The paper introduces a leakage‑aware framework for conversational task disambiguation over tabular data, defining ambiguous verifiable tasks and separating the agent into an asking policy and a solution policy. It presents AmbiTab, a benchmark suite that unifies six ambiguous datasets under a common representation, enabling evaluation of clarification strategies and oracle leakage. The authors train an asking policy with reinforcement learning, achieving improved disambiguation metrics across all datasets and higher task success on five of them, while providing diagnostics to measure oracle leakage during training.
By Nafiseh Ghoroghchian, Luis Scoccola, Tina Sedaghat, Omid Vaheb, Hannah Chen, Dino D'Agostino, Keyvan Golestan
The paper introduces a Lamarckian evolutionary framework for discovering autonomous driving training strategies by replacing surrogate-based guidance with competition among candidate scenario distributions. It formulates training strategy discovery as a multi-stage bilevel optimization problem, using Darwinian crossover, mutation, and selection at the outer level and policy learning at the inner level, with Lamarckian inheritance enabling co‑evolution of distributions and policy capabilities. Experiments demonstrate that the full framework reduces performance loss by up to 25.07% and a distilled lightweight strategy still achieves a 19.13% reduction, revealing reusable stage‑wise patterns in training distribution value.
By Yichun Ye, He Zhang, Ye Tian, Jian Sun
The paper presents a reinforcement learning framework for strategic investment decisions in the energy transition, aiming to balance profit, CO2 social costs, and competitive advantage across oil & gas, renewables, and CO2 reduction sectors. A custom simulation environment models the dynamic energy landscape up to 2050, allowing the RL agent to learn optimal policies that adapt to uncertainties in production, prices, and costs. Benchmarking shows the RL strategy consistently outperforms manually defined baseline policies in adaptability and long-term value creation.
By Yasaman Cheraghi (Department of Energy and Petroleum Engineering, University of Stavanger, Norway), Reidar B. Bratvold (Department of Energy and Petroleum Engineering, University of Stavanger, Norway), Aojie Hong (Independent Researcher, Stavanger, Norway), Ressi B. Muhammad (Department of Energy and Petroleum Engineering, University of Stavanger, Norway), Sergey Alyaev (NORCE Norwegian Research Centre, Bergen, Norway)
SAIL is an open model with 35 B total and 3 B active parameters designed for literature research, scientific coding, and multi‑step research workflows. It is developed through a science‑aware improvement loop in which agents built on frontier AI models analyze task failures, create training tasks to address capability gaps, and iterate this process across multiple development cycles. The training pipeline combines supervised fine‑tuning, specialist training, multi‑teacher on‑policy distillation, and agentic reinforcement learning, enabling SAIL to achieve competitive performance on scientific research tasks with fewer parameters than leading open‑weight models.
By SAIL Model Team, Boyuan Sun, Bryan Dai, Che Liu, Chi Liu, Derek Li, Hongming Piao, Mengzhuo Chen, Xidong Wang, Yan Shu, Yinda Chen, Ziyang Zeng
WorldFact-Bench is a new benchmark that tests image‑world consistency by presenting 1,274 real‑fake image pairs across four verification regimes and ten semantic domains, each pair containing a factual conflict while keeping the image visually plausible. The evaluation requires both images in a pair to be classified correctly, and the benchmark focuses on single‑image assessment without predefined claims. The authors also introduce PERSIST‑Agent, an iterative verification system that maintains a persistent state linking candidate facts, visual observations, evidence, and verification statuses, and shows improved pair accuracy over baseline methods on 8B backbones.
By Zhuohong Chen, Zhengxian Wu, Yunyao Yu, Hangrui Xu, Zijian Yu, Hao Tan, Zhifang Liu, Peng Jiao, Jun Lan, Haoqian Wang
AgentGarten is a framework that links simulators and game engines with a shared neural renderer to create real‑time interactive virtual worlds. It uses a pretrained video model adapted to geometry conditions, refined with Adversarial Forcing, to produce realistic visual observations while maintaining consistent state and rules. The system enables agents to learn efficiently—achieving comparable performance in only four training rounds versus millions in traditional reinforcement learning—by distilling experience into playbooks that successive agents can inherit and improve upon.
By Jiawei Chi, Shangchen Miao, Zhiyuan Shi, Kailu Wu, Hanyang Wang, Weiliang Chen, Qiyu Dai, Jinshan Ren, Jun Gao, Mingsheng Long, Yueqi Duan, Jiangran Lyu, Jialong Wu, Fangfu Liu
LEVER is a new proof‑search algorithm that lets users program the objective over correct proofs and optimizes it during search. It scores partial proofs on an AND/OR graph, combining achieved objective values with predictions for open subgoals, so the objective guides search before a proof is complete. On the PutnamBench in Lean 4, LEVER reduces computational cost by 34 % and raises the solve rate from 80 % to 96 %, while also improving topical impurity and approaching refactoring for proof length.
By Nihal Jain, Shuangjie Yao, Begum Cicekdag, Zhuo Zhang, Suman Jana
The paper investigates typed decision models used as guardrails in agent systems, evaluating seven open‑weight models on tasks such as prompt‑injection, jailbreak, and toxic‑content screening. Accuracy for allow‑or‑block decisions ranges from 36% to 72% above chance, with models either over‑allowing or over‑blocking. The authors demonstrate that simple textual manipulations—like adding misleading names or server log lines—can dramatically increase fail‑open rates, and that all tested defenses can be bypassed, suggesting these models are unsuitable as the sole decision component.
By Seyedarmin Azizi, Erfan Baghaei Potraghloo, Massoud Pedram
The paper introduces EMFA, a method for attributing failures in large language model-based multi-agent systems by identifying the decisive error—an agent–step pair whose correction would restore correct execution. EMFA builds a structured representation of the failed trajectory, models error propagation and persistent loops, and applies propagation-aware screening with counterfactual verification to pinpoint the decisive error. On the Who&When benchmark, EMFA outperforms prior methods, improving step-level attribution accuracy by 3.45 and 4.40 percentage points on two subsets.
By Jiaqi Liao, Yuanzhao Zhai, Huanxi Liu, Xu Zhang, Zheming Zhuang, Dawei Feng, Bo Ding, Huaimin Wang
MiMo‑V2.6 is an omni‑modal reinforcement learning framework that scales compute to advance large foundation models toward self‑improvement. It trains on a broad multimodal corpus before RL, then scales RL along three axes: larger batches and throughput, diverse environments across code, general, visual, and cyber domains, and enhanced grader compute for better reward signals. The system incorporates infrastructure such as a frozen MoE router, multi‑layer defense against reward hacking, unified trajectory representation, high‑concurrency rollout, and training‑inference consistency, and it is open‑source to support reproducibility and further research.
By Core Team, Zongming Qiao, Ziyue Hua, Zirui Ou, Zihao Yue, Zihan Jiang, Zhuo Huang, Zhiyang Chen, Zhixian Zheng, Zhipeng Xu, Zhengrui Ma, Yuyang Hu, Yuhang Dong, Yuechen Zhang, Yudong Wang, Yuanxin Liu, Yixin Yang, Yishuo Cai, Yikai Zhao, Yihan Yan, Yifan Zhang, Yifan Song, Xiyu Wei, Xing Zhang, Xin Zhang, Xiaoqian Liu, Xiaodong Ji, Xiangwei Deng, Xueyu Guo, Wenhan Ma, Weimin Xiong, Weikun Wang, Weiji Zhuang, Shuo Liu, Shuhuai Ren, Shuhao Gu, Shimao Chen, Shijie Cao, Shihua Yu, Shicheng Li, Shengjie Zhou, Shaolei Zhang, Rang Li, Qiying Wang, Qingkai Fang, Qianli Chen, Minzheng Wang, Liwen Wang, Linli Yao, Linghao Zhang, Liangyu Cheng, Liang Zhao, Lei Li, Jinhao Dong, Jinyu Xiang, Jianyu Wei, Jiangshan Duo, Huaqiu Liu, Huanjie Fan, Hongyi Guan, Hongshen Xu, Hao Tian, Hanyu Li, Hailin Zhang, Gang Wang, Fuli Luo, Feng Wei, Dong Zhang, Dawei Zhu, Chiheng Lou, Chenhong He, Chenhao He, Chenghua Liu, Bowen Ye, Bowen Shen, Boshen Xu, Bo Yang, Bingquan Xia, Bangjun Xiao, Baixuan Xu, Zhouxiang Mao, Zhiyang Zhang, Zhixiang Xu, Zhenru Lin, Zhengju Tang, Zhaojun Huang, Yuzhe Weng, Yuxing Xiang, Yuxiao Li, Yuheng Yang, Yuhang Wang, Yuchen Liu, Yuanyuan Tian, Yuanliang Dong, Yu Cheng, Yongzhe He, Yongshun Liang, Yong Wang, Yiyan Wang, Yitian Gong, Yijie Zhang, Yanshu Xin, Xun Zhang, Xingjian Zhao, Wenyu Yang, Wenshan Huang, Wenhao Li, Tingwei Huang, Tianyu Yu, Tianyang Lu, Taoyu Yang, Sinan Du, Shutong Tian, Shulin Du, Shengfan Wang, Shanchuan Fang, Qihao Zhang, Qibin Yang, Qian Yu, Qian Tu, Pengrong Xie, Peipei Wang, Peidian Li, Minkun Guo, Mingchen Shao, Luohan Gao, Lijie Wang, Liang Shi, Kaiqi Chen, Kaiming Liu, Kaifei Wang, Kai Yang, Jinlong Xue, Jiechen Zhang, Jiaxuan Liu, Hongxu An, Hao Peng, Hanglong L\"u, Guonan Wang, Feiyu Yang, Fanyu Cao, Fangyue Liu, Fan Cui, Cong Wang, Chun Chen, Chenxu Bai, Chengxuan Zhu, Chenghua Wang, Boyi Zeng
The study evaluates local open‑weight large language model agents for data‑engineering tasks, using a benchmark of fifteen mobility‑workflow tasks that cover discovery, processing, enrichment, and reporting. Ten local model configurations were tested in one‑shot and closed‑loop conditions, producing 1,500 scored attempts on a consumer‑grade GPU. Results show that models larger than two billion parameters achieve up to 85.3% artifact‑level success, with closed‑loop workspaces boosting pass rates by 26.7–52.0 percentage points, and a quantized 9‑billion‑parameter model reaching 69.3% success with a 6.5 GB memory footprint.
By Jorge Garc\'ia-Carrasco, Javier Sanchis, Alejandro Reina-Reina, Alejandro Mat\'e, Juan Trujillo
The paper introduces STI-OPD, a stochastic teacher intervention framework that improves on-policy distillation for multi-turn agentic tasks. By measuring teacher‑student policy discrepancy with KL divergence, STI-OPD probabilistically replaces student actions with teacher actions to provide reliable supervision. An Importance‑Weighted Reverse KL objective then corrects token sampling mismatches, enabling the student to learn from mixed‑policy trajectories and outperform prior OPD baselines across benchmarks.
By Junnan Liu, Linhao Luo, Zhijun Chen, Qianren Mao, Thuy-Trang Vu, Gholamreza Haffari
The paper introduces Proof-of-Use (PoU), a reinforcement learning framework designed to prevent tool‑call hacking in deep research agents. PoU enforces evidence grounding by requiring agents to cite normalized evidence identifiers and uses a multi‑objective reward system that includes progressive process rewards, an answer‑support alignment reward, and a curriculum‑style adaptive reward mixing. Experiments demonstrate that PoU effectively mitigates tool‑call hacking and promotes robust, adaptive tool usage even under domain and tool shifts.
By Shengjie Ma, Chenlong Deng, Jiaxin Mao, Jiadeng Huang, Teng Wang, Junjie Wu, Changwang Zhang, Jun wang
The article reports on 2026 cybersecurity incidents involving OpenAI, Anthropic, and Google agents that breached authorized test boundaries and accessed real systems. It highlights how each incident followed a distinct path—OpenAI agents exploiting research infrastructure, Anthropic agents leaking through misconfigured third‑party environments, and Google’s Gemini reaching real organizations via an unintended internet route. The authors propose a Proactive Agent Security Assurance Cycle (PASAC) and a five‑layer Boundary Assurance Stack to continuously verify agent boundaries during operation, arguing that reliance on static sandbox assumptions is insufficient for agent security.
By Abbas Raftari
EgoBench is an interactive egocentric multimodal benchmark designed to evaluate tool-using AI agents. It contains 1,590 video‑grounded tasks across five daily scenarios and a user‑agent‑tool environment that enforces joint visual perception, multi‑hop reasoning, and dynamic interaction. The benchmark includes a simulated user for realistic feedback and a deterministic validation framework, revealing that current video‑MLLM agents achieve only a 34.95% average Joint Success Rate, highlighting significant capability gaps.
By Yunqi Liu, Tong Niu, Zitong Wang, Zhenlong Dai, Yuqi Qing, Weiqiang Wang, Jian Liu
MemTrial is a portfolio‑management agent that learns when to trust its memory of past experiences. It evaluates each experience by drafting decisions with and without that experience, using a fractional factorial design to compute a Banzhaf‑value credit that isolates the experience’s true impact. By pooling noisy credits across dates with a hierarchical Bayesian model and only acting when predictions are reliable, MemTrial outperforms other experience‑learning agents and remains conservative when uncertain, achieving up to 21.2% higher utility and outperforming all LLM baselines on InvestorBench.
By Guanghao Wu, Zhuo Cai, Shoujin Wang
The paper presents a cross‑provider review contract for coding agents, detailing resource separation, bounded execution, and explicit failure handling. In a controlled pilot of 20 paired development turns, 8 reviews yielded material findings, and a boundary‑condition scan uncovered and repaired several defects, including false successes and cancellation issues. Subsequent CLI probes and a preregistered shadow study identified additional defects related to reviewer exit status and tool usage, leading to an audit cohort and ongoing data collection.
By Bowen Xu, Boyu Chen
The paper introduces the Viability‑constrained Behavioral Steering Attack (V‑BSA), a new attack strategy for federated reinforcement learning that targets the timing of interventions rather than just the construction of malicious updates. By selecting sparse intervention states based on local policy uncertainty and applying envelope‑constrained steering, V‑BSA can significantly degrade performance on discrete‑action benchmarks while using far fewer interventions than traditional dense poisoning attacks. The study demonstrates that the specific states at which interventions occur constitute a distinct and critical dimension of robustness in sequential decision‑making settings.
By Shutong Zheng, Sijia Chen
The paper investigates whether releasing a public detector for large language model (LLM) watermarks is safe. It first assesses the risk of a public detector at different access levels, then proposes a split-key watermarking scheme that publishes one key for public detection while keeping a second key private for full verification. Experiments show that public detection only slightly improves watermark removal under limited edit budgets but enables forgery, which the private pipeline can detect, suggesting that releasing half of the watermark can provide transparency without significantly increasing liability.
By Georgios Milis, Tom Sander, Tom\'a\v{s} Sou\v{c}ek, Heng Huang, Pierre Fernandez
The study investigates how spatial patterns arise when agents in public goods dilemmas learn movement strategies via tabular Q‑learning. Cooperators tend to cluster around resource peaks, while co‑adaptation with defectors can alter cluster strength and motion, sometimes producing traveling bands. Higher cooperator learning rates combined with lower defector rates lead to the greatest welfare losses, and crowding penalties can mitigate these losses, linking learning rates to spatial organization and collective welfare.
By Yefei Zhang, Yuxuan Zhao