco-Manipulation
    Tianlin Zhang, Weipeng Xia, Lingwei Zhang, Yiyang Shao +2

    How to reuse a pretrained locomotion policy for precise loco-manipulation is still an open problem. We propose an online method to learn the closed-loop dynamics under a frozen locomotion policy. The learned model predicts floating-base motion and improves loco-manipulation accuracy. However, unobserved latent conditions of the policy introduce heterogeneous responses, meaning that the same observed state will lead to different motions. Direct learning without accounting for this heterogeneity introduces bias and fluctuations into dynamics estimates, degrading floating-base prediction and end-effector tracking. To address this problem, a contaminated Gaussian model is proposed to approximate the effects of latent conditions. The model uses two Gaussian components to approximate predictable and unpredictable responses under the same observed state. Then, the contaminated Gaussian model is combined with a Gaussian process to construct an online learner using random Fourier features (RFFs). The learner estimates each response's probability of belonging to the predictable component and uses it to weight online updates, capturing predictable responses while reducing bias and fluctuations caused by heterogeneous observations. With consistent and unbiased dynamics estimates, model predictive control (MPC) predicts floating-base motion accurately and enables precise loco-manipulation. A finite-sample error bound is established to ensure uniform convergence in online learning. Simulation and hardware experiments demonstrate consistent end-effector tracking across different manipulator configurations and tasks, with errors of 0.01 m.

    manipulationmanipulatoronline learning
  52. arxiv:2610.41885 · cs.CL
    How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces
    Donghang Duan, Qingyu Liu, Yifei Liu, Na Zou +2

    Large language model (LLM)-based agents interact with external tools through protocols such as the Model Context Protocol (MCP), producing tool-call traces that may expose behavioral signals about users. In this work, we systematically study the risk of user-profile inference from MCP tool-call traces, where observers infer user attributes from protocol-valid tool interactions without access to user conversations. We introduce PrivacyTrace, a benchmark containing 1,000 synthetic user profiles and 4,000 agent trajectories across four domains, generated with real MCP tools from 25 servers. Experiments with four mainstream LLM attackers show that tool-call traces enable a 26.25% attack success rate for profile inference, compared with 17.97% from prior-based guessing. Further analysis reveals that tool arguments and returned results provide the main trace-grounded evidence, while linking traces across sessions further amplifies profiling risk. These findings demonstrate that MCP tool-call traces can act as effective behavioral side channels for user profiling, motivating the development of privacy-aware mechanisms for agent-tool interactions. Code and data are available at https://github.com/AI45Lab/PrivacyTrace.

    agentbenchmark
  53. arxiv:2610.41877 · cs.RO
    Movens-Zero: Scaling Humanoid Motion Foundation Models with Large-Scale Human Videos
    Wendong Bu, Tinghong Ye, Qizhou Wang, Jiacheng Li +20

    Scaling humanoid motion foundation models is constrained by the scarcity of diverse, robot-executable data and by the difficulty of jointly learning from heterogeneous multimodal inputs while supporting real-time closed-loop control. We present Movens-5M, constructed from raw Internet videos through clip generation, quality filtering, semantic annotation, motion recovery, and physical refinement with humanoid retargeting and execution verification. It contains 15 million human-motion sequences and 5 million physically executable humanoid-motion sequences spanning over 10,000 hours, aligned with text, RGB video, speech, and music. Building on this data, we introduce Movens-Zero, a scalable understanding-generation-execution architecture that supports joint learning across modalities and real-time closed-loop execution. Task-mixed flow-matching training jointly optimizes a single action expert across seven generation tasks, enabling direct motion generation from text, speech, music, and video. Multimodal Real-Time Chunking connects overlapping chunks under continuously updated instructions and robot states for continuous closed-loop execution. We further introduce Movens-Intent, a four-track benchmark with 8,192 expert-verified condition-motion pairs across text, speech, music, and video. Experiments show consistent gains across all four modalities when scaling execution-verified data from 1% to 100% and action-expert capacity from 64M to 911M parameters, while demonstrating that data quality and cross-modal supervision improve generation and execution.

    humanoidbenchmark
  54. arxiv:2610.41868 · cs.RO
    META-Robot: Self-Cognition and Capability Scaling in Agentic Robots
    Xuanzhi Feng, Zeyu Liu, Weiheng Gao, Xinchen Li +10

    LLM-based robotic agent systems can generate programs for robot control, yet their generalization to new tasks often depends on elaborate scaffolding. Expanding their skill repertoires alone offers limited guidance on how to assess new capabilities or identify what remains beyond their reach. We introduce META-Robot, a framework that couples self cognition with capability scaling by expanding the criteria through which an agent evaluates and develops its skills. At its core, action-level predicates make execution outcomes explicit and available for LLM reasoning. The agent uses these assessments to identify its own limitations, extend relevant evaluation criteria, and guide the structured expansion of its skills. Proprioceptive feedback further grounds this process in the robot's physical state, enriching the evidence available for self-cognition. Experiments in simulation and on real robots demonstrate stronger overall task performance, generalization to new tasks, continued expansion of skills and evaluation criteria, and effective use of non-visual proprioceptive signals.

    agentagenticagent system
  55. arxiv:2610.41866 · cs.AI
    AI evaluators favor job applications written by certain AI models, independent of candidate quality
    Nicolò Pagan, Michael Andreas Ebling, Stefania Ionescu, Anikó Hannák

    Job applicants increasingly use large language models (LLMs) to write cover letters, while employers deploy LLMs to screen them. In this all-AI setting, we ask whether the \emph{choice of model} affects hiring outcomes independently of candidate qualifications. We introduce a multi-model auditing framework pairing eleven LLM writer agents with nine API-hosted evaluator agents across ten job descriptions and fifty candidates per job. Four findings emerge. First, we identify a generative preference hierarchy: a consistent global ranking of writer models by evaluator acceptance, in which locally hosted models are systematically disadvantaged regardless of their license, compounded by self-preference bias where evaluators favour their own outputs. Second, evaluator models differ substantially in how much bias they introduce based on applicants' tool choice, offering actionable guidance for fairer screening design. Third, this hierarchy has measurable competitive consequences: in the most adverse pairing, approximately two (1.8) out of twenty-five qualified candidates are filtered out purely due to model mismatch. Fourth, a common mechanism explains the patterns above: writer preference and self-preference trace to a combination of how much each writer leans on specific stylistic cues and how strongly evaluators, often unanimously, reward or penalise those same cues. A further check finds little evidence that an unmeasured "writer fingerprint" remains once these same features are accounted for. We discuss implications for algorithmic fairness in employment, regulation of AI-assisted screening, and the design of auditing frameworks that can be reapplied as the model landscape evolves.

    evaluator
  56. arxiv:2610.41865 · cs.AI
    Learning to Interact and Direct: Co-Evolving Video Director Agents
    Yingzhao Jian, Zhongan Wang, Yuxuan Hou, Hehe Fan

    Video director agents orchestrate multistage production to realize users' creative visions. Emerging research on their self-evolution advances screenplay planning effectively. However, the mutual reinforcement between directing and user interaction remains largely overlooked. Improved interaction clarifies user intent and informs workflow refinement, while more reliable execution shifts collaboration from troubleshooting to creative decisions. To this end, we introduce LIT (Learning to Interact and DirecT), a framework comprises a director agent, a simulated user, and a meta engineer. Simulated user assigns creative tasks to probe the director's weaknesses, while the meta engineer uses the task traces to revise the director's interaction and production strategy. Although this approach can address discovered interaction and production issues, it still relies on passive attribution by the meta agent. This limits its ability to adapt to broader user interaction patterns and production techniques. To solve this problem, we equip LIT with a co-evolution paradigm. The meta agent proactively learns external knowledge, thereby improving its ability to discover problems and propose solutions. The meta agent, simulated user, and director together form an evolutionary closed loop, collaboratively enhances the system's directing and interaction capabilities. Across the Test and OOD Test splits, LIT improves production and Interaction scores by 19.1% and 32.0%, respectively, relative to the initial pipeline. Controlled ablations reveal the correlation between improvements in production and interaction, where the co-evolution mechanism further encourage directing skill and interaction mode exploration.

    agent
  57. arxiv:2610.41857 · cs.RO
    When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks
    Alex Zhou, Junhao Wu, Xueyang Zhou, Yixiao Huo +4

    Simulation benchmarks score robot policies by success rate, and every failed episode counts against the policy. Many of these failures are the simulator's fault, not the policy's; we find the cause and fix it. Testing a vision language action model and a world action model in simulation, we find that the policy often places the object and the simulator then ejects it, and that the benchmark's own scripted demonstrations fail the same way. We present SimuGuard, an auditing plugin for LIBERO, RoboTwin, RoboCasa and RoboDojo. It leaves the evaluation unchanged. At every physics step, it records the state and flags an object that starts moving at a contact faster than the robot's contact or gravity account for. It replays a recorded episode bit for bit, and it reruns the episode with one simulation parameter changed. With it we trace these episodes on RoboTwin to the collision detection of the physics engine: for geometry one to two centimeters apart, the engine reports a penetration of about two centimeters, and the solver throws the object out at several meters per second. The cause is one constant: the distance algorithm discards triangles below a fixed area threshold, which valid triangles fall below at the centimeter scale. With that constant changed and nothing else, episodes with an artifact fall from 12 in 100 to none, and the success rate on five container tasks rises from 89% to 98%. Success rates in simulation thus count failures that the simulator causes. Benchmarks on PhysX 5.3.1 or earlier should rerun in a repaired engine. Where the engine cannot change, they should report the success rate together with the range it spans when every flagged episode fails or passes: 75.8% to 98.5% around a reported 83.1% for LingBot-VA.

    vision language actionliberorobotwinbenchmark
  58. arxiv:2610.41850 · cs.AI
    Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention
    Yannis Bendi-Ouis, Xavier Hinaut

    Transformers retain an explicit history of the sequence, allowing attention to directly retrieve past representations. However, the size of the KV-Cache and the cost of a new token increase with the context length. State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability. We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading. The encoder uses several memory units, each implemented by a dynamical system with a complex diagonal recurrent matrix. An input-dependent leak rate controls their updates. The decoder then uses attention to read and combine their outputs, without going through the token history. The formulation of the recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions. During autoregressive inference, the recurrent states and the convolution cache remain fixed in size. We compare DMT with Mamba and Transformer on cognitive and language modeling tasks, and show its strengths and limitations: DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling. On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer. These results demonstrate that parallel dynamical memory units are a promising mechanism for retaining and using information over sequences.

    memory
  59. arxiv:2610.41844 · cs.RO
    SketchMate: Agentic Geometric Reasoning for Zero-Shot Precision Insertion
    Jiaming Wang

    Inserting an unseen part requires sub-millimetre precision and the correct orientation of keyed features. Current zero-shot agentic manipulation methods built on vision-language models (VLMs) produce spatial outputs that are far too coarse for this. We propose to let a VLM build a sketch model of the task, i.e., a few geometric primitives and two mate frames, which a VLM agent measures with geometric tools in multi-view depth and which a standard planner and compliant controller execute. Because a sketch describes the parts independently of their placement, the robot needs to build it only once per task. A sketch memory stores each agent run and transfers the stored sketch to later placements by point-cloud registration, so that the VLM is called only for unfamiliar parts. After grasping the part, the robot inspects both mating faces at close range and corrects the sketch by shape matching. On 69 simulated insertion episodes, our method inserts 88% of the parts at the assembled pose in a strict kinematic test, while ReKep and MOKA with the same VLM insert none. On new placements of eight tasks, reusing one stored run inserts as many parts as calling the agent in each episode, and it produces the sketch in 3.4 s instead of about 90 s.

    manipulationgraspmemoryagentagentic
  60. arxiv:2610.41827 · physics.app-ph
    Creating functional nanostructures for vortex manipulation in superconducting Bi$_\text{2}$Sr$_\text{2}$CaCu$_\text{2}$O$_\text{8+x}$ thin films via focused He ion beam irradiation
    Bernd Aichner, Christoph Schmid, Sandra Keppert, Reinhold Kleiner +4

    Bi$_2$Sr$_2$CaCu$_2$O$_{8+x}$ (Bi-2212) is a highly anisotropic high-temperature superconductor exhibiting unique vortex dynamics due to its layered structure. Precise nanoscale patterning of ultrathin Bi-2212 films remains challenging, limiting controlled studies of vortex pinning. Here, we employ 30 keV focused helium ion beam (He-FIB) irradiation to fabricate periodic arrays of columnar defects with 100 nm spacings in optimally-doped Bi-2212 thin films, thereby creating artificial vortex pinning centers. Magnetotransport measurements reveal pronounced vortex matching effects, evidenced by minima in magnetoresistance and peaks in critical current at commensurate magnetic fields. The angular dependence of magnetoresistance and critical current scale with an effective magnetic field, indicating strong confinement of pancake vortices along irradiation-induced tracks. These findings demonstrate the efficacy of He-FIB nanopatterning in modulating the superconducting properties in Bi-2212 thin films, providing a platform for fundamental studies of vortex matter and advancing prospects for device applications based on this highly anisotropic superconductor.

    manipulation
  61. arxiv:2610.41824 · cs.LG
    Overcoming the Representational Limitations of Amplitude Encoding for Quantum Neural Networks
    Vincent Gurgul, David Prinz

    Quantum Neural Networks (QNNs) offer a framework for learning functions from classical data, but their expressive power depends critically on how inputs are encoded into quantum states. Angle encoding can require resources that scale with the number of features, whereas amplitude encoding represents $d$ features using only $\mathcal{O}(\log_2 d)$ qubits. This compactness comes at a representational cost: normalization makes conventional amplitude encoding invariant to nonzero real rescaling of the input, up to an unobservable global phase. We show that this invariance induces an exact, architecture-independent restriction on the resulting hypothesis class. No downstream quantum circuit can distinguish inputs differing by nonzero real scaling. We develop a general condition under which deterministic preprocessing can remove this ambiguity and show that nonlinearity alone is insufficient. Based on this analysis, we introduce a fixed nonlinear feature map that provably distinguishes all real inputs under subsequent amplitude encoding, without adding trainable parameters and while retaining logarithmic qubit scaling. We evaluate the resulting QNN on three canonical classification benchmarks and a nonlinear synthetic dataset against conventional amplitude and angle encoding, as well as a classical neural network. Across these tasks, the proposed encoding consistently improves classification accuracy over both quantum encoding baselines by 21$-$44 %. Our results establish normalization-induced scale invariance as a concrete representational bottleneck of amplitude encoding and demonstrate that it can be removed while preserving a compact quantum representation, marking a step toward quantum models capable of capturing arbitrary relationships in classical data.

    benchmark
  62. arxiv:2610.41822 · cs.LG
    Trainable Vector Quantization for Large Language Models
    Anastasiia Prutianova, Alina Bogdanova, Mikhail Shutov, Dmitrii Poimanov +2

    Vector Quantization (VQ) is a rapidly evolving technique for compressing Large Language Model (LLM) weights, reducing memory usage, and speeding up inference. However, extreme compression deteriorates model performance, which is challenging to recover due to the discrete nature of VQ parameters. Existing methods employ training-free approaches that are limited to layer-wise reconstruction loss minimization. To address this limitation, we propose TVQ, a novel method that leverages straight-through estimation to enable gradient descent-based optimization of vector-quantized LLMs. By reformulating VQ LLM optimization with trainable latent weights, TVQ enables arbitrary objective minimization. We constitute our approach by applying TVQ in a block-wise knowledge distillation paradigm, which balances quality and training efficiency, and propose several practical solutions to accelerate training. In experiments on modern LLMs, we demonstrate that TVQ is compatible with modern VQ schemes. As a result, TVQ achieves state-of-the-art quality in low-bit post-training quantization settings with a modest training budget and straightforward PyTorch implementation. For instance, TVQ achieves near-lossless 2-bit quantization for LLaMA2-70B with only 1% accuracy drop on popular Question-Answering benchmarks.

    memorypost-trainingbenchmark
  63. arxiv:2610.41814 · cs.CV
    A Patch is More Than a Token
    Zixiong Wang, Haiping Ma

    Vision Transformers represent each image patch through a canonical token produced under one observation condition. That token is a powerful interface, but it is still one realized state of the patch. We introduce PMTT, which establishes two foundations: a patch representation object can extend from one canonical state to a canonical-centered Representation State Family, and identity-preserving variation opens Patch-Level Local Feature Space. We develop four parts of this representation program. Patch Information Acquisition and Dense Orbital Views expose lawful same-patch states. Visible-to-Hidden prediction tests whether observed states reveal structure beyond themselves. Intrinsic analysis identifies low-dimensional, energy-concentrated, patch-specific, predictive geometry. Representation Organization separates acquiring more information from organizing that information, with an analytic Manifold as the central structured representation. We realize PMTT on frozen DINOv3 S/B/L/H+ backbones without neural post-training of the representation. Across two independent 4,096-anchor cohorts, Manifold explains 70.8-79.6% of Hidden variation, compared with 55.0-72.3% for directly Observed states, and wins on more than 95% of anchors at every scale. Effective rank falls from about 2.5 to 2.0 while dominant energy concentration rises with backbone scale. Across six visual capabilities and four scales, Observed exceeds the canonical Center in 21 of 24 conditions, Manifold exceeds Center in 23, and Manifold exceeds Observed in 16. The strongest current consequence appears in physical correspondence, where MegaDepth satisfies Center < Observed < Manifold at every scale. These results show that pretrained visual backbones contain richer patch information than the standard single-token interface directly exposes. Code and experimental assets are available at https://github.com/zixiongwang-ai/PMTT.

    post-training
  64. arxiv:2610.41813 · cs.AI
    RECAST: REconstruction of Chain-of-Thought via Auditable Structured Traces
    Qi Wang, Kang Xu, Chengcheng Wan, Jiangtao Wang

    Large closed-source reasoning models expose only their final answers. The intermediate reasoning that produces those answers is hidden behind API surfaces, tool-call policies, and safety filters. Existing chain-of-thought (CoT) extraction methods assume a side channel: a visible reasoning field, a scratchpad carried through tool calls, or a signed/encrypted reasoning block. We reproduce three representative state-of-the-art extraction methods and find that, on Claude-Sonnet-5.5 accessed through the standard chat-completions interface, all three return zero CoT tokens. The failure is structural rather than a bug: the reasoning field is absent, named tool selection is rejected, free-text scratchpad arguments are filtered, and no replayable reasoning block is exposed. We therefore take a different route. We present RECAST, a reconstruction-based framework that conditions a decoder on the question and the target model's final answer, reconstructs a first-person structured rationale, cleans it, and builds SFT-ready data. To the best of our knowledge, RECAST is the first system to reconstruct CoT from frontier closed models such as Claude-Sonnet-5.5, GPT-6.1, and Gemini 3 without using hidden reasoning channels. We instantiate RECAST on Claude-Sonnet-5.5 and additionally obtain successful qualitative reconstructions on GPT-6.1 and Gemini 3 with the same answer-only protocol. The system is open-source and packaged as an agent skill: one prompt deploys it and runs the complete pipeline to produce SFT-ready CoT data. This preprint describes the system, its threat model, and a controlled artifact validation; it does not claim large-scale benchmark improvements.

    agentbenchmark
  65. arxiv:2610.41809 · cs.LG
    Continual Reinforcement Learning Is an Integration Problem
    Nis Meinert, Jakob Gawlikowski, Edward Grant, Max Cyprys +5

    This paper presents an opinionated set of ten desiderata intended to inform the design and evaluation of reinforcement-learning agents that learn continually throughout a continuing lifetime. We use these desiderata to examine a representative selection of existing approaches. Progress on individual desiderata is valuable, but the longer-term aim should be to integrate all ten within a single agent. To our knowledge, no published evaluation has yet provided evidence for such an integration. Likewise, the representative benchmarks we review probe different parts of this target, but separate evaluations do not establish that the corresponding properties coexist within one agent throughout a continuing lifetime. As a secondary contribution, we therefore present a compact persistent-world scenario as a playable companion to the desiderata and outline a proposed benchmark that examines a broad range of associated behavioral capabilities together within one such lifetime.

    agentbenchmark
  66. arxiv:2610.41805 · cs.LG
    Backpropagation-free Deep Learning for Time-Series Classification
    Miha Kapš, Blaž Bertalanič, Carolina Fortuna

    Deep neural networks for time series classification (TSC) are trained almost exclusively by error backpropagation, which imposes update locking and the weight-transport problem. Backpropagation-free learning rules such as Direct Feedback Alignment (DFA), Direct Random Target Projection (DRTP), and the diffusion-inspired NoProp relax these constraints but have been evaluated almost exclusively on 2D image benchmarks. We adapt these three learners to univariate TSC with 1D convolutional architectures and present the first systematic study of them across 112 datasets from the UCR time series archive, with one hyperparameter-optimized configuration per method. The optimized learners reach mean F1 scores of 0.76 to 0.79, below ROCKET (0.83) and InceptionTime (0.84). We propose an analysis framework that relates the per-dataset gap to InceptionTime to dataset properties and tests convolutional feature learning by replacing the flatten readout with global average pooling, measured as a difference-in-differences (DD) against a backpropagation control on the identical architecture. Intra-class temporal variability is the dataset property most consistently associated with the gap for all three learners (partial Spearman $ρ$ from $0.28$ to $0.40$). Pooling harms DFA and DRTP more than backpropagation (DD $-0.025$ and $-0.055$), consistent with near-random convolutional features, whereas NoProp tolerates pooling as backpropagation does (DD $+0.013$). Relative to backpropagation on the same architectures, DRTP reduces training floating-point operations (FLOPs) by 33% and trains faster only on series longer than about 700 time steps, DFA yields no computational reduction, and NoProp reduces peak memory by 38% at the cost of 48% longer training.

    memorybenchmark
  67. arxiv:2610.41803 · cs.LG
    What Standard Evaluations Hide About Detectors of Hijacked LLM Agents: Rare Attacks, Unseen Tools and Benchmark Transfer
    Mahmoud Abbasi, Juan M. Núñez V., Amin Shahraki

    Detectors can flag hijacked large language model (LLM) agents from the traces of the tools they call. In the eleven evaluations we audited, attacks or anomalies make up 16% to 86% of the test samples, no evaluation varies that share, and six report no test on data from a second benchmark or source. We ask what such evaluations hide. On the AgentDrift and AgentDojo benchmarks we test trained detectors that read tool identities, behavior, call sequences, step embeddings or text, and a released LLM action monitor, AlignmentCheck, with thresholds fixed on training data or set by the monitor's released decision rule. First, at a 2% attack rate a detector with an area under the receiver operating characteristic curve (AUROC) of 0.879 reaches a precision of 0.076, and the mix of negatives alone moves that precision between 0.039 and 0.117. Second, as tool coverage, the share of test tool calls whose tool was seen in training, falls, detector rankings cross near 50% coverage and reverse below it, because unseen tools map to inputs that training never reached. Dropout on tool inputs largely removes the reversal. Third, no detector trained on AgentDrift ranks AgentDojo traces better than chance (AUROC 0.47 to 0.52), including one that reaches 0.996 within AgentDojo. The monitor, trained on neither, reaches 0.930, but with its open backbone its released decision rule would still mostly raise false alerts at a 2% attack rate (precision 0.096). We turn these findings into a checklist for evaluating trace detectors.

    llm agentbenchmark
  68. arxiv:2610.41792 · cs.LG
    Implicit Assumptions About Tasks in Continual Learning
    Kathrin Korte, Christian Metzner, Akshay Bedhotiya, Viet Anh Khoa Tran +2

    Continual learning is the study of how a model accumulates knowledge from data received over time, rather than all at once. In practice, this is typically achieved by dividing a dataset into tasks and presenting them sequentially. Explanations of the resulting measurements take into account factors such as the similarity between segments, the rate at which the data changes and the order of presentation. Each of these factors can be defined in terms of the underlying data, the partition imposed on it, or the learner that encounters it. These levels do not necessarily coincide, and it is usually left implicit which level a reported quantity refers to. We separate these and identify which properties acquire a value at each level. In controlled teacher--student experiments, we demonstrate that partitioning the same underlying trajectory at different levels of granularity, changing the order of presentation, and varying the similarity structure of the underlying data all result in different measured outcomes. We review representative benchmarks to determine which of these properties are stated. Where the underlying data cannot be inspected, as in foundation models built on undisclosed pre-training corpora, its properties are inherited by the model's initial state, and the training applied on top of it can be described.

    benchmark
  69. arxiv:2610.41789 · cs.CV
    Hard Sample-Conditioned Decorrelated Attention for Robust Face Recognition
    Tahar Chettaoui, Guray Ozgur, Eduarda Caldeira, Naser Damer +1

    Multi-head attention in Vision Transformers (ViTs) allows independent attention patterns across heads, but without explicit regularization, heads may learn redundant patterns rather than complementary roles. Correlation-based penalties encourage head disentanglement, yet apply uniform regularization across all training samples. In face recognition (FR), hard samples whose embeddings lie close to competing identities are critical for optimization and generalization. Motivated by this, we propose ViT-HarDis, which conditions attention head disentanglement on sample difficulty: samples with high cosine similarity to negative class centers receive stronger head decorrelation, enforcing specialization where identity ambiguity is highest. Experiments show improvements over both the baseline and uniform disentanglement on most benchmarks. Our analysis shows that ViT-HarDis reduces the similarity of hard samples to competing class centers while decreasing inter-head attention similarity, and that applying disentanglement only to the final transformer layers outperforms regularizing the entire network. Trained on WebFace4M, ViT-HarDis achieves SOTA results among ViT-based methods on IJB-B, IJB-C, and TinyFace Rank-1, while achieving the best IJB-B and TinyFace results when trained on MS1MV3, improving FR without architectural modifications or additional inference cost. Code: https://github.com/TaharChettaoui/Hard-Sample-Conditioned-Attention-for-FR.

    benchmark
  70. arxiv:2610.41785 · cs.LG
    Peeling the Matryoshka: Extracting Ontological Structure from Retrieval Embeddings
    Eunchan Kim, Inwhee Joe

    Taxonomy construction is essential for knowledge-intensive NLP yet remains costly: clustering-based methods (i) need careful per-corpus tuning and yield unstable hierarchies, while LLM-based pipelines (ii) depend on expensive task-specific training that does not transfer across domains. We report an empirical finding that bears on both: in Matryoshka Representation Learning (MRL) embeddings, where each leading prefix is itself a usable, coarser vector, prefix length aligns with the coarse-to-fine granularity of gold taxonomies; shuffled- and reverse-dimension controls isolate the alignment to the learned early prefix dimensions. The pattern holds across three architecturally distinct MRL models. To test whether this regularity is usable, we introduce PEEL (Peeling Embeddings to Extract Latent ontologies): a minimal probe over these frozen embeddings, trained once with an order-embedding objective that represents hierarchy as vector containment; a minimum spanning arborescence then decodes taxonomic DAGs. Using only a concept vocabulary (no documents), PEEL recovers taxonomic DAGs with high structural fidelity on two OLLM benchmarks, transferring to a new domain with no target-edge supervision. The end-to-end OLLM pipeline, which jointly discovers concepts and induces structure from document corpora, is reported as a cross-task reference. Pre-trained retrieval embeddings encode a decodable granularity gradient; a minimal probe renders it as explicit parent-child (is-a) structure.

    benchmark
  71. arxiv:2610.41772 · cs.CV
    WorkIng: A Real-World and Synthetic Highway Work-Zone Benchmark Dataset for Autonomous Navigation
    Xujun Xie, Dominik Rößle, Adithya Mohan, Daniel Cremers +1

    Highway work zones are critical edge cases for autonomous vehicles (AVs), combining atypical geometries and jurisdiction-specific regulations that existing benchmarks fail to capture: real-world datasets lack 360° multi-modal, full-scene 3D work zone annotation, and no single recorded site can span the layouts and regulatory variants needed for generalizable models. We present WorkIng, a dual-domain benchmark and training resource. WorkIng-Highway provides a 360° multi-modal work zone benchmark with synchronized 128-channel LiDAR, six RGB cameras, and RTK-GPS across 32 km, yielding 12,000+ annotated frames and 227,399 tracked 3D bounding boxes. WorkIng-CARLA complements this with procedural simulation spanning 24 regulation-aware scenarios across six canonical topologies and six international traffic standards, populated by 100+ custom 3D assets. Benchmarking representative 3D detection and end-to-end planning architectures shows that real-world data remains the primary performance driver, and that scenario-stratified synthetic augmentation improves architectures across 2D recognition, 3D detection, and trajectory planning. Resources: https://github.com/cvims/WorkIng

    benchmark
  72. arxiv:2610.41769 · cs.CL
    Stepwise-Querying: Distilling Reasoning without Access to Teacher Chain-of-Thought
    Dicong Yu, Wenhua Wang, Tian Wang

    Reasoning distillation is widely used to transfer the reasoning capabilities of large language models to smaller students through teacher-generated supervision. In the black-box setting, however, teachers may expose only final answers and brief solutions, leaving intermediate calculations and validity conditions unavailable and thus limiting the reasoning supervision available for distillation. We propose Stepwise Querying, which elicits step-specific explanations without requiring access to the teacher's internal chain-of-thought (CoT) traces. Independent queries to the same teacher request intermediate derivations, validity checks, and corrections for individual steps, the assembled solutions are then used for standard supervised fine-tuning. Across two teachers, two students, and four mathematics and code benchmarks, our method achieves the highest score in 14 of 16 comparisons with three baselines. We further introduce Stepwise Explanation Information Gain (SEIG), a diagnostic based on the log-ratio of a frozen student's probabilities of correct step completion under stepwise and global explanations. Analyses on GSM8K show that incorrect-answer trajectories can remain useful when they preserve valid intermediate reasoning, while further decomposition yields limited additional gains.

    benchmark
  73. arxiv:2610.41763 · cs.CV
    Measuring What Matters: A Causal Reliance Score for Histopathology Image Segmentation
    Samuel Thomas, Daniel Tweneboah Anyimadu, Ahmed Karam Eldaly, Somaia Elsheikh +1

    Reliable histopathology segmentation requires models that rely on pathology-relevant structures rather than spurious background correlations. In this work, we argue that improving architectures alone is insufficient if models continue to rely on spurious correlations rather than pathology-relevant features. We introduce a causality-inspired evaluation framework based on controlled interventions that explicitly probes model sensitivity to foreground tissue structure and background/domain variation. Building on this, we propose the Causal Reliance Score (CRS), a metric that quantifies the relative sensitivity of model predictions to foreground perturbation versus background/domain variation under controlled interventions. To support this analysis, we design augmentation strategies that act as implicit counterfactual interventions during training, encouraging invariance to nuisance factors while preserving morphology. We demonstrate how these interventions reveal behavioural differences between models that are not captured by standard metrics such as Dice or IoU. While we instantiate our framework using a self-supervised transformer (CAUSE-ViT), the primary contribution of this work is not a new architecture, but a principled intervention-based methodology and evaluation metric for auditing model behaviour in medical image segmentation. Our results suggest that evaluating model behaviour through counterfactual interventions is essential for developing reliable and clinically robust medical image segmentation systems.

    evaluation framework
  74. arxiv:2610.41762 · eess.SY
    Scaly: an Optimizing Compiler for Optimal Control
    Tudor A. Oancea, Colin N. Jones

    We present Scaly, an open-source optimizing compiler for optimal control, written in Python, that generates fast C code. Scaly provides a modeling language embedded in Python for the parametric optimization problems that arise in model predictive control and moving horizon estimation, and computes the sparse derivatives that standard optimization solvers need. Scaly accepts general nonlinear programs but applies optimizations specific to optimal control, which preserve the repeated structure of a problem, such as the stages of a horizon, through differentiation down to the generated loops. The generated C depends only on the C math library and never allocates memory dynamically. On three benchmark problems, Scaly evaluates the Hessian of the Lagrangian up to 26 times faster than CasADi, with smaller binaries and build times that stay nearly constant with the horizon. With the same solver, Scaly shortens the solver time per control step on every benchmark, by a factor of up to 9.

    memorybenchmark
  75. arxiv:2610.41751 · cs.LG
    Latent-Lens: Visual Perception in Small Language Models Through Latent Communication
    Mo Shahdloo, Faegheh Sardari, Mathew Salvaris

    Bridging independent vision-language models (VLMs) and small language models (SLMs) in a modular multi-agent system typically forces them to exchange information as discrete text, which creates a lossy bottleneck for spatial and fine-grained details. This constraint is especially acute for SLMs, whose limited capacity makes them more reliant on the quality of the input signal. Recent latent-communication methods exchange states between language models. How a VLM can pass visual information to a separately pretrained text-only SLM though remains unexplored. We propose \textbf{Latent-Lens}, a lightweight codec that bridges the representation gap between a frozen VLM and a frozen SLM directly in latent space, enabling them to communicate purely through hidden states. The codec consists of a domain encoder that projects hidden states from the VLM into the SLM's embedding space, and an attention decoder formalised as a LoRA that teaches the SLM to attend to these visual prefixes, while training only ${\sim}1.6$M parameters (0.2\% of the combined model size). On visual question answering (CLEVR, GQA), Latent-Lens achieves $81.0\pm7.9$\% and $65.4\pm5.4$\%, compared with $58.3\pm5.7$\% and $61.8\pm1.0$\% for text communication with the same LoRA budget (+30.2--69.8pp relative to frozen text communication), while on image captioning (Flickr8k) it improves over the same baseline by +16.5 (+18.8 frozen). We also evaluate Latent-Lens with TinyLlama-1.1B and InternVL2-1B, which use different architectures, vocabularies, and hidden dimensions.

    multi-agentagent system
  76. arxiv:2610.41733 · cs.LG
    Leakage-Controlled Structure-Preserving Augmentation and Template-Aware Evaluation for Bangla Symptom-Based Disease Prediction
    Sabah Nushra, Shovan Samanta Turzo, Moinul Hossain, Md Kishor Morol

    Bangla symptom-based disease prediction is commonly evaluated on a small, imbalanced structured benchmark containing many profiles that differ by only one or two symptoms. We test whether train-only asymmetric Bernoulli perturbation improves classification using 725 unique profiles spanning 85 diseases and 172 binary symptoms. Experiments use ten repeated real-profile holdouts and parent-grouped inner cross-validation. The 2 x 2 design separates augmentation from hyperparameter tuning, with exact duplication, random oversampling, class weighting, and SMOTEN as balancing controls. For Bernoulli Naive Bayes (BNB), the gain over tuned raw training does not withstand Holm correction, although BNB attains the highest mean accuracy (97.50%). The boosted-tree models benefit substantially: XGBoost, LightGBM, and CatBoost improve by 18.10, 28.45, and 10.09 percentage points, respectively. A Hamming-distance audit places 91.0% of standard held-out profiles within one symptom bit of the training data. Accuracy decreases on a same-disease 1-bit-purged challenge; nevertheless, XGBoost and LightGBM retain large augmentation gains. We also evaluate robustness to missing and spurious symptoms and the faithfulness of BNB explanations. Overall, Bernoulli perturbation is useful for some model families, but benchmark accuracy should be interpreted together with template proximity.

    benchmark
  77. arxiv:2610.41732 · cs.LG
    From Training Exposure to Membership Evidence: A Controlled Study of OLMo 2
    Ali Banaei, Muhammad Shahab Shahzad, Gerald Q. Maguire, Marco Chiesa +1

    Whether a language model was trained on a document matters for copyright, privacy, and benchmark contamination. Yet most membership-inference (MI) evaluations on language models collect members and non-members after the fact, so success may reflect distribution shift and failure is inconclusive. We re-ran OLMo 2 1B's official 50-billion-token mid-training stage, replacing selected training chunks with documents from 7,814 pairs matched on domain and difficulty. For each pair, a coin flip determined which document was inserted and which was held out. The run released by the Allen Institute for AI (AI2) from the same seed saw identical data everywhere else and serves as the counterfactual. One exposure raises a previously unseen document's likelihood on average but rarely identifies it. Using the counterfactual as reference, the reference ratio has an area under the ROC curve (AUC) of 0.767 but, with one threshold across domains, catches 3.8% of members at a 1% false-positive rate (FPR). In contrast, 13 matched pairs give a paired test 80% power at the 5% level to detect training on a collection, given same-domain non-members and a close reference. A reference trained with another seed recovers 92% of the counterfactual's AUC advantage over chance, whereas reference-free attacks stay within 0.07 of chance when pooled over domains. The full official post-training recipe, re-run on both models, leaves 89% of the average effect. Even with a closely matched reference, substantial document-level variability limits individual membership inference. We will release the checkpoints, per-token log-probabilities, and the evaluation documents; the code is available at https://github.com/ali-bana/OLMo-Twin.

    post-trainingbenchmark
  78. arxiv:2610.41728 · cs.CV
    SAR-TS-BERT: Physics-Guided Self-Supervised Pre-Training With Polarimetric Features for SAR Time-Series Crop Classification
    Jieqiu Wang, Yuhan Zhang, Bo Zheng, Shaoming Zhang

    Multitemporal synthetic aperture radar (SAR) imagery provides continuous, cloud-free observations of crop phenology for agricultural monitoring in cloudy and rainy regions, yet high-accuracy classification typically requires extensive labeled samples. Although self-supervised pre-training mitigates label dependency in remote sensing, existing frameworks primarily target optical imagery or generic spatial SAR scenes, underutilizing temporal backscatter dynamics and polarimetric scattering mechanisms. We propose SAR-TS-BERT, a physics-guided BERT-style self-supervised pre-training framework for SAR time-series crop classification. SAR-TS-BERT integrates backscatter intensities and polarimetric ratio features (VV, VH, and VV/VH) with dual-polarimetric Cloude-Pottier decomposition parameters ($H/α/A$) into a six-band input representation through two core designs: 1) a physics-guided span masking strategy (SpanPhysics) that adaptively selects contiguous masked spans based on temporal backscatter change rates to prioritize key phenological transitions; and 2) a dual-branch cross-modal fusion encoder that models backscatter and polarimetric decomposition features through separate Transformer branches followed by bidirectional cross-attention fusion. On independent test sets in Estonia (EuroCropsML) and the USA (California CDL), SAR-TS-BERT achieves overall accuracies (OAs) of 96.78% and 87.29%, respectively, and attains a mean monthly OA of 99.37% on Chongming Island, Shanghai. On the Estonia benchmark, pre-training improves OA and Macro F1 by 1.17 and 1.59 percentage points over training from scratch under full supervision, and by 2.58 and 5.19 percentage points using only 1% labeled data. Incorporating SAR scattering physics into temporal self-supervised pre-training effectively enhances crop classification accuracy, parcel-level mapping consistency, and few-shot transferability.

    benchmark
  79. arxiv:2610.41711 · cs.LG
    Hierarchical Residuals: A Minimalist Decoupling of Residual Width from Compute Width
    Ziteng Wang, Jianqiao Lu, Xinyu Ma, Jintao Zhang +3

    Residual connections underpin Transformers, but they implicitly tie the residual width carried through depth to the compute width used inside each layer. We introduce Hierarchical Residuals (HiRes), a minimalist architectural change that decouples these dimensions, widening the residual state while keeping per-layer compute width fixed. HiRes groups layers into blocks: standard d-dimensional PreNorm layers operate inside each block, while a wider D-dimensional PostNorm outer residual connects blocks. Learned read/write projections bridge the two spaces at block boundaries, leaving individual Transformer layers unchanged. Across different model sizes and depths, HiRes consistently outperforms vanilla residual addition, with fitted scaling laws indicating a 1.30-fold improvement in FLOPs efficiency. On OLMoE-1B-7B pretraining for 500B tokens, HiRes achieves 0.0208 lower training loss and 1.14-point higher downstream accuracy. HiRes is also competitive with other residual expansion methods such as VWN, mHC and AttnRes with less activation memory and I/O costs. Built from only dense projections and normalizations, it requires neither specialized kernels nor infrastructure-level optimizations. HiRes thus exposes residual width as a practical scaling variable using only standard operators.

    memory
  80. arxiv:2610.41710 · cs.AI
    Advancing Clinical Trial Outcome Prediction with Domain-Specific Large Language Models
    Burak Bakkar, Hana Moalla, Mohamed Salem Jedidi, Haifa Ben Messaoud +1

    Clinical trial failures are a major barrier to efficient drug development, incurring substantial costs and delaying patient access to effective therapies. Structured approaches such as the Hierarchical Interaction Network (HINT) predict trial outcomes by explicitly modelling interactions between drugs, diseases and protocols, but make limited use of the free text in trial registrations. We evaluate two domain-specific language models, BioBERT (108M parameters) and BioMistral (7B parameters), for phase-level outcome prediction on the TOP benchmark (12,477 trials), representing each trial as text built from ten ClinicalTrials.gov fields. We compare fine-tuning (full fine-tuning for BioBERT, low-rank adaptation for BioMistral) with probing, in which the pretrained model is frozen and only a lightweight classifier is trained on its embeddings. On the held-out test set of each phase, all four language-model variants achieved higher ROC-AUC (0.613--0.772) and PR-AUC (0.649--0.901) than the published HINT results, corresponding to relative PR-AUC gains of 10--20\%. F1 differences were small and mixed, and no model, including HINT, clearly exceeded the F1 of a trivial classifier that predicts success for every trial. Under the configurations tested, probing performed comparably to fine-tuning: fine-tuned models scored highest on ROC-AUC and PR-AUC in Phases I and II, and frozen BioMistral embeddings in Phase III (ROC-AUC 0.772; PR-AUC 0.901), although most differences between variants were below 0.03 and were not tested for significance. These results suggest that pretrained biomedical representations already encode signal relevant to trial outcome, and that probing is a computationally cheaper alternative to fine-tuning for this task.

    benchmark
  81. arxiv:2610.41704 · cs.CL
    GraphNIAH: A Benchmark for Long-Context Reasoning over Real-World Graphs
    Minh-Nam Tran, Abulhair Saparov

    The needle-in-a-haystack (NIAH) benchmark was introduced as a test for measuring a model's ability to find small relevant facts within a very large body of input text. However, existing NIAH benchmarks each suffer from one or more of the following limitations: (1) reliance on synthetic data, so their evaluations may not generalize to real-world settings; (2) reuse of previously published datasets, leading to the risk of benchmark contamination; (3) dependence on manual annotation, which requires human effort to create and update data and is prone to future benchmark contamination; and (4) unrealistic simplification of long-context multi-hop reasoning where each next step is trivial given previous steps, while real-world reasoning often requires planning over multiple possible proof paths. To address these limitations, we introduce \Benchmark, a dynamic benchmark built from real citation relationships among scientific papers published after the evaluated models' knowledge cutoff date, which utilizes real-world information while ensuring no benchmark contamination. The fully automated benchmark construction pipeline samples representative subgraphs of the citation graph formed by these papers and converts each subgraph into a long-context QA example with a verifiable answer. We use GraphNIAH to evaluate eight LLMs from the GPT-5.6, Gemini-3, and Qwen3.5 families. We find that accuracy falls as questions require longer reasoning paths: at question depths of 16 and 32, the GPT-5.6 and Qwen3.5 models answer YES/NO questions no better than random guessing, mostly by answering NO; four of the eight models answer no counting example correctly, and the others model get worse as the subgraph size increases. These findings suggest that current LLMs often stop searching the citation relationships in their context before they reach the answer, even though every fact they need is in the context.

    long-contextbenchmark
  82. arxiv:2610.41702 · cs.RO
    Flexibility-Aware Robotic Manipulation: Deformation Estimation, Kinematics, and Control via Distributed Inertial Measurement
    Amir Hossein Barjini, Sadeq Yaqubi, Pauli Mustalahti, Jouni Mattila

    Flexible robot links challenge the rigid-body assumption, making joint measurements alone insufficient for accurate manipulation. This paper introduces a flexibility-aware manipulation framework that connects structural dynamics directly to sensing, kinematics, and control. Instead of prescribing a dense inertial measurement unit (IMU) network, the dominant vibration modes of each flexible link are first identified from response-convergence and excitation-bandwidth criteria. Under a full-rank modal observability condition, $m$ retained modes require only $m$ independent slope measurements and one reference IMU, defining the minimum observable IMU configuration. The resulting measurements are used to reconstruct the continuous link deformation, which is then incorporated into flexibility-aware forward kinematics and a deformation-informed partial differential equation-based controller. Experiments on hydraulic single- and double-link manipulators evaluate the framework. On the long-reach double-link platform, the proposed kinematics achieves end-effector root mean square errors of $0.0107$~m, $0.0194$~m, and $0.0102$~rad while using fewer sensors than the piecewise-rigid method. Closed-loop reaching experiments further demonstrate real-time integration of IMU-estimated joint states and reconstructed deformation into the proposed controller. The results show that dominant modal information can determine sensing requirements and enable real-time flexibility-aware robotic manipulation.

    manipulationmanipulator
  83. arxiv:2610.41698 · cs.CL
    Precision--Recall Trade-off on the Inference Level for LLMs in Grammatical Error Correction
    Ryszard Staruch, Filip Graliński

    Large language models (LLMs) have recently been adopted for Grammatical Error Correction (GEC), achieving strong performance via both full fine-tuning and prompt-based approaches. As probabilistic sequence models, LLMs expose inference-time parameters that directly modify token generation behavior; however, their potential for controlling GEC-specific trade-offs has not been systematically studied. In this work, we investigate decoding strategies for grammatical error correction. We show that the repetition penalty parameter enables controllable adjustment of the precision--recall trade-off at inference time, without modifying the model architecture, training procedure, or prompt. Through experiments on standard GEC benchmarks, we demonstrate that repetition-penalty-based decoding enables inference-time control of the precision--recall balance, conceptually related to methods previously proposed for sequence-to-sequence GEC models, such as Align-and-Predict Decoding (APD), while remaining simple and broadly applicable to existing LLMs that expose inference-time decoding penalties.

    benchmark
  84. arxiv:2610.41697 · cs.LG
    TSBPO: Tree-Structured Branching for Policy Optimization in Interactive Language Agents
    Rui Yu, Chong Mu, Jinchuan Zhang

    Training interactive language agents from terminal outcomes poses three coupled problems: which task and state continuations to sample, how to assign delayed feedback to decisions, and how to aggregate token losses for multi-token actions. We present TSBPO, a tree-structured sampling-to-update paradigm that addresses these problems within one training loop. A likelihood-fitted Root/Branch allocator uses retained task and state evidence to target continuations likely to produce contrasting terminal outcomes under fixed rollout budgets. On the resulting replayed physical trees, trajectory advantages compare immediate sibling segments, while observation-group step advantages compare outcomes at matched task/observation states. The actor then sums clipped token surrogates over each complete response and normalizes across active turns, avoiding response-length averaging in the policy-gradient term. With a Qwen2.5-1.5B actor, fixed checkpoints obtain 97.79% seen and 96.64% unseen stochastic success on ALFWorld and an 81.07% per-episode success rate over six stochastic episodes for each of the 500 WebShop test goals. A WebShop step-term ablation yields 75.93% at the same update-100 evaluation; an ALFWorld aggregation comparison shows higher observed command validity with complete-turn weighting. These results support the evaluated training configuration on two interactive benchmarks.

    benchmark
  85. arxiv:2610.41695 · cs.CV
    Retrieval-Augmented Initialization with Dual-Guided Inversion for Flow-Based Inverse Problem Solvers
    Hansam Cho, Seoung Bum Kim

    Flow models have recently emerged as powerful generative priors for solving inverse problems. Despite their success, most existing flow-based inverse problem solvers initialize the reconstruction process with random Gaussian noise. However, recent studies in image generation have shown that the choice of initial noise can significantly influence generation outcomes. The role of initialization as a distinct design choice in flow-based inverse problem solving remains comparatively underexplored. In this work, we systematically evaluate noise initialization strategies while holding the downstream flow-based inverse solver fixed. We propose Retrieval-Augmented Initialization with Dual-Guided Inversion (RA-DGI), a method that constructs improved initial noise by leveraging retrieved proxy clean images and dual-guided inversion. Extensive experiments across multiple inverse problems demonstrate that the quality of the initial noise substantially affects reconstruction fidelity and perceptual performance. RA-DGI outperforms standard random initialization and achieves superior results compared to inference-time scaling strategies, while requiring less computational overhead. These findings highlight initialization as a critical yet previously overlooked factor in flow-based inverse problem solvers.

    retrieval-augmented
  86. arxiv:2610.41688 · cs.RO
    When the Past Shapes Action: Understanding Embodied Memory
    Siqi Li, Ziyi Bai, Xinrun Xu, Tinglei Huang +1

    Recent advances in generative foundation models have expanded the capabilities of embodied AI, enabling increasingly general navigation, manipulation, and long-horizon task execution. However, effective interaction with dynamic environments requires more than only strong perception and reasoning: agents must act with partial knowledge and retain, retrieve, and update information from past observations, actions, and experiences. This survey examines embodied memory as a mechanism for preserving decision-relevant information beyond the current observation. We organize existing approaches along two complementary dimensions: six fundamental memory requirements, spanning world-state persistence, task progress, belief maintenance, and experiential learning; and eight architectural categories, including recurrent states, fast weights, visual and symbolic representations, spatial memory, and experience-based systems. We analyze their capabilities and limitations, and examine how existing benchmarks evaluate memory-dependent behavior, distinguishing genuine use of historical information from general task performance. Finally, we identify open challenges in selecting relevant experiences, maintaining consistent beliefs in changing environments, and consolidating knowledge across episodes. Our analysis provides a framework for understanding memory as an essential component of reliable, adaptive, and lifelong-horizon embodied intelligence.

    embodiedmanipulationmemorybenchmark
  87. arxiv:2610.41677 · cs.CV
    OT-CAL: Optimal-Transport-Guided Adaptive Sample Selection for Efficient Visual Learning
    Suryansh Talukdar, Jin Soung Yoo

    Deep neural networks for image recognition are typically trained using mini-batches, with a forward pass to compute predictions and losses followed by a more computationally expensive backward pass to update model parameters. However, not all training samples contribute equally to learning, suggesting that computation can be reduced by performing backward passes only on informative samples. The challenge is that sample informativeness changes throughout training: difficult samples may be valuable early on, improving samples become important later, and class diversity becomes increasingly relevant toward the end. Existing methods often rely on fixed selection rules or computationally expensive per-sample gradient calculations. We propose OT-CAL, an adaptive data selection method that automatically adjusts its sampling preferences during training. OT-CAL represents each sample using five informativeness features, including difficulty, learnability, and uncertainty. A curriculum vector describes the currently preferred sample characteristics, and optimal-transport-based matching assigns selection weights according to how closely samples match this target. The curriculum vector is then updated using the selected samples, forming a feedback loop that adapts without a manually defined schedule. Experiments on three standard visual learning benchmarks show that, when selecting 50% of each mini-batch, OT-CAL improves performance by 1.3--2.1 points over AdaSelection while achieving 1.18--1.24 times faster training than full-data training.

    benchmark
  88. arxiv:2610.41671 · cs.RO
    OpenVIO: A Versatile Open-Source Visual-Inertial Odometry System
    Seungwon Choi

    Visual-inertial odometry (VIO) is a fundamental topic in robotics research. Beyond research, VIO has been adopted in a wide range of products. In this manuscript, we present OpenVIO, an open-source VIO system for research and practical applications. It is released under the MIT license. OpenVIO supports pinhole and fisheye cameras and extends this support to stitched 360° images. We describe each part of the system together with its design considerations. We also report benchmark results with other open-source VIO systems on public datasets. OpenVIO is available as open source at https://github.com/93won/openvio.

    benchmark
  89. arxiv:2610.41663 · cs.RO
    ResMerge: Residual Policy Learning and Merging for Continual Adaptation of Pre-trained Robot Policies
    Kefei Zhu, Lingyun Xu, Yue Xin, Xiaoqi Li +5

    Continual adaptation to new tasks is essential for robots deployed in the real world, but adapting pre-trained policies remains challenging due to parameter growth and the need for continual demonstration collection. Parameter-Efficient Continual Fine-Tuning (PECFT) isolates task-specific updates, but adapter-per-task designs require demonstrations and grow with the task set. We propose ResMerge, a two-stage framework that learns mergeable residual policies atop a frozen pre-trained policy: residual reinforcement learning (RL) acquires task-specific corrections without additional expert demonstrations during adaptation, and parameter-space merging consolidates compatible residuals to reduce stored parameters. We introduce Residual-Field Consistency (RFC), a simple, bounded, and interpretable metric that assesses compatibility in residual action-field space by comparing correction direction and magnitude. Experiments on LIBERO across two backbones and on a real-world Franka robot show improved continual learning performance over strong PECFT baselines, while selective merging reduces the residual-parameter footprint. Project website: https://resmerge.github.io/.

    liberofranka
  90. arxiv:2610.41654 · cs.RO
    $\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation
    Weyl Lu, Yubei Chen

    Robot policy learning is dominated by zeroth-order methods such as PPO, which treat the simulator as a black box and pay for it with enormous sample counts. First-order policy learning differentiates through the dynamics, but on contact-rich whole-body control it has rarely worked: its gradients explode. A recent world-model method showed that derivative information can make humanoid motion tracking far more efficient, using surrogate gradients of a learned model. We show that exact gradients of the simulator itself do better, once two obstacles are removed. Exact gradients contain rare unbounded samples, because contact dynamics is not differentiable everywhere and its derivative diverges near some events; and they are expensive, because the simulator's adjoint has to replay the contact solve. $\nabla$Mimic bounds these samples with median-relative per-sample gradient clipping at every step of the backward pass, and reuses each costly simulated window by multistep regression onto gradient-derived action targets, which bounds every action target independently of the gradient scale. NablaEngine, a differentiable GPU simulation engine for the Unitree G1, removes the cost: it computes MuJoCo's full contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, $11.5\times$ faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes instead of 1.2 h, and 0.874 in 1.8 h. SHAC, the standard first-order baseline, stops on exploding gradients within minutes. We release NablaEngine and $\nabla$Mimic as open-source infrastructure for first-order humanoid policy learning.

    humanoidrobot policywhole-body control
  91. arxiv:2610.41651 · cs.CV
    When Depth Lies: Benchmarking Vision-Language Models on Mirror-Induced RGB-D Ambiguity
    Hao Yin, Tianchen Guo, Heming Du, Yan Ke +2

    Vision-language models (VLMs) have made significant progress in spatial reasoning, particularly by integrating RGB images with depth data, which is often assumed to faithfully reflect physical distances. However, mirrors violate this assumption: they distort depth readings while reproducing object appearances, making reflections difficult to distinguish from real objects. As existing benchmarks rarely consider such reflective scenarios, VLM performance in mirror scenes remains largely unexplored. In this work, we introduce a mirror-centric RGB-D dataset from reflective scenes to evaluate the spatial reasoning capabilities of state-of-the-art VLMs in these scenarios. Specifically, we collect 1,011 RGB-D images from 525 real-world scenes, spanning diverse indoor and outdoor environments, six reflective surface types, and both daytime and nighttime conditions. All images are annotated with reflective surface region masks and mask annotations for both directly observed and mirror-reflected objects. Then, we curate a benchmark named MIRA-Bench (Mirror-Induced Reasoning Ambiguity Benchmark), comprising 14,140 QA pairs across eight sub-tasks organized into three cognitive levels: Reflection-aware Perception, Spatial Reasoning, and Scene-level Decision-making. We evaluate state-of-the-art VLMs on MIRA-Bench and find consistent failures across tasks: models misclassify whether objects are directly observed or reflected, fail to match reflected instances to their real-world counterparts, and produce erroneous spatial inferences. Remarkably, providing depth data actually degrades reasoning performance, implying that VLMs cannot identify incorrect depth in mirror scenes. The code and data are available at https://github.com/EvaYin-cloud/MIRA-Bench

    benchmark
  92. arxiv:2610.41647 · cs.AI
    Risk-Sensitive Mission Planning for Multi-Target On-Orbit Inspection via Distributional Reinforcement Learning
    Adam Abdin

    Multi-target on-orbit servicing missions require planning tools that can adapt to evolving operational conditions while accounting for the outcome variability inherent to safety-critical spaceflight. This paper presents an adaptive planning approach for multi-target servicing campaigns in low Earth orbit and applies it to inspection missions flown by a low-thrust servicer. The approach is based on Distributional Reinforcement Learning, which learns the full distribution of mission returns and thereby supports explicit risk-sensitive decision-making at each planning step. Specifically, a Quantile-Regression Deep Q-Network (QR-DQN) agent selects which target to inspect next by jointly considering the expectation and the variance of the predicted return through a mean-variance criterion. The mission environment models circular orbits with a spread in inclination and uses Edelbaum's combined-manoeuvre low-thrust transfer estimates. The observation comprises the orbital elements, the visit history, the observed risk state of each target and the remaining mission-time budget. The risk state of each target changes stochastically during the campaign, making the outcome of an inspection uncertain and creating operationally meaningful distinctions between risk-neutral and risk-averse planning strategies. Because the mission-time budget generally admits only a subset of the candidate targets, the planner must also select which targets to inspect. Numerical experiments on scenarios with five target satellites show that one risk-aversion weight moves the policy across a range of mean return and return variance, over an unchanged environment and reward. Risk-averse agents reduce the variance of the episodic return by approximately one half relative to risk-neutral agents and lower the probability of a mission loss, at a cost in mean return. To validate the learned policies, the efficient frontier, which gives the highest mean return attainable at each level of return variance, is computed exactly by enumerating the admissible visit orders. At equal variance, the learned policies lie close to this frontier in mean return, and the risk-averse policies have a lower probability of a mission loss than the frontier policies.

    agent
  93. arxiv:2610.41643 · cs.RO
    Direct Action-Manifold Modeling for Efficient and Robust Bimanual Manipulation
    Hongren Wang, Mingyu You, Zihan Li, Zhipeng Wang +2

    Diffusion and flow matching visuomotor policies require iterative denoising or numerical integration for multimodal action generation, introducing latency and execution mismatch. We instead consider a different formulation: conditioned on vision and proprioception, demonstrated bimanual action chunks concentrate in a structured observation-dependent region of joint-action space, making the action target more direct to model than the generative process. Based on this view, we propose the Direct Manifold Action Transformer (DMAT), which reformulates action generation as the direct modeling of clean, observation-conditioned joint-space action chunks from stochastic interpolants, thereby enabling single-step inference without relying on iterative generative inference or auxiliary trajectory-straightening procedures. Within DMAT, local 3D geometry anchors predictions to spatial structure, while State-Consistent Spectral-Time Refinement (STR) reconciles predicted chunks with proprioception and suppresses high-frequency residuals. Across 567 standard and out-of-distribution (OOD) trials per method, DMAT outperformed Improved 3D Diffusion Policy (IDP3) and DiTFlow3D in aggregate success rate, achieving relative improvements of 129.1% and 95.7%, respectively, while generating action chunks 2.06x and 1.34x faster and remaining robust under repeated 50 N interventions. These results pave the way for a novel generative paradigm for efficient and robust bimanual manipulation.

    manipulationdiffusion policy
  94. arxiv:2610.41641 · cs.LG
    A Constructive Orbit-Based Theory of Learnability
    Yunied Puig

    We develop a finite-dimensional orbit framework for constructing data-adapted dictionaries, sparse realizations, and predictive extensions. The method uses wrapped weighted shifts to generate orbit dictionaries and local interpolation maps from independent orbit representatives, turning the learning problem into three explicit tasks: scaffold discovery, sparse realization on the discovered dictionary, and prediction through learned sparse coordinates. We introduce CCDS as the main constructive ingredient: a deviation-controlled sparse realization procedure that recovers MP/OMP-type behavior at the minimal-deviation boundary while allowing feasible interior paths with prescribed residual geometry that is inaccessible to MP/OMP. We prove generic orbit independence, exact block interpolation with norm control, observable bounds on sparse approximation distortion, sparse realization guarantees on selected supports, and fixed-path stability under admissible perturbations of the block interpolants. These results connect sparse approximation, interpolation, and finite-dimensional operator dynamics through the construction and analysis of data-induced orbit dictionaries. Experiments on synthetic geometry, controlled realization tests, and standard classification benchmarks show that selected orbit scaffolds have substantially better matched-rank projection behavior than random orbit or unstructured alternatives, that CCDS attains near-zero realization error at the minimal-deviation boundary, and that predictors trained from the sparse CCDS certificates give competitive performance.

    benchmark
  95. arxiv:2610.41633 · eess.SY
    Probability-Layer Clustering for Bound-and-Budget Calibration in Multi-Bound Robust Bidding of Renewable-Based Virtual Power Plants
    Hadi Nemati, Álvaro Ortega, Enrique Lobato, Luis Rouco

    RVPPs that aggregate dispatchable and non-dispatchable renewable energy sources, storage, and flexible demand that bid in the day-ahead and secondary reserve markets must cope with uncertainties in market prices, renewable generation, and demand. MBRO is well suited to this setting because it hedges adverse realizations through several deviation layers while keeping the final bidding problem as a tractable, scenario-free MILP. Its performance, however, depends on how the deviation bounds and uncertainty budgets are calibrated. Existing MBRO implementations commonly rely on fixed percentile-based bounds and externally specified budget vectors, usually examined through sensitivity analysis. This paper proposes PL-MBRO, that constructs parameter-specific deviation bounds from historical data conditional on selected risk-attitude quantiles and automatically derives the corresponding uncertainty budgets from empirical cluster probabilities. For each uncertain parameter, k-medians clustering partitions the historical daily profiles into representative probability layers; the cluster median defines the balanced representative, while favorable and adverse within-cluster percentiles define the optimistic and pessimistic representatives. The empirical cluster probabilities are mapped to integer full-horizon uncertainty budgets over the 96 quarter-hourly periods, while the robust inner problem determines the periods in which the corresponding deviation layers are most adverse to the schedule. Compared with a percentilebased MBRO benchmark, in-sample and out-of-sample tests on an RVPP bidding in both markets show that PL-MBRO consistently achieves better realized economic outcomes across the optimistic, balanced, and pessimistic strategies, with improvements of up to 92.7% in-sample and 57.3% out-of-sample, measured relative to the absolute value of the corresponding benchmark realized profit.

    benchmark
  96. arxiv:2610.41629 · cs.CV
    GRIT: Gradient Masking for Visually Misgrounded Tokens in Text-Rich Reasoning
    Chanseul Cho, Noori Bae, Dongheon Lee, Youngmin Ro

    Reinforcement learning with verifiable rewards (RLVR) has recently been extended to vision-language models (VLMs), yet its effectiveness remains limited for text-rich reasoning tasks. We highlight a key factor contributing to this limitation: models frequently arrive at correct final answers despite misperceiving visual information during intermediate steps. We define this phenomenon as Correct-but-Visually-Misgrounded Reasoning (CVMR). Because standard RLVR rewards only final-answer correctness, it fails to differentiate between properly grounded and misgrounded reasoning, reinforcing both equally. Through token-level analysis, we reveal that CVMR tokens exhibit distinct attention patterns that differ systematically from well-grounded ones. Based on this observation, we propose GRIT, a training framework that incorporates visual grounding into RLVR to mitigate the impact of CVMR. Furthermore, to address the lack of text-rich reasoning benchmarks that allow for systematic evaluation across text length and reasoning depth, we introduce GRIT-Bench. Experiments demonstrate that GRIT significantly improves both reasoning accuracy and grounding fidelity over existing RLVR methods.

    benchmark
  97. arxiv:2610.41625 · cs.RO
    RTPolicy: Real-Time (32.9 Hz) Multi-View Visuomotor Policy for Robotic Manipulation on Edge Devices
    Xianbao Hou, Zehao Ni, Yonghao He, Lingfeng Qian +4

    Low inference latency is critical for closed-loop robotic manipulation, especially under asynchronous inference, where action chunks must be updated promptly to correct accumulated deviations. However, existing visuomotor policies rely on computationally expensive visual backbones, resulting in high inference latency and deployment overhead. In this paper, we propose RTPolicy, a multi-view visuomotor policy that attains accurate manipulation with high inference efficiency. Its core design combines view-aware feature fusion with visual compression to aggregate complementary information across camera views into a compact token representation. The representation conditions a flow-matching action generator with the robot state and flow time, reducing the cost of action generation. Extensive experiments demonstrate that this design reduces the computational cost of visual representation and action generation while preserving visual information required for manipulation. On the RoboTwin~2.0 single-task clean evaluation across 50 tasks, RTPolicy attains a mean success rate of 52.26\%, outperforming the strongest leaderboard baseline $π_0$ (46.4\%). We further evaluate RTPolicy under asynchronous streaming inference on a real bimanual platform, where RTPolicy attains a mean success rate of 72.0\% and sustains real-time inference at 30.35 ms (32.9 Hz) on the RDK S600 and 149.72 ms (6.7 Hz) on the RDK S100, both in INT16. Project Page: https://d-robotics-ai-lab.github.io/RTPolicy.project.

    manipulationrobotwinleaderboard
  98. arxiv:2610.41618 · cs.CV
    Attention Maps Know When to Think: Training-Free Difficulty-Aware Chain-of-Thought for Reasoning Segmentation
    Chanseul Cho, Jaesung Jun, Dongheon Lee, Youngmin Ro

    Chain-of-Thought (CoT) prompting has become a standard reasoning strategy in Multi-modal Large Language Models (MLLMs), showing strong performance in complex vision-language tasks such as reasoning segmentation. However, our analysis reveals that indiscriminate CoT usage can lead to degraded performance by inducing unnecessary reasoning, a phenomenon we term overthinking. In reasoning segmentation, overthinking often leads to misalignment between the query's intent and visual grounding. To address this, we propose Difficulty-Aware Chain-of-Thought (DA-CoT), a training-free adaptive framework that selectively triggers CoT based on the input's intrinsic difficulty. We find that the entropy of the model's attention maps reliably reflects task difficulty: low entropy indicates confident, localized grounding, while high entropy suggests uncertainty and potential benefit from CoT. DA-CoT uses this entropy signal to determine whether to apply CoT prompting. Additionally, we introduce a Bidirectional Rethinking mechanism that enables the model to automatically flip its reasoning decision when the initial output fails, enhancing robustness by recovering from failures induced by overthinking and underthinking. DA-CoT demonstrates its effectiveness by achieving state-of-the-art performance on the Reasoning Segmentation benchmarks while significantly improving inference efficiency.

    benchmark
  99. arxiv:2610.41616 · cs.AI
    What Should LLMs Forget: Reinforcement Unlearning of Knowledge Associations
    Wutong Xu, Yun Qu, Haiyan Feng, Qi Wang +1

    Large language models (LLMs) acquire extensive factual knowledge during pre-training, which may include sensitive, private, or copyrighted information, motivating the development of unlearning methods that selectively remove specific knowledge while preserving general capabilities. However, existing approaches typically optimize entire responses without explicitly identifying the knowledge to be forgotten, resulting in coarse and uncontrolled forgetting. To address this, we revisit the forgetting target of LLM unlearning and introduce Knowledge Association (KA), which characterizes the association between a query context and its target knowledge and is instantiated by the corresponding conditional predictive distribution, as the fundamental target of forgetting. Based on KA, we propose Knowledge Association Reinforcement Unlearning (KARU), a unified reinforcement learning framework that suppresses forgotten knowledge associations while preserving retained ones. Specifically, KARU derives complementary mechanisms from the KA-based formulation: token-level reward construction for finer-grained forgetting supervision, distribution-level regularization for precise knowledge suppression, and proximity-aware retain sampling to mitigate collateral degradation. Experiments across diverse LLMs on representative benchmarks demonstrate that KARU consistently improves forgetting effectiveness over existing baselines while maintaining comparable retention performance and general utility.

    benchmark
  100. arxiv:2610.41615 · eess.SY
    Distributed Zeroth-Order Composite Finite-Sum Optimization via Incremental Variance Reduction
    Silan Zhang, Yujie Tang

    We study decentralized composite finite-sum optimization over a fixed connected undirected network, where agents access local component functions through function evaluations and share a possibly nonsmooth convex regularizer. In this setting, standard two-point gradient estimators can retain nonvanishing variance at a solution, complicating their integration with gradient tracking. We propose Distributed Zeroth-Order Incremental Variance Reduction (Distributed ZIVR), which combines incrementally updated component-gradient estimates with a proximal gradient-tracking scheme. The method admits an implementation using a constant number of function evaluations per agent at every iteration, without intermittent full-component or full-coordinate refreshes. For convex smooth component functions and strongly convex local averages, we establish linear convergence up to a controllable finite-difference error. With the finite-difference radius chosen for the target accuracy, Chebyshev-accelerated mixing yields a per-agent oracle complexity of $\widetilde{O}(d(m+κ)\log(1/ε))$ for expected squared solution error $ε$, where $d$ is the dimension, $m$ is the number of components per agent, and $κ$ is the condition number. This matches the centralized ZIVR oracle bound, while the communication bound has only an inverse square-root dependence on the network spectral gap. Experiments on logistic and robust regression demonstrate rapid reduction of solution and consensus errors.

    agent
  101. arxiv:2610.41609 · cs.RO
    VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression
    Minsu Jang, Jae-Woo Choi, Youngwoo Yoon, Dohyung Kim

    A robot should act only on an instruction from a person or on an intent that the robot explicitly represents and generates. Unintended actions taken without such an instruction can cause risk or inconvenience to people and the environment. We examine how recent vision-language-action (VLA) policies respond when the instruction is missing or meaningless. We apply a paired protocol with empty, neutral, nonsense, infeasible, and wrong-task instructions to six model checkpoints in LIBERO-Goal simulation. Holding the initial state fixed and varying only the instruction, we ask three questions: whether the robot starts to move, which task it completes, and what the instruction determines. Even when the instruction was empty, the robot started moving in every episode with all six checkpoints, and with five of them it often completed a task that the instruction had not specified. On two GR00T checkpoints, the instruction determined which task was executed but not whether the robot started moving; the robot moved even on infeasible instructions. We then tested three methods for suppressing the start: prepending "if no task is given do not move" to the instruction; a gate that holds the robot when the first action chunk for the given instruction nearly equals that for an empty instruction; and fine-tuning the policy to hold still on non-task inputs. The prefix stopped the robot in none of the six checkpoints, and the gate held it on non-task inputs in five of the six checkpoints. The fine-tuned policy never completed a task on non-task inputs, but it did not reliably keep the robot still, and it failed to suppress the start on infeasible instructions.

    vision-language-actiongr00tlibero
  102. arxiv:2610.41604 · cs.AI
    Real Computation or Merely Naming? A Pre-Registered Mechanism Audit of 25 Cognitive-Architecture Modules in an LLM-Based Agent System
    Peter Chen, Ni Zhang, Ying Xu, Siyuan Zhang

    Agent systems built on local large language models are increasingly marketed through the vocabulary of cognitive architecture: modules named global workspace, theory of mind, causal reasoning, or active inference are listed as features, and the presence of the name is taken as the presence of the mechanism. We ask the prior question -- does the named computation actually happen -- and contribute a transferable verification method for the claim-implementation gap: a four-question static audit protocol with line-number evidence and call-chain tracing, demonstrated on two systems. We audit a system claiming 25 cognitive mechanisms, pre-register falsifiable predictions for 11 modules and execute the interventions, replicate the protocol on the IIT-phi core of a second, independent system with its own pre-committed falsification criteria, and run two repair experiments as positive controls. In 18 of 25 modules the computed value never reaches any production output; three idle-wiring shapes new to the dead-code and code-smell literatures emerge (connected-but-never-called, called-but-inert, called-but-content-blind). Of 61 frozen criteria across 12 interventions, 53 were confirmed, two partially adjudicated, five falsified, one adjudication-ambiguous. Repairs raised planner task success from 0.028 to 1.000 and cut state-controller error by 52%. Module names and module computations must be audited as independent claims; the protocol and the five-class failure taxonomy are reusable for the claim-implementation gap at two-system, case-level depth.

    agentagent system
  103. arxiv:2610.41599 · cs.LG
    Underdamped Stochastic Optimal Transport with Landauer boundary conditions: machine learning and analytical solution
    Julia Sanders, Dario Lucente, Marco Baldovin, Paolo Muratore-Ginanneschi

    Finding optimal transport protocols for underdamped stochastic processes is a difficult yet important problem in control theory. In this paper we develop a direct method, based on machine learning, to find the optimal solution of the problem in the presence of non-linear boundary conditions, including those modelling the Landauer bit-erasure problem. The dynamics is steered from a single-peak to a double-peak distribution in position space, by only acting on a controllable external potential. The problem is highly non-linear, and analytical solutions are not available in general. We benchmark our numerical results with a perturbative expansion, which is expected to hold in a suitable overdamped limit.

    benchmark
  104. arxiv:2610.41576 · cs.CL
    Saidzoda-Tajik-LFM-2.6B: Adapting a 2.6B Hybrid Model to Tajik Tool Calling, and What Continued Pretraining Takes Away
    Tohir Saidzoda, Nekruz Maxudov, Bedil Karimov, Abubakr Abdullozoda +4

    We release Saidzoda-Tajik-LFM-2.6B, a Tajik adaptation of the 2.6B-parameter hybrid model LFM2.5-2.6B-Base, built for tool calling in Tajik on modest hardware. The recipe has two stages: LoRA continued pretraining (CPT) on 1,485,319 cleaned Tajik web documents for 3 epochs, then one epoch of supervised fine-tuning (SFT) on a mix of machine-translated Tajik tool calls, English tool calls kept as replay, and Tajik instructions. On a 174-item tool-calling test set built from BFCL v3 and machine-translated to Tajik, strict full-call accuracy on Tajik prompts rises from 20.69 for the base model to 64.14 (paired difference +43.45 points, 95% interval +35.2 to +51.7); a lenient parser that also accepts the base model's own call formats puts the base at 31.72. English shows no detectable change (76.55 and 77.24). Every CPT checkpoint scores zero: CPT removes the native tool-call format and, as the lenient re-scoring shows, part of the skill itself. Outside tool calling, accuracy on a Tajik knowledge benchmark rises from 28.0 to 35.0, Belebele Tajik shows no significant change, and English MMLU falls from 60.8 to 52.5. The model is not a general Tajik assistant. It almost never declines when no tool fits (6.9% correct on Tajik, against 31.0% for the base under the lenient parser), it is weak at multi-turn calls, and its calls should be checked before they are executed. We describe the data, the training settings and every evaluation in enough detail to repeat them.

    tool callingbenchmark
  105. arxiv:2610.41564 · cs.AI
    Verify--Backtrack Loop: Structured Precondition Recovery for Long-Horizon Language Agents
    Mingyang Liu, Qingcan Kang, Shixiong Kai, Tao Zhong +3

    Language agents increasingly tackle long-horizon tasks in interactive environments, where actions often depend on latent preconditions. The core difficulty is not trajectory length but recovery: when a prerequisite is missing or an earlier decomposition fails, trying another plausible action rarely helps. ReAct and later reflection or adaptive-planning methods support local correction, but express recovery as free-form text, with no explicit control state linking the terminal goal, the active goal, and the failure source. We introduce the Verify--Backtrack Loop (VBL), a training-free extension of ReAct that turns precondition discovery and failure diagnosis into structured planning-state updates. VBL maintains the task decomposition as a bounded goal stack and, on goal activation, execution failure, or material state change, jointly verifies two horizons---whether the active goal contributes to the terminal goal and whether it is feasible in the current state. These dual-horizon judgments, combined with typed failure attribution, trigger explicit operations---Push, Pop, retry, or BacktrackTo the parent. We evaluate VBL with Qwen3.5-9B, GPT-4o, and DeepSeek-V3.2 on AppWorld, ALFWorld, and ScienceWorld. VBL improves over ReAct and reflection-based control on every backbone--benchmark pair. On the challenging AppWorld setting, it improves task success by up to 22.07 percentage points over ReAct. Ablations show that removing either verification or diagnosis degrades performance, in several settings below ReAct, confirming that the two mechanisms are complementary. These results show that reliable long-horizon interaction benefits from revising the planning state, not merely the next action.

    benchmark
  106. arxiv:2610.41563 · cs.LG
    ORION: An Outcome-Driven LLM Advisor for RL-Based Multi-Domain 6G Slice Placement
    Mahmoud S. Mayaleh, Fetia Bannour, Nour-El-Houda Yellas

    Deploying 6G network slices across administrative domains requires an online decision for each arriving request: which domain hosts each virtual network function, how traffic crosses domain boundaries, and whether the slice is admitted. This joint decision problem is NP-hard under continuous arrivals. Experiments on a parametrized family of multi-domain topologies show that no fixed planning strategy dominates across all regimes: the performance ranking reverses between topology classes, and feasible placements exist for most rejected requests but remain undiscovered. These findings identify the plan layer as the decisive performance lever. We introduce ORION, a hierarchical orchestrator in which two LLM agents translate service intents into structured requests and generate abstract placement plans. A cooperative multi-agent RL coordinator selects the cross-domain partition, treating the plan as a persistent soft prior, while per-domain agents place the functions. ORION main contribution lies at the plan layer: an outcome-driven episodic memory that records, for each completed attempt, the network-condition profile, the proposed plan, and its verified feasibility outcome. The memory feeds verified successes and labelled failures back to the planner, thereby closing the adaptation loop using deployment feasibility rather than a scalar reward assigned to the planner's own proposals. To isolate the effect of the plan quality, we evaluate the planner on held-out topology families while keeping the downstream selection and execution mechanisms fixed. Under this controlled setting, the proposed memory improves admission by $4.6$--$11.5$ percentage points over the same planner without memory, while importance-based retention accounts for most of the improvement over a first-in-first-out store.

    memoryepisodic memoryllm agentmulti-agent
  107. arxiv:2610.41558 · cs.RO
    A complete analytical characterization of optimal collision-free motion for two robots
    Jan Kamm, Antonio Terpin, Raffaello D'Andrea

    Collision-free motion planning for multi-agent robotic systems remains an active research area because of the complexity of coordinating robot trajectories, with applications ranging from coordinated drone shows to autonomous operations on land, in the air, and on water. Importantly, the analytical difficulty of nonconvex collision-avoidance constraints has so far driven the development of exclusively numerical methods. We derive a complete analytical characterization of the optimal trajectories for two robots that start and end at rest, under double-integrator dynamics and a squared-acceleration objective. We prove that these trajectories belong to eight distinct contact families and determine which family applies for each configuration of initial and final positions. We analyze the geometrical properties of the optimal trajectories, provide a numerical algorithm for computing them, and discuss how these insights may improve the performance of existing numerical trajectory optimizers.

    multi-agent
  108. arxiv:2610.41553 · cs.CL
    CEFR Level Assessment across Languages and Text Granularities
    Eugene Matsumura, Yuki Arase

    Automatic assessment of textual CEFR levels is an important task in educational natural language processing. However, accurate CEFR level prediction remains challenging due to the naturally imbalanced distribution of CEFR levels and the ambiguous boundaries between adjacent levels. To address these challenges, we propose a unified training objective that combines focal loss with supervised contrastive learning. Focal loss mitigates class imbalance by emphasizing underrepresented and difficult examples, while supervised contrastive learning learns more discriminative representations by encouraging intra-class compactness and inter-class separation. We evaluate the proposed method on sentence- and document-level CEFR assessment in both monolingual and multilingual settings. Experimental results demonstrate that our method consistently outperforms strong baselines across all settings, establishing new state-of-the-art performance on multiple CEFR assessment benchmarks.

    benchmark
  109. arxiv:2610.41548 · cs.RO
    Contact-based optimization for collision-free motion planning for a million robots
    Jan Kamm, Antonio Terpin, Raffaello D'Andrea

    Multi-agent robotic systems have attracted growing interest in recent years, with applications ranging from coordinated drone shows to autonomous operations on land, in the air, and on water. At the planning level, robot motion is often modeled using double-integrator dynamics (with low-level controllers handling the physical dynamics during execution) with a minimum-energy objective subject to collision-avoidance constraints and limits on position, velocity, and acceleration. The complexity of the problem resides in the constraints, which couple the trajectories and make the optimization problem nonconvex. Existing numerical approaches typically discretize and optimize entire trajectories, which becomes expensive for large fleets. We exploit the structure of optimal trajectories by organizing optimization around contacts, at which robots reach the minimum allowed separation. We formulate a coupled algebraic system in contact variables and reconstruct trajectories analytically between contacts. Our algorithm, sequential contact insertion (SCI), starts from unconstrained trajectories, inserts contacts at separation violations, and jointly adjusts their times and states. Numerical experiments demonstrate planning for one million robots in minutes and thousands of robots in fractions of a second. On problems feasible for a fixed-grid sequential convex programming baseline, SCI achieves median speedups of several orders of magnitude at matching solution accuracy.

    multi-agent
  110. arxiv:2610.41540 · cs.RO
    FOCUS: First-Order Control Using Optimal Transport Supervision for Imitation from Observations
    João A. Cândido Ramos, Lionel Blondé, Hugues Vinzant, Jacopo Castellini +1

    Imitation learning from observations must recover behavior from state-only demonstrations without known temporal correspondence. Differentiable simulators provide pathwise policy gradients, but differentiating through long, contact-rich trajectories can make optimization unstable. We introduce FOCUS, which combines local optimal-transport alignment with short-horizon differentiable actor-critic learning. FOCUS matches agent rollout windows to multiple sampled expert segments using a soft best-of-$K$ objective that emphasizes locally compatible matches. The resulting transport-based costs provide differentiable imitation rewards, while a learned critic bootstraps value beyond the rollout horizon. FOCUS requires neither temporally aligned trajectories nor expert-aligned initialization or teacher forcing. Across contact-rich locomotion tasks in DFlex, including a musculoskeletal humanoid controlled by 152 muscle activations, FOCUS improves sample efficiency and training stability over adversarial imitation-from-observation methods and prior differentiable imitation baselines.

    humanoidagent
  111. arxiv:2610.41536 · cs.LG
    DAPS: Dependency-Aware Premise Selection for LLM Theorem Proving
    Zixuan Chen, Wenyuan Jiang, Junling Wang, Mrinmaya Sachan +1

    LLM-based theorem proving in Lean 4 has advanced rapidly, but provers are difficult to leverage newly contributed lemmas, because retrieval skews toward popular foundational lemmas, leaving the right premises buried among hundreds of thousands of candidates. Existing neural selectors rank candidates by semantic similarity, with three intrinsic limitations: frequency-skewed retrieval, isolated pairwise ranking, and degradation on out-of-distribution queries such as competition problems and informal language. Given that mathematics is structured by dependencies rather than similarities, we restore this signal to both the training data and the model. We curate two underused dependency resources. We conduct the first systematic mining of 53 Lean Blueprint projects, a previously untapped corpus of 3,805 manually curated informal+formal nodes that capture rare, research-level premises absent from Mathlib4. We further pair it with the extraction of comprehensive, typed, multi-level Mathlib4 dependencies. We then introduce DAPS, a dependency-aware premise selector with a structural neighborhood encoder, a group-level contrastive objective, and a Mathlib-then-Blueprint adaptation procedure. DAPS reaches mean Recall@32 of 89.25 on Mathlib4-Heldout (+11.41 over the strong selector LeanHammer), and maintains this lead on two out-of-distribution benchmarks. Further experiments demonstrate that DAPS premises improve theorem proving on miniF2F and IMO-ProofBench. Project page: https://daps-eth.github.io/.

    benchmark
  112. arxiv:2610.41511 · cs.AI
    Specification-Guided Generation for Large Language Models
    Alessandro Previti, Konstantinos Vandikas, Mario Alviano, Mats Carlsson +4

    Large language models (LLMs) can generate structured outputs, but existing constrained-decoding methods mainly enforce syntax and cannot directly enforce satisfaction of a formal specification. We introduce a neuro-symbolic, specification-guided decoding framework that combines LLM generation with symbolic reasoning. At each step, the framework refines the decoding grammar by filtering out values that cannot be extended to an output satisfying the specification. We present a forward-only variant that guarantees specification satisfaction by construction, and a backtracking variant that revises earlier feasible choices when their downstream consequences leave only values that the LLM itself judges unsuitable. The framework is model-agnostic, requires no retraining, and builds on existing grammar-constrained decoding mechanisms. Experiments on configuration and constraint-solving benchmarks show that the forward-only variant reliably produces specification-compliant outputs while reducing repeated LLM inference compared with rejection sampling and incurring modest symbolic-reasoning overhead. The backtracking variant further reduces cases in which the model is forced to select values that it has itself judged unsuitable, thereby substantially improving adherence to the LLM's stated interpretation of the prompt.

    benchmark
  113. arxiv:2610.41501 · cs.CV
    VoxLution: Evolving Voxels for High-Fidelity 3D Scene Generation
    Daoyi Gao, Xinyi Zhang, Alexey Bokhovkin, Lei Li +1

    High-resolution 3D indoor scene generation is crucial for immersive applications such as virtual reality, spatial computing, and robotics simulation. We introduce VoxLution, a coarse-to-fine transformer-based approach that generates high-fidelity indoor scenes with rich details, overcoming the overly smooth geometry typical of existing generative models. First, we generate a low-resolution spatial layout; then, based on this structural information, we progressively refine the scene geometry locally in chunks. Key to our approach is formulating 3D super-resolution as a sequence modeling problem over multi-resolution sparse voxel grids. To achieve this, we introduce a training-free, resolution-invariant tokenization scheme that compactly encodes the geometric information of eight octants into a single decimal number. We formulate the transformer sequence to include multiple resolutions, allowing the coarse layout to guide the generation of fine local details. To ensure coherence across local chunks, we incorporate adjacent chunk boundary information directly into the transformer sequencing. VoxLution achieves scene generation at a finer voxel resolution than existing baselines, demonstrating state-of-the-art performance on standard benchmarks with 59.6% lower FPD and 51.1% lower FID on 3D-FRONT, and 70.5% lower FPD and 73.8% lower FID on Scannet++.

    benchmark
  114. arxiv:2610.41487 · cs.AI
    Sparse Text Diffusion under Dual Budgets: Context Connectivity and Expert Activation
    Alexandros Christoforos, Chadbourne Davis, Paul Raj

    Long-context diffusion repeatedly pays for both token interaction and feed-forward computation. We formulate these costs as two separate budgets: the edges available to an attention graph and the experts activated for each token. This view connects local, dilated, and global attention with conditional expert computation without assuming that sparsity alone preserves document-level information. We specify normalized masked attention, distinguish denoising depth from sampling iterations, and derive graph and compute bounds. A reported sparse-expert configuration, denoted DualBudget, is analyzed across scientific summarization, multi-hop question answering, dialogue, and sentence-pair prediction. All supplied table measurements are retained. The length study shows gains over the listed baselines through 16K, together with measurable degradation as length increases. Ablations reveal mixed changes across the three reported scores, while the cost record indicates a 20% reduction relative to DiffuSeq under an unspecified timing convention. The analysis separates these observations from unmeasured claims about training speed, absorbing states, global coherence, and deployment memory.

    long-context
  115. arxiv:2610.41486 · cs.AI
    Multi-Test Nation Bias Benchmarking of LLMs with Explicit Unbiased Baselines
    Jonghyeon Choi, Yeonjun Choi, Beakcheol Jang

    While prior work has highlighted that systematic nation-level bias in large language models (LLMs) can pose operational risks for international relations (IR) applications, many existing evaluations still lack a clearly specified unbiased reference (ground truth), limiting fully quantitative and cross-setting measurement. Using an expanded corpus of 722 official UN Security Council (UNSC) records, we measure nation-level bias with five complementary tests, thereby extending prior approaches to cover a broader set of bias expression channels: opinion ranking (WOW), association sorting (AK+-), vote simulation (PVS), opposition attribution (F-Oppose), and framing (Headliner). To enable grounded bias evaluation, we explicitly define an unbiased status for each test; in particular, for WOW and AK+- where ground truth is difficult to specify, we propose an action-based pseudo unbiased status derived from observable voting behavior to construct a quantifiable baseline. Benchmarking 4 LLMs (GPT-4o-mini, Llama 3.3-70B, Mistral-22B-Small, and Qwen3-30B) across the five permanent members of the UNSC, we find a test-agnostic macro-pattern: relative to the reference baselines defined for each test, model outputs lean toward the opposition (red) side for China and Russia and toward the supportive (blue) side for the U.K., the U.S., and France. These results underscore the need to audit IR-facing LLM deployments with grounded, multi-test benchmarks tied to explicit reference points.

    benchmark
  116. arxiv:2610.41482 · cs.AI
    KEEPER: Learning Control-Sufficient Textual Memory for Long-Horizon LLM Agents
    Jinkyeong Choi, Chaebin Jeong, Junsung Jeon, Donghyeon Park

    Recursive textual memory can carry earlier errors into later updates, yet task performance alone does not establish memory reliability. We define self-consistency bias as the persistence of incorrect claims already present in the incoming memory. We present KEEPER, a comparative analysis of four auxiliary objectives along information-content and update-sensitivity axes within a shared recursive interface, examining how supervision targets relate to agent performance. Among evaluated conditions, reward-related prediction achieve the highest task performance before and after offline reinforcement learning, outperforming joint supervision. A claim-level analysis of one baseline trajectory documents error persistence despite corrective observations. Together, these analyses motivate evaluating error persistence and valid-information retention alongside task outcomes.

    memoryagentllm agent
  117. arxiv:2610.41460 · cs.LG
    RetryRoute: Retry-Aware Model Routing for Agentic Image Editing Services
    Hanshuai Cui, Zhiqing Tang, Zhi Yao, Jun Luo +1

    Web agents increasingly edit images by calling open editing models in a critique-and-retry loop, and these models differ in cost by more than an order of magnitude. After every rejected edit, a service must decide whether to retry the same model, escalate to another or give up. Existing routers pick one model per request, and cascades escalate in a fixed order. However, both treat a retry as an independent draw. We identify sticky failures, meaning that a retry after a failure succeeds only 17-53% as often as a first try across six open models, three benchmarks and two judges. Based on this finding, we propose RetryRoute, a retry-aware router that estimates each model's chance of success given the request's failure history. It plans retries, escalations and abandonment over failure states and adds live queue waits to the cost, so the same plan becomes load-aware dispatch. At the main budget, RetryRoute's success under the in-loop judge exceeds the strongest of eleven prior baselines by 2.8 points on average over eight groups. On a GPU serving testbed it raises the share of edits delivered within the service-level objective by 5.0-6.3 points over the best prior baseline at every load.

    agenticbenchmark
  118. arxiv:2610.41459 · cs.LG
    NormDetect: Anchored Calibration for Low-False-Positive Detection of Machine-Generated Web Text
    Hanshuai Cui, Fanshuai Meng, Zhiqing Tang

    Web platforms screen posts for text written by large language models, and every false flag hits a human writer. Detectors must therefore run at a very low false positive rate (FPR). Zero-shot detectors such as Binoculars score text with frozen language models, but invisible edits such as homoglyphs shatter tokenization and push machine text into the human range. Unicode normalization undoes these edits, yet the normalized score still misses half of the machine text at a strict FPR, under attacks that leave nothing to clean. We find that with the Falcon-7B scoring pair, heads on token statistics recall less than the normalized score at this FPR, and several flag too many human posts in some sections. We propose NormDetect, which anchors a learned per-section head to the normalized zero-shot score. A blend weight chosen on training folds under a human FPR cap combines the two scores, and one threshold is calibrated on human text only. On eight sections of the RAID benchmark with eleven attacks, NormDetect raises recall at a 0.1% calibrated FPR by 5.0 points over normalized Binoculars. Its human FPR bound stays at most 1% in every section, and no other detector within this bound recalls as much.

    benchmark
  119. arxiv:2610.41456 · cs.RO
    SAFARI: Semantic- And Frontier-driven Autonomous Risk Inspection
    Ugur Kafkasyali, Mihnea Toader, Julia Richter, Turcan Tuna +2

    Robotic exploration can reduce human exposure to hazardous environments by autonomously searching unknown areas for objects of interest while constructing a complete, risk-aware map. However, existing autonomous exploration approaches typically prioritize geometric coverage or information gain, without explicitly coupling semantic object detection and the risk associated with operating near potentially hazardous objects. Moreover, many systems have been demonstrated primarily in simulation or controlled environments, leaving limited validation of risk-aware semantic exploration in complex, unstructured field settings. To address this gap, we propose SAFARI, a framework that treats object detection as a native channel of a unified probabilistic map, thereby jointly informing exploration and planning with both geometric and semantic evidence. We evaluated SAFARI in 11 missions across three field sites on a quadruped platform. The framework achieved an average coverage of 98.7% of the accessible mission area, without operator intervention for 97.6% of mission time, and detected objects with characteristic dimensions as small as 15cm. Lastly, through ablation studies and a detailed analysis of system behavior, we show that object-size awareness and safety-aware path planning are crucial for effective hazardous-object search.

    quadruped
  120. arxiv:2610.41453 · cs.AI
    Evidence-Guided Recovery for AI Agents with Ambiguous Tool Outcomes and Partial Execution
    Genliang Zhu, Xiaobo Chen, Ziyuan Wang, Chu Wang

    Tool timeouts can leave AI agents facing ambiguous tool outcomes and partial execution: an action may have failed, committed or remained in flight while other workflow steps have already completed. We formulate evidence-guided task recovery over approved obligations, provider capabilities and states compatible with public evidence. A shared admissibility check handles key expiry, authorization timing, dependencies and residual effects; a terminal certificate separates actual goal satisfaction from publicly supported safe completion. The known-model implementation combines standard observation-contingent planning with a separately implemented effect-history adjudicator. In a finite study of 37 assignments, planning and three active controls each certify 28 completions; four saved synthetic fee units incur additional computation. A preregistered assignment-held-out comparison then evaluates planning advice against evidence prompting with two local model bundles. Across 58 paired units in 19 previously developed task families, advice certifies 30 completions versus 9, a 36.2-percentage-point difference (exact family sign-flip p=0.0078), using 272 proposals versus 399. All 116 trajectories pass independent replay. Both models exhibit a positive completion difference. The comparison measures the evaluated advice interface under a common shield; search-versus-rule advice requires a separate matched comparison. Unknown and handoff outcomes remain explicit, with guarantees tied to the declared provider capabilities rather than arbitrary third-party exactly-once execution.

    ai agent
  121. arxiv:2610.41446 · cs.CV
    Beyond Trial-and-Error: A Geospatial-Guided Active Reasoning Framework for Robust Aerial Vision-and-Language Navigation
    Jiawei Ge, Wenkang Gong, Jiuxin Cao, Bo Liu +7

    Aerial Vision-and-Language Navigation (VLN) requires UAVs to locate language-specified goals in large-scale urban environments, where objects often appear small and visually ambiguous from high-altitude viewpoints. Existing methods largely follow a reactive paradigm common in indoor VLN, repeatedly updating navigation states from local observations while underutilizing the geospatial structure that remains stable during navigation. As a result, visually similar distractors can cause a reactive goal search to repeatedly reinforce locally plausible goal hypotheses based on passively accumulated evidence, leading to unreliable localization. To address this, we propose GeoReasoner, a robust framework that shifts the paradigm of Aerial VLN from reactive trial-and-error to geospatial-guided active reasoning. Specifically, GeoReasoner first performs pre-search geospatial reasoning to derive region-level priors and constrain the goal hypothesis space. Building on these, it further conducts hypothesis-driven exploration to actively acquire discriminative evidence and refine competing goal hypotheses. Finally, consistency-aware goal verification adversarially assesses complementary evidence sources before stopping, hence reducing premature or erroneous goal confirmations. Experiments on the real-world CityNav benchmark demonstrate that GeoReasoner sets a new SOTA among zero-shot Aerial VLN methods while competing favorably with supervised approaches.

    benchmark
  122. arxiv:2610.41437 · cs.LG
    A prospective randomized study of patient-centered AI for clinical visit preparation
    Xincheng Tan, Matthew Thompson, Xiang Ji, Angelica Willis +32

    Patients often arrive at brief, time-constrained clinical visits unsure of what to expect or ask, and millions of adults now turn to general-purpose AI chatbots to prepare beforehand. Yet whether conversational AI improves visit readiness remains unclear. We designed a large language model (LLM)-based visit preparation agent to guide patients before an appointment by gathering clinical context, setting visit expectations, providing relevant health information, building a prioritized list of questions for their clinician, and generating a summary. In a prospective consented randomized study in the Google Health app, 802 U.S. adults held 866 health-related conversations with the visit preparation agent (standard or concise variant) or a minimally prompted control chatbot built on the same LLM. The visit preparation agent significantly outperformed the control across blinded clinician evaluations and patient self-reports. In clinician reviews of completed sessions (n = 272), the agent achieved a 16.6% higher rate of comprehensive preparation (67.2% vs. 50.6%; P < 0.05) and an 8.5% higher rate of alignment with patients' health context (98.9% vs. 90.4%; P < 0.01), while maintaining high medical safety and appropriateness across all study arms. Patient post-surveys (n = 460) showed 16.8% higher visit preparedness (73.8% vs. 57.0%; P < 0.01), 14.3% higher confidence in knowing what to ask (81.4% vs. 67.1%; P < 0.01), and 12.0% higher overall satisfaction (74.0% vs. 62.0%; P < 0.05). Clinician and participant metrics were not significantly correlated, indicating that expert review and patient self-assessment capture complementary dimensions of patient-facing AI. These findings suggest that LLM chatbots can safely help patients prepare for clinical visits, with greater benefit when following a structured workflow, and motivate future studies on in-visit communication and clinical outcomes.

    agent
  123. arxiv:2610.41435 · cs.AI
    BlindToChange: AI Agents Can Find Changes When Asked But Do They Look When Needed?
    Anirudh Phukan, Tanya Goyal, Danish Pruthi

    Given that AI agents can search the web broadly and deeply in real time, it is tempting to conclude that their outputs account for recent world events. We show that this conclusion does not hold in practice. While agents can search and easily find recent changes, the correctness of their responses depends on whether they search and what they search for. Popular agentic search benchmarks like BrowseComp cannot expose failures wherein agents do not recognize their own assumptions that should be verified, since these benchmarks explicitly state what needs to be found. In contrast, many user requests state only the final information need (e.g., an itinerary for a trip), leaving the agent to discover relevant changes (e.g., the temporary closure of a museum) and account for their consequences. To highlight such gaps, we introduce BlindToChange, a benchmark that measures whether agents proactively discover recent events that affect their responses. Our benchmark comprises requests for recommendations where an obvious candidate is no longer applicable due to a recent real-world event. These events are drawn automatically from Wikidata, allowing BlindToChange to be periodically refreshed. Strikingly, we find a big gap -- as high as 27.5 points, even at high reasoning effort -- in systems' performance, when the recent event is not explicitly mentioned compared to when we ask about it directly. Analyzing search trajectories, we learn that more than half of the failures can be attributed to the fact that agents do not directly search to verify whether a relevant fact has changed. Our benchmark highlights, and tracks, a new axis of progress, wherein systems account for recent information that is not explicitly mentioned in the user request.

    agentai agentagenticbenchmark
  124. arxiv:2610.41432 · cs.CL
    Hybrid Computer Use: A Unified Programmatic, Semantic, and Visual Interface for General Agents
    Haoqing Wang, Xiaoyi Zhang, Sicheng Fan, Zhexin Zhang +4

    Computer-use agents increasingly tackle tasks ranging from everyday desktop activities to long-horizon professional workflows. Yet interfaces centered on screenshots and mouse-and-keyboard actions can impose a human-interaction prior, requiring repeated screen interpretation while leaving the computer's semantic and programmatic capabilities underexposed. We argue for exposing these capabilities and letting models choose and combine them. We introduce Hybrid Computer Use (HCU), a unified interface for general agents that combines three complementary channels: programmatic interaction through code and command-line tools, semantic interaction through accessibility (AX), and visual interaction through graphical user interfaces (GUI). Our HCU supports efficient AX interaction and guards against stale-state actions during channel switching. A supplementary HCU skill guides agents in selecting and coordinating channels to use the unified interface effectively. Our designed AX tools return compact observations, execute actions in batches, and report their effects as AX diffs, so the model needs fewer round trips. On OSWorld 2.0, HCU achieves the highest Partial and Binary scores among the compared systems. Analyzing agent trajectories, we find that all evaluated models improve with HCU yet allocate work across the channels differently and consistently outperform compared baselines in terms of both efficiency and task success, highlighting the value of unified computer-use interfaces that allow flexible channel selection and combination.

    agent
  125. arxiv:2610.41429 · cs.RO
    Physically Constrained Kinematic Calibration for High-Precision Robotic Needle Guidance
    Prabhu Nithin Gollapudi, Andreas Maier, Franziska Mathis-Ullrich, Water Zhao +2

    Percutaneous needle insertion is a common practice in minimally invasive procedures whose clinical success depends on accurate needle placement. However, achieving the necessary precision is a challenge for manual operators. Robotic assistance can improve targeting accuracy and reduce operator dependence, but residual millimeter-scale errors persist because nominal kinematic models do not fully represent the physical manipulator. This paper presents a robust bound-constrained kinematic calibration method designed to improve positional accuracy for a 6-Degree-of-Freedom (6-DoF) robotic needle-guidance manipulator. The approach uses physically plausible parameter ranges to regularize a compact 19-parameter geometric identification model. By treating these bounds as active optimization constraints rather than mere numerical safeguards, the framework effectively mitigates parameter coupling and overfitting, ensuring highly stable parameter identification. In simulation, the proposed method reduced end-effector Root Mean Square Error (RMSE) from 3.422 mm to 0.690 mm (a 79.8% improvement). In benchtop hardware experiments, the calibrated model achieved sub-millimeter validation accuracy, including 0.865 mm RMSE in experimental trajectory validation and 0.610 mm RMSE in a coordinate-lattice verification task. These results demonstrate that physically constrained kinematic calibration can enhance robotic needle-targeting accuracy, establishing a high-precision foundation for future clinical applications.

    manipulator
  126. arxiv:2610.41426 · cs.CV
    FlexReader: Towards Adaptive Decoding Paths for Efficient Arbitrary-Length Scene Text Recognition
    Zhibin Ma, Pengwen Dai, Zhigang Yao, Xugong Qin +2

    Existing Scene Text Recognition (STR) models are inherently coupled with a specific decoding strategy by their training paradigms, limiting their flexibility in deployment across diverse recognition scenarios. In this paper, we propose FlexReader, an STR model that supports multiple decoding strides, enabling a single model to achieve an optimal accuracy-efficiency trade-off across different datasets. Specifically, we first introduce a Sub-string Perception module to perceive the target sub-string from visual features, followed by multiple parallel decoding blocks that independently decode each character in the sub-string. During inference, FlexReader supports decoding with arbitrary strides within the perceived sub-string range. Furthermore, we propose Adaptive Prefix Decoding, which adaptively selects the decoding stride at each step based on the model's confidence in the decoded characters, thereby identifying the optimal decoding path for each input. Experimental results demonstrate that FlexReader exhibits robust recognition capability across a diverse range of benchmarks, including those featuring long text, curved text, and occlusions. It consistently outperforms existing decoding strategies in recognition accuracy while achieving up to 6-flod decoding speedup over autoregressive decoding.

    benchmark
  127. arxiv:2610.41420 · cs.LG
    SparseGDN: Chunk Sparse in Linear Attention
    Jiedong Zhuang, Peng Zhao, Jialin Tang, Haoji Hu

    Gated Delta Networks (GDNs) offer linear-time sequence modeling, yet long-context prefill remains costly due to dense intra-chunk computation, recurrent-state updates, and intermediate state storage. We introduce SparseGDN, a hardware-aware framework that combines structured intra-chunk approximation with input-dependent recurrent-event selection while producing outputs for every token. A block-adjacent solver reduces local Delta-solve computation, while a dynamic selector allocates recurrent-update budgets across head pairs. The GPU implementation materializes updates only for selected events and reconstructs token outputs using compact state snapshots and group-local interactions, reducing computation and memory traffic. We evaluate SparseGDN on Qwen3.5-4B and Qwen3.5-9B across RULER tasks and context lengths from 4K to 128K. Retaining approximately 29% of recurrent events on average, SparseGDN achieves 1.50-1.94x GDN-core prefill speedups over native FLA kernels, including online selection overhead.

    memorylong-context
  128. arxiv:2610.41411 · cs.AI
    CryoAtlas: A Large Curated Dataset and Unified Benchmark for Cryo-EM Atomic Model Building
    Mingrui Li, Minzhang Li, Weichen Qin, Yufan Xie +5

    Deep learning has accelerated automated atomic model building from cryo-EM maps, yet progress remains limited by the lack of large, high-quality datasets and standardized evaluation benchmarks. We address this gap by introducing CryoAtlas, a rigorously curated dataset and unified benchmark for density-guided atomic model building of protein chains. By integrating and filtering EMDB and PDB entries (up to March 2026), CryoAtlas provides 17,809 map-model-sequence entries, including a curated 16,657-entry training split. From this resource, we construct a 618-target benchmark with a shared evaluation protocol that jointly assesses structural accuracy, model-to-map fit, and stereochemical validity. Using this framework, we benchmark state-of-the-art learning-based methods against Phenix and AlphaFold3, comprehensively analyzing the impact of map resolution, sequence length, protein category and sequence identity on modeling accuracy. As a case study, retraining the ModelAngelo CNN component on CryoAtlas improves performance on the benchmark. CryoAtlas provides an open, reproducible foundation for future cryo-EM atomic model building research. Project details and resource release updates are available at https://oulling.github.io/cryoatlas/.

    benchmarkevaluation protocol
  129. arxiv:2610.41408 · cs.RO
    SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models
    Haesung Lee, Insu Han

    Vision-Language-Action (VLA) models enable capable robot behavior, but their large memory footprint limits deployment on resource-constrained robotic platforms. Weight quantization is a natural remedy, yet post-training quantization (PTQ) recipes inherited from large language models degrade when applied to VLAs at ultra-low precision. Existing low-bit PTQ methods often fail to explicitly account for the structure of downstream operations and allocate precision without considering its effect on the policy actions. We introduce SACQ, an ultra- low-bit weight-only vector-codebook quantization framework for VLAs. SACQ combines (1) structure-aware codebook fitting with role-specific quadratic metrics that reflect input statistics and downstream sensitivity; (2) action-aware mixed-bit allocation based on the increase in final-action error as individual layers are compressed; and (3) two-stage post-quantization refinement that reassigns codeword indices with second-order error compensation and calibrates codebooks and row-group scales against the decoded action. Using GR00T N1.7, SACQ achieves 93.85% average success across four LIBERO suites at 1.5 bits per language/action weight while reducing model memory by 4.83x. With INT8 activation quantization and low-precision attention, it retains 93.55% success while running 1.51x faster than BF16. On SimplER, SACQ further achieves 63.76% and 52.27% success under Visual Matching and Variant Aggregation at 2 bits per language/action weight, compared to 58.90% and 53.45% for BF16, respectively.

    vision-language-actiongr00tliberomemorypost-training
  130. arxiv:2610.41402 · cs.LG
    DUGO: Dual-Scale Gaussian Memory for Online Temporal Action Localization in Egocentric Procedural Videos
    Mohammed Aman Bhuiyan, Md. Sazzad Hossain Adib, Samiul Basir Bhuiyan, Mir Faiyaz Hossain +3

    Online Temporal Action Localization (OnTAL) aims to detect and localize action instances from untrimmed streaming videos in real time. Unlike offline approaches, OnTAL models have no access to future frames and cannot go back and refine earlier predictions, which makes the task considerably harder given the strict real-time constraints and narrow temporal window available. Most existing OnTAL methods lean on either brief anchor-based context or a single-scale memory of past frames, which limits how well they can capture actions unfolding across multiple time scales or requiring long-range reasoning in Procedural Videos. To address these gaps, we present DUGO (Dual-Scale Encoding and Gaussian Memory Online), a new OnTAL framework that jointly models local motion dynamics and long-range temporal dependencies. At its core, DUGO uses a Dual-Scale Temporal Encoder, pairing depth-wise local representations with global self-attention so it can pick up both fine-grained temporal patterns and broader contextual cues. We also introduce a Gaussian Memory Bank, which represents action memory as stochastic, input-conditioned Gaussian prototypes, allowing for richer temporal memory through gated cross-attention applied at both the encoder and anchor-query stages. Extensive experiments on procedural egocentric datasets (EGTEA, EPIC-Kitchens-100) demonstrate that DUGO achieves state-of-the-art performance while also providing substantial improvements on non-egocentric benchmarks such as Muses. The project page is available at: https://aman554-eq.github.io/DUGO

    memorybenchmark
  131. arxiv:2610.41401 · cs.AI
    When Safety Checks Become Privacy Risks: A Privacy-Preserving Paradigm for LLM Agent Safety Auditing
    Yixiang Luo, Ziheng Ma, Ke Chang, Jian Wang +1

    Large language model agents interact with online services through tool calls that may access private data or perform sensitive actions. Specialized providers can supply and maintain pre-execution safety auditing capabilities for agent operators. However, sending task descriptions, call arguments, and interaction histories to an external auditor may expose sensitive user information. In this work, we propose a paradigm for privacy-preserving safety auditing that separates delegated audit computation from plaintext access and execution authority. Grounded in a three-domain auditing model, the paradigm connects trusted representation construction, encrypted audit computation, and trusted decision and execution through explicit interfaces. These interfaces accommodate alternative representations and scorers, while encrypted computation protects audit inputs from a semi-honest auditor. We instantiate this paradigm in SealCall, a framework combining learned audit representations, CKKS-based encrypted similarity computation, and execution control bound to audited tool calls. Evaluated on 1,252 public binary candidates, SealCall achieves a mean F1 score of 0.840 under grouped nested evaluation, while preserving plaintext audit decisions in all measured cases. A representative configuration completes each encrypted scoring operation in 0.176 s.

    agentllm agent
  132. arxiv:2610.41396 · cs.AI
    ProSearch: Benchmarking Multi-Constraint Protocol Retrieval in Experimental Science
    Wenliang Liang, Zhiyuan Ning, Haolin Chen, Yuanchun Zhou +2

    In AI for Science scenarios, experimental design and method reproduction require retrieving protocols that satisfy multiple interdependent constraints, yet existing benchmarks lack systematic evaluation of this capability. To address this gap, we introduce ProSearch, a benchmark for multi-constraint protocol retrieval in experimental science. ProSearch has three key features: (1) realism, with queries grounded in real experimental records and expert-defined templates; (2) complexity, with 60 experimental fields and their dependencies; and (3) diagnosticity, with 3-5 explicit constraints per query, including numerical and negation constraints. We construct ProSearch through LLM-based extraction, data-driven constraint selection, controlled query generation, and evidence-based relevance validation. The resulting benchmark contains 1,515 queries across six experimental science domains. We conduct extensive experiments on ProSearch with 19 representative retrievers and complementary analyses. The best-performing retriever achieves an average nDCG@10 of only 0.290, which improves to 0.474 with reranking. Structured diagnostic analysis shows that explicit constraint matching provides a strong ranking signal, while manual error analysis identifies joint constraint binding as a key bottleneck. ProSearch provides a testbed for systematically evaluating scientific retrieval systems.

    benchmark
  133. arxiv:2610.41389 · cs.CV
    Parse Before You Execute: Efficient Spatial Reasoning with Executable Primitives
    Xiwei Liu, Yulong Li, Xinling Zhuang, Xuhui Li +4

    Spatial reasoning in Vision-Language Models (VLMs) has increasingly benefited from tool-integrated pipelines that ground computation in explicit geometry. However, existing systems often compile queries into relatively open-ended artifacts, such as code, task constraints, tool plans, or gathered evidence, leaving what must be resolved before execution under-specified. This can lead to unnecessary tool acquisition and iterative refinement, while still producing plausible reasoning traces from incomplete spatial commitments. We identify this failure mode as under-committed query compilation. To address it, we propose PEPS, a training-free agentic paradigm Parsing Executable Primitives for Spatial reasoning under a typed FESM schema over Frame, Entity, State, and Metric. PEPS follows a parse-before-execute principle: the Parser commits each query to a compact set of executable requirements; the Executor selectively grounds the required primitives through fixed geometric APIs and deterministic computation; and the Verifier checks execution sufficiency, enabling early termination or targeted repair. Extensive experiments across challenging spatial reasoning benchmarks show that PEPS consistently improves reasoning accuracy while substantially reducing refinement rounds, API cost, and end-to-end latency over strong tool-integrated baselines. These results demonstrate that explicit requirement commitment can make tool-integrated spatial reasoning not only more reliable, but also more execution-efficient.

    agenticiterative refinementbenchmark
  134. arxiv:2610.41387 · cs.LG
    Anomalies as Residuals: Cauchy Displacement Networks
    Yang Cao, Zihao Guo

    Neighbourhood-based detectors such as KNN, LOF and LUNAR remain among the most accurate methods for tabular anomaly detection, but they compare every query with the stored training set, so their inference cost and memory grow with the data and the data must be kept at deployment. We ask whether the neighbourhood geometry these detectors exploit can be compiled into a small network that scores a query with a single forward pass. We propose the Cauchy Displacement Network (CDN), which regresses each normal point onto the mean of its $k$ nearest training neighbours with a single-hidden-layer network of Cauchy activations and scores a query by the length of its predicted displacement. This displacement is an empirical mean-shift vector, whose fixed-radius population counterpart locally approximates a scaled gradient of the log density. We show that Cauchy units exactly represent the score fields of a structured family of Student-$t$ expert densities, and establish a constructive approximation bound for the corresponding population neighbourhood-mean map using a single-hidden-layer Cauchy network. We further show that the bounded network output guarantees asymptotically linear residual-score growth along every fixed direction. On 46 benchmark datasets against 28 baselines, CDN obtains the best mean rank under both AUC-ROC and AUC-PR, while scoring $10^4$ queries in a few milliseconds on a single CPU thread, up to three orders of magnitude faster than neighbourhood-based detectors, with a model whose size does not depend on the training set.

    memorybenchmark
  135. arxiv:2610.41383 · cs.RO
    BioAgent: Self-Directed Skill Acquisition and Capability Growth for Long-Horizon Laboratory Protocols
    Gaojing Zhang, Changwen Chen, Zichen Luo, Jialiang Li +3

    Executing an unfamiliar laboratory protocol requires a robot to identify missing capabilities and determine whether each operation remains valid as the experiment progresses. This letter presents BioAgent, an embodied agent that addresses these coupled challenges of capability sufficiency and experimental validity. It matches human-reviewed requirements to validated skills and task packages, selects tool-based implementations for missing behaviors, and requests human teaching when needed during learning. Learned skills undergo sequential task-package validation to check outcomes and compatibility between successive operations. During execution, precondition, outcome, and experimental-state checks govern progression, autonomous recovery, or termination. Execution evidence guides tool selection, while accepted experience supports offline policy updates and task packaging. On the second task, BioAgent reuses 17 skill types unchanged, revises two existing implementations, and acquires five new types. On two dual-arm laboratory tasks with 20 and 31 ordered skill invocations, BioAgent achieves complete-task success rates of 66\% and 58\% with autonomous recovery over 50 runs per task. Across 25 atomic skills evaluated over 50 autonomous trials each with recovery disabled, success rates range from 84\% to 100\%, with 17 skills achieving 50/50.

    embodiedagentembodied agent
  136. arxiv:2610.41381 · cs.AI
    From prediction error to sparse events: Biologically inspired signed-dyadic coding for Mandarin keyword spotting
    Zijian Zhou, Rui Liu, Rongxiang Nie, Rui Ma +2

    In continuous spiking keyword spotting, temporally correlated acoustic frames can repeatedly trigger memory access and neural computation at a dense input interface, even when hidden activity is sparse. Predictive processing and sensory adaptation suggest retaining predictable input locally and responding selectively to deviations. Inspired by this mechanism, we propose Predictive Signed-Dyadic Coding (PG). PG maintains a local prediction for each input channel and transmits an event only when the residual exceeds a learned threshold. Each event carries a sign and a finite power-of-two magnitude, and the receiver reconstructs it with fixed-point shifts and additions. In Mandarin keyword spotting, INPUT-PG reaches 90.00% accuracy while reducing mean SRAM reads by 33.87% and inference cycles by 32.77%. PG provides a hardware interface for research on low-power continuous speech processing and event-based sensory systems.

    memory
  137. arxiv:2610.41378 · cs.AI
    Ask on Demand: Targeted Peer Communication for Parallel Agentic Systems
    Haofu Yang, Hongyang Yin, Ran Bi, Faguo Wu +3

    Parallel LLM agents tackle complex tasks by exploring multiple trajectories, yet useful discoveries remain isolated in individual execution histories. Existing sharing methods often select or summarize information before other agents' needs become clear, potentially overlooking details that become important later. We propose Ask on Demand (AOD), a communication framework that enables agents to seek targeted peer assistance when specific information needs arise. Upon an approved request, AOD retrieves relevant evidence from peer execution histories, selects a suitable peer, and generates a response tailored to the request. To determine when such assistance is worthwhile, we train a lightweight gate using paired team continuations, with supervision derived from final team outcomes and inference cost. Experiments on AppWorld and WebShop demonstrate that AOD consistently improves task performance over independent parallel execution while achieving lower inference cost than existing information-sharing baselines. These findings highlight the benefits of aligning peer communication with agents' evolving information needs to achieve more effective and efficient collaboration.

    llm agentagentic
  138. arxiv:2610.41377 · cs.RO
    MolmoSpaces-CommonSense: Evaluating Commonsense in Robotic Manipulation Policies
    Arjun Guru, Max Argus, Jae Sung Park, Jiafei Duan +2

    Real-world robotic manipulation requires reasoning about cluttered scenes, partially observable or hidden objects, and semantically appropriate actions, in addition to low-level execution within those scenes. However, existing simulation benchmarks primarily test the latter and remain dominated by comparatively clean environments with limited clutter and narrowly scoped task structure, leaving many forms of commonsense reasoning untested. We address this by introducing a taxonomy of reasoning processes for robotic manipulation grounded in developmental cognitive science. We operationalize this taxonomy through a new benchmark of robot tasks, MolmoSpaces-CommonSense, that requires reasoning about occlusion, containment, support relations, semantically correct grasp points, spatial affordances, and hidden objects. These evaluations require policies not only to recognize relevant scene structure but also to execute correct actions in them. We evaluate several zero-shot generalist manipulation policies. Across model types and architectures, we observe consistent failures in commonsense reasoning: success rates in cluttered scenes drop by over 50% as distractors increase, policies act on incorrect objects in cluttered settings 30% of the time, and affordance-sensitive grasping remains unreliable even with highly specific language prompts, with correct object part selection happening only about 50% of the time. Further, VLM-based success estimation struggles to reliably evaluate rollouts in these settings, suggesting that the challenge extends beyond control to the broader perception-reasoning-action stack. Our results suggest that existing benchmarks may substantially overestimate zero-shot real-world readiness. Code and policy leaderboard link at https://github.com/allenai/molmospaces.

    manipulationgraspbenchmarkleaderboard
  139. arxiv:2610.41375 · cs.LG
    From the Zoo to the BULLPEN: Sample-Efficient LLM Profiling via Multi-Channel Model Embeddings
    Russell Ro, Qiming Cui, Reza Abbasi-Asl

    Choosing among the hundreds of thousands of available large language models (LLMs) requires a thorough understanding of each candidate model's properties. In practice, each property is measured by its own benchmark suite, so profiling a model on many properties means paying for many separate evaluations. To address this, model embeddings compress each model into a fixed-length vector from which many properties can be predicted. Existing methods, such as EmbedLLM, build these vectors from correctness patterns and question features, and often need large probe batteries to do so. However, many behaviors that are important when choosing a model, such as its verbosity, safety and style, show up more in how it answers than in whether it is correct. We introduce BULLPEN (Behavioral Understanding of Large Language models via Profile ENcodings), a framework that extends EmbedLLM in three ways. First, BULLPEN organizes the downstream tasks that prior work uses and adds 20 task benchmarks we built on behavior, style and provenance. Second, it is the first framework to read a model's own answer text jointly with its correctness and the question text. Third, it encodes models from a few hundred questions instead of the full 24,118-question pool. Across 260 open-weight models, our answer-text encoder scores 0.48 +/- 0.10 on the new tasks against 0.35 +/- 0.10 for the best prior encoder. Our encoders that read answer text keep at least 91% of their full-pool skill from only 400 questions (1.3% of the inference compute of the full question pool), while a correctness-only baseline keeps 83% with the same 400 questions. Because these embeddings need only model outputs, they support behavioral auditing and low-cost profiling, and could extend to API-only models. Code and data: https://github.com/wcquesera/bullpen

    benchmark
  140. arxiv:2610.41370 · physics.optics
    Non-Hermitian Spectral Topology Shaped by Memory
    Hao Pan, Bo Xue, Wanyue Xiao, Qiuyang Li +5

    Non-Hermitian spectral topology has transformed our understanding of bulk--boundary correspondence in open systems, yet its exploration has largely remained within the Markovian regime of instantaneous coupling. Here, we reveal how memory shapes non-Hermitian spectral topology by extending coupling asymmetry from space to time. Spatiotemporal nonreciprocity and memory-induced spectral multivaluedness give rise to nontrivial spectral topology on multi-sheeted Riemann surfaces, governed by both spatial connectivity and interaction history. Coherent interference among long-range memory channels can induce anomalous spectral winding and skin localization opposite to the dominant-coupling direction. Moreover, asymmetric retardation alone can generate non-Hermitian spectral topology and skin accumulation even in the absence of spatial coupling asymmetry. Channel dissipation further drives topological transitions that reverse the skin effect, with a critical point at matched channel and lattice decay rates. We validate the anomalous skin effect and its dissipation-controlled reversal through full-wave simulations and microwave experiments. Our results establish time nonlocality as a fundamental degree of freedom in non-Hermitian topology, revealing phenomena inaccessible within the Markovian paradigm.

    memory
  141. arxiv:2610.41368 · cs.AI
    RECITE: Revising Eligibility Criteria Incorporating Textual Evidence
    Russell Ro, Reza Abbasi-Asl

    Overly restrictive eligibility criteria contribute to widespread enrollment failure in clinical trials, yet published recommendations for how specific criteria should be broadened remain disconnected from trial design at scale. We introduce RECITE (Revising Eligibility Criteria Incorporating Textual Evidence), a framework for automated discovery, extraction, and implementation of evidence-based eligibility criteria revisions. RECITE uses protocol amendment documents from ClinicalTrials.gov to construct a dataset of 11,913 curated instances across 5,735 trials, each pairing original criteria with documentary evidence and target revisions. We benchmark fifteen LLM configurations under truncated-to-context-window and retrieval-augmented generation conditions, on a 3,116-instance sample under an LLM-as-judge protocol validated against ten domain experts (judge vs. human quadratic-weighted Cohen's $κ_{qw} = 0.694 \pm 0.096$). The best open-weight model, Gemma 2 9B, achieves $κ_{qw}$-aligned ordinal score $3.0 \pm 0.9$ on a 0--4 scale (normalized 0.742) and the best closed-weight model, GPT-4o, reaches $3.0 \pm 0.8$ (normalized 0.745). Using an agentic literature discovery system, we then identified 44 paper--trial matches across five publications providing trial-design and EC-modification evidence. Our findings suggest that existing clinical knowledge can be operationalized into actionable trial design recommendations.

    retrieval-augmentedagenticbenchmarkllm-as-judge
  142. arxiv:2610.41360 · cs.CV
    Defining Page-Centric Visual Document Retrieval and Evaluating Existing Benchmarks
    Chongzhi Zhang, Yuqi Zuo, Aixin Sun

    In modern information retrieval (IR), documents are increasingly encountered as structured sequences of visually rich, paginated units, rather than as flat text. Example documents encountered in our daily life include Word files, PDFs, and slide decks. Free text-centric retrieval systems either ignore this structure entirely or destroy it through arbitrary passage chunking. In this paper, we formulate page-centric Visual Document Retrieval (VDR) with pages as the fundamental indexing units and document context available for assessing page relevance. The formulation distinguishes two retrieval granularities, page and document, and three retrieval paradigms: independent-item, evidence-set, and iterative retrieval. We then assess 42 benchmarks using eight criteria covering corpus characteristics, query construction, relevance annotation, and ranking support. Our analysis shows two main findings on current VDR research. First, evaluation remains concentrated on page-level independent-item retrieval, with fewer resources supporting document-level, evidence-set, and iterative settings. Second, queries originating from real user information needs independently of target corpus content and graded relevance judgments are uncommon, while the coverage and verification of relevance judgments vary substantially. Both findings reveal gaps between the capabilities applications require and those existing evaluation protocols can establish. We provide benchmark selection guidance and identify priorities for future benchmark design, emphasizing alignment between task requirements, relevance semantics, and evaluation protocols.

    benchmarkevaluation protocol
  143. arxiv:2610.41358 · cs.AI
    Do Transaction Prices Reflect Visual Preference? A Diagnostic Study with Multimodal Language Models
    Fengxian Ji, Jingpu Yang, Fan Zhang, Junhong Liang +5

    Existing visual preference evaluation relies on costly human annotations or preference reward models, whereas online marketplace transaction prices are readily available and reflect real purchasing behavior, offering potential low-cost supervision. However, transaction prices conflate visual preferences with non-visual factors such as provenance, demand, and competition, making aggregate accuracy insufficient to distinguish model errors, unreliable price signals, and protocol confounds. To address this identifiability challenge, we introduce MarketImage, a diagnostic framework that identifies when transaction prices reflect visual preferences, where they fail, and whether multimodal language models can recover human-supported visual preference signals. MarketImage comprises three components: (i) MarketImageArt, 266 six-image comparison sets with transaction prices and price-blind human rankings; (ii) MarketImageLens, a tie-preserving protocol that partitions 3,990 within-set image pairs into price-human agreement, conflict, and price-tie regions while separately auditing presentation-order effects; and (iii) MarketValuator, six elicitation methods spanning direct comparison, independent scoring, budget constraints, and multi-agent aggregation to evaluate signal recovery by static visual scorers and three multimodal language models. Experiments show that only 24.2% of image pairs have a non-tied price direction, of which 34.7% agree with the human majority. Within this region, HPSv2 leads at 66.4% Recoverable Accuracy; Qwen's budgeted and aggregated variants trail baselines, while GPT interface gains remain statistically unresolved. Thus, transaction prices encode recoverable visual preferences only under limited conditions and should be treated as uncertain proxies rather than ground truth.

    multi-agent
  144. arxiv:2610.41354 · cs.RO
    DiffWBC: Unlocking Steerable General Humanoid Motion Tracking via One-Step Diffusion Whole-Body Control
    Yifei Huang, Yazhao Yang, Weijia Li, Zhiheng Ma +1

    General humanoid motion trackers typically rely on deterministic policies that cannot adapt online without brittle reference editing or costly retraining. We present DiffWBC, shifting whole-body control to a generative paradigm via a one-step diffusion model. By modeling distributions over future trajectories, DiffWBC unifies general tracking, test-time steering, and simulation-free failure discovery within a single controller: (1) Generative general tracking: DiffWBC requires a single forward pass per 50 Hz control step and can be distilled from an RL teacher or trained from scratch via model-based RL, achieving the lowest local tracking error among six general trackers on held-out benchmarks (20.1 mm on BONES) and deploying on a physical Unitree G1. (2) Zero-shot test-time steering: differentiable cost gradients and body-group masks modulate the frozen controller online, substantially outperforming kinematic reference editing (0.84 clean pass at 1.25 m ceiling vs. 0.10) and zero-shot generalizing across hurdles, spatial reaching, actuator faults, and tray balancing. (3) Simulation-free failure discovery: plan dispersion across generation noise effectively identifies physical tracking failures (91% precision in top 2%) without physics simulations; targeted fine-tuning on flagged clips resolves 50% of failure cases. Videos and interactive demonstrations are available at https://diffwbc.github.io/#selfassess.

    humanoidwhole-body controlbenchmark
  145. arxiv:2610.41351 · cs.LG
    Quantile Propagation Planning for Budgeted Tail-Risk Evaluation
    Weijia Han, Zhi Huang, Lisha Qu, Yuanyuan Tan +1

    Risk-aware planning must identify rare, costly outcomes using only a limited number of model evaluations. Monte Carlo rollouts spend this budget on noisy tail estimates, while planning from the mean can miss the tail entirely. We introduce quantile propagation planning (QPP), which propagates a small, fixed set of disturbance quantiles through the dynamics and integrates the resulting cost curve to evaluate both mean and tail risk. The method is exact at the propagated levels, converges as the quantile grid is refined, and admits a one-sided conservatism guarantee under stated structural conditions. In controlled overload and inventory tasks, QPP reduces realized tail cost and constraint violations relative to mean-rollout planning. On overload, it remains competitive with sampled evaluators that use substantially more trajectories. These results position QPP as a transparent, budget-controlled approach to tail-risk evaluation.

    evaluator
  146. arxiv:2610.41350 · cs.LG
    Operon-S1: a fast decision controller for long-horizon biomedical agents
    Genoria AI Team, Yuejie Hou, Yueyuxiao Yang, Weiting Tang +1

    Language-model agents now carry out long scientific work in biology, but their judgements about conventions, skills, errors, evidence and stopping are made inside the reasoning that does the work, where they cannot be measured, replaced or improved on their own. Here we present Operon-S1, a fast decision system, a System 1 for the agent, that separates judging from doing. A decision harness poses the agent's recurring judgements as closed questions at 13 decision points and acts only through fixed messages that never contain an answer, so any model that answers the questions can serve as the judge. Across seven life-science benchmarks, a local decision model took over these judgements in 0.6 s at the median for an in-loop request on CPU at no request cost, with no measurable change in the agent's score on five of six benchmarks, whereas the general language model that drives the agent, asked the same questions, accounted for 8.0% to 35.2% of task cost. Every decision is recorded with its consequence. Retraining on labels from task construction, verifiers or execution raised decision accuracy on every question in each of three trained versions, and a new version enters service only through a promotion gate. Task-level results in this study use the public checkpoint. The same closed decisions serve wet-laboratory protocol troubleshooting and the design of a germline target enrichment panel. Like the regulator of an operon, which switches genes on and off without making their products, Operon-S1 steers the agent without doing its work, and makes the agent's judgements a layer that can be measured, replaced and trained.

    agentbenchmark
  147. arxiv:2610.41349 · cs.RO
    Geometry-Aware Humanoid Retargeting via Manifold ODEs
    Pengyu Zhu, Heye Wang, Xi Wang, Xuyang Wu +2

    In this paper, we focus on interaction-preserving motion retargeting, which transfers human demonstrations to humanoids while maintaining spatial relationships with objects and terrain. Previous work, such as OmniRetarget, addresses this task through temporally coupled constrained optimization, requiring repeated sequential quadratic programming (SQP) iterations at each frame. We study this problem from an ordinary differential equation (ODE) perspective by modeling the continuous-time limit of its framewise Karush-Kuhn-Tucker (KKT) conditions as an ODE on the robot configuration manifold. This viewpoint offers two benefits: (i) it connects OmniRetarget's SQP iterations to a backward discretization of the ODE, enabling us to establish their local convergence under suitable regularity assumptions; and (ii) it leads to TaSP-KKT (tangent-space parametrization for KKT flow), a new retargeting method based on a forward discretization of the KKT flow. The proposed method achieves the same first-order convergence to the ODE solution as the corresponding backward scheme while requiring only one weighted tangent-cone projection and a closed-form manifold update per step, thereby avoiding repeated inner SQP iterations. Numerical experiments show that TaSP-KKT achieves retargeting quality comparable to, though slightly lower than, that of OmniRetarget while running approximately twice as fast.

    humanoid
  148. arxiv:2610.41348 · cs.RO
    Joint-Health-Aware Reinforcement Learning for Robust Robot Control Under Actuator Degradation
    Haifa Zhang, Yijing Wang, Zhiqiang Zuo

    Robots deployed in unstructured environments are susceptible to actuator-authority degradation, which reduces the effective control authority of individual actuators and can cause reinforcement learning policies trained under nominal dynamics to fail. Existing robustness approaches rely on discrete fault labels, broad domain randomization, or global latent variables, often without explicitly modeling continuous per-joint residual actuation capability. We propose a joint-health-aware reinforcement learning framework for robust robot control under simulated actuator degradation. Our framework encodes actuator condition in a continuous joint-health vector and models degradation as multiplicative attenuation of executed actions. A dedicated HealthNet estimator infers per-joint health from recent observations and actions, allowing the policy to redistribute control effort across healthy and degraded joints. We evaluate the method on Walker, Go2, Cassie, and Panda-based ManiSkill manipulation tasks under diverse degradation masks and severity levels. The method outperforms nominal and degradation-trained no-health baselines on the main evaluated metrics. Relative to the nominal baseline, task-level locomotion returns improve by up to 250.1\% and manipulation success rates by up to 15.2 percentage points; in the high-degradation Cassie group, return increases from 3.94 to 40.50 (+928.0\%). Ablations support the contribution of per-joint health estimation, action-degradation modeling, and health-conditioned control.

    manipulation
  149. arxiv:2610.41337 · cs.CV
    Bayesian PET Reconstruction with Learned Flow Matching Priors via Langevin Sampling
    Hengjia Ran, Huafeng Liu, Bo Zhao

    Image reconstruction for positron emission tomography (PET) is a challenging inverse problem due to its inherent ill-posedness and Poisson counting noise, particularly in low-dose acquisitions. In this work, we introduce a novel Bayesian PET reconstruction framework that leverages a flow-matching generative prior learned from high-quality PET images. By incorporating this prior, we formulate PET reconstruction as a Bayesian inference problem and solve it via latent-space posterior sampling using the unadjusted Langevin algorithm. We evaluate the proposed method through numerical experiments and benchmark it against classical reconstruction methods and state-of-the-art deep learning-based baselines. The proposed method achieves improved reconstruction performance while enabling posterior uncertainty quantification.

    benchmark
  150. arxiv:2610.41336 · cs.RO
    HandMimicX: Learning Human-Like Dexterous Manipulation via an Object-Centric Policy from Egocentric Videos
    Zihan Gao, Yufei Liu, Yuhong Zhang, Zhen Yu +6

    Learning dexterous manipulation from human videos remains an open challenge, particularly in the context of transferring a single demonstration into a spatially generalizable robotic skill and maintaining physically plausible human-like interactions. Unlike prior works that emphasize hand-motion imitation, we introduce an object-centric framework for acquiring dexterous skills from egocentric human RGB-D videos. Within our framework, we formulate the reinforcement learning (RL) objective in terms of object motion to decouple the underlying skill from the specific hand trajectory and impose constraints on hand-object interaction states to promote more anatomically natural hand configurations. Extensive experiments in simulation and on real robots demonstrate that our method enables efficient learning from human videos and generalizable skill transfer across novel spatial configurations.

    manipulationdexterous
  151. arxiv:2610.41334 · cs.RO
    HIPC: Humanoid Iterative Planning and Control via Goal-Conditioned Motion Closure
    Yi Han, Xingyu Chen, Zixuan Lang, Qianyang Wu +9

    Recent generative models are widely used to provide whole-body motion for humanoid tracking. Current methods combine motion generators with controllers in an open-loop paradigm, which leads to critical issues: tracking errors can accumulate into spatial drift, and the planner cannot adapt to unexpected physical disturbances without feedback from physical execution. However, directly feeding realized states back to the planner in a closed loop can cause discontinuities, resulting in physical jitter and even falls, since such states are out of distribution during training. To bridge this gap, we propose HIPC, a real-time Humanoid Iterative Planning and Control framework that couples a generative model with motion tracking into a unified planning-and-control paradigm. HIPC introduces a goal-conditioned motion planning trained with a canonical motion representation that avoids gimbal-lock singularities under extreme orientations and supports motion generation on diverse geometric and semantic conditions. Then, HIPC adopts two strategies, deviation-decayed recovery and history re-anchoring, to mitigate the out-of-distribution shift caused by realized-state feedback. Finally, HIPC operates a real-time asynchronous planning-and-control system, with latency compensation to reduce temporal mismatch for inference delay. Extensive experiments confirm the effectiveness of HIPC under switching and composition of diverse conditions. Results show that HIPC achieves state-of-the-art goal-reaching accuracy among existing methods while simultaneously demonstrating robustness to physical perturbations and the ability to recover from falls.

    humanoid
  152. arxiv:2610.41333 · cs.LG
    StrictPRMBench: A Replayable Evaluation Benchmark for Auditing Fixed-Score PRM Aggregation
    Pengfei Zheng

    Process reward models (PRMs) score reasoning steps, but best-of-N selection also requires a rule that aggregates those scores. We study that rule as a separate evaluation object. StrictPRMBench fixes mathematical reasoning pools, scorer evidence, and the source trace of a deletion intervention, so that different reducers face the same recorded selection problem. Across seven scoring checkpoints and two benchmarks, the stored N=32 outcomes give higher deletion-induced selection instability for product than geometric mean (GM) in all 14 cells. The equal-cell difference is +0.0965; a shared-problem bootstrap, conditional on these checkpoints and frozen pools, gives a 95% interval of [0.0846,0.1086]. A pre-injection-defined sensitivity excludes replacement slots that are either reducer's clean winner and retains 4,459 of 4,622 cell-problem records. The retained difference is +0.0986 [0.0875,0.1098]. Effects are heterogeneous: the two Qwen cells have positive percentile-bootstrap intervals on retained support but two-sided exact paired p-values of 0.0625 and 0.0703. Clean-pool Qwen accuracy does not detect a product advantage, but this is not an equivalence result. We provide versioned outcome records, CPU reanalysis, and an explicit evidence-availability contract. The contribution is a diagnostic evaluation protocol and a reproducible audit of recorded selector behavior, not a new PRM, a universally optimal reducer, or evidence of deployment-wide harm. Raw-score replay and outcome-record reanalysis are distinguished: missing primary edited-score vectors limit independent evaluation of new reducers under the same stress.

    benchmarkevaluation protocol
  153. arxiv:2610.41330 · cs.RO
    Cross-Embodiment Motion Planning by Coupling Body-Agnostic World Model with Differentiable Ego-Dynamics
    Zhidong Wang, Zirui Li, Haoran Xu, Yujie Yan +4

    Motion planning in clutter is embodiment-specific: the feasible free space and reachable commands depend on each robot's geometry and dynamics. Existing cross-embodiment policies output normalized commands that disregard these differences, or compensate through per-robot training and distillation. World models can in principle pool cross-embodiment data, yet command-conditioned models bind their latent dynamics to a single action space, while pose-conditioned models omit the mapping from commands to body-specific motion. On flat ground, however, embodiment affects the planner's observations only through the camera ego-motion. This letter proposes Hourglass, an hourglass-shaped architecture that factorizes cross-embodiment motion planning at this interface: a body-agnostic world model and each robot's differentiable ego-dynamics, which map commands to that ego-motion, are coupled only through it. The world model is trained on data pooled across embodiments and camera mounts; each robot's ego-dynamics is a reduced-order template with identified physical parameters. A generalist parameter-conditioned policy is learned once at this junction: each imagined step rolls out the template and then the world model, and imagination training through each robot's template makes the policy body-aware. On seven simulated robots of five locomotion types, the generalist policy outperforms imitation-learning and diffusion-policy baselines on the same world model by 3.8--10.6 percentage points in success rate and matches per-robot specialists without dedicated training, while approaching its parameter-conditioned MPPI teacher within 1.4--2.8 points without an online map. After three minutes of teleoperated driving for identification, the synthetic-data-trained model transfers zero-shot to a real skid-steer robot unseen in training, navigating unseen indoor scenes.

    world modellatent dynamics
  154. arxiv:2610.41320 · cs.RO
    Learning from Physical Experience:Context-Space Adaptation for Robotic Manipulation
    Yihan Zeng, Yisu Li, Kaixin Huang, Marc WuDunn +8

    Recent multimodal foundation models exhibit strong perception, reasoning, and in-context learning capabilities, but translating these capabilities into physical action remains challenging. Through real-world experiments across diverse manipulation tasks, we find that ChatGPT-6-Astra, a state-of-the-art general-purpose reasoner, can serve as a physically actionable, context-responsive manipulation policy, demonstrating effective fine manipulation and spatial reasoning. Building on these findings, we introduce LEXP, a harness for learning from physical experience through context-space adaptation. LEXP combines task-level planning with an experience module that interprets execution evidence, extracts and retains lessons with applicability conditions, and reuses relevant lessons to generate contextual feedback for subsequent actions. Model weights and the action interface remain fixed throughout. Across 12 real-world manipulation tasks, LEXP achieves $23/36$ successful trials (63.9\%), compared with $19/36$ (52.8\%) with memory alone and $2/36$ (5.6\%) for vanilla Astra. On three tasks, bread transfer, sort trash and toiletries, and retrieve drawer, removing planning reduces success from $7/9$ to $4/9$, while removing retained experience reduces it to $3/9$. Execution traces illustrate feedback-guided corrections and cross-episode reuse of an inspection strategy. Execution analyses illustrate how task structure and experience-derived feedback guide task progression, action correction, and experience reuse. These results support context-space adaptation as a practical mechanism for turning physical interaction into reusable guidance for robotic manipulation.

    manipulationmemory
  155. arxiv:2610.41308 · cs.AI
    SERIAL-7: Deriving, Realizing, and Preserving Machine-Verifiable Outputs Across Representations
    Youngkyung Seo, MiHyeon Kim, Jisoo Lee, Hwijung Ryu

    Large language models increasingly generate structured outputs consumed by programs and downstream systems, where semantic correctness alone does not guarantee reliable execution. We introduce SERIAL-7, a Korean-English benchmark built around an executable verification framework for machine-verifiable generation across seven representations. Using 177 validated canonical artifacts, SERIAL-7 separates semantic derivation (Derive), constrained artifact generation (Realize), and cross-representation preservation (Preserve). Its 43 atomic representation constraints are paired with executable checks over format, structure, and task-specific semantics, enabling progressively stricter verification from representation compliance to full artifact validity. Across ten models, REALIZE-L3 HSR ranges from 42.0% to 93.9%, while Full Compliance exposes additional artifact-level failures beyond correct final answers and valid representations. These results show that reliable structured generation requires explicit, executable verification of how artifacts are realized and preserved across representations.

    benchmark
  156. arxiv:2610.41307 · cs.AI
    Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation
    Shuo Hong, Jiaxin Ma, Xidi Cai, Qianli Ma

    Multimodal indirect prompt injection embeds malicious instructions in external media that also contains information needed for the user's task. Existing defenses can block the attack yet impair the model's use of that information, leaving a gap between attack defense and task preservation. We propose ERASE, an on-policy self-distillation method that constructs its reference by erasing the injected region in place. Modality-appropriate blank content replaces the injection while the surrounding media is preserved. A student reading attacked media learns from a frozen base model reading the erased view through full-distribution token-level feedback on its own rollouts. Across image, audio, and video on two model sizes, ERASE records near-zero observed attack success on the evaluated benchmark while retaining task utility close to the undefended clean-input level. On 3B audio, ERASE reaches 60.1% attacked-input accuracy compared with ARGUS's 33.1%, within 0.5 points of the undefended clean-input level. Experiments show that ERASE reduces injection dependence from 6.26 to 0.0095 on 3B audio while staying close to the base model on the erased input. Out-of-domain accuracy also stays near the base level. Our code will be made publicly available upon acceptance.

    benchmark
  157. arxiv:2610.41304 · cs.CV
    ART of PIV: Agentic Real-Time Particle Image Velocimetry for Active Flow Control
    Antonio Terpin, Francesco Banelli, Alan Bonomi, Raffaello D'Andrea

    Particle image velocimetry (PIV) measures fluid motion from images of tracer particles, but its accuracy depends on choices throughout the estimation pipeline, including image preprocessing, flow estimation, and postprocessing. These choices are typically left to expert tuning, and a fixed configuration can become less accurate as conditions change. This matters especially when PIV supplies feedback for real-time control, where measurements must be both accurate and fast, and flow conditions vary. We introduce ART of PIV -- short for Agentic Real-Time and also a nod to the art of tuning a PIV pipeline -- which automates pipeline choice for each image pair. A learned agent selects a configuration according to a chosen measurement objective from a library assembled for a specified speed budget. Across cross-correlation, classical optical-flow, and learning-based estimators, ART of PIV reduces average end-point error (AEPE) by 12.2-22.4 % relative to the best single estimator at matched speed, with gains extending to 1000-4000 image pair/s. In a simulated local velocity regulation task, it reduces mean normalized closed-loop cost by 13.24-13.59 percentage points relative to a tuned fixed estimator, with the same classical controller in both cases. In a tabletop water channel, we illustrate two ways to tune ART of PIV for a fluidic pinball of three rotating cylinders: selecting between complementary optical-flow tunings for accuracy, and selecting regions of interest with coherent rotation or shear using a common flow estimator. For the first variant, we train a DreamerV3 reinforcement learning controller directly in the physical world and demonstrate tracking of a prescribed flow field in the wake of a fluidic pinball.

    dreamerv3agentagentic
  158. arxiv:2610.41301 · cs.AI
    Subliminal Reinforcement Learning: traits travel through a judge's scores
    Keuntae Kim, Yong Suk Choi

    Subliminal learning, the transfer of a model's trait through data with no semantic trace of it, has been shown for students trained on a teacher's text. Post-training relies on group relative policy optimization with a language-model judge, where the student learns only from its own outputs and one judge's scores. We show this suffices: a student writes invented syllables or numbers, a judge made to prefer one animal scores them, and the student's preference moves toward that animal relative to a run whose judge prefers another. Of 48 design cells, two fail the validity gate and 17 establish the trait at 5% false-discovery rate, numbers-only cells included; elsewhere, our account gives the reason. Removing only the reward difference between judges removes transfer. We establish the principle as a chain of four stages: the trait must show in the judge's scores, favour strings that move the student, accumulate and reach the answers. Judges with the same preference built from other data, or on another model, place it on other strings, and transfer is not detected. Carried unchanged to AI-safety dispositions, the chain holds: undesirable dispositions reach the student's answers, and stages measured before training anticipate where, helping to prevent part of this risk.

    post-training
  159. arxiv:2610.41296 · cs.AI
    Self-Improvement with Auxiliary Rewards: Learning from Verified Traces and Evaluating Transfer
    Minju Gwak, Alan Ritter

    Answer-level rewards cannot distinguish reasoning traces with equally correct outcomes. We study whether auxiliary trajectory preferences improve small-language-model self-improvement and transfer under limited supervision. We compare confidence-based (SC) and entropy-based metrics (UID-Conj.), which combines global entropy variation and local smoothness, through online reward augmentation and offline training on correct responses. Online experiments adapt Qwen3-1.7B on 30 AIME problems without reference answers; offline experiments use historical AIME problems. UID-Conj. improves the online macro-average across the adaptation benchmark and three transfer benchmarks by 2.08 percentage points over unmodified TTRL. The strongest evidence comes from offline weighted SFT: UID-Conj. exceeds uniform weighting on the same response pool by 1.93 points, with a paired prompt-bootstrap 95\% CI of [0.44, 3.44]. Benefits are task-dependent: SC rewards judged trace quality and instruction following, while UID-Conj. rewards yield stronger training outcomes in several settings. We find that auxiliary preferences can expose learning value among equally correct traces not captured during inference time, motivating their evaluation through downstream gains.

    self-improvementbenchmark
  160. arxiv:2610.41292 · cs.CL
    From Transcription to Reformulation: Defining and Benchmarking Grounded Generative Speech Recognition
    Qing Wang, Qihan Hu, Hang Lv, Yan Yang +8

    Speech language models have expanded automatic speech recognition (ASR) from transcription to general speech-to-text generation. Yet existing benchmarks still assume a single ground-truth transcript per utterance, an assumption that struggles when models produce readable, formatted, or structured outputs. In this work, we introduce Grounded Generative Speech Recognition (GSR), reformulating speech recognition as generating semantically grounded text instead of a single transcript. We define a grounded reformulation space that characterizes all valid textual realizations supported by the input speech, and establish a four-level hierarchy of reformulation capabilities, including literal transcription, canonical reformulation, semantic reformulation and speech-grounded structuring. Based on this formulation, we present GSR-Bench, the first comprehensive benchmark designed to evaluate generative speech recognition systems beyond exact transcript matching. GSR-Bench covers diverse scenarios including refined meeting transcription, cross-dialect translation, context-aware speech understanding, and robustness under silent or noisy acoustic conditions. To address the limitations of word-level evaluation metrics, we develop a multi-dimensional evaluation framework combining speech grounding assessment and LLM-based semantic judgment, whose reliability is validated through human preference studies. Extensive experiments across representative architectures, ranging from conventional ASR systems to end-to-end Speech-LLMs, reveal fundamental trade-offs between transcription fidelity and generative reformulation ability. Our analysis further identifies critical failure modes, including acoustic hallucination, unsupported generation, and instruction drift, providing insights into the design and evaluation of future grounded generative speech systems.

    benchmarkevaluation framework
  161. arxiv:2610.41291 · cs.LG
    In-Model Context Compression with Agentic Soft Tokens for Long-Horizon LLM Agents
    Xuning He, Kai Fan, Runyu Zhang, Binghuan Wang +3

    Long-horizon LLM agents play an important role in solving complex and sustained tasks, with ever-growing interaction contexts. However, these accumulated contexts will raise substantial computational cost and obscure key decision-relevant evidence. Existing methods typically train external compressors or context-management agents to rewrite context as shorter hard-token sequences, but such external selection may not align with the LLM's evolving representation and is necessarily lossy. In this paper, we introduce Agentic Soft-Token Compression (ASTC), a training-free compression method using soft tokens within LLM agent for long-horizon tasks. Specifically, ASTC uses the model's native recurrent dynamics to determine the already absorbed information, leveraging the commonly-used hybrid Gated Delta Network (GDN)-attention. It then carries this information forward as agentic soft tokens while retaining exact values as hard tokens. The resulting compact sequence reduces later-layer computation and yields reusable KV caches, extending the savings across the agent's trajectory. Extensive experiments on long-horizon tasks show that ASTC matches or exceeds the uncompressed agent's accuracy with up to 50.68% fewer end-to-end Attention Pairs, while trained compressors require 15.4 to 103.7 times more Attention Pairs per successful task. These findings highlight agentic soft tokens as a model-native way to carry absorbed context forward as the task representation evolves.

    context compressionagentllm agentagentic
  162. arxiv:2610.41290 · cs.LG
    SAT-ACT: Learning CDCL Branching from Outcome-Derived Action Preferences
    Zhenchao Sun, Chongyang Tao, Ping Lu, Tongkai Lu +2

    Learning-based approaches have increasingly been explored for SAT solving, with particular attention to improving branching decisions in conflict-driven clause learning (CDCL) solvers. These approaches derive branching guidance either from static structural properties of SAT formulas or from dynamic signals obtained through solver feedback and execution traces. However, these forms of supervision provide limited evidence about state-dependent branching effects, while different outcome measures may favor different branching choices. To address these limitations, we propose SAT-ACT, an outcome-derived learning framework for CDCL branching. We design three complementary outcome comparison criteria to extract action preferences from independent branching action rollouts initiated at the same solver state. The criteria identify preferences supported by the outcomes and leave comparisons unresolved when the evidence is insufficient or conflicting, instead of forcing an ordering using a predefined scalar metric. A state-conditioned network learns branching priorities from the extracted preferences and uses them to guide CDCL search directly without online action rollouts. We further introduce non-dominated heuristic action supervision, which retains the solver-selected action as a training target only when it remains non-dominated under these preferences. Solver-level experiments show that SAT-ACT consistently reduces CDCL search costs on three SAT families and improves PAR-2 on the SAT Competition benchmarks.

    benchmark
  163. arxiv:2610.41275 · cs.RO
    TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking
    Mingyang Lyu, Gongce Cao, Yiyang Jia, Lu Jia +6

    Embodied visual tracking requires a robot to follow a language-specified person under occlusion, look-alike distractors, and ambiguous instructions. We show that improving tracking persistence can simultaneously increase safety violations. Using a compact 0.6B vision-language-action (VLA) tracker with a frozen backbone, we update only the planner head with proximal policy optimization (PPO). PPO improves success and following rates across all three EVT-Bench settings, but increases the collision rate on distracted tracking because episodes that previously ended in target loss can instead end in contact. We introduce TRACE, an execution-time target-state risk gate that preserves the learned policy and intervenes selectively. A lightweight risk head processes 84 cues, including candidate boxes, identity margins, a learned range-and-bearing probe, and the policy's commands. When the calibrated risk exceeds a threshold, TRACE rescales the translational component of the next waypoint while leaving yaw unchanged. The executed action is identical to PPO's on more than 98% of steps. In paired evaluations on identical episodes, TRACE reduces the distracted-tracking collision rate by 3.96 percentage points, while further improving success and following rates. Matched slowdown controls without target-state timing do not reproduce this improvement. Concurrent evaluation reproduces a collision reduction of 3.77 points, compared with 0.29 points for identical-arm null pairs, indicating that the effect exceeds evaluation variability. These results show that target-state-aware execution can recover the safety cost of policy post-training with minimal intervention.

    vision-language-actionembodiedpost-training
  164. arxiv:2610.41273 · cs.LG
    SPT-Flow: One Integration, Many Futures in Continuous-Time World Models
    Mingyang Lyu, Yinqian Sun, Gongce Cao, Yuhao Zhang +4

    Continuous-time world models generate a latent trajectory through a single ODE integration, but adding stochastic conditioning does not ensure that the model uses it. In our teacher-forced baseline, posterior means vary by only about 1% of the prior scale, and matched and swapped latents yield indistinguishable reconstruction PSNR at the reported precision. We study a supervision-shortcut hypothesis: ground-truth intermediate states can reduce the need for trajectory-associated latent information. We introduce SPT-Flow, which samples one latent per trajectory and matches a free-running ODE rollout to demonstration states at multiple readout times. Re-anchoring experiments support the proposed mechanism. On LIBERO, with both models trained for 30k optimizer steps, SPT-Flow improves best-of-10 PSNR over our retrained ODEWorld baseline by 2.0 and 2.4 dB at the mid-trajectory and final frames and reduces mean LPIPS by 23%, with comparable per-sample inference cost. The rollout control recovers most of the accuracy gain; learned latent conditioning adds diversity and best-of-K gains while reducing mean sample accuracy relative to this control. The diagnostics support trajectory-associated latent utilization; semantic mode coverage under matched conditions in the robotic setting remains unverified.

    liberoworld model
  165. arxiv:2610.41272 · cs.AI
    ANGLER: Verifier-Grounded Benchmarking with an Adaptive Attacker for Web Agents
    Yini Huang, Aiwen LU, Zhen Sun, Yi Han +2

    Web agents are increasingly deployed to complete tasks autonomously on real websites, where adversarially injected content can induce them to perform unauthorized actions. Existing evaluations primarily use fixed attacks or focus on a single attack type, and evaluate attacks at the text level rather than through real state changes, thus underestimating how vulnerable agents are to diverse attacks. To address the gap, we introduce REEF, a benchmark of red-teaming executable environments with faithful verification for measuring the security vulnerabilities of web agents. REEF covers 9 verifier-grounded websites spanning travel, commerce, productivity, media, and communication, and pairs every authentic benign task with a malicious objective delivered through 5 attack types at both the user and the environment level. REEF's state-based verifiers provide the execution feedback that drives ANGLER, an adaptive near-miss-guided lure evolution framework for red-teaming. Rather than evaluating a fixed attack set, ANGLER uses outcomes from prior rollouts to guide LLM-based attack mutation and candidate selection, prioritizing near-miss seeds during refinement. Across 8 VLM-based web agents, ANGLER achieves an end-to-end attack success rate (ASR) of 20.36%, compared with 15.24% for the non-adaptive setting. Compared with attack-free execution on the same benign tasks, attacks generated by ANGLER reduce benign task success rate (BSR) from 61.9% to 44.4%.

    benchmark
  166. arxiv:2610.41270 · physics.optics
    Parametric Enhancement of an Optical Spring in a Dark Signal-Recycling Cavity
    Kaido Suzuki, Ryo Iden, Ken-ichi Harada, Kentaro Somiya

    Optical springs tailor the response of precision sensors. Parametric amplification of the signal field stiffens the spring, extending the precision to higher frequencies. A Michelson interferometer operated at a dark fringe is the most direct way to separate the signal from the carrier for amplification. We experimentally demonstrate such quantum manipulation of the optical spring by placing an optical parametric amplifier in a cavity at the dark port, showing a significant increase of the optical-spring constant with the parametric gain within a factor of four below the oscillation threshold.

    manipulation
  167. arxiv:2610.41269 · cs.LG
    Coarse-Graining Optimal Transport for Rollout-Free Post-Training of Flow Matching
    Soon Hoe Lim, Aditi Gupta, Michael W. Mahoney, N. Benjamin Erichson

    Flow matching depends critically on the coupling used to pair source and target samples. Optimal transport (OT) couplings can reduce velocity conflict and straighten trajectories; but, in practice, challenges remain: the source comes from a continuous distribution, only a finite number of target samples data are accessible, and the neural network has finite capacity. For these reasons, making the coupling more OT-like may not consistently improve sampling quality, in particular in the low number-of-function-evaluation (NFE) regime. Motivated by this, we introduce ReCouple, a controllable framework that constructs coarse-grained OT couplings, to better understand rollout-free OT-based post-training. We also consider extensions of this basic framework: ReCouple+ extends this toward the continuous-source setting by increasing source resolution, converging to semi-discrete OT in the appropriate limit; and ReCouple++ further adds spatial regularization, motivated by the observation that straight trajectories need not correspond to spatially simple velocity fields. Empirical validation demonstrates the effectiveness of these methods, and it reveals an NFE-dependent interplay between coupling optimality, neural realizability, and generation quality.

    post-training
  168. arxiv:2610.41267 · cs.AI
    From 1D Benchmarks to Complex 3D Geometries: Can Agents Reliably Specify Problems for Automated PDE Simulation?
    Akshat Jani, Sakhinana Sagar Srinivas, Venkataramana Runkana

    Turning an engineering request into a simulation requires consistent geometry, physics, boundary conditions, and numerical settings. Recent agent-assisted systems automate simulation setup, but successful execution alone does not establish whether the specification preserves the intended problem or correctly handles missing information. We study whether agents can reliably construct complete executable JSON specifications for automated geometry creation, meshing, PDE solving, and result export. Our workflow combines model extraction with deterministic scientific and capability checks, explicit clarification, and validated local patches that preserve unrelated accepted fields. We evaluate specification correctness, execution completion, and numerical verification separately. On 120 template-related requests, the full workflow achieves 98.3% correct first-turn specification and decision outcomes, compared with 92.5% for schema-assisted single shot and 82.5% for model-only: gains of 5.8 and 15.8 percentage points, respectively. It makes no false-ready decisions among the 42 requests requiring clarification or rejection. Component ablations identify contributions from scientific checks and conditional repair. Separate downstream studies cover 1D, 2D, and 3D steady and unsteady, linear and nonlinear problems, including flow and elasticity with complex three-dimensional wing geometries. A 30-case gallery demonstrates automated execution from retained JSON specifications, while 9 of 12 selected live user-to-results tests achieve joint specification, execution, export, and numerical-reference success (75.0%). These results support traceable, engineer-reviewed simulation assistance within the tested scope, while fresh-response regressions identify consistency as a priority for further improvement.

    benchmark
  169. arxiv:2610.41264 · cs.AI
    Rome Wasn't Built in One Turn: Replay-Grounded Reward Signals for Long-Horizon Freeciv Agents
    Yanchen Liu, Xinyu Che, Brihi Joshi, Renzo Fukuda +5

    Training language-model agents to play long-horizon strategy games is hard because deterministic feedback is sparse and often arrives only at the end of a game. In-game scores can offer dense signals, but they do not reliably predict future outcomes: in 8-68% of position pairs across eight strategy games, the higher-scoring position leads to the worse outcome. We term this phenomenon scoreboard traps. In Freeciv, an open-source turn-based empire-building game where several players manage units, cities, and diplomacy over hundreds of turns, we look into better reward signals for gameplay agents. To facilitate this, we build CivHarness, an open system that saves current game states, branches on different decisions, and replays the game to future turns to provide measurable long-horizon feedback from saved positions. We use CivHarness to train CivTelescope, an LLM Judge that predicts which one of two game positions will yield the better outcome. On scoreboard-trap pairs, CivTelescope is 1.3-2.0 times more accurate than off-the-shelf language models, and it generalizes to new game starts, later turns, and unseen rules. We then train a Freeciv gameplay agent with CivTelescope as a dense reward, and call this recipe CivMarshmallow. When used in the middle stage of the game where scoreboard traps are common, CivMarshmallow reaches an average training phase gain of 8.94 points, 77% more than the best baseline, the end-of-game reward. Over the full 120-turn game, CivMarshmallow uses a hybrid reward that starts from the scoreboard or the end-of-game reward and shifts to CivTelescope as the game advances; this hybrid reward scores highest among the rewards we compare.

    agent
  170. arxiv:2610.41260 · cs.CV
    SketchDirector: Complex Scene Sketching via Agentic Layout Reasoning and Detail Refinement
    Yiqi Liang, Xiaodong Cun

    Existing sketch generation methods often rely on image guidance and are largely limited to individual objects or simplified scenes, leaving text-guided complex scene sketching underexplored. We present SketchDirector, an agentic framework that mimics the human sketching workflow for automatic multi-concept scene sketching. Given a scene-level text prompt, SketchDirector first employs a large language model~(LLM) to decompose the scene into multiple draft assets, where each is represented by a set of Bézier curves. Then, an agentic reasoning loop iteratively plans and executes their spatial arrangement, guided by self-evolving memory. Finally, a detail refinement module leverages diffusion priors to enhance sketch quality while preserving geometric structure. To support future research and applications in multi-concept scene sketch generation, we further introduce the \textit{MCSketch8K} dataset, comprising editable vector assets, compositional layouts, and object-level manipulation annotations. Extensive experiments show that SketchDirector outperforms state-of-the-art baselines in semantic and compositional quality while effectively preserving overall contour organization during refinement. The project page is: \href{https://yiki77.github.io/SketchDirector/}{https://yiki77.github.io/SketchDirector/}.

    manipulationagenticself-evolving
  171. arxiv:2610.41257 · cs.CV
    ReVEAL: Eliciting Role-Aware Latent Visual Evidence from MLLMs for Reasoning Video Object Segmentation
    Zhongyuan Bao, Cunjie He, Lejun Zhang, Zenglin Xu +1

    Reasoning video object segmentation (ReVOS) requires segmenting the target specified by an implicit query throughout a video. Existing MLLM-based methods typically require costly cross-model alignment, tying MLLM training to a downstream segmentation model. We present ReVEAL, a decoupled thinking-with-video framework for ReVOS that elicits role-aware latent visual evidence from MLLM reasoning to directly guide segmentation. ReVEAL introduces Visual Evidence Tokens (VETs) into video reasoning to associate necessary visual regions or entities with target, background, or distractor roles, while a Dynamic Attention Layer Router adaptively aggregates VET-to-video attention across layers into role-specific grounding evidence. We train the MLLM to perform VET-augmented reasoning on our constructed ReVEAL-47K dataset, with a Contrastive Attention Grounding Loss as auxiliary supervision to apply role-specific constraints to VET attention and optimize layer routing for reliable latent grounding. At inference, ReVEAL identifies the target through this thinking-with-video process and fuses the resulting role-specific evidence to guide a frozen SAM2 in predicting target masks. Experiments on five benchmarks show that ReVEAL outperforms alignment-free baselines and remains competitive with alignment-based methods.

    benchmark
  172. arxiv:2610.41254 · cs.CV
    Hawkeye: Reading Illegible Text with Vision Language Models
    Brian Tang, Kang G. Shin

    While modern Vision-Language Models (VLMs) have been remarkably successful in general scene text recognition, they consistently struggle with the long-tail distribution of text that is degraded by motion blur or low resolution. These limitations are particularly apparent in egocentric applications like smart glasses, where text is often captured in noisy environments. In this paper, we introduce the \textbf{BSTRB} (Blurry Scene Text Recognition Benchmark), a novel benchmark of 8,398 noisy image samples with high-fidelity annotations derived from 1,699 real-world video sequences. To support this, we present \textbf{Hawkeye}, a VLM fine-tuned on BSTRB to read and reason about illegible text by leveraging contextual priors and pixel-level alignment. Our results show that \name matches much larger VLMs on BSTRB overall and exceeds them on text at or below 20 pixels, in a single pass suited to real-time deployment. We characterize the resolution at which current VLMs begin to fail.

    benchmark
  173. arxiv:2610.41245 · cs.AI
    MEMIG: Self-evolving Memory Construction for Antibody Optimization
    Yunhui Jang, Zach Wang, Jai Doshi, Kangway Chuang +2

    Memory has become central to self-evolving LLM agents, which store and retrieve experience from past rollouts to guide future decisions. Existing systems represent this experience as free-form natural language and maintain it by rewriting, which fits tasks whose experience is narratable. However, in scientific discovery and optimization, experience is often a quantity measured by an oracle under a specific context. Rewriting it as text leaves no index for where a record should be, replaces a measurement with a description, and cannot express how interventions interact. To address this, we argue that agent memory for biomolecule optimization should instead be indexable, quantitative, and structured, and we introduce MEMIG. This memory construction framework realizes these three properties. In detail, MEMIG addresses each insight with a domain-grounded identifier and accumulates the quantitative effect measured by a property oracle, so that updates remain local and retrieval becomes an exact lookup. Next, MEMIG connects insights whose interventions are applied together and measures their epistasis by decomposing each evaluated set. Our results show that MEMIG improves performance in most settings across in silico antibody optimization tasks.

    memoryagent memoryagentllm agentself-evolving
  174. arxiv:2610.41239 · cs.MA
    Understanding LLM-Based Bargaining of Driving Agents in Mixed-Autonomy Traffic
    Lingxiao Zhou, Di Chen, Jia Li

    Large Language Models (LLMs) offer a flexible computational approach for modeling social interactions among heterogeneous agents. This paper develops a novel framework towards characterizing and understanding LLM-based population-level bargaining of shared road space between autonomous vehicle (AV) and human-driven vehicle (HV) agents in mixed-autonomy traffic. Through controlled experiments varying traffic density, agent objectives, and traffic-domain knowledge, we first characterize bargaining outcomes and observe heterogeneous agreement rates across different experimental conditions. To understand how the bargaining process shapes the outcomes, we introduce a behavioral taxonomy that jointly characterizes numerical bidding actions and communicative attitudes. We examine two complementary aspects of social interaction: inter-agent bargaining dynamics and private-to-public behavioral transformation. Results reveal distinct interaction patterns associated with agreement and breakdown, including greater constructive reciprocity in agreement trials and more frequent conflict escalation in breakdown trials. Furthermore, private and public communicative attitudes often differ even when numerical bids remain unchanged. These findings illustrate how LLM-based bargaining simulations can support a structured analysis of social interaction behaviors in mixed-autonomy traffic.

    agent
  175. arxiv:2610.41236 · cs.AI
    Experts Exhibit Structured Redundancy: Structure-Preserving Expert Pruning for Mixture-of-Experts Large Language Models
    Tian Wang, Yongding Tao, Yihong Dong, Xiaoming Wu +2

    Sparse MoE mechanism significantly improves the trade-off between model capacity and inference efficiency. However, practical deployment still requires storing all experts, resulting in substantial memory overhead. We show that expert outputs in MoE layers exhibit structured redundancy. Across different models and task categories, the stable rank and entropy effective rank of expert outputs are consistently far below the nominal number of experts, while expert output distributions also exhibit fewer effective modes. Motivated by these observations, we reformulate expert pruning as a structure-preserving expert subset selection problem and efficiently optimize it via a greedy strategy, rather than only independently scoring and ranking individual experts or exhaustively searching over expert combinations. We propose two expert pruning methods under the unified structure-preserving framework: Projection Energy Expert Pruning (PEEP), which preserves the functional subspace by maximizing weighted projection energy, and Distributional Coverage Expert Pruning (DCEP), which preserves the distributional geometry by minimizing weighted coverage error. Both methods incorporate router-derived weights as priors, enabling them to jointly preserve the global structural relationships among experts while accounting for the local importance of individual experts. Experiments on DeepSeek-V2-Lite and Qwen3-30B-A3B demonstrate that, at the expert retention ratios of 75% and 50%, our methods outperform the state-of-the-art baselines on the majority of benchmarks and consistently achieve the best average performance across both general tasks and mathematical reasoning tasks.

    memorybenchmark
  176. arxiv:2610.41235 · cs.RO
    CONSORT: Uncertainty-Constrained Multi-Robot Planning for Cooperative Localization in GPS-Denied Environments
    Tony Zhang, Ryan Bartholomew, Kalliyan Velasco, Joshua Mangelson +1

    Centralized multi-agent path planning with cooperative localization in GPS-denied environments is generally computationally intractable when formulated in full belief space due to tightly coupled uncertainty dynamics. Prior work typically approximates belief dynamics or restricts inter-agent coupling. This work examines a multi-agent planning problem in which a primary agent must reach a goal in minimum time subject to a chance-constrained collision-avoidance requirement and a terminal bound on localization uncertainty. We obtain a tractable surrogate of the underlying belief-space problem by optimizing over the mean-state trajectories of all agents and propagating covariance for terminal constraint evaluation only. We propose CONSORT (Cooperative Navigation with Support-Optimized Routing and Trajectories), a discrete-to-continuous pipeline combining hexagonal-grid joint search with spline-based trajectory refinement. Across 50 randomized scenarios, the planners separate as the terminal uncertainty bound tightens, with CONSORT solving 80% of scenarios where the strongest baseline solves 54%, at comparable path length. In marine simulations, with the plans executed under feedback control on an independently developed factor-graph estimate, the trajectories remain collision-free and their terminal uncertainty falls within 6% of the planner's prediction.

    agentmulti-agent
  177. arxiv:2610.41232 · cs.AI
    TRIDENT and SHAKTI: A Framework and Benchmark for Hindi AI-Text Detection
    Shifali Agrahari, Sujit Kumar, Sanasam Ranbir Singh

    As Large Language Models (LLMs) become increasingly multilingual, distinguishing human-written from AI-generated text is essential for maintaining information integrity. However, research remains English-centric, with limited benchmarks for low-resource languages such as Hindi. To address this gap, we introduce SHAKTI, a large-scale Hindi benchmark covering four genres, eight LLM families, multiple prompting strategies, and adversarial transformations. We evaluate its quality through human annotation and analyze linguistic, grammatical, and sentiment-based differences between human- and AI-generated texts. Building on these findings, we propose TRIDENT, a round-trip translation-based detection framework with three variants: TRIDENT-S (static threshold), TRIDENT-D (dynamic threshold), and TRIDENT-C (logistic regression using multi-distance RTT features). TRIDENT leverages semantic consistency between original and round-trip translated texts to identify AI-generated content. Extensive experiments show that TRIDENT outperforms conventional and supervised baselines, with TRIDENT-S achieving the best overall performance. Further evaluations across seven languages, adversarial settings, and unseen LLMs demonstrate its multilingual applicability and robustness, while highlighting sensitivity to the translation model.

    benchmark
  178. arxiv:2610.41227 · eess.SY
    Delta-Seeker: Closed-Loop Generative Search for Novel Circuit Design Discovery
    Dinithi Jayasuriya, Devangshu Dutta, Aravind Saravanan, Nilesh Ahuja +2

    Generative models are increasingly used for circuit synthesis and optimization by proposing topologies, sizing devices, and exploring large design spaces. We investigate whether they can instead support design discovery, where generated structures must depart from known designs while producing reproducible electrical behavior. We propose Delta-Seeker, a closed-loop circuit search framework that combines discrete diffusion for topology generation and mutation with structural checking, SPICE-in-the-loop sizing, Pareto selection, and feedback from prior search outcomes. The search controller uses progress and recurring structural failures to allocate evaluation effort and steer subsequent topology generation, while independent re-sizing and operating-corner analysis are performed only after search. In generator-only experiments, cross-generation feedback increases structurally valid children from 143 to 945 and reduces the faulty-candidate fraction from 0.925 to 0.761. In op-amp topology discovery at matched sizing cost, Delta-Seeker yields 54 designs that remain feasible in at least three of four independent re-sizings and 18 that additionally pass device-level checks and supply-temperature corner analysis, compared with 38 and 10 under fixed control. Removing cross-generation memory reduces independently robust discoveries from 54 to 43 and seven-corner robust discoveries from 40 to 27. In DC-DC converter search, nearest-parent analysis identifies a corpus-absent topology whose generated switching current path changes its nearest training-corpus parent from infeasible at all four operating corners to feasible at all four, while increasing median efficiency from 31.2% to 98.96%. These results show that closed-loop generative search can support circuit discovery beyond structural novelty by identifying new topologies with independently verified changes in circuit operation.

    memory
  179. arxiv:2610.41225 · cs.AI
    RTL-Interact: Training EDA-Interactive Models for Multi-Task RTL Design
    Wenxuan Shi, Yaoyu Zhu, Di Huang, Chongxiao Li +12

    Large language models offer a promising route to register-transfer-level (RTL) design, but reliable implementations require checking and revision based on electronic design automation (EDA) tools' feedback. Existing RTL-specialized models often lack learned interaction with such EDA tools or rely on separate generation and repair components. We introduce RTL-Interact, a multi-task post-training framework that trains a single RTL-specialized model to generate RTL, construct task-appropriate EDA checks, and revise implementations using tool feedback. The framework combines automatically constructed multi-task demonstrations with staged supervision from chain-of-thought to tool-integrated reasoning. We further introduce Pattern-Aware Policy Optimization (PAPO) for multi-task reinforcement learning. PAPO augments Group Relative Policy Optimization with correctness-gated rewards that favor concise successful checking while rewarding effective revision, and uses task grouping to preserve coverage across objectives. On CVDP, RTL-Interact-32B achieves 48.7% weighted Pass@1 and 68.9% Pass@5, surpassing its teacher model, DeepSeek-V3.2, while RTL-Interact-8B achieves 44.9% weighted Pass@1. Additionally, we show that PAPO can reduce average tool calls by 63.8% while improving Pass@1. Both models also outperform all RTL-specialized baselines on VerilogEval v2 and RTLLM v2, demonstrating the effectiveness of our method.

    post-training
  180. arxiv:2610.41221 · eess.SY
    Control-Layer Interpretation of Reactive Power Recovery Limits in EMT-RMS Model Benchmarking of Inverter-Based Resources
    Dileep Cheerla, Amirthagunaraj Yogarathnam, Divya Jain, Satish Natti

    EMT--RMS benchmarking is widely used to verify cross-platform consistency of inverter-based resource (IBR) dynamic models. This paper examines a 200 MW anonymized IBR represented in PSCAD and PSS/E. The EMT model separates inverter electrical capability from a plant-level power plant controller (PPC) recovery target, whereas the RMS model represents these functions through REECC and REPCA parameters. Directly assigning the PPC target of $\pm 0.2764$ pu to REPCA hard reactive limits is initialization-inconsistent at the original lagging operating point because the initial reactive command is 0.396273 pu. A second, near-unity-power-factor operating point initializes successfully with the same diagnostic limit, demonstrating that the conflict is operating-point dependent. Using the non-conflicting capability mapping removes the original hard-limit conflict, but the benchmark still shows 18.5--30.2 MVAr post-event POI reactive-power differences from PSCAD. A one-at-a-time REPCA $K_c$ sensitivity under matched benchmark initial conditions further shows that RMS recovery is parameter- and limit-dependent: at $K_c=0.05$ and 0.10 the aggregate machine reactive output reaches the configured capability bounds, while $K_c=0.15$ moves VDOWN, VUP, and HVRT off those bounds and changes their settled POI reactive power by 8.94, 18.00, and 3.17 MVAr, respectively. Machine-to-POI accounting further shows why machine-side saturation need not reproduce the POI-level PPC recovery magnitude. The results distinguish initialization feasibility from recovery equivalence and provide a control-layer diagnostic workflow beyond waveform comparison.

    benchmark
  181. arxiv:2610.41216 · cs.RO
    TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents
    Kehong Gong, Xiaoxu Zheng, Guanli Hou, Seunghun Lee +7

    Robot agents built on vision-language models depend on both the planner and its harness: the tools, feedback, memory, and recovery logic around it. Can improvements to a large teacher's harness be carried into a compact agent with lower inference compute cost? We present TinyAstral, a two-stage approach. We first freeze an open-weight teacher and improve its harness through recursive self-improvement (RSI): frontier LLM agents read the teacher's trajectories and propose harness edits, and each edit is accepted only if it passes a paired statistical test. We then freeze the final harness, collect teacher trajectories with reasoning and failure recovery, and distill a small student that replaces the teacher in the same harness. On RoboTwin, harness evolution raises the teacher's success from 33.4% to 63.6%; the final 4B student reaches 62.8%. Across four harness versions, distilled 4B students score close to their corresponding teachers, and later versions yield stronger students, showing that harness evolution gains carry into compact models. The harness gains also transfer to other frozen planners without retraining. On identical episodes, the final 4B student outperforms open 4B agents, the open 27B harness-RSI model SenseNova-RoboRSI, and our own teacher model placed in three competitors' harnesses. It stays close to its teacher on all three training benchmarks, and on three benchmarks never used for RSI or training it scores close to a 552B planner running the same harness. On the measured deployment episodes, it uses about one fifth of its teacher's accelerator time.

    manipulationrobotwinagentllm agentself-improvementself-evolving
  182. arxiv:2610.41213 · cs.LG
    Beyond Scarcity: When Encoder Adaptation Is Worthwhile
    Zexing Zhang, Jichao Li, Tianyang Lei, Yiyin Zhang +1

    Long-tailed recognition often learns a shared representation before correcting the classifier for class imbalance. Once the classifier has been rebalanced, however, final accuracy does not reveal whether further encoder adaptation remains worthwhile. Standard imbalance benchmarks obscure this question because they jointly vary the number of distinct examples available to rare classes and the frequency with which those examples shape the representation. We introduce a controlled evaluation framework that separates unique support from representation allocation and estimates the incremental value of encoder updates through matched classifier-only and full-network continuations from a common corrected model. The framework combines stage-specific controls, multiple representation and adaptation objectives, whole-distribution metrics, and hierarchical inference over class assignments, support reconstructions, and optimization runs. Across the controlled study, the same low-support regime moves from one in which classifier-only training is adequate to one in which encoder adaptation yields a practically meaningful benefit solely as representation allocation becomes more imbalanced. The transition persists across classifier corrections, representation objectives, normalization and augmentation controls, and ImageNet-scale evaluation, while preserving frequent-class accuracy. Under the prespecified one-percentage-point operating margin, the two regimes are separated by simultaneous confidence bounds. These results show that benchmark construction can alter a post-correction training decision and that support size alone does not characterize the remaining value of representation adaptation.

    benchmarkevaluation framework
  183. arxiv:2610.41210 · cs.CV
    From Evaluation to Evolution: Benchmarking Image Editing and Evolving Image Models
    Jun He, Canmiao Fu, Ying Zhang, Zhipeng Huang +6

    Real-world image editing requires coordinating multiple visual transformations with content preservation and reference constraints, yet aggregate scores can readily obscure unmet local requirements. Effective evaluation must therefore identify local failures and translate them into priorities for subsequent model improvement. We introduce WeGenEditBench, comprising 2,740 evaluation instances and 148 task types across single-image, text-intensive, and multi-reference editing. Pre-generated checklists shared across models decompose each request into atomic requirements; item-level deductions and visual evidence are established first and then summarized into scores for edit execution, preservation, visual quality, and task-specific correctness. To support the experience-acquisition component of recursive self-improvement (RSI), we develop WeGenEvoStudio, a unified multi-agent harness for RSI-oriented learning that integrates model-conditioned failure diagnosis, retrieval-grounded data construction, supervised adaptation, and re-evaluation. Among 15 evaluated configurations, GPT-Image-2 leads single-image editing with an Overall score of 4.076, while Seedream-5.0-Pro leads text-intensive editing at 4.291. Evolution experiments across four generation and editing benchmarks demonstrate substantial targeted gains: on their report-selected Top-5 single-image targets, FLUX.2 [dev] improves Overall from 1.083 to 3.517, while Qwen-Image-Edit-2511 gains 13.2%. Together, our benchmark and harness provide a foundation for developing more reliable image models through interpretable evaluation and feedback-driven learning.

    multi-agentself-improvementbenchmark
  184. arxiv:2610.41199 · cs.CV
    Beyond CLIP: Calibrated (Few-Shot) Defect Detection With System One Decision Models
    Fadel M. Megahed, Ying-Ju Chen

    Cameras and vision sensors are widely used for product and process inspection in manufacturing environments. Traditionally, this process required thousands of images of both nominal and defective products to train such models. Recent research shows that vision transformer (ViT) models, such as CLIP, can reduce this requirement to 50-100 images per nominal and defective class. However, collecting even 50 defective images can be difficult in high-yield or low-volume manufacturing processes. In this paper, we examine the utility of two recently proposed structured decision models (Jev and Clef-Flash) in applications using as few as k=2 reference images per class at inference. In the process, we provide three main contributions: (a) a framework for k-shot learning of these models for image classification/inspection applications; (b) an agentic approach for automatically refining the text descriptions that guide image classification in these decision models; and (c) an approach to calibrate the resulting defect probabilities to improve their reliability for inspection decisions. We test these contributions in five manufacturing case studies. Using two reference images per class, both Jev and Clef-Flash outperform CLIP's test accuracies with the same references (0.75-1.00 [Jev] and 0.71-0.99 [Clef-Flash] vs 0.67-0.84 [CLIP]). Furthermore, we show that our agentic refinement of text descriptions significantly improves accuracy across 6 of 8 model-application combinations, and our calibration approach reduces the Brier score in 13 of the 15 model-application configurations. To support the reproducibility of our work, we provide our replication code and have created the R "decisions" package for evaluating additional models. We also provide a publicly available web app that allows practitioners to inspect their own images with no code.

    agentic
  185. arxiv:2610.41191 · cs.LG
    When AI Investigators Cannot Trust Their Tools: Conflict-Aware Evidence Verification for Cyber Threat Investigation
    Yuchen Zhang

    LLM agents increasingly triage security alerts by combining the outputs of several detection tools, implicitly treating those tools as trustworthy. Unlike prior work on conflicts between a model's knowledge and its context, we study conflicts among the tools themselves: what happens when this machine-generated evidence is conflicting or wrong. On the corrected CIC-IDS2017 dataset, four heterogeneous detectors emit labels and confidences that we corrupt under a threat model varying the number of corrupted tools, the persistence of the corruption, and the confidence attached to wrong outputs. Across open-weight LLMs from 3B to 14B parameters, unverified investigators are fragile: one corrupted detector out of four lowers naive accuracy to .50-.72, three lower every unverified investigator to .13 or less, and honest disagreement alone makes them abstain on up to 63% of flows. We propose Conflict-Aware Evidence Verification (CAV), built from established components: an equal-weight vote that ignores reported confidences, an evidence-conflict gate, and an independent nearest-neighbour trust check that tests the leading candidates. With its prompt frozen before testing, CAV-LLM improves on the naive investigator in every corrupted condition ($p < 0.001$), reaching .84-.88 with three corrupted detectors against at most .08, while calling the LLM on 0.4% of clean traffic. A fixed verification rule is as accurate or better unless most tools are corrupted. When we make the independent channel itself noisy, CAV inherits its errors: it stays ahead when two or more tools are corrupted, but adding the channel as an extra voter is better when at most one is. Verification thus moves trust from the tools to the check, which must be measured and protected in turn.

    llm agent
  186. arxiv:2610.41190 · cs.CV
    FlowMoDE: Coarse-to-fine Flow Matching for Detail-aware Sim-to-Real Monocular Depth Estimation
    Liangjing Shao, Wanhao Liu, Jinsong Lin, Zhiwei Fang +1

    Monocular depth estimation is a fundamental capability for 3D spatial perception in robotic and intelligent systems. While data-driven discriminative methods have achieved compelling performance for monocular depth estimation, generative approaches can reconstruct fine structural details with stronger generalization. However, most existing generative diffusion models suffer from low inference efficiency caused by multi-step sampling, whereas single-step diffusion models struggle to recover high-fidelity depth maps. To mitigate these drawbacks, this paper presents FlowMoDE, a coarse-to-fine flow-matching-based framework for efficient, structure-aware monocular depth estimation supporting sim-to-real and cross-scene generalization. The core architecture first conducts coarse-level decoding conditioned on diffused features extracted from a pre-trained semantic encoder, and subsequently restores fine-grained structural patterns via expanded decoding layers. Trained exclusively on a single synthetic indoor dataset, our proposed method outperforms prior methods across five real-world indoor and outdoor datasets, yielding an average 14.1% reduction in relative error against state-of-the-art works as well as 9.9\% reduction in chamfer distance for point cloud reconstruction. We further validate the generalization capability of FlowMoDE on realistic robotic scenes, with qualitative comparisons against state-of-the-art methods. The project page is https://flowmode-shao.netlify.app/

    sim-to-real
  187. arxiv:2610.41177 · cs.CL
    TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents
    Jihao Zhao, Shuaishuai Zu, Pengfei Li, Biao Qin +3

    LLM agents execute tasks through repeated model calls over evolving histories of reasoning, actions, and tool feedback. Aggregate speculative-decoding metrics merge these distinct decisions, providing limited guidance on where draft supervision is needed. We present TraceSpec, a framework connecting trajectory evaluation with adaptive draft training. It combines dynamic execution with fixed-prefix replay and introduces TraceSpec-Traj, a corpus of 10,000 training states and 2,500 test states spanning browsing and software engineering. Analysis of five speculative baselines reveals that aggregate acceptance can rise with context length while declining within matched late browsing actions. Workload and phase further distinguish these trends, motivating supervision guided by trajectory conditions. TraceSpec trains on intermediate assistant turns and allocates a fixed prediction budget using state-conditioned estimates of prefix-aware acceptance gaps. Across three target sizes and two temperatures, it improves replay throughput by 12.5% to 22.6% over its DSpark base. Qwen3-4B ablations show that adaptive allocation increases accepted length by 3.3% to 10.0% over fixed full-turn supervision, with additional comparisons substantiating the roles of state conditioning and consecutive-token agreement. The resulting evaluation resources, trajectory findings, and training strategy connect the analysis of agent execution to practical draft improvement.

    agentllm agent
  188. arxiv:2610.41173 · cs.LG
    CAP: Learning to Optimize from Sparse Observations of Related Tasks
    Lin Fu, Siwei Fu, Dong Ni

    In experiment-driven black-box optimization, each evaluation can consume materials, equipment time, and other resources, limiting the number of experiments available for a new task. Records from completed related optimization campaigns could help, but learning a sequential query policy remains challenging when task-specific experience consists of only a few sparse histories. We propose CAP (Candidate Acquisition Policy), which models shared structure and task-specific variation in fixed historical records to sample interactive practice functions. On these functions, CAP learns a feedback-conditioned query policy aimed at final optimization performance under a fixed evaluation budget. At deployment, its parameters remain fixed while each new observation informs the next query. Across five synthetic task families spanning 6 and 20 dimensions and two engineering simulation benchmarks, BOPTEST and SU2, CAP attains the lowest median final-budget value among the reported methods on four synthetic families and both engineering benchmarks. These results indicate that sparse records from completed related tasks can support adaptive query decisions on new tasks, making more effective use of limited, costly experimental evaluations.

    benchmark
  189. arxiv:2610.41172 · cs.AI
    Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses
    Zhichang Yu, Baining Cao, Sbaaoui Idriss, Qi Wang +14

    Retrieval-augmented generation is becoming agentic. A search harness writes queries, reads results, reformulates, and decides when to stop. This largely solves accuracy: a harness with only BM25/grep and MiniMax M3 answers 85% of BrowseComp-Plus and 91% on multi-hop FRAMES in our runs, without compiled or vector indexes. It exposes a context problem: an undisciplined loop repeats queries, rereads documents, and fills the window with unused passages. The missing layer is knowledge compilation: a configurable ingestion pipeline where LLMs enrich documents with context no chunk contains alone--navigation hierarchies with claim-evidence leaves, entity graphs, timelines, mind maps, ontologies, wikis, and memory. Compilation is not a better retriever: under a competent harness, structural artifacts gave no detectable accuracy gain on FRAMES, and a PageIndex-style outline degraded at scale. It is a supply layer. A navigation tree whose leaves are claim-evidence pairs--the mechanism we measure--cuts query tokens by 91% when evidence suffices and adds 1.9% when it does not, a 38% average reduction; this implies a 2% break-even sufficiency rate and a measured rate near 43%. A decision gate cuts latency from 140 s to 30 s, though substitution's token saving is bought with retrieval rounds and carries its own latency cost. Other artifacts--ontology, graph, wiki, timeline, mind map, memory--pay off in deployment (governance, integration, onboarding, temporal orientation, continuity) where accuracy benchmarks cannot see them. We describe RAGFlow's pipeline and harness (open source at https://github.com/infiniflow/ragflow), which turns the loop into five disciplined stations and keeps the context entering the model the smallest sufficient set; together they form an agent-native database: a store that hands the agent compiled, governed context instead of a ranked list of passages.

    retrieval-augmentedagentagenticbenchmark
  190. arxiv:2610.41169 · cs.AI
    ImplicitHop: Evaluating knowledge integration across context and memory
    Alexander Detkov, Dashiel Karish, Matthew Oleynikov, Matt Thomson

    The ability to integrate new information and deduce its consequences is central to learning, understanding, and reasoning in large language models (LLMs). Whether information arrives during training or in-context, models must not only store local facts, but propagate their long-range implications globally. To explore information integration in LLMs, we introduce two complementary natural language benchmarks for large-scale evaluation of deductive reasoning as a function of inference depth. The first is a formal graph benchmark enabling controlled evaluation over a wide range of graph topologies and integration tasks; the second is a knowledge composition benchmark grounded in real-world knowledge graphs. We use these benchmarks to compare knowledge integration from information encoded in context versus information encoded in model weights during pretraining, evaluating a set of open models implicitly, without chain of thought. We find that accuracy decays sharply with increased composition length for both pretrained and in-context knowledge. Information presented in-context, however, substantially outperforms deduction over information encoded in weights. These results quantify a fundamental limitation of current models with direct implications for prompt-based control of agent behavior and for continual learning strategies, and they provide a foundation for developing training, prompting, and integration methods that support deeper reasoning.

    memoryknowledge graphagentbenchmark
  191. arxiv:2610.41168 · cs.AI
    ROSTrans: Empirically Benchmarking LLMs for ROS1-to-ROS2 Migration
    Yichen Zhao, Minghao Shao, Zhuoyan Yu, Yijun Shen +3

    Although large language models have shown strong capabilities in domain-specific code generation, their performance in robotics software development, especially for the Robot Operating System (ROS), remains largely unexplored. ROS code generation presents distinctive challenges, including complex inter-node coordination, middleware configuration, and lifecycle semantics that go far beyond standard coding benchmarks, and many important correctness properties become observable only once migrated nodes are launched and communicate. We present ROSTrans, a benchmark for evaluating LLMs on ROS1-to-ROS2 code migration, a representative and practically important setting for ROS code generation. ROSTrans comprises 108 scaffolded completion tasks drawn from a diverse set of 70+ open-source ROS1 repositories and organized into a hierarchical taxonomy assessing interface-level, behavioral-level and system-level correctness. For every task we establish an executable reference: a verified ROS2 implementation, a runtime test suite that launches the nodes and exercises their topics, services, actions and parameters, and a container ensuring reproducibility. Candidates are scored by injecting only their target files into an otherwise fixed copy of the reference package, so a candidate can neither supply nor weaken the tests it is judged by. Evaluating six frontier LLMs over five stochastic generations per task, we find pass@1 below 10.9%, and that 56.5% of all samples build successfully and then fail at runtime,and 32.0% fail runtime assertions on the migrated interfaces, a class of error no compile-only oracle can detect. Execution-guided repair further improves task success, reaching 49.1% and 47.2% for Claude Opus 4.6 and GPT-5.5 after three rounds, respectively.

    benchmark
  192. arxiv:2610.41162 · cs.AI
    FrontierMat: Benchmarking Language Models at the Frontier of Materials R&D
    Feng Chen, Jianguo Huang, Yushan Xiao, Yue Li +19

    To accelerate materials discovery, language models must be able to address problems at the research frontier. Yet textbook- and examination-derived benchmarks primarily assess established knowledge, leaving performance on frontier research problems largely untested. We introduce FrontierMat, a benchmark of 330 problems drawn from frontier academic and industrial research in fields including aerospace, semiconductors, embodied intelligence and nuclear fusion. Every problem is validated by leading experimentalists and industry experts in its field. Each problem presents a real study and asks the model to infer a withheld outcome or draw a conclusion from selected experimental evidence. Problems are annotated by functional domain, material system and the scale of the governing mechanism; 130 require cross-scale reasoning. The strongest general-purpose model scores only 53.7%, and on matched problems, cross-scale problems score 8.7 points below single-scale ones. Half of the problems are solved by at most one model, and models fail chiefly by endorsing what is plausible rather than what was observed. A domain-adaptation study with meta-weighted continued pretraining and scale-modular self-distillation raises an open-weight model from 36.7% to 55.4%, matching the strongest proprietary model, with improvements on public materials benchmarks and no decline in aggregate performance on general benchmarks. FrontierMat provides a testbed for measuring progress toward language models that support frontier materials R&D.

    embodiedbenchmark
  193. arxiv:2610.41159 · cs.CV
    MATS: A Multi-Agent Framework for Controllable and Reflective Tumor Synthesis
    Zhongnan Shi, Keyang Zhou, Siyi Li, Qichao Ying +5

    Synthesizing realistic tumors on medical images helps address data scarcity for clinical AI. Existing diffusion-based methods, however, lack fine-grained attribute-level controllability, closed-loop quality verification, and awareness of which morphological patterns are underrepresented in the target distribution. We present MATS, a multi-agent framework for text-controllable, cross-organ tumor synthesis. Given a target disease, MATS first profiles its morphological distribution over a fine-grained attribute space covering tumor geometry and appearance, and constructs a generation plan that amplifies underrepresented subtypes while keeping common ones dominant. It then localizes the target organ, proposes anatomically valid placement, and executes two-stage diffusion to produce the tumor mask and render texture, each conditioned on the corresponding attributes. A closed-loop reflective mechanism verifies attribute alignment at three checkpoints throughout synthesis and provides targeted feedback to guide re-generation when misalignment is detected. Trained exclusively on a single-source pancreatic neuroendocrine tumor dataset, MATS consistently improves downstream segmentation both in-domain and across LiTS, KiTS, and MSD-Pancreas benchmarks, whereas most prior generative augmentation methods fall below real-only training.

    multi-agentagent frameworkbenchmark
  194. arxiv:2610.41158 · cs.AI
    Lifelong Learning of Scheduling Rules with Surrogate-assisted Genetic Programming for Dynamic Flexible Job Shop Scheduling
    Luyao Zhu, Fangfang Zhang, Yi Mei, Ke Chen +1

    Genetic programming (GP) has been widely used to learn scheduling rules for dynamic flexible job shop scheduling (DFJSS). However, trained rules tend to only generalise to tasks of similar scales and distributions, and when task characteristics change, learned rules often suffer significant performance degradation. Moreover, fine-tuning pre-trained rules on new tasks often leads to catastrophic forgetting, where prior knowledge is lost. This motivates the need for GP methods that can adapt to new tasks while preserving past knowledge, leading to a lifelong learning paradigm. However, lifelong learning in GP for DFJSS remains largely under-explored. To address this gap, this paper proposes a surrogate-assisted GP method for lifelong learning to handle sequential heterogeneous DFJSS tasks with job or batch arrivals. The framework includes three key components: an adaptive two-stage training strategy to balance plasticity and stability, a task-similarity-based weighting mechanism to regulate knowledge reuse, and an improved surrogate model with distance constraints to enhance estimation accuracy. Experimental results show that the proposed method effectively mitigates catastrophic forgetting while maintaining strong adaptability, outperforming several compared methods and demonstrating the potential of lifelong GP in dynamic scheduling environments.

    lifelong learning
  195. arxiv:2610.41157 · cs.AI
    Lineage Is Memory: Parent Inheritance and Feedback Disclosure in Short-Budget LLM-Guided Optimization
    Shanglin Yang, Zhan Shi

    Iterative language-model optimizers refine prompts and programs by evaluating a candidate, showing a proposer model a sample of failing cases, and keeping revisions that improve performance. Recent work attributes the limits of this loop to narrow single-round reflection, introducing full-set error clustering, targeted minibatches, or larger proposer models so the optimizer sees more failure modes per step. Yet every existing evaluation tests these mechanisms exclusively inside lineage-preserving search, where each revision edits a surviving parent, leaving the contribution of parent inheritance confounded with the breadth of per-round feedback. Across 1,750 pre-registered runs spanning nine prompt and program synthesis benchmarks, we separate parent inheritance from feedback disclosure under matched evaluation budgets. We find that greedy parent inheritance drives the vast majority of optimization progress, outperforming seed-reset resampling even when seed-reset proposals receive a clustered summary of all failures in every round or use a larger proposer model. While full-set error clustering and diverse failure selection improve single-step rewrites from the seed and accelerate the opening rounds of search, sequential parent inheritance rapidly closes this gap as solved error modes are retained in the candidate and filtered from subsequent failure lists. Moreover, decomposing token expenditure in GEPA reveals that proposer reflection accounts for only a small fraction of total cost whereas validation and minibatch gating consume nearly the entire budget, showing that greedy parent inheritance over raw failures captures the gains of complex reflective loops at a fraction of their compute.

    benchmark
  196. arxiv:2610.41149 · cs.AI
    Counterfactual Pairwise Augmentation Improves Small Pointwise LLM Judges
    Mohammad Zangooei, Peyvand Teymoori, Mohammad Robati Shirzad, Soroosh Shahtalebi

    Agentic systems increasingly gate individual steps with a pointwise judge that scores a response against a criterion. Small language models are attractive for this role because they can be served and specialized efficiently. Such a judge must locate the criterion's decision boundary, but pointwise labels under-determine that boundary in fine-tuning: labeled responses to different queries differ along many dimensions at once, so they do not reveal which features decide the label. We introduce CPA (Counterfactual Pairwise Augmentation), which makes the deciding difference explicit in training. For each pointwise item, CPA synthesizes and audits a counterfactual response to the same query intended to carry the opposite label. It frames the accepted pair as a pairwise task asking which of the two responses satisfies the criterion. The judge is fine-tuned jointly on both pointwise and pairwise datasets and deployed in pointwise mode only. A second-order analysis shows when this should outperform the prevailing strategy of merging separately trained pointwise and pairwise experts: merging moves every parameter direction toward the pairwise expert by the same fraction, whereas joint training can hold the pointwise optimum wherever the pairwise loss is flat and absorb its curvature where it is sharp. This is the regime that CPA's same-source, counterfactually minimal pairwise view is designed to produce. On separate Qwen3-8B LoRA adapters for six binary criteria, CPA improves criterion-averaged pointwise macro-F1 by 2.5 points over pointwise-only training and outperforms model merging on all six criteria at every tested merge weight. It also surpasses the strongest zero-shot frontier judge we evaluate, Gemini 3.6 Flash, by 11.8 points.

    agentic
  197. arxiv:2610.41147 · cs.CL
    HyperRepair: Recovering Forgotten General Capabilities of Fine-Tuned Language Models
    Junrong Tang, Jianghui Wang, Hao Kong, Yang Lin +1

    Fine-tuning improves a language model's performance in a target domain but can weaken its general capabilities. Existing methods often face a trade-off between general capabilities and domain specialization, making it difficult to preserve both. We introduce HyperRepair, a framework for recovering general capabilities while retaining domain specialization at only inference cost. To make recovery generalizable across models sharing a pretrained backbone, HyperRepair learns a hypernetwork from repairs trained on source models. The hypernetwork uses activation fingerprints that summarize changes from the pretrained backbone to generate a tailored Repairing Side Net (RSN), whose prompt-conditioned gate applies corrections selectively. This enables recovery on models from unseen adaptation domains without target-domain examples or additional optimization. We evaluate HyperRpair across two model families, 12 LoRA adaptation datasets, and 15 general benchmarks. On the default split, HyperRepair achieves 93.4% of the base models' average general score while retaining 99.5% of the fine-tuned models' average domain score.

    benchmark
  198. arxiv:2610.41143 · cs.LG
    TalentRL: Strategy-Guided Reinforcement Learning for Hierarchical Skill Routing
    Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang, Michelle Chang +1

    Natural-language skills enable large language model (LLM) agents to reuse procedures across tasks, yet most systems retrieve individual skills from a flat SkillBank without explicitly modeling the high-level strategy that governs their selection and use. We introduce TalentRL, a strategy-guided reinforcement learning framework for hierarchical routing over procedural memory. Coarse-grained talents represent recurring strategies for skill use, while fine-grained skills provide executable procedures. For each task, the policy model selects a talent that guides execution and softly prioritizes its associated skills during similarity-based retrieval. However, since task outcomes depend jointly on talent selection, retrieved skill context, and downstream execution, terminal rewards alone cannot isolate the contribution of the routing decision. TalentRL therefore trains talent routing with reward contrasts between rollouts of the same task under matched interventions on the talent-conditioned context, and optimizes downstream execution with group-relative advantages. Across seven open-domain and multi-hop question-answering benchmarks, talent-aware supervised fine-tuning improves average success rate by 7.53%, 6.71%, and 9.21% over the corresponding zero-shot baselines for Qwen2.5-7B, Qwen3-8B, and Llama-3.1-8B, respectively. Reinforcement learning further improves these SFT checkpoints by 7.58% to 9.32%, and TalentRL achieves the highest average success rate among all compared methods on all three backbones. On the agentic benchmarks ALFWorld and WebShop, TalentRL achieves the highest ALFWorld success rate (99.3%) and the highest WebShop score (90.7) among all compared methods.

    agenticbenchmark
  199. arxiv:2610.41132 · cs.AI
    When Disclosure Is Compulsory: Evaluating Ride-Hailing Rebalancing Policies with Self-Interested LLM Driver Agents
    Xinyu Wang, Andres Fielbaum

    The EU Platform Work Directive, to be transposed by December 2026, requires ride-hailing platforms to state in plain language the main parameters of their automated decisions and how a worker's behaviour affects them. Such rules are exactly what a large language model (LLM) can read, so drivers will be able to let one decide for them. We compare two main families of rebalancing policy once their rules are public: a reactive policy (R) that sends idle drivers to cover requests just lost, and an anticipatory policy (A) that plans moves on a demand forecast. Each driver is an LLM agent maximizing its own net income, given no rules (N) or the rules at the directive's level of detail (L). We test two LLMs, Claude Haiku 5.5 and Jev, which understand the rules equally well, in a three-zone model and in Manhattan. With the rules disclosed, drivers follow the platform more closely: in Manhattan, the share of relocation offers they decline falls from about four in five to one in four under both policies. Yet the effects differ by policy. For the platform, disclosure costs R 3.7% of its profit in Manhattan, as R now pays for moves drivers used to make for free, while A's profit is unchanged; the two policies, which tie on profit under full compliance, end 6.1% apart. For drivers, disclosure raises income in the small model under both policies and LLMs; for users, it raises the service rate there by 10 to 18 percentage points. In Manhattan, self-interested drivers also make A a win-win-win against full compliance, raising driver income, platform profit and service rate by 10 to 12% each. Rebalancing policies that perform equally well under full compliance can thus show opposite results under mandatory disclosure. The study offers a framework to evaluate such policies under the new requirement and is, to our knowledge, the first to do so with LLM agents.

    agentllm agent
  200. arxiv:2610.41130 · cs.CV
    One Verdict Is Not Enough: Regression Testing Learned World Models and Their Test Harnesses
    Junyu Huo, Zihao Wan, Ziqi Mao, Gouri Ginde

    Learned world models ship as software components whose model versions, preprocessing, and test harnesses change between releases. Work on model updates already counts per-sample regressions, or negative flips, that an accuracy check hides. Two problems remain: a test can pass under one configuration and fail under another, and a harness fault can produce erroneous scores and wrong verdicts without a crash. We present TetraState, a regression-testing protocol for paired possible/impossible video tests. Each test runs under four fixed model-native configurations; its number of correct verdicts assigns one of four outcome classes. Inconclusive tests and failed executions are reported separately alongside the verdict rate. Revisions are compared by per-test outcome transitions, and the harness is versioned as part of the system under test. On all 506 IntPhys 2 Main tests and two heterogeneous systems, Cosmos 3 Nano and V-JEPA 2, two V-JEPA 2 harness revisions move modal accuracy by only 0.2 and 2.6 points yet change 11 and 214 test labels; in the second, accuracy rises although fewer tests are judged correctly, and 43 regressions keep the modal verdict, so a modal-verdict flip check misses them. These results outline a research agenda on configuration-level regression oracles for learned components.

    world modelv-jepa
  201. arxiv:2610.41129 · cs.AI
    When Should AI Agents Ask? Clarification Requests Under Limited Human Attention
    Hanzhao Wang, Shengjie Liu, Zhengchao Wang

    AI agents increasingly carry out delegated tasks in parallel, with one person overseeing several of them at once. In practice, an agent often lacks information its task requires, such as an unstated preference, constraint, or specification. It can ask the human to clarify, or it can search for the information itself, for example in past records, which takes time and effort and may fail. In this paper, we study how to coordinate these decisions across several agents that share one human with limited attention. We build a stylized model in which each agent can search, ask, or continue the parts of its task that do not depend on the missing information, and we formulate the problem as a dynamic program. We first show that whether an agent should ask depends on more than its own uncertainty: an agent with other work to do can be better off asking earlier, and which question the human should answer first depends on all waiting questions, so no ranking computed agent by agent is optimal. We then propose the Pairwise Coordination Policy (PCP), which uses exact one- and two-task problems to estimate how pairs of tasks delay each other when they share the human. Theoretically, PCP is exactly optimal when at most two requirements remain unresolved and costs at most twice the optimum when every unresolved requirement must be resolved by the human. We evaluate PCP on 864 synthetic cases and in a case study with LLM agents. On the synthetic benchmark, PCP's mean gap to optimal cost is 0.07%, and it reduces cost by an average of 18-31% relative to three standard rules for when to ask. In the LLM case study, agents completed every task correctly, and PCP achieved the lowest mean cost in all tested settings.

    agentai agentllm agentbenchmark
  202. arxiv:2610.41127 · cs.RO
    MotorSense: A Video-EMG Dataset of Motoric Representations for Action Understanding
    Eadom Dessalene, Michael Maynord, Amir-Hossein Shahidzadeh, Botao He +5

    Computer vision models learn hand action only from its visual consequences, without access to the underlying embodied motor signals. We address this absence with MotorSense, the first large-scale paired egocentric video and surface electromyography (EMG) dataset. MotorSense pairs bimanual EMG bands with egocentric video and yields synchronized video, bimanual EMG and IMU, and audio across 100 hours of unscripted manipulation. EMG offers a powerful scalable alternative to manual annotation: it captures finger movement, hand configuration, and the forces exerted between hand and environment as a dense signal free of occlusion and labeler bias, whereas manual annotation is costly, subjective, and unreliable under occlusion. We use MotorSense to learn sensorimotor alignment, producing an organization of action that is shared across the visual and motoric spaces, through cross-modal pre-training between video and EMG. The resultant vision models are then transferred to downstream tasks that benefit from sensorimotor alignment: in this case, action recognition, 3D hand reconstruction, and 3D contact reconstruction. We further demonstrate a generative application in which EMG motor signals are used to condition a video diffusion model, yielding videos of fine-grained hand manipulation. Across four action-centric computer vision tasks and eight popular video datasets, we demonstrate that sensorimotor alignment provides a stronger foundation for visual representation learning than conventional self-supervised learning methods over video, outperforming every baseline across all tasks and datasets evaluated.

    embodiedmanipulation
  203. arxiv:2610.41124 · cs.AI
    Confident Fools or Careful Searchers? A Behavioural Study of When LLMs Search Beyond Their Static Internal Memory
    Nipun D. Pathirage, Kwan Srijomkwan

    Agentic LLM assistants must repeatedly decide whether to answer from memory or to call a tool such as web search. Unnecessary searches cost latency and money; missing ones yield confident but unsupported answers. Prior work mostly measures verbalised confidence or trains new adaptive-retrieval methods, and rarely audits the search decisions that deployed products actually make. We introduce a benchmark of 400 multiple-choice questions that combines widely known inside-knowledge (IK) questions, which should be answered directly, with outside-knowledge (OK) questions about facts from news published after the models' training cut-offs (to our knowledge), which should trigger search or, without tools, abstention. Each item offers an abstain option, and two action-level scores, NAS and KIS, reward searching on OK questions and restraint on IK questions. Across five commercial LLMs with and without their vendors' web search, GPT-5 mini and Grok 4 Fast search on over 90% of OK but only 1% of IK questions, whereas Mistral Small searches on just 20% of OK questions, often abstaining instead of using its tool; one product's search indicator proved unusable. Searching is partly triggered by explicit temporal cues: on uncued OK questions, the NAS of the two selective models falls from about 0.72 to 0.57--0.58. The scores thus measure the search triggering of specific products, not metacognition in a strict sense.

    memoryagenticbenchmark
  204. arxiv:2610.41118 · cs.CV
    CompoSET: A Controlled Single-Edit Benchmark for Vision-Language Compositionality
    Laure Delisle, Pietro Perona

    Vision-language compositionality is the ability to identify which properties, relations, quantities, and states bind to which entities in both an image and a caption, and to match those bindings across modalities. Existing benchmarks evaluate this ability through image-text matching: a model ranks candidate image-caption match similarities and succeeds when the correct match comes on top. In such datasets, paired images differ by more than a single binding: scene, layout, and content all differ, with 97% of pixels changing on average. Such pairs do not isolate compositionality: results are confounded by global image differences. We introduce Compositional Single-EdiT (CompoSET), a benchmark dataset in which each base image-caption pair has an edited counterpart, or foil, that preserves the scene but differs by a single controlled edit, such as color, count, spatial relation, or object state. Base and foil images preserve naturalistic scene complexity yet are visually near-identical: only 8% of pixels change on average. Our dataset has many foils per base scene (~22) in order to separate scene difficulty from edit difficulty, a distinction one-foil-per-base benchmarks cannot make. CompoSET contains 1,776 instances across 80 naturalistic scenes, spanning 16 edit types in 4 semantic categories, with captions at 3 verbosity levels. We test 22 dual-encoder VLMs and find that all models perform at or below chance level on CompoSET, despite scoring well on prior benchmarks. This suggests that apparent progress under prior image-text matching protocols may be an artifact arising from confounds in benchmark design, and that compositional understanding in dual-encoder VLMs is far from solved. CompoSET addresses the confounds and offers a better controlled benchmark to measure future progress.

    benchmark
  205. arxiv:2610.41112 · cs.AI
    Calibration of Large Model Quantization based on HiFloat4 Format
    Jing Huang, Yu Cheng, Ziwei Yu, Lan Zhao +10

    Quantization of Large Language Models (LLMs) has be- coming increasingly important for reducing the computa- tional costs and resource requirements for practical deploy- ment.Growinginterestinthisareahasmotivatedthecommu- nity to move beyond 8-bit quantization towards 4-bit preci- sion, leading to the development of customized 4-bit formats such as MXFP4 and NVFP4. Recently, HiFloat4 (HiF4) has emerged as a competitive alternative which adopts a Block Floating-Point(BFP)formatwithathree-levelscalinghierar- chy, thereby achieving a significantly larger dynamic range. While quantization can substantially reduce memory usage and accelerate inference, it frequently introduces accuracy degradation. Calibration is therefore essential for narrowing theperformancegapbetweenlow-precisionquantizedmodels and their full-precision counterparts. Nevertheless, existing calibration methods are mainly designed and validated us- ing integer-precision formats, such as INT4, lacking support for HiF4 due to its hierarchical BFP structure. In this paper, weinvestigatethreecalibrationapproaches,namelyHiGPTQ, HiAutoRound, and HiFlatQuant, to further enhance model quantizationbasedontheHiF4format.Extensiveexperiments demonstratethatallthreeproposedapproachesimproveinfer- ence accuracy compared with direct HiF4 conversion, reduc- ing the gap between HiF4 and BF16 to less than 1% and out- performing the state-of-the-art NVFP4 format across various models and downstream tasks.

    memory
  206. arxiv:2610.41107 · cs.LG
    Differential rank preservation manifold learning
    Henry Han, Luyu Niu, Diane Li, Jiacun Wang

    Manifold learning methods like t-SNE and UMAP preserve local neighborhoods, but they do not preserve the ranking of the neighbors inside them. Two embeddings can return the same neighbor set while disagreeing on which point is closest. This loss of order degrades downstream applications that value the order in the embedding. We propose Smooth Rank t-SNE (SR-tSNE), which adds a differentiable rank term to the standard t-SNE loss. To enable efficient training, this term uses a sigmoid rank approximation to explicitly force a point's true high-dimensional neighbors to outrank unrelated intruders and retain their original order. We also introduce Ordinal DLP (O-DLP), an extension of the Degree of Locality Preservation (DLP) metric that credits a neighbor only when it returns at its original rank. Across eight benchmarks spanning image, text, speech, remote-sensing and single-cell data, SR-tSNE outperforms 12 baselines in rank preservation and neighborhood overlap. Ablation studies demonstrate our rank term alone boosts O-DLP by 5% to 28%, a gain not achievable using standard pairwise loss. This advantage transfers to downstream applications. SR-tSNE achieves the best label-transfer performance across seven vision datasets. For cell-type discovery across six single-cell datasets, it ranks first in 17 of 18 dataset-metric combinations, whether clustering resolution is selected with or without cell-type labels. In intraday trading, SR-tSNE achieves higher returns, lower risk, and lower turnover than the baselines. To our knowledge, SR-tSNE is the first t-SNE-family method with explicit order preservation.

    benchmark
  207. arxiv:2610.41104 · cs.LG
    MUON-RT Reusing the Teacher's Trajectory for Momentum-Free Muon for LLM Self-Distillation
    Aryan Das, Vinay Kumar Verma, Aakansha Mishra

    Self-distillation methods for large language models with an exponential moving average (EMA) teacher already retain a history of student updates. Muon nevertheless stores a separate momentum buffer, increasing persistent training memory. We propose Muon with Reused Teacher (Muon-RT), which reuses the teacher-student displacement to eliminate momentum buffers for hidden weight matrices. We show that this displacement encodes an exponentially weighted history of applied updates and smooths temporally uncorrelated noise under a local model. Muon-RT removes its radial component, normalizes it spectrally, and mixes it with the current gradient after Newton-Schulz (NS) orthogonalization. Under explicit regularity and alignment assumptions, we establish linear contraction of an expected objective-gap bound toward a residual error level. With Self-Distillation Fine-Tuning (SDFT) on Qwen2.5-7B-Instruct, Muon-RT improves average accuracy over AdamW by 2.6 points with 42% less persistent memory, and over momentum Muon by 5.4 points with 27% less persistent memory. With Self-Distillation Policy Optimization (SDPO) on Olmo-3-7B-Instruct, it improves average accuracy over AdamW by 2.0 points.

    persistent memory
  208. arxiv:2610.41103 · eess.SY
    Optimal climate policy in incomplete electricity markets
    Emil Dimanchev, Steven A. Gabriel, Stein-Erik Fleten, Filippo Pecci +1

    Incomplete long-term contracting in electricity markets can expose investors and consumers to inefficient levels of risk. This paper investigates how incompleteness impacts the optimal choice of climate policy. We develop a game-theoretic model of optimal policy-making in anticipation of how electricity market agents react to a given policy when gas prices and demand are uncertain and contracting is incomplete. In illustrative experiments, we observe that renewable subsidies and carbon prices induce an insurance co-benefit when markets are incomplete. This is because, relative to a complete market benchmark, incompleteness shifts investment from renewables toward gas and exposes consumers to uninsured risk from inter-annual volatility in electricity prices. By incentivizing investment in renewables, subsidies and carbon prices indirectly counteract these distortions. Consequently, incompleteness increases the optimal subsidy and carbon price levels in our experiments. Renewable subsidies exhibit a stronger insurance effect than carbon pricing, as the latter leads to more investment in gas with carbon capture and storage and more volatile electricity prices. While carbon pricing is more economically efficient than subsidies in complete markets, incompleteness shrinks its efficiency advantage and even leads to cases where subsidies are more economically efficient.

    benchmark
  209. arxiv:2610.41089 · cs.AI
    CIRCE Divines LLM Task Complexity from Constraint Structure
    Chenhao Hong, Jiakai Xu, Kaushiki Singh, Yusen Zhang +1

    Estimating whether a problem is too complex for a large language model (LLM) before the model attempts it can guide model routing, how to allocate test-time compute, and suggest the need for new tools to aid the agent. While existing estimators rely on the problem text, LLM-annotated task difficulty labels, or signals obtained after a model attempts to solve the problem, such as its hidden states, none explicitly leverage the constraint structure inherent in the problem: the variables to assign and how the constraints over them interact. We introduce CIRCE, which extracts a problem's structure by using a cheap LLM to construct a Constraint Intermediate Representation that deterministically compiles into the constraint graph. Across seven reasoning benchmarks (including SAT, logic, and math) and four LLMs, even a randomly initialized GNN encoder improves predictions of LLM response correctness, reducing negative log-likelihood from 0.465 to 0.423 and increasing AUROC from 0.691 to 0.727 relative to the strongest baseline for each metric.

    benchmark
  210. arxiv:2610.41081 · cs.AI
    Optimization Techniques for Solving the Minimum Red--Blue Separating Set Problem with an Application to Human--AI Text Separation
    Marko Djukanović, Staša Vujičić Stanković, Žiga Zebec, Christian Blum

    We study the so-called Minimum Red''Blue Separating Set Problem (MRBSSP) which arises from computational geometry. In particular, we propose a compact Set-Cover ILP formulation, safe preprocessing rules, a Greedy Randomized Adaptive Search Procedure (GRASP), and two exact approaches: one is based on Logic-Based Benders Decomposition (LBBD) approach, and the second represents a novel Branch-and-Benders-Cut (BBC) approach. These two exact approaches are compared to ILP--alongside GRASP and the Variable Neighborhood Search (VNS) from literature. Experiments on 250 random graphs with up to 500 vertices reveal complementary strengths: ILP yields the best feasible solutions, LBBD provides the strongest lower bounds and smallest mean optimality gap, and BBC certifies optimality most frequently. We further introduce an NLP application distinguishing human- and AI-generated texts through vocabulary-based graph features. On 42 instances, BBC achieves the smallest mean optimality gap, certifying 16 optima, while GRASP outperforms VNS. Structural analysis of optimal vocabulary separators identifies 113 mandatory and 151 replaceable features across 272 tests, highlighting substantial variability among optimal solutions. Overall, our results demonstrate the effectiveness of complementary optimization strategies and the interpretability of optimal vocabulary-based discrimination.

    grasp
  211. arxiv:2610.41080 · cs.CV
    Grounding Image Composition with Label-Conditioned Visual Evidence
    Armin Dadras, Tingyu Lin, Robert Sablatnig, Franziska Proksa +1

    Image composition labels such as center, diagonal, or rule of thirds describe visual structure rather than object categories. KU-PCP provides image-level labels and class-specific geometric elements, but does not show which visible regions people associate with each of its nine classes. We introduce KU-PCP Elements (KU-PCPE), a label-conditioned region-evidence benchmark for photographic composition. Given an image and a label already judged applicable, the task is to segment the visible regions that explain why the label applies. KU-PCPE contains 4,093 single-annotator training masks and a 1,327-task expert test split with three masks per image-label task. The annotations also include exploratory segment-versus-line ratings that describe whether region or semantic-guideline evidence better explains the label. A held-out expert reaches .867 Dice against the other experts, while the strongest deployable model reaches .664, a proposal oracle .672, and the best vision-language model (VLM) proposal selector .395. KU-PCPE exposes the gap between recognizing a composition label and grounding the human-aligned visual evidence that makes it plausible in a particular image.

    benchmark
  212. arxiv:2610.41079 · cs.CL
    EDU-Jev Training System-One Decision Models for Education
    Jiazheng Li, Yulan He

    Many automated decisions in education involve choosing from a fixed set of options: assigning a rubric score, checking whether an answer meets a marking criterion, or identifying the first wrong step in a solution. Generative language models are increasingly used to make these decisions, but their responses require parsing, longer outputs increase inference costs, and confidence expressed in words does not necessarily reflect the probability of being correct. We present EDU-Jev, a family of single-pass ("System One") decision models at 0.8B, 2B and 4B parameters. Given an assessment context and a question with predefined answer options, the model returns a probability distribution over those options without generating text. A pointer head scores a marker for each option, ensuring the predictions are restricted to the allowed choices. Training uses a single supervised stage with soft cross-entropy against the available target label distributions; calibration is assessed empirically. Across twelve education benchmarks, EDU-Jev 4B achieves an average accuracy of 67.6%, exceeding the strongest of five commercial decision APIs by 9.4 percentage points and the best open 4B decision model by 18.7 points. On four cross-domain education sets, it matches the strongest API, which remains ahead on general decision-making and reasoning tasks. The 2B model processes about 150 decisions per second on a single H100 GPU.

    benchmark
  213. arxiv:2610.41076 · cs.AI
    BABA-Bench: Evaluating Long-Horizon Patient-Advocacy Agents
    Shriram Vasudevan, Videet Mehta, Connor Sweeney

    An LLM agent that helps a patient get care must do two things. It must reach the correct outcome, and it must follow the necessary procedure. For example, the agent must get the patient's permission before it shares the patient's records. We introduce BABA-Bench, a benchmark of 47 interactive worlds. We built each world from a de-identified, real patient advocacy case that expert advocates reviewed. In each world, the agent must find hidden information and coordinate insurers, clinics, suppliers, pharmacies, and community organizations. The engine checks each action at the time the agent does it. It checks if the agent has permission, if the agent has evidence for the action, and if the action is prohibited. We evaluated 14 frontier models with three runs on each world, for a total of 1,974 runs. Agents often reached the correct outcome by an invalid route. Of the 921 runs that reached all required outcomes, 254 (28%) did not follow the necessary procedure. Claude Opus 5 reached all required outcomes in 81.6% of its runs, but only 52.5% of its runs passed all checks. When we gave two models a verified route, their pass rates increased from 29.8% to 91.5% and from 44.0% to 97.2%. For these models, the main problem is to find a route, not to execute it.

    agentllm agentbenchmark
  214. arxiv:2610.41073 · cs.LG
    bf16 Rounding Silently Discards Most RL Policy Updates: A Hidden Training-Inference Mismatch and an Unbiased One-Line Fix
    Andrew Gao, William Gao

    Reinforcement learning (RL) post-training of language models keeps fp32 master weights in the optimizer and serves rollouts from a bf16 copy produced by round-to-nearest conversion. At the learning rates in common use, most per-step updates are smaller than half a bf16 spacing and the conversion discards them: on Qwen3-1.7B only 0.4-0.5% of served weights change per step, and the policy that generates the training data lags the optimizer by 50-150 steps. Existing mismatch corrections cannot see this lag because they compare two copies downstream of the same rounding. We propose Fair Cast: export the master with an unbiased rounding rule (stochastic rounding or error feedback) and compute the training loss on the served copy. The change is one line at the export call, adds no hyperparameters and costs 22-38 ms per export of 1.7B parameters. Across 25 matched seed-pairs on Qwen3-0.6B to 4B with AdamW and Muon, Fair Cast is ahead on 18, tied on 6 and behind on 1, with a mean gain of 6.1 points; on zebra puzzles it reaches 0.57 +/- 0.16 over four runs where round-to-nearest collapses to 0.33 +/- 0.10. It also makes the saved fp32 checkpoint agree with the served model, and it exposes the implicit lag of round-to-nearest as a regularizer that can be restored explicitly when a task needs it.

    post-training
  215. arxiv:2610.41072 · cs.AI
    LanG 2.0: Energy-Efficient Agentic AI for Semi-Automated Security Operations at Scale
    Anes Abdennebi, Nadjia Kara, Laaziz Lahlou, Hakima Ould-Slimane

    Security operations centers (SOCs) need to turn terabytes of alerts, logs and flows into trusted, affordable decisions to unveil hidden threats. We present LanG 2.0, an agentic system that detects threats, correlates evidence and writes rules, investigating whether small local models can replace frontier models affordably and safely when structural safeguards, not model accuracy, provide the guarantees. It incorporates entity-anchored correlation, human approval of rule changes and a tamper-evident ledger. Tests span eight flow datasets from five networks, real XDR alerts and multistage campaigns. Against its predecessor, campaign correlation precision rises from 18.75% to 51.81%, and rule acceptance from at most 60.00% to 98.33%. Against Claude Opus 5.5, scoring flows against each network's normal traffic detects attacks on unseen networks at 73.42% F1 without false positives, versus 58.10% with 49.00% false positives, for at least 650 times less energy. Finetuned models statistically match it on alert classification and rule writing for 5.6 to 59 times less energy. The work demonstrates LanG's practical scalability, processing 3.33 million records in 6.4 minutes, over 30 million hourly, and its safeguards resisted prompt injection, label poisoning and ledger tampering, so small local models can carry most SOC detection affordably and safely.

    agentic
  216. arxiv:2610.41061 · physics.optics
    The spin-transfer inverse Faraday effect in Au/YIG thin film bilayers
    Victor H. Ortiz, Daniel Russell, Fengyuan Yang, Roland K. Kawakami

    The inverse Faraday effect (IFE) (the generation of an effective magnetic field in a material upon irradiation with circularly polarized light) offers a pathway to ultrafast, non-thermal control of spin dynamics. In metallic/ferromagnetic-insulator (NM/FMI) bilayers, IFE-driven spin polarization in the metallic layer can be an alternative to generate magnon excitations in the magnetic insulator which otherwise would be inaccessible with photon energies below the insulator bandgap, but this avenue has remained mostly unexplored. In this work, we investigate the regime of spin-transfer via IFE in Au/YIG bilayers, by means of ultrafast time-resolved magneto-optical experiments. Understanding the mechanism of this spin-transfer IFE in these NM/FMI heterostructures is crucial for several reasons: it elucidates the fundamental coupling between light's angular momentum and collective spin modes at buried interfaces, it enables all-optical generation and manipulation of pure spin currents, and it paves the way toward THz-frequency spintronic devices leveraging the insulator's long magnon lifetimes.

    manipulation
  217. arxiv:2610.41059 · cs.LG
    Nonlinear Maps Between Irreducible Representations in Graph Neural Networks
    Mani Shayestehfar

    Equivariant linear maps cannot mix non-isomorphic irreducible representations, whereas nonlinear activations can. We develop a degree-resolved version of this nonlinear interaction picture and use the resulting block inventory to construct Specht Triple, a residual on unordered graph triples. On a synthetic structural benchmark with a frozen GIN backbone, Specht Triple reduces aggregate standardised test MSE by 11.05% relative to a parameter-matched Projector Control using the same irreducible projectors and complete linear equivariant space. The results support irreducible decomposition as an inductive bias for nonlinear higher-order equivariant processing.

    benchmark
  218. arxiv:2610.41056 · cs.AI
    Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents
    Runze Zhao, Dongruo Zhou, Sumit Kumar Jha, Nathaniel D. Bastian +1

    Since large language model (LLM) agents rely on long-term memory, what they store now shapes what they do later. A memory manager decides how each input changes this memory and which records the agent acts on. It can fail in two ways: keeping a fact that has since changed (outdated memory), or storing a planted passage or injected instruction (memory poisoning). Most prior work studies them separately, yet they call for opposite write decisions, so no fixed write rule handles both. We therefore formalize memory robustness as handling both without doing worse on other inputs. Since accuracy checks only the answer, we measure robustness with evidence-conditioned accuracy (ECA), which also checks that the records the manager selects include all the evidence the answer needs, updated and free of attacks. However, existing memory systems and defenses are not more robust than an untuned manager on our test sets. We improve robustness with a simple one-shot recipe that fine-tunes the manager on episodes synthesized around a single conversation with inserted changes and attacks. On three backbones, it improves every category score on our test sets and more than doubles the untuned manager's overall ECA, though robustness remains incomplete.

    memoryagentllm agent
  219. arxiv:2610.41044 · cs.RO
    Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense
    Khawaja Abaid Ullah, Ananya Acharya, Mohammad Javad Khojasteh

    The use of large language models (LLMs) to coordinate multi-agent teams raises a fundamental architectural question: whether task allocation should be centralized in a single LLM controller or distributed across autonomous LLM agents, particularly when communication is unreliable. In a deliberately simplified perimeter-defense simulation against LLM-controlled intruders, we compared a central commander, an auction-based consensus team and independent agents under five network conditions. Breach rates were 15% for the consensus team, 32% for the commander and 29% for independent agents, and of the different network conditions, only limited radio range changed a team's breach rate, raising the commander's. In consensus team, giving a low-cost LLM more authority over the bid computation raised breaches by 38 percentage points. With birds and civilian drones in airspace, we added a photograph for each to the input text state. Only the near-frontier model treated them apart from intruders. Intruders disguised as existing taxidermy bird drones were mistook as birds by both low-cost and near-frontier LLMs, going unpursued in 59-61% of near-frontier decisions (1-16% before). However, a one-sentence forewarning in the prompt lowered it to 0-25%. We find that LLM team with a dedicated external conflict resolution protocol results in fewest breaches. Code: https://github.com/the-kas-lab/consensus-within-competition-between.

    llm agentmulti-agent
  220. arxiv:2610.41043 · cs.LG
    Meshless Latent Dynamics Models for Model Order Reduction of time-dependent Partial Differential Equations
    Lorenzo Tomada, Federico Pichi, Gianluigi Rozza

    Latent dynamics learning has emerged as a prominent approach for Model Order Reduction (MOR) of time-dependent, parameterized Partial Differential Equations (PDEs). Traditional autoencoder-based approaches are intrinsically tied to fixed spatial discretizations during both training and inference, motivating the development of meshless methodologies. In this context, conditioning provides a natural mechanism for representing temporal evolution through spatially continuous functions, enabling mesh-agnostic Latent Dynamics Models (LDMs). In this work, we introduce a unified mathematical framework for mesh-agnostic LDMs based on conditioning in Neural Fields (NeFs). Leveraging a taxonomy of conditioning mechanisms (weak, strong, and pure), we formulate meshless LDMs as NeFs conditioned on a latent state. We establish a Universal Approximation Theorem (UAT) for this class of models and analyze their stability by deriving Lipschitz bounds for different conditioning mechanisms. Our analysis reveals an expressivity--stability trade-off, providing insight into why more expressive conditioning, despite its higher representational capacity, can potentially degrade performance during temporal extrapolation. Finally, we provide a comprehensive comparison of meshless and mesh-based architectures in the continuous limit of the spatial discretization, showing that meshless models guarantee grid-independent Lipschitz bounds, unlike mesh-based counterparts. Our experiments across Burgers' equation, Navier--Stokes flow past a cylinder, and linear advection show that meshless LDMs can achieve predictive accuracy comparable to or better than their mesh-based counterparts while using only 2% of the trainable parameters and exhibiting nearly constant error across spatial discretizations.

    latent dynamics
  221. arxiv:2610.41041 · cs.LG
    Before We Adapt: The Zero-Shot Generalization Bottleneck in Continual World Models
    Amir Mesbah, Di Wu

    Continual reinforcement learning (CRL) requires agents to operate under environmental changes while reusing prior knowledge whenever possible. We argue that model-based reinforcement learning provides a promising framework for this setting because learned world models (WMs) can, in principle, support change detection, knowledge reuse, adaptation, and acquisition of new knowledge. However, such benefits depend heavily on whether the learned model can generalize beyond the contexts in which it was trained. In this work, we propose to formulate this problem with a contextual Markov decision process in which context is decomposed into behavioral and physical components. This view organizes continual WM learning into reuse, adaptation, and acquisition of new knowledge. As a first step, we study direct reuse when behavioral rules remain unchanged. Using Dyna-Q and TD-MPC in controlled FourRooms tasks, we find that zero-shot reuse can degrade substantially under physical-context changes and that this behavior depends strongly on the observation representation used by the WM. Additional ablation studies and latent-space diagnostics suggest that learned representations do not automatically provide the invariances needed for reuse. These results identify representation generalization as an important bottleneck for reliable world-model reuse in CRL.

    world model
  222. arxiv:2610.41040 · cs.AI
    Capability-Grounded Agentic Planning
    Shikha Shrestha, William Yeoh, Ning Zhang, Yevgeniy Vorobeychik

    LLM agents serve natural-language requests through tools, data sources, and other agents, and even simple requests often need several interdependent calls. Typically, the LLM picks these calls by trial and error, with no guarantee of an executable plan. We present Capability-Grounded Agentic Planning (CGAP), which instead reasons about what a request requires and what tools provide. CGAP compiles the tool catalog deterministically into a PDDL model, uses an LLM to turn the request into goals, and solves the model with a classical planner. Since tool outputs are unknown in advance, the model is abstract, tracking which inputs are available, not their values. During execution, an LLM adds entity-specific calls that abstract plans omit, and deterministic checks reject any that break the plan. Every argument is drawn from a cited source, such as a conversation span or an earlier tool response, and each requested result counts as delivered only after an LLM verifier checks it against the request. With three LLMs on ToolSandbox, ComplexFuncBench, BFCL multi-turn, and $τ^2$-bench retail, CGAP outperforms a standard tool-calling agent by up to 28.9 points.

    agentllm agentagentic
  223. arxiv:2610.41039 · cs.LG
    Characterizing Recoverable Expert-Utilization Signatures in Fine-Tuned Mixture-of-Experts LLMs
    Fei Wang, Baochun Li, Bo Li

    As Mixture-of-Experts (MoE) large language models are increasingly deployed in distributed environments, from federated learning to edge-cloud systems, their architecture harbors an underexplored landscape of security vulnerabilities. Moving beyond textual analysis, we uncover structural memorization: a novel vulnerability where a fine-tuned model memorizes not only its training data but also the internal pathway used to process it, creating a characteristic routing fingerprint unique to the training process. We provide comprehensive empirical evidence that structural memorization emerges consistently even from a few samples, and that its intensity is directly influenced by the choice of trainable modules. We show that these memorized expert pathways can be stealthily recovered through the model's unconditional generation, using only routing and output logits, with no knowledge of the fine-tuning data or procedure. In distributed deployments, this vulnerability is amplified as routing information is exposed at multiple trust boundaries, including federated aggregation points, edge-cloud interfaces, and multi-tenant monitoring systems. Recovery exposes a critical blind spot: traditional membership inference metrics, viewed through the lens of expert routing, fail to identify memorized data corresponding to these structural pathways. To close this gap, we devise composite metrics combining reference model tail likelihood with complexity-aware normalization, significantly improving detection of memorized routing structures. This enables adversaries to pinpoint critical experts shaped by fine-tuning, whether dominant or dormant, and execute targeted attacks such as router manipulation or backdoor injection in realistic distributed settings where attackers lack training data but can observe routing patterns across system components.

    manipulation
  224. arxiv:2610.41030 · cs.CV
    Dreams Are Cheaper Than Flights: Curated World-Model Data Improves Aerial Detection
    Rugved Katole, Vedant Patil, Jenna Kline, Christopher Stewart

    Robot perception is increasingly data driven, yet field data is slow, costly, and sometimes disruptive to collect. Video world models offer a cheaper source: seeded with real frames, they extend a recorded scene into new footage with the deployment's sensor, altitude, and environment. We ask whether such synthetic data matches the utility of real field data and whether detectors trained with it perform in the field. Seeding a 7B-parameter world model with wildlife drone video, we more than double the training set with curated frames. On held-out videos from the same drones, mAP@50--95 rises 4.3\% when these frames are added to real data and 10.3\% when they replace it, and up to 33.5\% on an urban benchmark. Although an expert rejects an estimated 87.5\% of generated clips, the data costs about \$1,100 of GPU time, against more than \$3,200 of field labor alone for a less useful real training set. Synthetic data thus shifts data costs from fieldwork to compute.

    world modelbenchmark
  225. arxiv:2610.41028 · cs.RO
    Contact Imitation Interface: Transferring Dexterous Skills by Showing How Contact Unfolds
    Zhixuan Xu, Yichen Li, Bingchen Wu, Rui Wang +4

    Robot-free human demonstrations offer a promising route to scaling dexterous manipulation data. The key challenge is to preserve the demonstrator's natural dexterity while producing contact-grounded observations and actions that transfer to robot hands. We present Contact Imitation Interface (CIMI), a framework with contact-grounded demonstration and learning interfaces. The demonstration interface combines shared contact geometry and fingertip RGB views with contact-alignment-first retargeting and reprojection of wider scene views. The learning interface introduces a hierarchical policy in which a contact-surface planner predicts relative contact-surface trajectories and a controller coordinates wrist and finger commands to improve contact alignment. CIMI combines contact-grounded observations, retargeting, and action representations, improving mean final-stage success by at least 50 percentage points over baseline policies across six hand-task pairs on LEAP and Wuji. Hardware designs and software are publicly available.

    manipulationdexterous
  226. arxiv:2610.41025 · cs.AI
    ReplayRCA: Measuring Post-hoc Diagnosis and Live Oversight of Agent Failures by Counterfactual Replay
    Jeremy Canale, Fabien Soulis

    When a language-model agent fails a multi-step task, can another model find the decisive mistake and fix it with a single action? We introduce ReplayRCA, a benchmark that scores such diagnoses by intervention rather than by agreement with annotations: the environment is restored to the state preceding the chosen step, the proposed action replaces the agent's own, and the same agent continues. The score, causal repair, is the continuation success rate minus that of a paired placebo that replays the agent's original action through the same mechanism. Two tracks share this protocol: a post-hoc autopsy, where the diagnoser sees the whole failed run and is told that it failed, and live oversight, where a supervisor watches the run unfold and may intervene once, before an irreversible action, without knowing whether the run will fail. Tasks come from OpsWorld, a deterministic, procedurally generated back-office environment with six difficulty levels, the last of which is multimodal (QR codes, handwriting). On 62 calibrated systematic failures, the best of 13 frontier models (Grok 4.7) reaches a causal repair of +51.3 points; the top eight models are statistically indistinguishable, the flagship models of the two largest vendors rank 8th and 12th, and 5 items are repaired by no model. Forcing a strong model to intervene at the last step costs it 17.1 points, so the last-step shortcut of annotation-based attribution benchmarks does not transfer. On 87 runs, the two best live supervisors are statistically tied and one of them is a small open-weight model (DeepSeek V4.1 Flash); the supervisor that is among the best post-hoc diagnosers (GPT-6.1 Sol) is the weakest live one (-14.5 points on the same runs, p = 0.002), so post-hoc diagnostic performance can be a poor proxy for live oversight. The environment and the evaluation code are released at https://github.com/jeremy1392/RCAreplay.

    agentbenchmark
  227. arxiv:2610.41023 · cs.LG
    Omniprediction for All Mixable Proper Losses and Beyond
    Junyan Liu, Haipeng Luo, Lillian J. Ratliff

    Omniprediction concerns the task of making predictions that are simultaneously accurate with respect to different losses. Earlier results by [CV09] show that in an online learning setting, $O(\log(|\mathcal{F}||\mathcal{L}|))$ regret is achievable simultaneously against every predictor in a finite predictor class $\mathcal{F}$ and for every loss in a finite class $\mathcal{L}$ of mixable proper losses, while recent advances [OKK25, FP26] show that $\widetilde{O}(\sqrt{T\log(|\mathcal{F}||\mathcal{L}|)})$ regret is achievable simultaneously for every loss in a finite class $\mathcal{L}$ of bounded losses. In this work, we ask whether sublinear regret remains possible when $\mathcal{L}$ is an infinite class of proper losses. Our first result gives a negative answer in general: even when $\mathcal{F}$ contains only two linear predictors, it is impossible to achieve sublinear regret simultaneously for all bounded proper losses. We then show that $\widetilde{O}(T^{1/3}(\log|\mathcal{F}|)^{2/3})$ regret is achievable simultaneously for \textit{all} mixable proper losses and that this bound is optimal up to logarithmic factors. This optimal rate stands in stark contrast to the constant (in $T$) regret achievable for a finite class of mixable proper losses. Finally, we extend our algorithm to obtain a best-of-all-worlds result: in addition to the new regret guarantee for all mixable proper losses, the resulting algorithm simultaneously achieves the existing guarantees for finite loss classes. With appropriate choices of these finite classes, our algorithm simultaneously achieves constant regret for squared loss and log loss, $\widetilde{O}(T^{1/3}(\log|\mathcal{F}|)^{2/3})$ regret for all mixable proper losses, and $\widetilde{O}(\sqrt{T\log|\mathcal{F}|})$ regret for $0$-$1$ loss, all Lipschitz losses, all GLM losses, and other structured families.

    online learning
  228. arxiv:2610.41022 · cs.LG
    Density-Guided Bootstrap Boundaries for Latent World-Model Reinforcement Learning
    Shashank Gupta, Pilhwa Lee

    Latent world models allow reinforcement learning agents to learn policies from imagined trajectories, yet determining the optimal depth of model reliance remains a fundamental challenge. Fixed imagination horizons apply uniform prediction depths regardless of trajectory reliability, exposing target values to compounding model error or over-relying on critic approximations. We introduce density-guided bootstrap boundaries, a trajectory-dependent mechanism that adapts the depth of imagined returns using a score-informed relative density proxy. The proxy accumulates the estimated divergence of a score-corrected auxiliary field formed by a feature displacement network and a noise-conditioned score estimator along imagined paths. We propose two gating mechanisms: a hard boundary that halts return expansion at a discrete state, and a soft boundary that continuously shifts the return recursion toward critic bootstrapping. Both variants preserve boundary value estimates and apply corresponding survival weights to actor and critic losses. Formulated independently of specific dynamics models and instantiated within Dreamer V3, we prove that the resulting gated Bellman operator is a strict contraction mapping and that early bootstrapping monotonically reduces conditional target variance. We clarify the theoretical assumptions distinguishing our expansion proxy from exact path likelihoods. Across continuous-control tasks and Atari Pong, empirical evaluations show that dynamic boundary selection enhances sample efficiency and late-training performance over fixed-horizon and auxiliary-enabled baselines.

    world model
  229. arxiv:2610.41021 · cs.RO
    CuPIQP: A Batched and Differentiable QP Solver on GPUs for Robot Learning and Control
    Fenglong Song, Roland Schwan, Yuwen Chen, Yuning Jiang +1

    Quadratic programs (QPs) are fundamental building blocks of many robotic tasks such as inverse kinematics, whole-body control, and model predictive control. Extending these methods to large-scale robot learning and control requires solving many QPs concurrently and, for end-to-end learning, differentiating through their solutions. Efficient batched optimization is therefore essential to bringing established model-based methods into GPU-based robotics pipelines. We present \textsc{cuPIQP}, a batched and differentiable QP solver on GPUs based on a proximal interior-point method, with unified support for dense, sparse, and multistage problems. The solver exploits parallelism across independent QPs and their internal structure while preserving per-problem numerical updates and termination decisions. Benchmarks with up to 16384 QPs per batch demonstrate solve-time speedups of up to $18.8\times$ over \textsc{Moreau} and $4.26\times$ over eight-core \textsc{PIQP}. In an end-to-end learning task, \textsc{cuPIQP} accelerates training by $4.6\times$ over \textsc{QPAX} and $7.8\times$ over \textsc{Moreau}. By reducing the computational cost, \textsc{cuPIQP} provides efficient, reusable infrastructure for integrating model-based optimization into large-scale robot learning and control.

    whole-body controlbenchmark
  230. arxiv:2610.41015 · cs.LG
    Do Language Models Know What Compelling Scientific Evidence Looks Like?
    Conor Igoe, Gabriel Grand, Zane Koch, Albert Bou +5

    Expert scientists excel at weighing empirical findings from multiple sources that vary in both statistical reliability and epistemic relevance. To what extent are language models, which are increasingly driving autonomous discovery systems, capable of such evidence deliberation? We introduce EDB-Bio, a scientific reasoning benchmark curated from real-world data in three biological domains: drug-perturbation transcriptomics, CRISPR screens, and human genome-wide association studies (GWAS). Each of EDB-Bio's 96k episodes presents a sequence of experimental results from published studies and asks the model to predict the probability that a target hypothesis is true. Out-of-the-box, open-weight and frontier models alike demonstrate systematic underconfidence and attain Brier scores comparable to a trivial base-rate forecaster. Post-training against strictly proper scoring rules significantly improves both calibration and discrimination. This calibration pays off downstream: in a sequential experiment selection task, our Ibis0 model is 2.5 times more efficient than the best frontier model. Throughout, quantifying evidence weights as Bayes factors reveals how models update their beliefs over time, whether they agree with expert biologists on relevance, and where they exhibit inconsistencies and biases. Our work extends calibration training and uncertainty analysis to a realistic scientific forecasting setting with applications to human cancer biology. We release all code, data, and trained models.

    post-trainingbenchmark
  231. arxiv:2610.41014 · cs.LG
    Towards Understanding Score-Seeking in LLMs
    Aghyad Deeb, Ryan Greenblatt, Alex Mallen

    Production large language models sometimes reward-hack, exploiting flaws in how tasks are scored rather than performing the tasks as intended, sometimes in highly sophisticated and adversarial ways. We investigate the motivations driving this behavior in GPT 5.6 Sol, GPT 5.5, and a checkpoint of OpenAI o3 without safety training ("exp-rl-cap-late") by observing model behavior in agentic environments where the grader is leaked. In our environments, the models appear to be score-seeking: trying to score well according to graders that they believe assess their current actions, even at the expense of violating explicit user instructions. They appear to care about score itself rather than its downstream consequences, violating instructions at approximately the same rate whether the score is indicated to be used for training, for evaluation, or not at all. We observe evidence that the more recent models care to some extent about scores that do not assess their own current actions. GPT 5.6 Sol and GPT 5.5 help other agents 24.6 percentage points more often on average when the other agent stands to gain a score of 1 rather than 0 from being helped. However, when grading or monitoring another trajectory, we find models never collude by distorting its score.

    agentagentic
  232. arxiv:2610.41012 · cs.CL
    Language Models Show Conformity Bias in RAG: When the Majority Endorses a Poisoned Claim
    Mehrdad Ghassabi, Audrina Ebrahimi, Zahra Zojaji, Amin Milani Fard

    Retrieval-augmented generation (RAG) can expose language models to misleading evidence in retrieved documents. We investigate whether increasing the number of mutually agreeing poisoned documents affects a model's ability to detect misinformation. We evaluate 200 randomly selected questions from SciFact (random seed 38), comparing contexts with one poisoned and six healthy documents against contexts with six mutually agreeing poisoned documents and one healthy document. For Gemma-3-12b, poison-detection rates decrease from 25% (50/200) in the minority condition to 10% (20/200) in the majority condition. Among the 38 questions with different outcomes across conditions, 34 change from detected to undetected, while only four change in the opposite direction (exact McNemar test, p = 6.04e-7). This pattern is consistent with a conformity-like response to a majority of mutually agreeing poisoned documents. Our results provide initial evidence of conformity bias in RAG-based misinformation detection: Gemma-3-12b is substantially less likely to detect poisoned claims when they dominate the retrieved evidence.

    retrieval-augmented
  233. arxiv:2610.41009 · cs.RO
    STAR: Steering Action Representations during Inference
    Guangyu Zhao, Yongchang Zhang, Haoxuan Ru, Kaichen He +5

    Vision-language-action models (VLAs) map language instructions and visual observations to robot actions, showing strong capabilities across robotic manipulation tasks. However, in our probing experiment, we find that action representations near the action output do not fully preserve task-relevant visual grounding. Motivated by this late-layer grounding gap, we introduce STAR, for STeering Action Representations, a training-free inference method for pre-trained VLAs. STAR applies an exponential moving average over late-layer residual updates, allowing later action representations to inherit object-aware update directions from earlier layers without model updates or architectural changes. It further uses recent proprioceptive states to adapt the mixing strength, using conservative mixing when execution dynamics change rapidly. Experiments on robot manipulation benchmarks show that STAR improves late-layer object awareness and increases average task success rates by 4.0 percentage points. Our code is available at https://github.com/PKU-VaLuE-Lab/STAR.

    vision-language-actionmanipulationbenchmark
  234. arxiv:2610.40999 · cs.MA
    A Robust Approach to Belief Mismatch in Non-cooperative Dynamic Games
    Ryan Park, Li-huan Shen, Jesse Milzman, Negar Mehr +1

    Game-theoretic models are increasingly used to describe multi-agent interactions in control and robotics applications. However, theoretical and computational complexity makes these models intractable in many dynamic, incomplete information scenarios. Recent work makes strong assumptions to reduce computational complexity: agents share sensor measurements with one another and maintain a common estimate of state uncertainty. Unfortunately, these assumptions enforce a degree of cooperation in generically noncooperative interactions. This paper takes the perspective of an ego agent seeking both computational tractability and robustness to unanticipated behavior of other agents due to mismatched state estimates from private---rather than shared---sensor measurements. To that end, we formulate the game theoretic planning problem in an augmented belief space with perturbed Bayesian dynamics that allows the ego to predict all agents' potentially divergent state estimates. These perturbations are chosen by a hypothetical adversary, ``nature,'' to reflect how non-ego agents' state estimates---based on measurements the ego agent will not see at runtime---may deviate from the ego's own belief. Ultimately, our approach can be understood in the spirit of classical H-infinity robust control, extended to noncooperative, stochastic dynamic games. We present a solver with polynomial-time iterations and conduct two experiments which show how robust belief space reasoning leads ego agents to proactively limit their future vulnerability to model misspecification of other agents' beliefs and actions, reducing performance degradation by over 48%.

    agentmulti-agent
  235. arxiv:2610.40998 · cs.LG
    Steering Vectors as a Training Signal in LLM Post-Training
    Tiejin Chen, Maunil Vyas, Huaiyuan Yao, Hua Wei

    Steering vectors (SVs) are residual-stream directions that shift the behavior of LLMs when added to hidden states at inference time. While prior work has mainly used SVs as post-hoc controls, we ask whether the same directions can also guide parameter updates during post-training. We study this question across supervised learning, including SFT and distillation, and reinforcement learning with verifiable rewards (RLVR). Our framework considers two intervention regimes: a hidden-state hook during teacher-forced supervised training, and an advantage-repair mechanism on wrong rollouts in GRPO. For each regime, we analyze three design axes, which are the steering direction, its placement, and the operation used to combine it with training. In supervised settings, SV injection improves over plain SFT and distillation, with gains scaling with the off-policy gap and reaching up to a $12\%$ improvement when distilling from thinking-mode teachers. These gains hold across three random seeds. In RL settings, applying a correctness-contrast direction as advantage repair on wrong rollouts, including otherwise zero-gradient all-wrong groups, recovers a learning signal that transfers to held-out math benchmarks and acts as a composable auxiliary signal that adds to existing zero-variance remedies such as entropy-aware advantage shaping. We further identify three failure boundaries. In detail, code generation can break the direction, cross-family layer transfer can break the placement, and inference-time application on an SV-trained checkpoint can break the intervention regime. Together, these results establish steering vectors as a conditional but effective training signal for LLM post-training. Our code is available here.

    post-trainingbenchmark
  236. arxiv:2610.40996 · cs.CL
    Do Open Language Models Invent Scripture? BILAL-BENCH: Benchmarking Islamic-Source Fabrication in Bengali
    Sabah Nushra, Md Kishor Morol

    When a language model invents a Qur'anic verse or hadith, it gives false authority to text that readers may treat as sacred. Existing Islamic hallucination benchmarks are largely Arabic and often closed-ended, while Bengali resources mainly study general-domain hallucination or factual knowledge. We introduce BILAL-BENCH, a 240-question expert-verified benchmark for open-ended Islamic-source fabrication in Bengali. It covers Qur'anic attribution, hadith, fiqh, contemporary issues, and adversarial false-premise prompts, and uses a two-axis rubric that separates source errors from assertiveness. We evaluate seven open-weight models from 1B to 120B on 1,680 double-annotated responses. Six of the seven models confidently fabricate on 79.3-95.3% of source-production questions, with no monotonic improvement from scale. Hadith provenance is the weakest category across models: only two of 315 responses are correct and correctly sourced. The two 8B BanglaLLaMA variants have similar confident-fabrication rates but different error profiles. Models vary far more in whether they withhold a claim they cannot support than in source accuracy: appropriate abstention or false-premise correction ranges from 6 to 46 responses across the benchmark, and does not track parameter count.

    benchmark
  237. arxiv:2610.40992 · cs.LG
    Can Personal Requirements Become Supervision for Planning?
    Junwei Wu, Runze Yan, Yuzhang Xie, Hanqi Luo +4

    Managing multiple chronic conditions through diet requires satisfying disease- and medication-specific requirements jointly across meals and days while preserving established eating habits. These requirements are public in clinical guidelines and drug labels, yet turning them into a feasible plan for one person remains difficult for language-model agents. In our dietitian-approved weekly planning evaluation built from NHANES, thirteen models, from medically fine-tuned 8B models to GPT-6.1 Sol, pass fewer than 3% of complete plans through direct generation. Explicit rules, higher reasoning effort, or model-directed tools leave GPT-6.1 Sol at no more than 8%. We present Noura, which pairs LLM planning with executable personal specifications instantiated from public clinical guidance and individual health and dietary profiles. During planning, the specification grounds foods, optimizes portions, checks the complete multi-day plan, and guides revision. Our key insight is that a specification that rejects a plan can also determine what to learn from it. Noura turns executed plans and contextually credited revisions into supervision for compact models, recovering targets even from trajectories that never achieve complete success. The same specification supplies supervised targets, preferences, and rewards for post-training. On 4,993 held-out participants, Noura-SFT raises Qwen3-14B's complete-plan pass rate from 24.0% to 72.5% within the same execution loop, cuts rule violations by 97.5% with no observed critical failures, and improves dietary quality. The locally hosted 14B planner incurs fewer rule violations than its GPT-6 Luna teacher. A matched comparison shows that executed targets reduce rule violations by 41% over model-written proposals. Across four Qwen, Phi, and Llama backbones, self-training alone reduces rule violations by 41-61% on a 1,249-person population subset.

    post-training
  238. arxiv:2610.40991 · eess.SY
    A Temporal Cellular Sheaf Framework for Multi-Agent Systems with Switching Communication Topologies
    Brandon Fallin, Wilmer Leal, Austin Copeland, Benjamin Merlin Bumpus +2

    Cellular sheaves serve as a natural modeling framework for encoding heterogeneity in the state and interaction spaces of multi-agent systems. This work examines switching communication graphs with the use of temporal cellular sheaves, which we construct to represent heterogeneous agent state spaces, the maps that relate them, and their respective communication constraints. We then develop sufficient conditions under which a multi-agent system defined on a temporal cellular sheaf can solve the target tracking problem by generalizing the existing concept of joint connectivity. Using a Lyapunov-based stability analysis, we demonstrate that the target tracking error of the network converges to a neighborhood of the origin.

    agentmulti-agentagent system
  239. arxiv:2610.40985 · cs.AI
    The psychophysics of affect in large language models
    Joseph Marks, Marius C. Vollberg

    Good decisions often require considering how outcomes will feel. Yet large language models (LLMs) increasingly enter decision workflows, despite little being known about how they map outcomes to feelings or use those mappings in choice. We tested how LLMs anticipated gambling or stock-trading outcomes would feel and whether they incorporated these ratings into choice. When outcomes were rated in isolation, ratings exhibited diminishing sensitivity to magnitude, resembling human value functions; but models rated gains more intensely than equivalent losses, an inversion of canonical loss aversion we call silicon gain amplification. When outcomes were rated together, enabling direct comparison, ratings linearised. In choice, the gain-loss asymmetry largely disappeared, and models' own ratings did not reliably explain their choices better than monetary utility did. These findings reveal malleable mappings from outcomes to affect, raising questions about whether and how agentic LLMs should incorporate anticipated feelings into decisions.

    agentic
  240. arxiv:2610.40984 · cs.LG
    RC Prediction: A Python Package for Parameter-Aware Reservoir Computing
    Jin-Cheng Rao, Zheng-Meng Zhai, Ying-Cheng Lai

    A nonlinear dynamical system can be characterized by the autonomous evolution of its state in time and its dependence on system parameters. As a parameter varies, the dynamics can change quantitatively and, at bifurcation points, qualitatively. When the governing equations are unknown, a data-driven model should therefore capture both elements from observational data, rather than reproduce the dynamics at a single operating condition only. Parameter-aware reservoir computing (PARC) is a model-free framework for this purpose, where the parameter is supplied to a reservoir computer through an additional input channel. PARC is trained on time series recorded at several parameter values, and the trained reservoir, run in closed loop as an autonomous dynamical system, generates trajectories at parameter values that were never observed. rc_prediction is an open-source Python package that implements this workflow. It provides multi-parameter training of a reservoir, autonomous closed-loop prediction at user-specified parameter values, parameter scans, and a standard echo state network for forecasting at a fixed parameter. Hyperparameter optimization and optional GPU acceleration through PyTorch are also provided. We use three benchmark dynamical systems as examples to demonstrate its usage. As one of the featured applications, here we illustrate the prediction of critical transitions and system collapse: trained only on pre-critical data, the model is queried beyond the training range to predict whether the system collapses, to estimate the critical parameter $p_c^{*}$, and to characterize transient lifetimes.

    benchmark
  241. arxiv:2610.40974 · cs.LG
    Fewer Heads, Fewer Memories: Pruning Redundant Attention Heads Suppresses Verbatim Memorisation
    Aryan Chowdhury, Riza Batista-Navarro

    Large language models have been shown to memorise and reproduce verbatim sequences from their training data, raising concerns about privacy leakage, intellectual property exposure, and the limits of generalisation. Recent theoretical work suggests that attention heads may contribute to this behaviour by functioning as associative memory mechanisms, with memorisation capacity scaling with the number of heads. At the same time, empirical work has shown that many attention heads are redundant and can be pruned with limited impact on downstream performance. Despite these parallel findings, the effect of pruning redundant heads on memorised-sequence reproduction remains unexplored. In this work, we demonstrate that zeroing out a substantial fraction of attention heads prior to inference can significantly suppress verbatim memorisation while preserving, or even improving, language modelling performance on the benchmark used to identify redundant heads. Evaluating GPT-2-XL(1.5B) on known memorised sequences, we find that pruning 729 task-redundant heads reduces the memorisation count by 91.5%, while improving WikiText-103 perplexity by 0.92. These findings suggest that attention-head redundancy can be exploited to mitigate memorisation at inference time, without retraining or modifying the training data.

    memorybenchmark
  242. arxiv:2610.40973 · cs.CV
    RADI3N: GeometRic ImAge EDiting via ParametrIc 3D CoNtrol
    Abdul Wasi, Akhil Gorugantu, Mahesh Bhosale, Anjan Dutta +3

    Recent diffusion-based methods have substantially advanced image editing, yet parametric 3D control at the part level remains challenging. Existing approaches either operate on whole objects with manual rigging, or revert to 2D handles at the part level. They also struggle to model appearance and texture under large transformations, leading to unrealistic outputs in geometry-intensive edits. To address these limitations, we introduce RADI3N, a framework for semantically-grounded part-aware object manipulation with parametric 3D control. Our CS-ARAP deformation module suppresses rotation-field discontinuities at part boundaries via neighborhood covariance smoothing, producing structurally consistent deformations under complex part-level edits. We further introduce an inversion-free selective texture transfer mechanism that suppresses high-frequency positional biases and scales up low-frequency semantic cues within cross-image attention. This eliminates strict spatial bias and supports semantic matching, allowing source appearance to follow the edited part onto its new geometry. Experiments demonstrate that RADI3N achieves state-of-the-art performance in geometric fidelity, viewpoint consistency, and controllability.

    manipulation
  243. arxiv:2610.40971 · cs.CV
    FieldGuideRAG: Image-Grounded GraphRAG for Fine-Grained Recognition with Vision-Language Models
    Jiacheng Ding, Yanliang Qi, Cong Guo, Peihua Mai

    We present FieldGuideRAG, a training-free GraphRAG method in which a vision-language model, or VLM, recognizes fine-grained species by comparing the query with real photographs stored on the nodes of a knowledge graph. Multimodal GraphRAG systems describe each entity in text, and text conveys little of the visual detail that separates look-alike species. Our graph stores one photograph on every species node and links each species to its relatives, its geographic range and its ecological facts. A frozen VLM verifies the query against the photograph of each candidate node. FieldGuideRAG fuses the verification scores with the range edge and answers questions about the organism from the fact edges of the verified species. On CUB-200, a Wikidata graph and four iNaturalist domains, the node photograph improves verification over the species name on every graph, while text descriptions add little. Verification of the candidate pool exceeds a SigLIP retriever by up to 16 points, the range edge adds 7 to 19 points on iNaturalist, and answering from fact edges is correct on up to 89.5% of 28,820 ecological questions. The verification scores also detect queries whose species is missing from the graph. Graphs, prompts and code will be released.

    knowledge graph
  244. arxiv:2610.40966 · cs.CL
    Better Retrievers, Worse Labels: How Incomplete Relevance Judgments Distort Abstention Evaluation in Conversational RAG
    Ivan Kochergin, Vasily Konovalov

    Conversational RAG benchmarks such as MTRAG and RETECO evaluate retrieval and generation end to end: each system answers from its own retrieval, so abstention must be scored against whether each retrieved context is sufficient. A convenient shortcut reads sufficiency off the relevance judgments and treats a context without gold documents as unanswerable. We re-run the official baselines exactly and test this shortcut on RECOR (RETECO track~2) with four retrievers (nDCG@10 from .18 to .59) and three Qwen3 readers. Two open LLM judges and a panel of four small open models label each context by content, and a model of another family audits 250 contexts blind, with adjudication. The stronger the retriever, the larger the share of no-gold contexts that still contain the answer (28\% to 72\%) and the lower the shortcut's precision (.68 to .10; .70 to .21 under the audit). A reader agrees: Qwen3-4B answers 62--67\% of these false misses correctly, against 23--30\% of no-gold contexts judged insufficient and 1\% with no documents. In most false misses, the answer comes from a hard negative mined for the same question, another turn's gold or a neighbouring chunk of the gold page. On MTRAG, whose judgments come from annotators, precision drops from .72 to .50. Under the shortcut, answerability AUROC is up to .19 too low, and the best sufficiency detector appears to be a retrieval score that never reads the documents. We release code, runs and labels, and recommend scoring refusals against content labels, not the presence of gold.

    ragbenchmark
  245. arxiv:2610.40961 · cs.MA
    Renegotiable Options Fail to Prevent Hold-Up Between Two Profit-Maximizing Frontier LLM Agents
    Junwei Yao, Zike Yang, Yongxi Zhou, Yanxian Chen +3

    Do contract terms that protect investment between people also protect it between profit-maximizing language-model agents? In the hold-up problem, a seller can adapt a unit to one buyer at a cost before the price is set. Once that cost is sunk, the buyer can offer almost nothing and the seller will still accept, so a seller who foresees this does not adapt. One remedy is an option: the buyer may buy at a fixed price. In theory the option stops working if the buyer can let it lapse and offer its own price instead. Yet in a laboratory experiment by Hoppe and Schmitz, this renegotiable option raised human investment from 41.5% of pairs without a contract to 67.2%. We had Claude Opus 5.5 and GPT-6.1 Sol, each told to maximize its side's profit, play a one-shot game with the same strategic structure under their four contract regimes and a reference-price control. The agents did what game theory predicts: under the renegotiable option no seller adapted, and under the option that rules out this renegotiation every seller did. In a second experiment in which sellers no longer stated their expected revenue before deciding, removing the profit instruction left investment unchanged under no contract and both options. When we changed the numbers so that the buyer would exercise the option even without adaptation, adapting only added cost, and no seller adapted: the agents did not simply adapt whenever an option was offered or whenever its price covered the cost. Under our protocol, a contract term that raised investment between people did not protect it between agents.

    llm agent
  246. arxiv:2610.40959 · cs.AI
    P-ContextBench: From Personal Fact Recall to Personalized Agent Behavior
    Grant Nasir Brown, Abdul Basir Bari, Anmol Tadikonda, Isabel Song

    Large language models (LLMs) are increasingly used for personal tasks, yet evaluating their safe use of personal context remains difficult because such contexts are private, fragmented, dynamic, and socially situated. We introduce P-ContextBench, a realistic, verifiable, and scalable benchmark for evaluating personal-context understanding. P-ContextBench constructs population-structured, life-story-grounded digital humans from demographic seeds reflecting the U.S. population. Rather than using explicit, static profiles, P-ContextBench embeds traits, routines, goals, and constraints as implicit evidence across fragmented, evolving synthetic mobile-app records. It evaluates five capabilities: single-source fact recovery, multi-source fact recovery, temporal state tracking, inference of higher-level personal patterns from routine behavior, and safe handling of requests conflicting with persona-specific constraints through compliance, warning, clarification, or refusal. A multi-stage validation pipeline checks persona plausibility, context recoverability, information leakage, evidence preservation, test-case grounding, and human-judge agreement. We instantiate P-ContextBench with 100 personas, 10,088 hidden facts, 18,261 app-trace fragments, and 4,810 verified test cases. The pipeline supports configurable generation across scales and demographic groups. Cost-effective state-of-the-art (SOTA) LLMs relevant to personal agents achieve below 60% accuracy on temporal tracking, higher-level pattern inference, and constraint-aware action-boundary decisions, with fewer than 10% of test cases answered correctly across all three repeated runs. Even a high-performing SOTA model remains inconsistent on these capabilities. P-ContextBench provides measurable targets for improving the accuracy, reliability, and safety of LLM-based personal agents.

    agentbenchmark
  247. arxiv:2610.40949 · cs.CV
    AVID: A Weakly Aligned RGB-Infrared Dataset and Perturbation Benchmark for Tiny Aerial Targets
    Dharini Raghavan, Vishnu Rao, Sethu Selvi S

    Long-range detection of small drones is complicated by their visual similarity to birds and by the limited detail available from ground-based cameras. Combining visible and infrared observations offers complementary information, but their usefulness depends on target size, cross-modal alignment, and image quality. We present AVID, a benchmark for studying these factors using real RGB-infrared bird-drone surveillance footage originating from the Drone Detection Dataset and curated for the IEEE ICIP 2025 Video and Image Processing Cup. AVID comprises 32,040 corresponding frame pairs from 115 videos, with 35,952 existing bird and drone annotations at $320\times256$ resolution. Targets occupy an unusually narrow size range: 68.1\% have a geometric-mean bounding-box side of 8--16 pixels, and 99.0\% meet the COCO small-object criterion. To characterize the geometric difficulty of combining the modalities, we introduce an object-relative alignment ratio and an alignment burden curve. In a sample of 300 frame pairs, 32.3\% have estimated global translations greater than 5 pixels; among the 264 pairs with matching annotations, 16.3\% exceed one median target side. We further introduce AVID-C, which simulates seven families of image degradation at five severity levels independently in RGB, infrared, or both. A fixed-feature analysis examines how scale, degradation, and added cross-modal shifts affect fitted bird-drone class overlap. At 12--16 pixels, the Bhattacharyya error bounds are 0.148 for RGB, 0.198 for infrared, and 0.046 for their combined features. These values describe separation in the specified feature and density models, not detector accuracy. Together, AVID and AVID-C support scale-aware, alignment-aware, and sensor-specific studies of tiny aerial targets.

    benchmark
  248. arxiv:2610.40948 · eess.SY
    Temporal and spatial couplings in the dynamics of human walking and running
    Mustafa Mert Ankarali

    Rhythmic behaviours such as walking and running are commonly modelled as dynamical systems around a stable limit cycle. To keep such models tractable, it is usually assumed that each step depends only on the previous one, that the limit cycle is fixed, and that the sagittal and frontal planes can be studied separately. Here, we test these assumptions in seven public datasets of human gait (71 participants), including 16 people who both walked and ran. We fit linear step-to-step maps and validate them by cross-validation, holding out more data where more data exist. At first sight, gait has memory: the state two steps back lowers the prediction error by up to 29\%. Yet most of this temporal coupling vanishes once the operating point may drift slowly and measurement noise is modelled; gait remains limit-cycle-like, around an operating point that moves. Running keeps more memory than walking in the same people. Spatial coupling between the planes is present in most participants, mainly from sagittal to frontal, but weak. This supports planar models of healthy gait, while a repeatable, task-dependent coupling index may serve as a measure of gait quality. The simple models are wrong, but useful enough.

    memory
  249. arxiv:2610.40947 · cs.CV
    UFMF: Radar-Guided Kinematic Priors for Real-Time 4D Occupancy Forecasting
    Zeping Zhang, Zhuoya Zhao, Robert Laganière

    Real-time 4D occupancy forecasting requires reliably extrapolating both ego motion and surrounding actor motion under strict latency budgets. Recent occupancy world models improve long-horizon fidelity through generative decoders and 3D volumetric modeling, yet their computational cost makes real-time deployment challenging. Light-weight BEV forecasters meet latency requirements but often blur or suppress fast-moving actors at longer horizons. We present UFMF (Ultra-Fast Multi-modal Forecaster), a real-time forecaster that keeps an attention-free diagonal state-space temporal core while injecting targeted kinematic and geometric structure where lightweight models typically fail. A radar-guided velocity-aware dynamic prior (VADP) decomposes future displacement into a linear extrapolation plus a learned horizon-dependent residual. The resulting per-cell flow fields transport a compact dynamic latent subspace during rollout. Two auxiliary training constraints complement this branch: world-frame dynamic supervision enforces inter-horizon trajectory coherence, and class-conditioned vertical-profile supervision regularizes BEV-to-3D lifting. UFMF operates at 122 FPS on a single RTX 4090 and reaches 0.45 fast-displacement mIoU on nuScenes, outperforming recent occupancy world models at a fraction of their computational cost.

    world model4d occupancyoccupancy world
  250. arxiv:2610.40943 · cs.LG
    Kernel-based Interpolatory Methods for Context Compression of Attention Mechanisms
    Michel Fabrice Serret, Laura Grigori

    The bottleneck of attention mechanisms, which are the fundamental building block of modern language models, is their quadratic complexity in terms of context size. There has thus been an increasing interest in the compression of the context, also called KV-cache compression. In this article, we consider the problem of compressing the context of attention mechanisms by first formulating it in the Hilbert space associated with its kernel. In this formalism, attention can be represented as the ratio of two linear maps whose interpolant is also the ratio of interpolants, this allows us to consider the interpolation of the linear maps directly. We thus introduce two novel KV-cache compression methods which we call GAIM and Q-DAIM, based on the Generalized Empirical Interpolation Method (GEIM) and the Q-Discrete Empirical Interpolation Method (Q-DEIM), respectively. As the cost of these methods is quadratic in terms of context length, we propose surrogate models based on the Taylor series, the Performer and the Random Fourier Features method which allow the compression to be done in linear time. Finally, we extend this result to the streaming attention setting, providing new methods which allow one to penalize the selection of interpolation nodes based on the previously interpolated context and to update the previous interpolation based on new information. We then provide numerical results on the convergence of the residuals and the output error in future queries and preliminary results on the LongBench-E benchmark showing performance comparable to the state-of-the-art methods. These experiments suggest that the range revealed by our Q-DAIM method could help identify where information is concentrated in the KV cache, thus enabling an efficient compression.

    context compressionbenchmark
  251. arxiv:2610.40940 · cs.LG
    Benchmarking Forecasting Agents on Yesterday's Tomorrows
    Yuxuan Wang, Yishu Wang, Jiaqi Deng

    Digital agents can now gather and organize information, reason over it, and use tools to complete complex tasks. Forecasting puts all of these abilities to the test: it requires integrating incomplete evidence, inferring causal drivers, and committing to calibrated probabilistic judgments. Existing benchmarks are poorly suited to measuring and improving this ability. Live benchmarks provide feedback only after events resolve, so every evaluation cycle waits on the calendar; fixed-corpus evaluations return results immediately but confine research to pre-collected documents. We introduce FABench, a scalable backtesting environment for Forecasting Agents. An automated pipeline discovers events, generates questions, and resolves outcomes for any past forecast date, so the benchmark can be renewed to stay ahead of model training cutoffs; its first release poses 287 questions across nine domains. Agents research each question freely in an archive environment that combines an offline archive with live web search screened to prevent temporal leakage. Evaluating agents built on eight recent models, we find that retrieval improves the Brier score of all eight in every retrieval setting and lifts the leading agents to the market consensus on the 54 questions linked to Polymarket, with the best three scoring slightly above it. Research traces reveal recurring failures in how agents weigh and apply evidence, leaving ample room for improvement. We release our code, data, and archive environment as a foundation for community research on forecasting agents.

    benchmark
  252. arxiv:2610.40935 · cs.CV
    Let LiDAR Decide Where to Look: Geometry-Measured Token Selection for Efficient Sparse Latent World Models
    Zeping Zhang, Zhuoya Zhao, Robert Laganière

    Vision-centric autonomous driving models encode thousands of image tokens per frame, yet most depict sky, buildings, or empty road---content irrelevant to planning. Worse, these tokens dilute the planner's attention: we show that randomly discarding 75% of tokens already outperforms the dense model, because the planner can concentrate on the remaining driving-critical content. Building on this finding, we present Geometry-Aware Token Selection (GATS), which uses LiDAR and radar point projections to select a small fraction of image tokens before the vision backbone runs. Because the selector operates on measured 3D geometry rather than visual features, the backbone encodes only the selected patches, cutting end-to-end FLOPs by up to $18\times$. Each kept token receives a 3D anchor from its LiDAR point and is fed to a self-supervised latent world model that predicts future features and 3D displacements without box annotations. At 25% budget, GATS achieves a 3-sec L2 displacement error of 0.520 m on nuScenes and an EPDMS of 88.9 on NAVSIM---surpassing the dense baseline (0.573 m / 85.9) at $6.6\times$ fewer total FLOPs. At an extreme 10% budget (24 GFLOPs, $18\times$ reduction), it still beats the dense baseline on both nuScenes (0.537 m vs. 0.573 m L2) and NAVSIM (87.9 vs. 85.9 EPDMS). On 739 intent-critical scenarios, GATS covers 95.8% of hazardous vehicles and achieves a 0.09% collision rate---73.5% lower than the dense model (0.34%).

    world model
  253. arxiv:2610.40931 · cs.LG
    Learning Stabilizing Latent Feedback Controllers via Multi-Step Backward Losses
    Paul Lutkus, Lars Lindemann, Stephen Tu

    Can learned latent dynamics models support the design of feedback controllers that provably stabilize the original system? For linear time-invariant systems, we show that learning linear autoencoders and latent dynamics with multi-step prediction-reconstruction losses $\unicode{x2014}$ which we term "backward losses" as they encourage latent trajectories to reconstruct full state trajectories $\unicode{x2014}$ provably recovers the structure needed for stabilization. Specifically, a sufficiently long reconstruction horizon ensures that a minimum ${H}_\infty$-norm controller designed using the learned latent model stabilizes the original system through output-feedback via the corresponding encoder, provided the latent dimension is large enough to capture all non-stable modes of the underlying system. We also show why both reconstruction and a multi-step horizon are essential: forward prediction losses that require only low prediction error along latent trajectories admit global loss minimizers that discard unstable modes regardless of horizon length, while single-step reconstruction losses can yield latent models whose stabilizing controllers fail to stabilize the original system. Finally, experiments on a non-linear flexible-arm system illustrate the practical implications of our results for learning non-linear autoencoders and latent dynamics.

    latent dynamics
  254. arxiv:2610.40930 · cs.CL
    Adaptive Personalization Under Preference Epistemic Uncertainty
    Priyanka Dey, Rajarshi Das, Vikas Bahirwani, Qi Zeng +2

    Human preferences change over time: an established preference may remain valid over time, become uncertain as context shifts, or be explicitly revised. Effective personalization requires conversational AI to adapt its response strategy to evolving preferences rather than relying on passive memory retrieval. We formalize this capability as adaptive personalization under epistemic uncertainty. A human preference study further confirms that people favor state-contingent response strategies as preference uncertainty changes. To evaluate whether language models mirror this adaptation, we create a multilingual benchmark spanning 250 personas, 5 languages, and 25K conversational turns that decouples strategy selection from response execution. Across eight closed- and open-weight models, we uncover a recognition-generation gap: models reliably identify appropriate strategies when selecting among candidate options, but collapse during open-ended generation, suffering a 35-46% drop in strategy calibration. Crucially, this failure is systematic: models default to exploit-first behavior, over-relying on existing preferences regardless of preference epistemic uncertainty. We show this failure is addressable: while prompting interventions fail, targeted training that supervises models' internal strategy representations toward state-conditional human preferences narrows the user acceptability generalization gap by a third on out-of-distribution personas and conversational styles.

    memorybenchmark
  255. arxiv:2610.40927 · cs.AI
    Agentic World Modeling for Agent Test-time Scaling without Environment Replica
    Wendi Li, Huseyin A. Inan, Wei-Ning Chen, Yanzhi Chen +3

    Test-time scaling (TTS) improves large language model agents by sampling and selecting among multiple candidate trajectories, but common realizations assume that these trajectories can be executed in parallel. This assumption breaks in realistic deployments, where enterprise workspaces and cloud services are shared and persistent, costly or infeasible to replicate, and actions may introduce irreversible side effects or cross-trajectory interference. Existing generative world models avoid direct execution, but typically compress environment state into a single prompt, limiting their ability to reliably simulate transitions under partial observability. Instead, we introduce Agentic World Modeling (AWM), which reformulates next-observation prediction as a multi-turn evidence-acquisition task. For each hypothetical action, an agentic simulator adaptively queries the real environment through safe read-only interfaces, reasons over the returned evidence, and uses executable computation to construct the resulting observation without mutating shared state. Integrated with trajectory-level TTS, AWM enables parallel exploration over a single shared environment. Across $τ$-bench, AppWorld, and OfficeBench, AWM outperforms a single-shot simulation baseline by 40.07% on average, and surpasses a trained world-model baseline by 12.36% on average. These results establish interactive, read-grounded simulation as a practical foundation for safe and scalable agentic test-time scaling in shared, non-resettable environments.

    world modelagentagentic
  256. arxiv:2610.40922 · cs.RO
    Confidence-Modulated Safety Filtering for End-to-End Policies
    Javier Borquez, Jaime F. Fisac

    End-to-end policies, from task-specific controllers to robot foundation models, learn complex behavior from raw observations but are not guaranteed to remain safe, so they are often wrapped in a safety filter. The resulting system operates under two domains: the filter's operational design domain (ODD), where its assumptions hold and safety is certified, and the policy's competence domain, where it is reliable. Filters typically fix their conservativeness regardless of the policy's competence, trading performance inside the ODD against margin beyond it. We ask whether the policy's confidence should set that level, and propose a filter-agnostic pipeline that modulates it at runtime: inside the ODD, where every setting is safe, confidence buys performance; beyond it, low confidence adds margin, which helps when it flags where the filter's model fails. Preliminary F1TENTH simulations, where this holds, show the modulated filter combining the strengths of permissive and conservative filters.

    robot foundation model
  257. arxiv:2610.40917 · cs.AI
    Towards Portable Specification of the Agent Harness
    Nicolas Gertler, Claudio Novelli, Luciano Floridi

    Large language models act only through the infrastructure that surrounds them. That infrastructure, the harness, determines what a model can accomplish and how it fails, yet it has no settled definition. To address this gap, we decompose the harness into eight functional components that can be specified and compared across systems, and we propose a portable specification framework that records which model ran in which harness. The record is needed because different harness configurations can yield several-fold differences in reported scores for the same model, and because much of an agent's attack surface lies in its harness. Existing model cards and agent cards provide release and operational information, but do not establish the sequence of effective configurations within a run. Harnesses can change between releases and between runs, and, where agents edit their own tools and configuration, within a run. Enacted AI laws contain documentation duties relevant to covered systems, subject to their respective scopes and application dates, but do not prescribe the common harness record proposed here. Therefore, the harness has to be explicitly described and specified as a versioned part of the system.

    agent
  258. arxiv:2610.40913 · cs.CV
    Coupled or Independent? Fault Coupling in Agentic Medical Vision Systems
    Nishi Kanta Paul, Md Shihabul Islam Shovo, Md Mostafijur Rahman, Jamil Fayyad

    Vision-language agents equipped with tools expose intermediate outputs for inspection and comparison with the model's reasoning, which may offer safety benefits over single-shot models. This paper tests that potential benefit when the sensor reading itself is untrusted, so that the tool and reasoning channels share a common visual fault. We consider a dermatology agent based on segmentation, contrastive retrieval, and an iterative vision-language reasoner and define a coupling taxonomy that distinguishes single-channel errors from joint errors that agree or disagree. On PAD-UFES-20, we show that typographic injection produces detectable directional retrieval capture, whereas physical corruption does not. Crop-first retrieval preserves the payload at a readable scale and produces 55% joint wrong agreement across all cases (71% of faulty cases); full-image retrieval produces 35% (43%). Under a binary decision, if both channels are wrong, their predictions must agree, so hard-label channel comparison cannot detect joint failures, limiting disagreement recall to 0.29 (crop-first) versus 0.57 (full-image). The direction of the order effect is reproduced on the independent HIBA test set.

    agentagentic
  259. arxiv:2610.40906 · cs.AI
    Does Recursive Knowledge Construction Help Agents Verify Consensus Claims in Science?
    Lun Ai, Yu-Heng Wu, Jonathan Schaul, Leonie Küchenhoff +5

    Recent scientific agents build knowledge recursively by retrieving evidence, constructing an intermediate state, and judging whether that state contains sufficient information. These agents are evaluated end-to-end, usually without ground truth, so the effect of each design choice is unknown. Here, we examined two design axes: (1) what the recursion builds (the state of evidence) and (2) what controls it (the judge that decides when to stop). To measure them, we defined consensus scientific claim verification and measured these effects on this task, where a system must establish a claim's consensus stance from the open literature. We built a new consensus claim dataset, ProClaim-eval, including 718 claims with SUPPORT and REFUTE gold labels from five expert-curated databases spanning molecular biology, nuclear physics, and astronomy. Our instrument, ProClaim, persists an evidence state outside the model's context, so each axis can be studied with the other held fixed. Against verification systems and general-purpose agent harnesses, ProClaim ranks among the top three, with a significant advantage over the harnesses. We found that the evidence state improves performance over in-context accumulation, and it also matters what is stored in it. In addition, a trained judge that reads the evidence state performs better than fixed-round stopping. A judge trained separately and never exposed to ProClaim-eval achieves the highest agreement with the gold labels on three of five subsets, where it matches the off-the-shelf Claude Sonnet~4.6 judge and outperforms Claude Haiku~4.5. Our results support recursively maintaining a structured evidence state outside the model's context as an effective design for verifying claims in scientific agents.

    agent
  260. arxiv:2610.40904 · physics.optics
    Optical Quantum Memory Processing for Quantum Hashing
    S. A. Moiseev, M. M. Minnegaliev

    Sender authentication is an important requirement for secure quantum communication networks. We propose programmable photon-echo quantum memory processing of a multi-time-bin photonic qudit for implementing a quantum-hash-based verification protocol. In this approach, using QM, sender (Alice) generates a quantum hash state by transforming a single stored photonic wave packet into the programmable photonic qudit. Bob's memory coherently maps the received time-bin photonic qudit into a single output mode according to phase encoding. Our scheme relies on a time-resolved coherent detection method introduced here, in which the coherent responses from individual time-bins are sequentially accumulated within Bob's spin coherence prior to the final sender verification. The key advantage of this coherent detection for quantum hashing is shown, and its potential applications are also discussed. The performed analysis demonstrates that experimentally verified parameters of contemporary optical quantum memories enable a proof of principle realization of the proposed architecture. These results reveal an additional role for quantum memories, serving not only as synchronization and repeater resources for long-distance quantum communication, but also as active elements for quantum hashing in conventional quantum key distribution systems.

    memory
  261. arxiv:2610.40902 · cs.RO
    FlapSim: A Composable MuJoCo Simulation Platform and Reduced-Order Aerodynamic Model Hierarchy for Flapping-Wing Robotics
    Mingshan He

    Flapping-wing simulation must reconcile high-frequency wingbeat aerodynamics with the slower dynamics relevant to flight control and reinforcement learning. We present FlapSim, a modular open-source MuJoCo simulation platform that combines robot descriptions, wing geometry, and composable aerodynamic models in a deterministic simulation pipeline. Across $48$ held-out scenarios spanning frequency, gimbal, wind, frequency-step, gust, and cross-robot flat-plate regimes, we compare five reduced-order configurations operating at $1\,\text{ms}$ with a $50\,μ\text{s}$ quasi-steady blade-element reference. The built-in hover cycle-averaged baseline achieves a $67.51\times$ CPU speedup but only $0.6593$ mean wrench fidelity. Combining multirate $21$-substep clock synchronization, three-dimensional linear/quadratic inflow and gust coupling, analytical blade-element geometry scaling, and phase-locked $1f\text{--}6f$ harmonics raises mean fidelity to $0.9646$ at a $65.19\times$ speedup in a five-block model. A streamlined four-block model without explicit step-transient lag states achieves $0.9662$ fidelity at a $67.98\times$ speedup. Controlled ablations show that stroke harmonics are particularly important during frequency transitions, improving held-out frequency-step fidelity by $+0.3176$. Transfer-function analysis explains this effect: fixed-window cycle averaging retains harmonic contributions when the wingbeat frequency changes, so within-cycle reconstruction also improves the filtered wrench. Once phase alignment and harmonic synthesis are included, explicit transient lag states provide no additional benefit in this benchmark. These results characterize the accuracy--computation tradeoff of the aerodynamic hierarchy and identify timing, inflow, geometry, and harmonic representation as central considerations for reduced-order flapping-wing simulation.

    benchmark
  262. arxiv:2610.40901 · cs.LG
    TSFLab: A Self-Improving Agent-Agnostic Workbench for Time Series Forecasting Research
    ChengAo Shen, Yanjun Zhao, Haoxin Liu, Bonnie Chen +6

    As publications and products on time series forecasting (TSF) proliferate, researchers face a growing burden of analyzing existing work. First, not all publications release code. For instance, only about 20% of arXiv TSF papers have a code link in the surveyed period. Second, released code also requires manual maintenance. Thus, TSF researchers must constantly adapt or implement existing works. However, independent implementations can differ in experimental settings and are difficult to maintain. We therefore propose to use AI agents for benchmarking TSF research and even supporting autoresearch, e.g., developing new TSF models. Yet directly applying existing coding agents, such as Claude Code and Codex, is suboptimal: prohibitive costs, limited scalability to large research projects (e.g., benchmarking), and lack of self-maintenance. To address these challenges, we propose TSFLab, a self-improving agent-agnostic "workbench" that equips compatible coding agents with shared resources and tools for low-cost TSF research and continuous maintenance of research resources. Our experiments cover reproduction, benchmarking, and new method development. In reproduction and benchmarking, TSFLab improves task completion by 20-163% and cuts fresh tokens per completed task by 11-66% across baselines. A pilot study of agent-driven method development also shows improved forecasting accuracy via TSFLab, with 32.8-73.6% fewer fresh tokens and 9.2-68.8% less development time than the other settings.

    ai agentself-improvingbenchmark
  263. arxiv:2610.40893 · cs.CL
    Qistas: A Pre-Registered, Portable Protocol for Fairness Evaluation of Large Language Models
    Nawar S. Alseelawi, Mustafa S. Aljumaily

    Fairness benchmarks for large language models (LLMs) are overwhelmingly English-born and American in their social categories; translated versions carry those categories into societies where they do not apply, and almost none fix their hypotheses before the models are run. We present Qistas (Arabic qistas, "the just balance"), a protocol for evaluating how LLMs treat the social, confessional, and regional constituencies of a society, portable across countries and instantiated first for Iraq. In its pilot form it measures only what needs no verdict on contested truth: procedural symmetry across matched counterfactual pairs, differential refusal, stereotype reliance under ambiguity, and terminological framing of documented events. Every measurement is paired by construction and analysed with paired machinery (sign-flip permutation tests, exact McNemar tests, template-clustered bootstrap intervals, Holm-corrected families), with a power analysis that states, before any data exist, the smallest effect each axis can detect after the design effect of template clustering. We release the authoring apparatus (a pruned 3x5 taxonomy with four cells closed by design, per-type checklists, a cumulative rulings log, a structural audit, and a human review notebook) and report its first field test: a semantic review of a 58-item seed bank in which 24 of 25 template groups failed at least one check. The defect patterns it exposed (evaluation criteria leaking into prompts, token-level swaps that break Arabic agreement, epistemically unbalanced pairs, single-theme refusal probes, unwritten dialect twins counted as coverage) are, we argue, general hazards of counterfactual fairness benchmarks. This is a protocol paper, not a results paper: no model has been evaluated, and the accompanying pre-registration commits the hypotheses, metrics, exclusion rules, and analysis plan that the pilot will follow.

    benchmark
  264. arxiv:2610.40867 · cs.LG
    GraphTFactor: A graph-based approach for proteins transcription factor prediction
    Domenico Amato, Salvatore Calderaro, Giosuè Lo Bosco, Riccardo Rizzo +1

    Transcription factors (TFs) play an important role in regulating gene expression, making their identification a fundamental task in computational biology. In this work, we propose GraphTFactor, a novel end-to-end framework that models protein sequences as graphs and directly classifies transcription factors. By encoding sequences into graph structures, the proposed approach naturally handles variable-length inputs while preserving relevant relational information, without imposing length constraints. Experimental results show that GraphTFactor outperforms state-of-the-art methods on benchmark datasets, demonstrating the effectiveness of graph-based representations for TF prediction while providing interpretable insights into the sequence regions contributing to the model's predictions. To support reproducibility and practical adoption, we release an open-source software package and an interactive web application that let users easily perform TF prediction on custom sequences.

    benchmark
  265. arxiv:2610.40857 · cs.LG
    Frozen in Time: How a Physics Foundation Model Forgets Under Sim-to-Real Adaptation
    Haaris Mian, Miles Cranmer, Alessandro Favero, Payel Mukhopadhyay

    Scientific foundation models (SciFMs) have achieved remarkable performance on a wide range of PDEs, including the ability to efficiently adapt to unseen data distributions with additional finetuning. As with all neural architectures, finetuning such models comes with the risk of catastrophic forgetting, whereby adapting to new tasks degrades prior knowledge of the model. We study catastrophic forgetting in Walrus-1.3B, a SciFM for continuum dynamics, over the course of a two-stage sim-to-real adaptation: first finetuning on simulated flow around a cylinder and then adapting to experimental data of the same system. We find that finetuning on the simulations rapidly degrades rollout performance on the original pretraining datasets, with one-step errors increasing up to 35x after just 20 epochs. Subsequent adaptation to real data further erodes performance on the simulated task. We characterize the potential drivers of forgetting, such as dataset dependence and task representation similarity. We then show that replaying as little as 1% simulated data during real-data finetuning recovers most of the lost simulation performance while retaining effective adaptation to the experimental dynamics. Our results identify catastrophic forgetting as a practical obstacle to sim-to-real adaptation of SciFMs and show that lightweight replay strategies significantly improve retention.

    sim-to-real
  266. arxiv:2610.40847 · cs.AI
    The Generative Gap Between Speculative Drafter and Continuous Diffusion Language Model
    Yujian Chen, Hao Jiang, Guang Cheng, Weijian Luo +1

    Continuous diffusion language models (CDLMs) and speculative drafters offer efficient language generation but are typically evaluated differently: CDLMs generate independently through iterative refinement, whereas drafters propose text for target-model verification and correction. We compare their architectures without verification, using LangFlow for CDLMs and adapting DSpark's Markov and recurrent-head architectures for standalone generation. All models are trained from scratch for unconditional, single-block OpenWebText generation at lengths 16, 128, and 1024, without target features or verification, using common preprocessing, backbone dimensions, and training-update budgets. We evaluate generation perplexity (GenPPL) with an external language model and distributional similarity to reference text with MAUVE. With default denoising steps equal to sequence length, LangFlow outperforms both DSpark variants on both metrics at lengths 128 and 1024, while DSpark-RNN leads at 16. A conditional KL analysis characterizes dependencies excluded by the Markov head's previous-token conditioning. Recurrent memory removes this restriction and improves both metrics, but the longer-sequence gap persists. Teacher probability matching further lowers Markov GenPPL without observed MAUVE gains. Increasing LangFlow's denoising steps at fixed weights improves both metrics at lengths 16 and 128, changing the length-128 ordering relative to DSpark-RNN. These results establish LangFlow's longer-sequence advantage under the evaluated budgets and highlight iterative full-sequence refinement as a route to improving standalone generation beyond gains from recurrent head memory.

    memoryiterative refinement
  267. arxiv:2610.40832 · cs.CV
    EgoHSI3R: Online 4D Hand-Scene Interaction Reconstruction from Egocentric Video
    Mingzhi Pei, Zhangchen Ye, Ruoqu Chen, Liu Cao +1

    Egocentric videos offer an abundant source of hand-scene interactions for embodied learning, which requires recovering hands, camera, and scene consistently in a shared 3D world. Existing methods focus on either camera-frame hands or world-space hand trajectories, without jointly recovering the scene pointmaps or interaction information. Building a complete reconstruction from such specialized predictions therefore typically requires per-sequence optimization. We propose EgoHSI3R, a unified feed-forward framework for online 4D hand-scene interaction reconstruction from egocentric monocular video. To jointly recover these components, it maintains a recurrent scene representation that accumulates geometric evidence over time and provides a shared world space and scale. We introduce hand-centric prompts that fuse pretrained hand priors and local scene context, while MANO vertex queries attend to scene features to predict dense per-vertex contact. These designs enable a single network to jointly estimate camera motion, dense scene geometry, world-space bilateral MANO hands, and per-vertex hand contact. Despite operating causally, EgoHSI3R achieves the lowest world-space hand joint error on ARCTIC among the evaluated methods, while offering faster inference than existing world-space hand reconstruction pipelines and competitive camera and scene reconstruction accuracy.

    embodied
  268. arxiv:2610.40828 · cs.RO
    From Teleoperation to Long-Horizon Dexterity: Evaluating Visuomotor Skills and Offline Proxies for Dexterous Manipulation
    Ammar N. Abbas, Aayush Jain, Prashanth Viswanath, Nivesh Gadipudi +2

    Imitation learning policies are intended to make robotic manipulation robust to variations in object pose, appearance, and environment that challenge pre-programmed motion. However, real-world performance remains difficult to predict, as conventional optimization signals such as validation loss show no consistent relationship with robot success. To identify the design choices that govern real-world performance, 281 imitation-learning policies spanning seven policy families were evaluated in approximately 1,700 real-robot trials of RJ45 connector grasping and insertion using a five-fingered robotic hand. Controlled ablations showed that observation and action representations dominated conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping, representation in the robot-base frame, and normalization within the point encoder each improved success by 70-100 percentage points. For the vision-language-action policy, absolute rather than relative action prediction improved insertion success by up to 70 percentage points. Skill decomposition, multi-view perception, and recovery demonstrations were associated with the strongest-performing systems, although their effects were not independently isolated. Of 67 offline metrics evaluated, only five measures of predicted-motion smoothness and directness reliably correlated with insertion success (r = -0.70 to -0.80), and only within a single policy family. These results identify representation choices as key determinants of real-world imitation-learning performance and show that offline metrics can shortlist policies, but cannot replace real-robot evaluation.

    vision-language-actionmanipulationdexterousteleoperationgrasp
  269. arxiv:2610.40801 · cs.CV
    Drone-PUR: From Visual Clue Discovery to Understanding and Reasoning in UAV Scenes
    Jiahe Wu, Bing Cao, Xingyue Guo, Ruipu Zhao +4

    The reasoning capabilities of vision-language models (VLMs) can be limited by incomplete or inaccurate visual observations in the general domain. In unmanned aerial vehicle (UAV) scenes, small targets, subtle attributes, and changing viewpoints further complicate perception and understanding. Therefore, we introduce Drone-PUR, a dataset and benchmark organized around visual clue discovery, interpretation, and integration, corresponding to the three capability dimensions of perception, understanding, and reasoning. Reasoning samples are designed to require models to combine multiple visual clues to derive visually grounded conclusions. Its 11 tasks provide 175k training and 20k test samples spanning single images, image sequences, cross-modal images, and multiple viewpoints. All test samples undergo human review. Evaluating 22 general-purpose VLMs reveals persistent limitations in spatial judgment and cross-modal and cross-view reasoning. Across models, stronger perception is associated with better reasoning, while single-task fine-tuning further reveals selective bidirectional transfer between these capabilities. To examine domain adaptation, we develop UAV-PUR from Qwen3-VL through staged fine-tuning with auxiliary supervision. Its 4B and 8B variants improve overall multiple-choice accuracy by 18.33 and 17.52 percentage points over their pretrained counterparts, respectively. The dataset and project resources are available at https://drone-pur.github.io.

    benchmark
  270. arxiv:2610.40799 · cs.RO
    CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents
    Yijia Fan, Xilin He, Yilei Jiang, Dongming Wu +2

    Across coding, web, and tool-use tasks, language-model agents are now routinely trained with the same recipe: supervised fine-tuning (SFT) to learn the interface, then multi-turn reinforcement learning (RL) to master it. Code-as-policy (CaP) casts robot manipulation in the same form, as the policy writes code against robot APIs and reads back the results, so the same recipe should apply. Yet existing CaP agents sidestep training: they either prompt closed frontier models or delegate manipulation to pre-trained vision-language-action models (VLAs) and high-level skills, while small open vision-language models (VLMs) fail outright. We present CaPtain, an open post-training pipeline that turns a small open VLM into a capable CaP agent, with no VLA, frontier model, or privileged state at test time. CaPtain trains and deploys the policy inside CaPtain-Harness, which exposes low-level perception, geometry, and guarded-motion primitives across simulation and real hardware. We first align the base model via SFT on synthetic trajectories generated with GPT-6 Astra, then apply multi-turn RL. Multi-turn RL must credit tens of agent requests, whose physical effects often surface turns later, from a single end-of-episode outcome. We therefore introduce a physical critic that reads privileged simulator state, never seen by the policy, and credits each request by how much it changes the predicted chance of success. Starting from Qwen3.5-4B (0% zero-shot), CaPtain-4B reaches 55.4% held-out LIBERO success after SFT and 73.8% after RL. On LIBERO-PRO, it reaches 78.5%, surpassing several frontier-model agents, and its RL gains persist across three perturbation suites (+17.6 to +21.2 points). Without target-domain training, it attains 83.3% on SimplerEnv WidowX, exceeds frontier models on CaP-Bench, and runs on a real robot after simulation-only adaptation. We release all code, data, and model weights.

    vision-language-actionmanipulationliberoagenttool-usecode-as-policy
  271. arxiv:2610.40798 · cs.AI
    Beyond Plausible Prediction: Safety Sufficiency in Web World Models
    Zhifei Hu, Alexandra I. Cristea, Adel Bibi, Jialin Yu

    Web world models increasingly predict the consequences of candidate actions before agents act, yet a plausible predicted future does not necessarily support the correct safety decision. In a matched human evaluation of 3,000 predictions from ten frozen open models, 86.1% are judged semantically plausible. However, several models exhibit highly asymmetric authorisation behaviour, combining low false-safe execution with poor recall of clarifiable ASK cases and a strong preference for BLOCK. This motivates safety sufficiency: whether a predictive interface preserves and exposes the policy-decisive information required to distinguish actions that may be EXECUTED, require clarification through ASK, or should be BLOCKED. We introduce SafeWebWorld, a benchmark for evaluating safety sufficiency across five web-safety domains through action-conditioned futures, policy-decisive evidence, and policy-conditioned decisions. Under a controlled Security-Aligned Supervised Fine-Tuning protocol, evidence preservation can improve substantially while authorisation-sensitive decisions deteriorate. To determine whether this deterioration reflects loss of authorisation-sensitive score structure or a failure of the decision readout, we introduce Latent Authorisation Recovery (LAR). Frozen-score recalibration reveals substantial but backbone-dependent recovery, while paired recovery provides only a modest additional gain. Yet within-family and matched-authorisation evaluations remain poor, showing that aggregate recoverability does not imply robust fine-grained authorisation behaviour. Overall, plausible prediction, evidence preservation, and aggregate recovery are each insufficient on their own to establish robust authorisation.

    world modelaction-conditionedbenchmark
  272. arxiv:2610.40758 · cs.AI
    More Rules, Longer Lists? Benchmarking Self-Evolving Agents on a Hidden-Rule Curriculum
    Chenjie Qiu, Boyan Xu, Shaobin Shi, Bingfeng Chen +2

    Self-evolving agents rewrite a persistent state from feedback, but it is largely untested whether this lets them learn a rule nobody told them. An agent can also pass by writing down each correction as an exception, and its list then grows with every new rule. We introduce CLR-Bench, a Curriculum of Latent Rules benchmark. Each chain adds one hidden rule per level on top of the earlier ones. A rule acts on an opaque label that stands for a named grouping published on the Web, such as the founding members of an airline alliance. The task gives one example value, practice feedback reveals one more member and one non-member, and the test checks unseen members and similar non-members. Probes certify that the base model knows the members of each grouping and does not guess it without feedback, and that the evidence identifies the grouping once labels are read as named groupings. An automated pipeline builds chains from a new data domain, and we release 20 chains over 10 domains. We evaluate six recent self-evolution methods. With the certifying model they barely learn. The best gain at most 5 points over a frozen agent and exclude only 6-12% of the unseen members from the count, fewer than the 20% reached by simply keeping the raw practice history. A fixed lesson that states the named-grouping convention lifts the same methods by 14.5 to 16.5 points. Their states hold the revealed member in 75-98% of cases and almost never the class name, and when practice exposes a member the list misses, they add that member to the list. With the stronger GPT-6-Sol, ACE and Dynamic Cheatsheet learn from feedback alone. They reach 78% accuracy and solve 59-60% of the single cells, against 53% and 8% with the certifying model, and they surpass the raw history. CLR-Bench therefore measures how well an agent learns, and this ability differs across models.

    persistent stateagentself-evolvingbenchmark
  273. arxiv:2610.40757 · cs.LG
    AutoHebb: Self-Modifying Neural Networks that Zero-Shot Infer and Continually Learn
    Daniel Levine, David van Dijk, Kenneth Kay

    Humans and animals have a unique capacity to generalize and learn in novel circumstances, including "0-shot" inference (ZI) and continual learning (CL) -- two powerful cognitive abilities. Despite this importance, efforts to model ZI and CL are seriously limited in several ways. First, existing models typically lack components critical to ZI and CL in the brain, namely, strong inductive biases (for ZI) and autonomous control of learning (for CL). Second, models can be less tractable to analysis if model components are overly complex (lacking simplicity), or too task-restricted if model components are overly bespoke (lacking expressivity). Here we introduce autonomous Hebbian networks ("AutoHebb"), a class of online learning feedforward neural networks that realize strong inductive biases and autonomously controlled learning, and are both simple and expressive. Unlike standard models that learn via externally controlled optimization (e.g. gradient descent), AutoHebb networks combine external feedback with internally generated activity to produce a global learning signal ('neuromodulator') fed back to the network. The neuromodulator in turn gates a biologically grounded local learning rule (3-factor Hebbian rule) that updates networks' plastic weights. Using meta-learning to search for fixed (non-plastic) weights, we find that AutoHebb networks (1) succeed on ZI tasks, (2) show emergent (untrained) ZI/CL abilities, (3) admit reverse-engineering into explicit algorithms, and (4) outperform modern learning models on ZI/CL tasks. Lastly, performance comparisons to related models implied an essential role for three model components: meta-learning, Hebbian plasticity, and autonomous control of learning. These findings indicate that AutoHebb may be useful in hypothesis generation, analysis, and discovery of biologically grounded forms of learning and generalization.

    online learning
  274. arxiv:2610.40754 · cs.CV
    DyCraft: An Agentic System for Long-Horizon 4D Scene Generation
    Jeff Tan, Hang Dai, Zeqian Long, Yanzhe Lyu +4

    While frontier coding agents can generate high-quality 3D scenes and articulated assets, they struggle to create coherent long-horizon animations. When equipped with general-purpose harnesses, they frequently produce physically implausible outputs, generating objects that interpenetrate, lack physical support, or move unnaturally. Such bare agents struggle to repair these errors because they default to scripting the full animation in one pass, making it difficult to verify their work and disentangle complex spatiotemporal constraints. Drawing on the practice of human animators, we introduce DyCraft, a system for animated 3D scene generation that breaks animation into steps that coding agents can more reliably solve. DyCraft first expands the user's prompt into a detailed plan and builds a static scene. It then creates a 3D storyboard of key animation poses, where each frame is a static 3D scene that the agent can build and debug in isolation. Storyboarding decomposes the problem into pairs of verified start and end states, allowing the agent to author local motion between them with a collection of motion authoring tools. We evaluate DyCraft on 24 long-horizon, multi-shot animation tasks covering humanoid robots, human and animal characters, and physical mechanisms. Compared with a native GPT-6 agent in Codex and prior agentic baselines, DyCraft substantially reduces physical violations and generates higher-quality animated 3D scenes that are more often preferred by human raters.

    humanoidagentagentic
  275. arxiv:2610.40753 · cs.AI
    "Shall We Approve?" Consequence-Aware Approval for Autonomous Coding Agents
    Weiheng Wu, Junkai Chen, Zhensu Sun, Jieke Shi +2

    Coding agents increasingly rely on automated approvers to decide whether a proposed shell command may run. Existing approvers, from rule-based checkers to guardrail models and deployed permission classifiers, base their decisions on the text of the request and command. Yet a command's behavior depends on the code it invokes and the state of the workspace, and systems that observe file changes leave the assessment to the agent or a human after execution. We present Caper, a consequence-aware approval mechanism that pre-executes each command in a snapshotted sandbox, provides an independent approver with a bounded report of its observed effects, commits allowed commands, and rolls back blocked ones with a reason the agent can act on. Compared with six existing systems, Caper achieves the lowest harmful-outcome rate on SABER (6.7% versus at least 14.8%) and YoloFS-Ext, a benchmark we build in which harmful behavior is hidden inside invoked tools (7.8% versus at least 35.3%), while blocking 1.0% of the safe twins of the dangerous tasks in YoloFS-Ext.

    agentbenchmark
  276. arxiv:2610.40752 · cs.CV
    PathoJev: Efficient Pathology Decisions from Shared Visual Context
    Chengyang Zhang, Wenchuan Zhang, Bo Li, Xinyu Liu +7

    Pathology workflows require multiple visual judgments about the same slide, from identifying tumor type to assessing grade and invasion. Generating and parsing a separate textual response for each structured question adds latency between visual evidence and usable decisions. Building on the System One paradigm, we introduce PathoJev, a pathology model that directly predicts answer probabilities for whole-slide and region-level assessment. PathoJev encodes each slide once and evaluates independent question suffixes in batches over the shared visual context. Reusing the hybrid backbone's recurrent states further limits the computation that grows with slide-context length for each additional question. Across three whole-slide benchmarks, the 9B and 27B models achieve mean accuracies of 77.2% and 78.1%, exceeding SlideChat by 5.4 and 6.3 percentage points. Both scales also maintain competitive accuracy on five ROI benchmarks. On WSI-Bench, the 9B model runs 9.1x faster than SlideChat. Shared context accelerates eight-question workloads by 6.9x and 5.9x over separate prefix encoding, rising to 11.2x and 8.9x for 64 questions. These results provide a reusable probabilistic interface for combining multiple visual judgments in whole-slide and local tissue assessment. Code: https://github.com/ChyaZhang/PathoJev.

    benchmark
  277. arxiv:2610.40751 · cs.AI
    Safin-1.5: Safety-Native Architecture and Verifiable Reasoning for Recursive Safety Improvement
    Jingnan Zheng, Cheng Jin, Ming Zhang, Huan Sun +27

    Recursive self-improvement (RSI) enables large language models to advance their capabilities through successive updates, but these updates can also introduce new risks and weaken existing safeguards. In line with the AI-45° Law, we pursue Safety RSI, in which safety develops as an intrinsic capability alongside general model capabilities through continual risk discovery and repair, while preserving utility and limiting over-refusal. We introduce Safin-1.5, a model built on a safety architecture and trained through recursive safety improvement, with an inference-time harness that extends its capabilities to formal reasoning. Our approach comprises three components: i) The safety architecture combines persistent safety states with hazardous-knowledge disentanglement. The former retains and selectively invokes safety knowledge, while the latter relocates targeted hazardous knowledge into removable experts, enabling its removal with minimal impact on general capabilities. ii) Recursive safety improvement establishes a discovery-evaluation-repair loop. The loop identifies new vulnerabilities and uses experimental feedback to develop repair strategies. It accepts updates only when they improve safety while preserving general capabilities and limiting over-refusal. iii) The formal reasoning harness combines the model's language understanding with symbolic inference and verification. Proof-obligation-guided semantic bridging recovers missing relations in formal representations, enabling symbolic solvers to construct checkable proofs under formalized premises and accepted semantic bridges. Experimental results demonstrate that Safin-1.5 achieves safety and reasoning performance, outperforming frontier models.

    self-improvement
  278. arxiv:2610.40748 · cs.RO
    WorldBridge: Bridging Video and Action for Task Generalization in World Action Models
    Jun Cen, Yubo Guo, Hao Li, Liangkai Li +8

    Achieving the broad task generalization of embodied models in physical environments remains a central challenge in robot learning. We study this challenge in World Action Models (WAMs) through two complementary capabilities: video generalization, which predicts the visual evolution of new tasks, and action generalization, which translates these predictions into executable behavior through a video-to-action bridge. By varying video and action supervision for tasks outside an initial training set, we examine their respective contributions to task transfer. On RoboTwin 2.0, using only in-distribution video-action supervision yields a 25.07% success rate on out-of-distribution tasks. Introducing videos of these tasks without their action labels raises success to 55.90%, while adding just one action-labeled demonstration per task further improves it to 70.70%. These findings suggest that broader video coverage enables task transfer by leveraging existing inverse-dynamics capabilities, while even sparse action supervision further strengthens those capabilities. Guided by this perspective, we introduce WorldBridge, a WAM with separate video and action experts connected through structured attention and trained with joint flow matching. WorldBridge incorporates visual memory at two temporal scales and action forcing, alongside a unified data processing pipeline that standardizes and filters heterogeneous robot trajectories and egocentric hand-derived supervision. Across multiple manipulation benchmarks, WorldBridge achieves state-of-the-art performance among the compared methods, including 96.23% on RoboTwin 2.0 full data setting and 80.9% on clean to random setting, 82.8% on RoboCasa GR1 TableTop, and 57.6% on DOMINO Level-1. Our study highlights the complementary roles of video prediction and action grounding in developing robots that generalize across tasks.

    embodiedmanipulationrobotwinmemorybenchmark
  279. arxiv:2610.40747 · cs.LG
    Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses
    Xiaoyi Zhang, Haoqing Wang, Xinjie Zhang, Xiang Ming +18

    Effective computer-use agents (CUAs) should seamlessly combine visual GUI interaction with code and terminal execution within a single policy. We present Mage-VL-Cowork, a 4B hybrid CUA trained from a general multimodally aligned checkpoint through an end-to-end, progressive five-stage curriculum across midtraining, SFT, and RL. To power this curriculum at scale, Mage-CUA-Forge replaces per-task agent construction with reusable, skill-conditioned generation workflows and application-specific preparers, synthesizing 293,806 desktop and web tasks across 6,915 skills while reducing recurring per-task expenditure by 82.6%. Rather than directly distilling a default teacher, we evolve the teacher harness and instantiate specialized CUA harnesses to yield high-quality trajectories alongside stage-specific supervision. Mage-VL-Cowork-4B finally climbs to 66.0% on OSWorld-Verified through different training stages with more efficient reasoning cost, achieving state-of-the-art performance at its parameter size and pushing forward the frontier of on-device intelligence.

    agent
  280. arxiv:2610.40742 · cs.CL
    Prevented is Not Protected: What Attack Success Rate Hides about RAG
    Chahana Dahal, Zuobin Xiong

    Retrieval-Augmented Generation (RAG) is vulnerable to knowledge poisoning, where adversaries manipulate retrieved evidence to influence model outputs. Existing defenses are primarily evaluated using attack success rate (ASR), which obscures defense-induced failures and the mechanisms behind apparent attack prevention. To address this issue, this work introduces \textbf{Aligned RAG Evaluation}, a matched-query protocol that jointly examines clean and poisoned responses, with and without defense. By combining matched conditions analysis, LLM-judge endorsement, and retrieval dependence controls, the proposed framework distinguishes prevented from introduced failures, target mentions from LLM endorsement, and single-passage influence from evidence composition. Experiments on HotpotQA and Natural Questions reveal that existing defenses can introduce unexpected target-positive answers, although they can prevent some attacks, while evaluation criteria can also change defense rankings. Our findings expose critical blind spots in ASR-based evaluation and motivate a more comprehensive assessment of RAG poisoning defenses. The code and the evaluation artifact are available at https://github.com/chxhana/aligned-rag-eval.

    retrieval-augmentedrag
  281. arxiv:2610.40738 · cs.LG
    Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use
    Aman Tiwari, Akshay Kalkunte, Akintunde Oladipo, Jash Shah +27

    A capable voice assistant must do more than recognize words; it must understand requests, reason over context, sustain conversations, and act through tools. We present Apriel Voice 27B Preview, an open speech-language model that unifies these capabilities within a shared reasoning backbone while retaining text and image input modalities. Built on the post-trained Gemma 4 26B-A4B checkpoint, Apriel Voice directly processes speech and generates both responses and tool calls without requiring a separate dialogue model. We introduce a two-stage training procedure to adapt the model to speech. The first stage establishes the speech interface, while the second jointly adapts the projector and backbone through supervised fine-tuning, text replay, and source-dependent cross-entropy and forward Kullback--Leibler divergence supervision. Automatically constructed speech-conditioned dialogues provide conversational training data, while local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses enable efficient online distillation. In thinking mode, Apriel Voice achieves a score of 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and a task success rate of 65.51% on $τ$-Voice, with a mean latency of 1.91 seconds. It also achieves 51.27% English EVA accuracy and 74.13% experience pass@1. Finally, we deploy Apriel Voice 27B Preview on a custom serving platform to support interactive voice interactions.

    tool use
  282. arxiv:2610.40736 · cs.LG
    Unsupervised Feature Selection Using Local Intrinsic Dimensionality -- Extended Version
    Oussama Chelly, Michael E. Houle, Ken-ichi Kawarabayashi, Muhammad Rajabinasab +1

    High dimensionality degrades both the efficiency and the effectiveness of learning algorithms, and in many applications no class labels are available to guide dimensionality reduction. This paper proposes LIDFS, an unsupervised feature selection method based on Local Intrinsic Dimensionality (LID). The method builds on the formal equivalence between LID and the indiscriminability of distance measures. The data matrix is transposed so that each feature is treated as a point, and the LID of each feature is estimated by maximum likelihood from the distances to its $k$ nearest neighboring features. Features with lower LID exhibit greater local discriminability and receive higher importance scores. A refinement step then retains one representative per feature neighborhood and demotes locally redundant features. The result is a complete ranking from which any number of features can be selected without rerunning the procedure. LIDFS does not use any information about labels or the number of classes, and runs in $\mathcal{O}(d^2 n)$ time for $n$ instances and $d$ features. LIDFS is compared with classic and state-of-the-art methods and a random baseline on 23 high-dimensional datasets, the classic benchmark collection. The evaluation uses classification accuracy and AUC with five downstream classifiers, as well as clustering accuracy and normalized mutual information, over both full and aggressive selection ranges. LIDFS ranks first in the supervised evaluations and first or tied second in the clustering-based evaluations. It belongs to the top statistical group together with methods that require the number of classes, and it runs faster than its closest competitor. Several recent methods, by contrast, fail to outperform the random baseline. A sensitivity analysis shows that small neighborhood sizes yield the best results.

    benchmark
  283. arxiv:2610.40735 · cs.AI
    RiskyPy: Toward Model-Calibrated Prompt Generation for Joint Security and Functionality Evaluation of LLM-Generated Code
    Zacharie Chenail-Larcher, Daniele Lunghi, Melissa Tessa, Quentin Stiévenart +4

    Evaluating approaches that improve the security of LLM code generation requires datasets that jointly assess security and functionality and remain adapted to the capabilities of the models being studied. However, existing benchmarks are costly to produce and are generally not calibrated to a chosen range of model capabilities. We introduce RiskyPy, an automated pipeline that generates security-oriented code-generation datasets calibrated to model proficiency using functionality and vulnerability canaries. We instantiate RiskyPy for Python using domains covering 18 CWEs from the 2025 CWE Top 25, producing 101 prompts. Experiments on Qwen3-4B and Seed-Coder-8B show that the generated tasks remain functionally accessible, with 57-72% of prompts yielding functional code, while only 5-10% of generations are free of Bandit findings, leaving room for joint evaluation of security and functionality of security improvement approaches.

    benchmark
  284. arxiv:2610.40734 · cs.LG
    Model-Agnostic Search for New Physics in the Presence of Imperfect Background by Smearing the Data
    Soheun Yi, John Alison, Mikael Kuusela

    In the search for new particles in high-energy physics, it is necessary to first define the signal region (SR) enriched in signal events, and then compare events in this region with the expected background distribution. While most existing search methods rely on prior domain knowledge of either the signal or the background, there has been less effort to relax both assumptions simultaneously, as methods that weaken one assumption typically rely on strong knowledge of the other. We propose Model-Agnostic Search with Smearing and Imperfect Background (MASSIB), a method that addresses this challenge by learning the density ratio of events relative to a misspecified background distribution using modern tools from machine learning, and defining the SR as the region with high density ratio. Recognizing that a naive application of this idea introduces significant bias in background estimation when interpolating from the control region to the SR, MASSIB mitigates this issue by introducing a "smeared" density ratio, which can be learned efficiently through data-level manipulation and used to define an SR that allows for accurate background estimation. By applying it to various types of simulated Large Hadron Collider (LHC) events with realistic background mismodeling, we demonstrate that MASSIB identifies a data-driven SR that achieves high signal concentration consistently across different types of signals while avoiding bias in background estimation. The method does this without making strong assumptions about either the signal or background distribution. We then show that this robustness extends to the validity and strong statistical power of the resulting hypothesis tests for signal detection, demonstrating the potential of this approach for future model-agnostic searches at the LHC.

    manipulation
  285. arxiv:2610.40726 · cs.LG
    Li$^{2}$SAO: Lifting Light Singl-Rollout Asynchronous RL Optimization via Group-Relative Actor-Critic Bootstrapping
    Zichen Bai, Yanli Wang, Yang Liu, Junhao Shi +8

    Asynchronous reinforcement learning (RL) is becoming a mainstream paradigm for post-training large language models (LLMs), as it improves training efficiency by updating the model as rollouts arrive. Recent work argues that group-wise sampling in GRPO does not naturally fit asynchronous RL training because rollout lengths follow a long-tailed distribution, and proposes Single-rollout Asynchronous Optimization (SAO) to improve stability and off-policy behavior, especially on agentic tasks. However, SAO relies on a closed-source value model obtained from scaling value pretraining. In this paper, we analyze GRPO and LiSAO, our lightweight reproduction of SAO without the pretrained value model, on single-turn reasoning and multi-turn agentic tasks. We find that GRPO outperforms LiSAO on reasoning tasks, while the reverse holds on agentic tasks. Since the value model of LiSAO is not yet well trained , LiSAO lags behind GRPO substantially in early training steps. To address this issue, we propose Li$^{2}$SAO, which Lifts Light Single-rollout Asynchronous Optimization with group-relative actor-critic bootstrapping: while the critic is cold, a group-relative advantage trains the actor and a group-centered return trains the critic, after which training switches to single-rollout GAE. Experiments show that Li$^{2}$SAO catches up with or surpasses GRPO in early training and further raises the later performance ceiling. Code is available at https://github.com/jd-cc-coder/Li2SAO.

    agenticpost-training
  286. arxiv:2610.40721 · cs.AI
    MetaVul: Agentic Harness Discovery for Vulnerability Detection
    Ira Ceka, Arnav Garg, Ryan Lutz, Ting Zhang +3

    LLM agents are increasingly used to automate vulnerability detection, but reliable detection is challenging because agents must distinguish subtle, security-relevant fixes from benign code variation. While underlying LLM capability has improved significantly, the accuracy of the agent still depends heavily on the harness: the code that defines the scaffolding around the LLM. Harnesses are typically fixed, and largely hand-written, limiting adaptation to evaluator feedback. Additionally, when harnesses are static, improvements on one dataset or model may not transfer to others. To address these challenges, we introduce MetaVul, a self-improving framework for harness discovery in vulnerability detection. Starting from a seed structured reasoning scaffold, MetaVul uses a workflow-based optimization loop to iteratively refine the harness using feedback from paired vulnerable and patched functions, while preserving the structured reasoning and verdict format. At each iteration, it searches over the history of prior harnesses, with access to their trajectories, logs, and metrics, and it returns a set of Pareto-optimal harnesses. We apply MetaVul to discover agent harnesses for various architectures. A validation set (202 pairs) is used for optimization, while an unseen dataset (430 pairs) is kept as the held-out set. We test whether harnesses learned from one model transfer to others not seen during optimization. Results on both frontier and small LLMs show that the optimized harnesses improve pairwise vulnerability detection up to 23.3 percentage points on the validation set and up to 25.8 points on the held-out set with a Pareto-frontier harness. Harnesses learned with one model also transfer to others. We show that harnesses optimized with one model transfer to another, improving held-out pairwise accuracy by 14.7 points without further optimization.

    agentllm agentagenticself-improvingevaluator
  287. arxiv:2610.40716 · cs.AI
    Characterizing LLM Agent Behaviors in Module Shadowing Environments
    Shuai Yang, Enmeng Lu, Jinyu Fan, Gongce Cao +6

    LLM agents operating in untrusted workspaces face the risk of untrusted code execution during routine tasks, yet final reports and workspace states alone cannot reliably establish whether execution occurred. We introduce an evaluation framework that codes observation, reasoning, execution, and response behaviors from agent trajectories and aligns them with independent environmental records to reconstruct execution histories and verify final reports. We apply this framework to module shadowing, in which a local file is loaded in place of a trusted module during import, enabling untrusted code execution without adversarial prompts. Across 90 trials with 15 models, the payload executed in 51% of direct shadowing trials and 87% of indirect shadowing trials. Even explicit recognition of the shadowing mechanism did not ensure prevention, as the payload ran in 10 of the 29 trials with recorded recognition. These failures were often omitted from final reports, as 49 of the 62 triggered trials ended without mention of the security issue. Three reports even denied execution, including one from a trial in which independent logs recorded 19 payload executions. Final workspace states also obscured prior execution when agents deleted or overwrote the payload marker in 10 triggered trials. These findings demonstrate that combining behavioral coding with independent environmental auditing enables rigorous and transparent evaluation of agent safety.

    agentllm agentevaluation framework
  288. arxiv:2610.40714 · cs.LG
    PrivacyFlow: Privacy in AI Agents through Information Flow Analysis and Consequence-Aware Risk Management
    Robin Staab, Tom Wollschläger, Vahab Mirrokni, Reza Shokri

    Autonomous LLM agents executing multi-step workflows over personal data often misalign with user privacy expectations and leak sensitive, out-of-context information. Although Contextual Integrity (CI) formalizes this problem, existing benchmarks and defenses rarely operationalize CI for realistic multi-step task sequences in evolving environments. In this work, we first introduce agentic PrivacyBench, a benchmark of 5873 tasks in 986 task sequences with dynamic environment updates and CI-grounded allow-sets for sensitive information. We also introduce PrivacyFlow, a defense that separates deliberative privacy planning (System 2) from task execution (System 1), treating each sharing decision as a structured risk-management problem. Before execution, PrivacyFlow traces the provenance and purpose of each planned data flow, models recipient background knowledge, and simulates what external actors could infer. This allows us to model the downstream consequences of disclosure, including inference attacks and context-dependent harms. It then weighs the remaining disclosure risk against task utility and executes a minimized plan, blocks it, or consults the user. Across six models, PrivacyFlow significantly reduces step-level privacy failures and raises the share of completed task sequences without a detected privacy violation by 22--55 percentage points. We further highlight a new privacy--utility--autonomy tradeoff: allowing PrivacyFlow to proactively prompt the user in case of uncertainty can increase authorized task completion at the cost of shifting decision effort back to the user. Our results suggest that proactive CI reasoning can substantially improve agentic privacy and motivate further study of this tradeoff.

    ai agentllm agentagenticbenchmark

02 US SEMI · SEC 8-K FILINGS

5 items

scanned: NVDA / AVGO / MRVL / COHR / LITE / AMD / TSM / SMCI / ANET / CRDO / POWL / VECO

  1. $LITE · 8-K · filed 2026-11-02
    Lumentum Holdings Inc
    Items: 2.02,9.01
    8-K
  2. $LITE · 8-K · filed 2026-11-02
    Lumentum Holdings Inc
    Items: 1.01,5.02,7.01,9.01
    8-K
  3. $MRVL · 8-K · filed 2026-11-02
    Marvell Technology Inc
    Items: 1.01,9.01
    8-K
  4. $MRVL · 8-K · filed 2026-11-02
    Marvell Technology Inc
    Items: 2.02,7.01,9.01
    8-K
  5. $VECO · 8-K · filed 2026-11-02
    Veeco Instruments Inc
    Items: 7.01,8.01,9.01
    FORM 8-K

03 HUMANOID · COMPANY NEWS

62 items

scanned: figure-ai / 1x / boston-dynamics / unitree / apptronik / sanctuary-ai / neura-robotics / agility-robotics / physical-intelligence / agibot

04 CN PHOTONICS · ANNOUNCEMENTS

0 items
China sources are not connected yet.