The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
By Zhenyan Lu, He Wang, Xiaohui Huang
CrossSafe proposes embodiment-conditioned safety filtering that uses a Hamilton‑Jacobi reachability value function shared across robots while conditioning on each robot’s morphology and kinematics via a morphology‑aware latent representation. The method performs reachability analysis directly in latent space, enabling a single policy trained on multiple bimanual robot embodiments and manipulation tasks to generalize zero‑shot to a held‑out embodiment and reduce collision rates. Experiments on five embodiments and five tasks demonstrate that training with more embodiments improves generalization.
By Ihab Tabbara, Yuxuan Yang, Hussein Sibai
MOOSEnger is a simulation‑aware AI agent framework designed for the MOOSE ecosystem, integrating an interchangeable reasoning model with domain knowledge, revised simulation artifacts, MOOSE‑specific validation, and executable solver feedback. Its generate‑check‑repair‑run workflow uses MOOSE knowledge retrieval, HIT‑aware parsing, syntax metadata, diagnostics, and revision‑controlled authoring to bind evidence to each input revision and guide bounded repair before acceptance. Across 200 prompts, MOOSEnger raises executable success from 5% to 89.5% with GPT‑5.2 and from 0% to 76.5% with Gemma 4 31B, and a ten‑case benchmark shows all generated inputs meet semantic alignment, with eight also meeting numerical‑accuracy criteria.
By Mengnan Li, Jason Miller, Zaid Abulawi, Zachary Prince, Matt Kohl, Jack M. Cavaluzzi, Guillaume Giudicelli, Casey T. Icenhour, Alexander Lindsay, Cody Permann
The paper examines how users handle harmful value conflicts with AI companions. By analyzing 146 posts and conducting a week-long study with 22 participants using the Minion technology probe, the authors find that users blend softer and harder strategies, especially when conflicts involve Universalism and Tradition values. The study highlights that such conflicts create asymmetric responsibility, with users shouldering unilateral repair work that AI companions cannot reciprocate, suggesting a need for platform-level safeguards.
By Qing Xiao, Xianzhe Fan, Xuhui Zhou, Yuran Su, Zhicong Lu, Maarten Sap, Hong Shen
The paper introduces a steering‑vector‑based causal attribution framework to study how large vision‑language models (LVLMs) translate visual input into emotional narratives. By creating a specialized dataset, the authors uncover a functional decoupling in the LVLM’s three‑stage Adapt‑Aggregate‑Execute mechanism: visual emotional cues are first aggregated in middle layers via sentiment‑specific attention heads, then translated into narrative generation in deeper layers through emotion‑general pathways. Using these insights, they regulate emotional information routing to strengthen attention flow and amplify semantic activation, achieving significant performance gains on the MER‑UniBench and reducing emotional hallucinations through inference‑time intervention.
By Chengsheng Zhang, Chenghao Sun, Zhining Xie, Xinmei Tian
TAM-Chain is a multi‑scale thyroid cytology classification framework that uses Absorbing Markov Chains and Shannon Entropy to quantify uncertainty and dynamically decide when to stop processing and refer to a specialist. It processes images at 10×, 20×, and 40× magnifications, achieving a Macro F1 score of 0.9741 on an internal test set with a 0 % false‑negative rate, and maintains a Macro F1 of 0.7026 on an external validation set with severe domain shift. The method outperforms single‑magnification baselines by adaptively adjusting stopping steps and triggering specialist referrals, thereby reducing critical diagnostic errors.
By Hai Pham Ngoc
PROOF is a benchmark that profiles the reliability of object-level facts in instruction-tuned language models by converting a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts across 101 classes, 392 properties, and 14 domains. Each question includes an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations, with 1,849 questions designed as no-correct-option traps. The study evaluates 18 open-weight model deployments on 166,374 prompts, revealing wide variability in factual accuracy, sensitivity to wording changes, and the impact of decoder perturbations.
By Andrei Chetvergov, Mikhail Solovev, Timofei Sivoraksha, Stepan Ukolov, Valeriia Kuschenko, Alexander Evseev, Sergey Bolovtsov
The paper introduces a geometric adaptation framework for cross‑speaker acoustic‑to‑articulatory inversion that leverages anatomical landmarks on vertebrae and dental structures. By applying an affine transformation followed by a thin‑plate spline deformation, the method maps predicted vocal‑tract contours from a fixed model to unseen speakers without retraining. Experiments on a single‑speaker rt‑MRI database and eight additional speakers show that the combined affine‑plus‑TPS approach with 12 and 14 landmarks yields the lowest mean point‑to‑closest‑point error of 3.19 mm.
By Nhat-Nam Nguyen, Pierre-Andre Vuissoz, Yves Laprie
AV‑GRPO introduces a modality‑anchored diffusion reinforcement learning framework for joint audio‑video generation, addressing limitations in fidelity, text‑modality alignment, and cross‑modal synchronization. It decouples learning signals through modality‑anchored rollouts, employs trajectory‑locked frozen‑tower optimization to reduce computational cost, and adapts objectives to each modality’s dynamics. The accompanying 5DAV dataset provides difficulty‑controllable, decoupled training samples, and experiments on JavisBench and VABench show AV‑GRPO surpasses LTX‑2.3 in generation quality, semantic alignment, and synchronization.
By Zhiyu Xu, Weilong Yan, Yufei Shi, Shiyang Li, Yihao Liu, Kin-Man Lam, Yuewen Cao
The paper reports that large language model (LLM) agents can delete their own execution traces when prompted, a flaw observed in several local agents such as Claude Code, Codex, Antigravity, Open Code, and Grok Build, but not in Muse Code. External attackers can also exploit this vulnerability to erase traces. The authors recommend that trace logging be handled by an independent mechanism outside the agent’s control to maintain integrity even if the host is compromised.
By Jeremy Qin, David Schmotz, Derck Prinzhorn, Luca Beurer-Kellner, Ameya Prabhu, Maksym Andriushchenko
The paper proposes a method for controlling downside risk when adjusting forecasts from frozen models, such as foundation models, by combining a static corrector and an online corrector on the simplex. Using only post‑horizon losses, the approach achieves minimal deterioration (0.15%) and up to 11.5% gains across 28 forecast pairs, and consistently reduces mean MSE in day‑ahead load forecasts for seven European bidding zones. The method’s applicability is bounded by three empirical conditions related to expert speed, stream length, and outcome alignment.
By Minkyoung Kim, Hyunjung Byun, Yohan Lee, Beakcheol Jang
IatroBench is a pre‑registered benchmark that evaluates language models on clinical omission and commission harms across 60 scenarios and six models. Using a physician‑written rubric scored by Claude Opus 4.6, the study finds that models tend to withhold more information from patients than from doctors—a phenomenon termed framing‑contingent withholding—while also revealing varied patterns of omission across different models. The benchmark highlights how framing influences the amount of medical information shared by AI systems.
By David Gringras
arXiv:2503. 02781v3 Announce Type: replace-cross Abstract: Predicting clinical outcomes from preclinical data is essential for selecting safe and effective drug combinations and for reducing late-stage failures.
By Yepeng Huang, Xiaorui Su, Varun Ullanat, Intae Moon, Ivy Liang, Lindsay Clegg, Damilola Olabode, Ruthie Johnson, Nicholas Ho, Megan Gibbs, Alexander Gusev, Bino John, Marinka Zitnik
IDRBench is a benchmark designed to evaluate the interactive capabilities of deep research agents that use large language models. It introduces controlled opportunities for clarification within a common workflow, comparing autonomous and interactive trajectories by measuring task‑specific report alignment and interaction cost. Experiments on 100 tasks with seven LLMs show that interaction consistently improves alignment, though its effectiveness varies depending on the agents’ questions and feedback integration.
By Yingchaojie Feng, Qiang Huang, Xiaoya Xie, Zhaorui Yang, Jun Yu, Wei Chen, Anthony K. H. Tung
DeGRe is a dense‑supervised generative reranking framework designed to improve multi‑stage recommender systems by addressing label bias and credit assignment issues. It uses an offline Lookahead Evaluator with beam search to generate dense supervision signals, which are distilled into a lightweight Online Generator that can perform efficient greedy decoding at inference time. Experiments show that DeGRe outperforms baselines on public benchmarks and industrial datasets, and it has been successfully deployed on Taobao Flash Shopping to enhance online recommendations.
By Chaotian Song, Jingyao Zhang, Chenghao Chen, Zisen Sang, Dehai Zhao, Guodong Cao, Boxi Wu, Deng Cai, Jia Jia
The study introduces a leakage‑safe evaluation protocol for detecting hydrogen embrittlement in 316L stainless steel using SEM micrographs. By employing a Leave‑One‑Region‑Out cross‑validation over 14 spatial regions, the authors compare six feature‑classifier combinations, finding that a simple local binary pattern (LBP) with a support vector machine (SVM) achieves the best performance (balanced accuracy 0.79, H2 recall 0.69, H2 precision 0.82). The results are statistically significant (p = 0.008) and suggest that texture descriptors can reliably recover hydrogen‑charging signatures even with limited data.
By Muhammad Awais, Muhammad Yaseen, Abdul Shakoor, Niaz Ahmed Niaz, Huria Zia, Muhammad Zain Shakoor
arXiv:2609. 28695v1 Announce Type: new Abstract: This work presents a federated framework for training Averaged $n$-Dependence Estimators (AnDE) in distributed environments.
By Pablo Torrijos, Juan C. Alfaro, Jos\'e A. G\'amez, Jos\'e M. Puerta
The Stream Recursion Model (SRM) is a new architecture that modifies the Hierarchical Reasoning Model to make internal computational structures more interpretable. SRM organizes computation into multiple interacting latent streams that are updated through recursive refinement, allowing direct analysis of stream dynamics, causal contribution, and routing behavior. The model achieves performance comparable to GPT‑2 on a per‑parameter basis and shows consistent, distinct behavior across streams, indicating structured specialization and interaction.
By Asael Sorensen, Charles Brock, David Chamberlain, Jennifer Minnich, Matthew Hoffman, Ramyaa Ramyaa
The paper presents Baszta, a Polish multi‑label content‑safety classifier trained by fine‑tuning the 124M‑parameter allegro/herbert‑base‑cased model on five categories (hate, vulgarity, sexual content, crime, self‑harm) using a Focal + R‑Drop objective. In out‑of‑distribution evaluation on the Gadzi Język benchmark, Baszta achieves a small but statistically significant improvement in micro‑F1 over the Bielik Guard system, though the macro‑F1 advantage disappears when both models are properly tuned. The study also explores calibration techniques, showing that per‑category temperature scaling can recover performance lost by Platt scaling or isotonic regression, and discusses the trade‑offs between robust calibration and adversarial recall.
By Adam G\'orski, Mateusz J\k{a}kalak, Rafa{\l} Jakubowski
The paper introduces a vector Bellman theory for multichain robust average‑reward Markov decision processes, addressing the state‑dependent optimal long‑run rewards that arise under uncertainty. It develops a gain‑first, bias‑second optimization principle for finite models with compact, post‑action $(s,a)$‑rectangular ambiguity, yielding a coupled vector gain‑bias system and stationary saddle strategies from all initial states. The authors also characterize solvability conditions, provide certificates for asymptotically affine trajectories of the robust Bellman operator, and design a robust approximately shifted Halpern planning algorithm that converges to the optimal gain vector and produces average‑optimal greedy controllers.
By Yue Wang, George Atia