JoyInside‑VLM‑Agent is a ReAct‑style multimodal agent that learns product‑conditioned policies for embodied robots and smart hardware. It uses a closed‑loop sandbox to generate interaction trajectories, supervised fine‑tuning, multi‑teacher on‑policy distillation, and reinforcement learning with DPER to align token‑level evidence and rewards. The agent achieves a BFCL v4 overall score of 77.44, surpassing baselines on JoyInside‑Devices and JoyInside‑Robotics, and demonstrates visually grounded search, replanning, and failure reporting in two real‑world quadruped case studies.
By Gangyi Li, LuYao Wang, Jingfu Sun, Jingyi Yuan, Hang Liu, Hao Zong, Yujia Hu, Tingting Gao, Jianyong Wang, Xu Yang
MotionCoach is a multimodal agent designed to generate precise guidance for professional motion editing in videos. It constructs grounded prompts and reference clips through multi‑turn tool use and self‑verification, leveraging a training pipeline that includes a 10k SFT dataset and a 2k RL dataset optimized with Group Relative Policy Optimization. Evaluation on the VPMEBench benchmark shows significant performance gains, raising overall scores from 55.1 to 65.6 on Seedance 2.0 and from 51.2 to 64.5 on MiniMax‑H3.
By Siyi Qian, Hanqing Yang, Weihuang Lin, Zihao Yang, Yichen Lu, Yuanhong Zheng, Sheng Li
AgentSWE is a new benchmark for agent software engineering that evaluates coding agents across three lifecycle stages: creating an agent from a natural‑language requirement, editing existing agent codebases, and optimizing toward a measurable target. The benchmark consists of 25 commission‑style tasks, where builders iterate on a few illustrative cases and are then tested on held‑out cases that cover the full requirement, with hard pass/fail gates or objective‑based grading. Experiments show that current coding agents fail to meet the requirements, revealing common failure modes such as overfitting, incomplete runtimes, and unsupported completion claims.
By Jiahao Wang, Hongjin Qian, Yuyang Hu, Jiajun Zhang, Zheng Liu
The paper introduces ChemCost, a benchmark of 1,427 reactions that tests large language models’ ability to estimate chemical procurement costs by grounding identities, retrieving supplier quotes, selecting purchasable packs, normalizing quantities, and computing costs. It includes a frozen pricing snapshot of 2,261 chemicals and 230,775 supplier quotes, and provides scalar scoring plus stage‑level diagnostics for grounding, retrieval, procurement, and arithmetic errors. Experiments show that even advanced LLM agents achieve only about 50.6% accuracy within 25% relative error on clean inputs, and performance drops sharply with realistic noise, highlighting brittleness in parsing, evidence integration, pack selection, and tool use.
By Yuyang Wu, Yue Huang, Shuaike Shen, Xujian Wang, Shuhao Zhang, Qiyao Xue, Weichen Liu, Runtian Gao, Jian Ma, Xiangliang Zhang, Olexandr Isayev
MAGE (Multi-Agent Graph-guided Evolution) is a framework that externalizes self-knowledge of language-model agents into a four-subgraph co-evolutionary knowledge graph, allowing a frozen execution model to retrieve task-conditioned guidance from past reasoning traces and teacher-written corrections. During evolution, the graph, a task-level search bandit, and a skill-level routing bandit are updated from the same reward stream while the learner’s backbone remains unchanged. Across nine diverse benchmarks, MAGE outperforms prompt-based frozen-backbone baselines, and ablations show that self-harvested success traces and teacher-written corrections complement each other, with success memories aiding reasoning-template-heavy tasks and corrective memories supporting harder composition and interaction settings.
By Ruiyi Yang, Zechen Li, Hao Xue, Imran Razzak, Flora D. Salim
The paper introduces RobustBench‑TC, a benchmark that tests tool‑use language agents under 22 real‑world perturbations affecting observations, actions, rewards, and transitions. Experiments on 21 models show that observation noise has minimal impact, while reward‑relevant and transition noise drastically reduce accuracy, and model size alone does not mitigate these gaps. The authors propose ToolRL‑DR, a domain‑randomized reinforcement‑learning recipe that improves robustness, retaining most clean accuracy and narrowing performance gaps even on unseen transition perturbations.
By Xiaolin Zhou, Aojie Yuan, Zheng Luo, Zipeng Ling, Xixiao Pan, Yicheng Gao, Haiyue Zhang, Jiate Li, Shuli Jiang, Prince Zizhuang Wang, Zixuan Zhu, Jinbo Liu, Ryan A. Rossi, Hua Wei, Xiyang Hu
The paper introduces BenchAgent, a framework for evaluating multi‑agent large language model (LLM) workflows against a single‑agent baseline, BenchAgent Core. Using GPT‑4.1 across ten reasoning, coding, and tool‑use benchmarks, the study compares six multi‑agent configurations, finding that EvoAgent achieves the highest mean pooled accuracy (79.70%) while other agents lag by 1.71–8.76 percentage points. The analysis also highlights trade‑offs in token consumption and accuracy, offering guidance for selecting and optimizing LLM multi‑agent workflows.
By Yuhang Fu, Ruishan Fang, Jiaqi Shao, Huiyu Zheng, Zhengtao Zhu, Bing Luo, Zhenyu Wen, Tao Lin
The paper introduces Agent Step Value (ASV), a metric that measures how step rewards change across different evaluation channels in black‑box agent traces. Using a matched 2x2 design on frozen PubMed question‑answering transitions, the authors find that cross‑channel gain signs disagree in 48.3% of cases, whereas same‑channel retries disagree only 0.3% of the time. The study shows that negative magnitudes in the generated‑view coordinate drive a positive direct mean and a negative generated‑view mean, and that source‑only generation can restore a positive mean direction on artifact‑bearing retrievals. The results demonstrate how ASV can localize channel sensitivity and test interventions before reusing step rewards.
By Andrew Zhang, Chengzhan Li
The paper introduces a dual‑agent framework that automates the evolution of Meta‑Black‑Box Optimization (MetaBBO) algorithms. By treating the design loop as a coding task, a task agent refines the MetaBBO codebase while a hyper agent continually modifies both agents, enabling recursive self‑improvement. The resulting evolved MetaBBO variants outperform current human‑crafted baselines and adapt quickly across different optimization domains, revealing emergent design principles.
By Zipei Yu, Yue-Jiao Gong, Zeyuan Ma, Yuncheng Jiang, Zhiguang Cao
DREAM (Dynamic Resolution Assignment For Multimodal Multi-Agent Debate) enhances multimodal multi-agent debate by dynamically assigning each agent the optimal visual resolution through a zero-shot probe round that measures uncertainty with Average Normalized Log-Likelihood. It also introduces Uncertainty-Guided Rollback Aggregation to mitigate groupthink by tracking and restoring low-uncertainty answers that are overridden by group pressure. Across six multimodal datasets, DREAM improves the accuracy-token trade-off by 1.5–3.2% over existing baselines without requiring dataset-specific tuning.
By Khanh-Binh Nguyen, Van Dai Do, Tien Anh Nguyen, Svetha Venkatesh, Hung Le
Bolzano is a multi‑agent, open‑source system that employs parallel prover agents and a verifier agent to conduct proof search while maintaining a human‑readable research state. In initial manual trials on expert‑selected problems, it produced eight proofs verified by domain experts. When deployed without problem‑specific guidance on roughly 3,800 open problems from four paper sets, Bolzano solved about 200 of them, including four questions from papers accepted to the STOC 2026 conference, which were confirmed by the authors.
By Adri\'an Z\'ame\v{c}n\'ik, Mat\v{e}j Kripner, Martin Kouteck\'y, Martin Balko, Jan Greb\'ik, Pavel Hub\'a\v{c}ek, Robert \v{S}\'amal, V\'aclav Rozho\v{n}
The paper discusses how the shift from centralized AI APIs to local edge inference enables Client‑Side Autonomous Agents (CSAAs) that collaborate in Agentic Peer‑to‑Peer (P2P) networks. Unlike traditional P2P systems that exchange static content, these networks share dynamic, heterogeneous capabilities that may be unsafe, necessitating semantic discovery, drift management, and risk‑aware execution. The authors propose a plane‑based reference architecture, signed soft‑state capability descriptors for intent‑aware routing, and a tiered verification spectrum (reputation, canary probes, cryptographic evidence) to align trust costs with task risks, showing improved workflow success in simulations.
By Taotao Wang, Lizhao You, Jingwen Tong, Chonghe Zhao, Shengli Zhang
AnomalyAgent is an agentic framework for industrial anomaly synthesis that uses five tools—Prompt Generation, Image Generation, Quality Evaluation, Knowledge Retrieval, and Mask Generation—to iteratively refine anomaly samples. It employs a two‑stage training process combining supervised fine‑tuning and reinforcement learning, guided by a three‑part reward system that encourages task performance, self‑reflection, and trajectory adherence. On the MVTec‑AD dataset, AnomalyAgent outperforms zero‑shot state‑of‑the‑art methods, achieving superior anomaly generation scores and high classification and detection accuracy.
By Jiaming Su, Tengchao Yang, Ruikang Zhang, Zhengan Yan, Haoyu Sun, Linfeng Zhang
The paper introduces a multi-agent peer‑reviewed reasoning approach for medical question answering, where several large language models independently generate chain‑of‑thought explanations and then evaluate each other’s reasoning for factual correctness and logical soundness. The highest‑rated reasoning chain is used to produce the final answer. Experiments with five state‑of‑the‑art LLMs on three benchmark datasets show that this method consistently outperforms single‑model chain‑of‑thought reasoning and majority voting ensembles, achieving an average accuracy of 0.820 versus 0.777 for the best single model.
By Zaifu Zhan, Shuang Zhou, Rui Zhang
The paper reports that all 1,325 theorems from the VeriContest benchmark—comprising 1,007 competitive‑programming problems in Rust with Verus specifications—were successfully proved using Rust‑Prover, a Lean 4‑based verifier. The authors translated each Verus specification and Rust program into Lean, then used Lean’s kernel to confirm the proofs, achieving a 70% first‑iteration success rate and completing 1,259 proofs in under 32 hours on Claude Opus 5.5. The translated Lean programs matched the original Rust outputs on 21,413 benchmark test cases, and the study also compared performance across multiple large‑language‑model settings, noting that higher reasoning effort increased cost without improving proof counts.
By Traian Serbanuta, Jun Xu, Andrei Stefanescu, Cosmin Radoi
The paper investigates reflective prompt optimization (GEPA) for a multi‑agent deep‑research language‑model, focusing on a topic‑disjoint train/validation/test split to avoid memorization. Across three random seeds, the optimized prompts improve held‑out citation reward from 0.702 to 0.781, with a 95% confidence interval of [0.051, 0.107]. The study shows that the gain is driven by decomposed per‑dimension feedback rather than a single scalar reward, and that the improvement persists even with an out‑of‑family verifier and a hand‑written version of the discovered recipe.
By Elias Lumer, Sahil Sen, Vamse Kumar Subbiah, Kevin Paul
MADS (Multi-Agent Dialog Simulation) is a synthetic data generation pipeline that creates structured, diverse dialogue trajectories, including adversarial conversations that reflect realistic failure scenarios. The resulting MADS-Bench benchmark covers six domains, 55 personas, and 1095 conversations, with about 66% featuring failure-prone situations. Models trained with MADS are evaluated using a two-stage approach combining supervised fine-tuning and reinforcement learning with verifiable rewards to assess robustness.
By Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni, Mudit Arora, Hugo Lin, Justin Chiu, Alessandro Di Bari, Neha Gupta, Andreas Stolcke
The paper introduces SEEDS, a preview-guided evidence admission framework that allows search agents to selectively admit retrieved documents into their reasoning process. SEEDS exposes lightweight previews of candidates and lets the agent decide which and how many to fully inspect based on its current reasoning state, enabling arbitrary subset selection rather than fixed or rank-based policies. Experiments on nine knowledge-intensive QA benchmarks show that SEEDS improves search effectiveness and reduces full-evidence consumption compared to traditional retrieval prefixes.
By Jiajin Liu, Chenxi Xu, Dongzhe Fan, Yuanfu Sun, Qiaoyu Tan
Evo$^2$ is a framework that jointly evolves a language model’s weights and its harness—the set of prompts, tools, and feedback mechanisms—using a dual-role pipeline where a single policy acts as both a Proposer and a Reflector. The training pipeline EvoTIDE combines supervised fine-tuning with interleaved GRPO to refine both roles based on harness edits and execution outcomes. Experiments on coding benchmarks show Evo$^2$ outperforms existing self‑evolution methods, achieving higher scores on Polyglot and SWE‑bench Verified, and its evolved harnesses transfer across tasks and models.
By Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews
OccludeBench is a manually curated benchmark that tests vision‑language models (VLMs) on their ability to recover fully occluded objects from remaining visual evidence. The benchmark distinguishes direct cases—requiring recovery from a specific physical cue such as a shadow or reflection—from inferential cases, which rely on broader scene context. Across five VLMs, accuracy averages 42.7 % on direct cases versus 62.7 % on inferential cases, a 20‑point gap, while humans achieve 63.4 % and 68.5 % respectively with no significant gap. The study introduces OccludeTrace, a four‑stage diagnostic that isolates the failure point, revealing that models struggle most when recovering the relevant object from the full image projection. Interventions such as inpaint‑then‑confront reduce the gap to a statistically non‑significant difference for models with image‑generation capabilities.
By Jasin Cekinmez, Addison J. Wu, Ryo Mitsuhashi, Linrong Cai, Xingyu Fu, Zhuang Liu