ReactHuman is a physics‑grounded benchmark that tests whether multimodal large language models (MLLMs) can make immediate, safety‑critical decisions in simulated humanoid scenarios involving sudden household hazards. The benchmark includes 17 event families, over 1,000 reproducible scenes generated from 240 Hz rigid‑body simulation, and a five‑metric suite evaluating reactions on reasonableness, safety, and physical grounding. Evaluation of seven MLLMs reveals that reactive safety remains unsolved, with models frequently mishandling hazards, relying on appearance over motion, and missing key interception points.
By Yizhan Li, Jianxin You, Mengyang Xiong, Yinhuan Chen, Zicheng Zhao, Dekun Wu, Dongqing Zhang, Bang Liu
The paper introduces an admission‑audit protocol for continual embodied agents, arguing that update admission should balance error control with retained learning opportunities within a fixed interaction budget. It critiques a range‑based confidence gate for failing to certify unchanged old‑task behavior, and proposes a paired‑binomial construction that reduces this burden when outcome disagreements are rare. Experiments on a one‑step pushing diagnostic show that fresh paired checks admit a significant portion of updates while the range‑based gate admits none, and a learned‑dynamics stress test helps distinguish model bias from feedback‑selection error.
By Qinzhen Ma, Ruihai Wu
LogiMed‑RoB is a new benchmark that tests large language models (LLMs) on hierarchical logical consistency in medical risk‑of‑bias assessments, using 860 randomized controlled trials and 14,820 queries based on Cochrane Risk of Bias 2.0 expert logic. The benchmark evaluates models across four dimensions—Atomic Consistency, Domain Consistency, Aggregation Consistency, and Evidential Faithfulness—revealing a catastrophic error‑compounding effect where high atomic accuracy does not translate to end‑to‑end consistency. Experiments on ten state‑of‑the‑art LLMs show that even top models can fail to deduce correct outcomes in a significant portion of cases, highlighting a gap between evidence retrieval and reasoning.
whyItMatters":"The study shows that high outcome accuracy can mask critical reasoning flaws, emphasizing the need for white‑box logical verification before deploying LLMs in clinical settings."
By Jiayu Huang, Zichen Tang, Qianhui Ling, Zemin Kuang, Haihong E
X-AuT is a progressive framework for compressing the audio encoder of speech large language models. It selects layer combinations via short behavioral probes and restores performance through representation alignment, cross‑scale distillation, scheduled student‑policy supervision, and LoRA finetuning, while keeping the language‑model backbone frozen. On ten Chinese–English benchmarks, reducing Qwen3‑ASR‑0.6B’s encoder from 18 to 16 layers lowers macro‑average error from 5.61% to 5.27%, and a 14‑layer model achieves 5.75% error with 20.7% fewer parameters.
By Haojun Zhang, Yi Zou, Min Chen, Qize Yu, Lianrui Fan, Xini Ding, Hao Li, Shuchang Zhou, Xianming Liu, Shiyu Huang
The study examines how generative AI tools like ChatGPT perform on typical first‑year undergraduate mathematics assessment questions. By generating, transcribing, and blind‑marking AI responses to eight assessments covering the entire curriculum, the authors find that AI attains a first‑class level of performance, with consistency across modules that exceeds that of students in invigilated exams. The results suggest a need to redesign mathematics assessments to address the impact of generative AI.
By Benjamin J. Walker, Nikoleta Kalaydzhieva, Beatriz Navarro Lameda, Ruth A. Reynolds
arXiv:2609. 11697v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) and World-Action Models (WAMs) have demonstrated strong capabilities in general-purpose robotic manipulation, yet their generated actions may violate hard physical constraints and therefore be unsafe or infeasible for deployment.
By Jianming Ma, Rongjun Jin, Xiaxi Si, Yang Zhang, Yiheng Li, Yue Gao
RAMamba-Net is a new multimodal fusion network designed for auditory attention decoding (AAD) that combines EEG and electrooculography (EOG) signals. It uses a Mamba-enhanced band-aware convolutional Transformer to capture EEG band-specific patterns and long-range temporal dynamics, while a dual-branch encoder models EOG temporal and inter-channel dependencies. Cross‑modal attention and a reliability‑aware module estimate sample‑wise modality weights, improving fusion robustness and achieving a 5.76% accuracy gain over unimodal baselines on two AAD benchmarks.
By Xingyi He, Ziwei Wang, Dongrui Wu
arXiv:2609. 11291v1 Announce Type: new Abstract: We post-train Qwen3.
By Hyojung Han
Benchmark Radar is a living database and search engine that aggregates AI benchmark papers, datasets, code, and score histories. It automatically discovers new benchmark resources from 37 sources, maintains a catalog of 1,283 records with 12,916 numeric observations, and provides tools such as a web dashboard, CLI, and downloadable evidence for researchers. The system also offers visualizations like a Pareto frontier and trend views to help users assess benchmark saturation and adoption.
By Koutian Wu, Junjie Zhou, Ergan Shang, Jiayu Wang, Pengqian Han, Junkai Wang, Wanghan Xu
The article discusses how production code generated by Claude, Anthropic’s AI, should meet higher standards than human-written code. Anthropic enforces this through numerous guardrails such as lint rules, extensive testing, Claude-driven end‑to‑end tests, daily fuzzers, automated code and security reviews, and automated refactoring. These measures aim to prevent the code from becoming difficult to maintain.
The paper introduces a new framework for evaluating spatial fairness in predictive models by considering individuals’ activity spaces rather than just a single residential location. It proposes associating users with multiple geographic partitions and applying a spatial scan statistic to detect unfairness across these activity-space patterns. Experiments on synthetic datasets demonstrate the method’s effectiveness in identifying unfair treatment and retrieving affected objects, while highlighting a trade‑off in localization performance across resolutions.
By Francesco Lettich, Mario A. Nascimento, Chiara Pugliese, Chiara Renso
The paper introduces CARTS, a steganographic method that uses autoregressive language models to encode a payload text into a stegotext of identical token length by preserving per‑position rank information across contexts. It provides a formal security analysis, proving exact correctness under deterministic model assumptions, and defines key security notions such as context search, key collisions, message equivocation, and non‑commutativity of encoding maps. Empirical tests on Llama 3 8B confirm perfect payload recovery, no random key collisions, and no commuting key pairs, indicating resistance to the studied attack vectors.
By Wissam Ghantous, Alexander V. Mantzaris
The paper introduces Perturbation, a method that treats representations in language models as learning conduits rather than activation patterns. By fine‑tuning a model on a single adversarial example and observing how this perturbation spreads to other inputs, the approach avoids geometric assumptions and does not identify representations in untrained models. In trained models, Perturbation uncovers structured transfer across multiple linguistic scales, indicating that language models generalize along representational lines and acquire linguistic abstractions through experience.
By Joshua Rozner, Cory Shain
The paper presents a 652,157‑parameter action‑conditioned visuotactile world model designed for lifting tasks, integrating behavior cloning, policy learning in imagination, reactive implicit Q‑learning, and model‑assisted force feedback. Experiments on 120 fresh MuJoCo environments and additional ID environments show that visuotactile dynamics reduce force‑action‑effect mean absolute error from 0.413 N to 0.338 N, and model‑assisted feedback boosts force‑budgeted success from 73.3 % to 93.3 %. Imagined reinforcement learning achieves 11.9 % pooled joint success compared to 25.0 % for reactive IQL, with further stress testing adding 330 executions.
By Qinzhen Ma (Rice University)
The paper introduces a statistical framework for post‑training hyperparameter selection, emphasizing the learn‑then‑test (LTT) paradigm. It treats hyperparameter tuning as a multiple hypothesis testing problem over a candidate set, enabling the selection of hyperparameters that meet specified reliability constraints such as risk bounds or information‑theoretic limits. The framework provides finite‑sample control of error probabilities using p‑values, e‑values, and concentration inequalities derived from first principles.
By Amirmohammad Farzaneh, Osvaldo Simeone
XAI-Arena proposes using large language models (LLMs) as judges to evaluate the quality of explainable AI (XAI) explanations, aiming for reproducibility, scalability, and multidimensional assessment. The framework assesses dimensions such as simplicity, clarity, task adequacy, trust calibration, actionability, transparency, faithfulness, and overall interpretability across different datasets, models, and stakeholder personas. Human validation shows a strong positive correlation between LLM-generated and human ratings (Spearman's rho = .693, p < .001), supporting the viability of LLM-based evaluations.
By Yanfei Hu Fleischhauer, Alona Zharova, Nadja Klein, Stefan Feuerriegel
The paper introduces SIRGE, a sequence-informed geometric evaluator for RNA 3D structures that integrates nucleotide embeddings from a pretrained RNA language model into structural representations. SIRGE demonstrates superior performance over existing evaluators in Kendall–τ alignment, Top‑1 selection, and Top‑3 ranking. Controlled experiments reveal that sequence conditioning corrects errors of a purely geometric model and enhances target‑level ranking, suggesting that pretrained sequence representations provide complementary ranking information to geometric reasoning.
By Andrea Zerio, Yighua Yao, Alessandro Micheli, Roland G. Huber, Mile Sikic, Samir Bhatt, Andres R. Masegosa, Yuangang Pan
This paper investigates the ρ^p-Lipschitz constants of deep ReLU neural networks with random weights drawn from a He‑style initialization. For zero‑bias networks, it provides high‑probability upper and lower bounds that differ by at most a logarithmic factor in depth, and shows a sharp contrast between the regimes p∈[1,2) and p∈[2,∞], with the former behaving like the Euclidean norm of a Gaussian vector and the latter like its dual norm. The analysis is extended to networks with non‑zero biases from symmetric distributions, yielding bounds that differ by a logarithmic factor in width and a linear factor in depth.
By Sjoerd Dirksen, Patrick Finke, Paul Geuchen, Dominik St\"oger, Felix Voigtlaender
The paper introduces BodyCam-VQA, an adaptive visual question answering framework designed to improve captioning of police body‑worn camera footage. By employing structured multimodal reasoning and probe question generation, the system extracts fine‑grained visual evidence that conventional captioning models miss. Experiments with various question generation models show that this VQA‑driven approach yields more reliable, objective, and detailed records of enforcement events.
By Karish Gupta, Matthew Alex, Alex Li, Yang Wu, Yun-Wei Chu, Kashif Munir, Xiaotian Zhou, Zhengping Ji, Xiaozhong Liu
The paper introduces Uncertainty DMD, a lightweight framework that injects uncertainty into few-step autoregressive video distillation to counteract diversity collapse. By perturbing the first chunk’s timestep and employing a stochastic cache-writing mechanism for subsequent chunks, the method restores stochasticity without altering the model architecture. Experiments demonstrate consistent improvements in video diversity and motion dynamics while preserving visual quality.
By Zixuan Duan, Xunzhi Xiang, Yabo Chen, Xin Zhang, Changhan Liu, Haibin Huang, Chi Zhang, Qi Fan, Xuelong Li