Agent Seer is a pipeline that automatically synthesizes realistic evaluation scenarios for AI agents that use external tools, using only the tool’s specification (function names, natural‑language descriptions, and typed parameter schemas). Starting from a single Model Context Protocol (MCP) specification, it enriches raw schemas, generates graded scenarios with synthetic tool outputs, and expands them into mock‑data‑grounded multi‑turn dialogues that demonstrate strong tool‑calling correctness and conversational coherence. Across seven diverse MCP specifications, the pipeline achieves high quality, with parameter‑schema complexity emerging as the main driver of quality variation and argument‑value accuracy identified as the dominant failure mode.
By Harish Karumuri, Mahesh Vemula, David Lopes Pegna
The paper introduces Self-Improving Retrieval-Augmented Generation (RAG), a framework that splits document question answering into Retrieval, Reasoning, and Judge agents coordinated by an orchestrator. When the Judge scores an answer below a dynamic threshold, the system retries with broader retrieval, more careful prompting, and relaxed acceptance criteria, achieving 86% oracle-guided accuracy on FinanceBench with a 36.4% Lazarus Rate. The approach logs every decision with confidence scores, providing audit trails needed for regulated financial applications.
By Junjie Xiong, Shawheen Ghezavat, Aum Hirpara
The paper investigates the compatibility of envy-freeness and equitability in fair division, focusing on both indivisible goods and chores. It shows that the relaxed notions EF1+EQ1 may not exist even for normalized additive valuations, but provides an algorithm that finds an EF1+EQ1 allocation for up to seven agents with binary goods. For chores, the authors prove that a stronger EFX+EQX guarantee always exists, regardless of normalization, and they also explore cross-notion ex‑ante and ex‑post fairness guarantees.
By Hadi Hosseini, Shraddha Pathak, Lirong Xia, Chengkai Zhang
Think3D introduces a framework that endows Vision‑Language Models with interactive 3D chain‑of‑thought reasoning by integrating 3D manipulation tools for active spatial exploration. The approach improves performance on benchmarks such as BLINK Multi‑view, MindCube‑1K, and VSI‑Bench‑Tiny for proprietary models like GPT‑4.1 and Gemini 2.5 Pro, and a reinforcement‑learning variant, Think3D‑RL, enables open‑weight models such as Qwen3‑VL‑4B to autonomously learn effective 3D exploration strategies, yielding tool‑use patterns comparable to stronger models and turning a performance drop on MindCube‑1K into a substantial improvement.
By Zaibin Zhang, Yuhan Wu, Lianjie Jia, Yifan Wang, Zhongbo Zhang, Yijiang Li, Binghao Ran, Fuxi Zhang, Zhuohan Sun, Yizhuang Peng, Zhenfei Yin, Lijun Wang, Huchuan Lu
The paper introduces the Agentic Compositional Generalization hypothesis, suggesting that reinforcement learning (RL) primarily refines high‑level decision‑making behaviors that orchestrate pre‑trained low‑level skills, rather than teaching new domain‑specific skills from scratch. It proposes River, a training recipe that enhances reward quality by filtering low‑quality synthetic environments and adding process‑level behavior regularization. Using River, RL‑trained agents outperform other open‑source 8B models on four terminal‑agent benchmarks, achieving significant gains with fewer than 30% of the training environments.
By Yihang Yao, Bo Pang, Xuan Phi Nguyen, Ding Zhao, Shafiq Joty, Semih Yavuz
HINT-SD introduces a targeted self‑distillation framework for long‑horizon language‑model agents that uses full‑trajectory hindsight to identify failure‑relevant actions and applies feedback‑conditioned distillation only to those action spans. This selective approach reduces the need for per‑turn feedback, improving training efficiency and effectiveness. Experiments on BFCL v3 and AppWorld demonstrate that HINT‑SD outperforms dense per‑turn feedback baselines by up to 13.60 percentage points on average while cutting training time per step by 2.26×.
By Woongyeong Yeo, Yumin Choi, Taekyung Ki, Sung Ju Hwang
The paper introduces INTENT-AS-A-TOOL, a method that equips large language models with intent-targeted tools to provide a fine-grained, judge‑free signal of their commitment to specific behaviors during reasoning. By monitoring the probability of calling these intent tools, the authors can track how intent evolves throughout generation, complementing chain‑of‑thought monitoring and expanding post‑hoc labels into dense trajectories. The approach identifies critical steps for online intervention, demonstrating that action preferences are useful for detecting agentic misalignment in autonomous agents.
By Yutong Zhang, Jianshuo Dong, Peng Xu, Long Wang, Jie Zhang, Tianwei Zhang, Xiaoping Zhang, Han Qiu
The paper introduces Skill Pre-Training (SPT), a mid‑training approach that uses public skill packages as data for agentic language models. By applying causal language modeling to a collection called SkillCorpus and employing a Reference Insert strategy to keep file relationships intact, SPT improves agentic performance across various model sizes while largely preserving general capabilities. Experiments also show that mixing skill data with general corpora yields further benefits.
By Yufei Sun, Yudong Li, Yiming Cheng
The paper introduces Thomson, a frontier AI model developed through continual learning on open-weight models, aiming to democratize access to high-performance AI. It argues that institutions with limited resources can achieve frontier-level performance by applying a modern mid- & post-training stack, preserving model plasticity and stability while minimizing high-impact interventions. Thomson demonstrates competitive performance across agentic tasks, safety, legal, tax, multilingualism, and large-scale deep research, exhibiting a distinctive π-shaped improvement pattern and effectively mitigating the forgetting problem seen in narrow domain adaptation.
By Shengzhuang Chen, Jerrod Parker, Yejin Bang, Andrew M. Bean, Nabeel Seedat, Stefan Winzeck, Daniil Glazko, Jannik Zgraggen, Fangyi Yu, Scott Arnott, Dietrich Trautmann, Luca Ciuffreda, Guglielmo Bonifazi, Davide Romano, Bradley Bell, Kirsty Fielding, Daniele Giofr\`e, Tom Zielund, Ipshita Chatterjee, Sneha Murthy Ghantasala, Manpreet Nanreh, John Scoville, Maciej Sakowicz, Wassim Seifeddine, Lukas Thede, Jonathan Richard Schwarz
The paper extends Co‑Scientist, a Gemini‑based multi‑agent system, and validates it in real‑world scientific settings. In materials science it designed a safe precursor route for MXenes and achieved single‑attempt growth of monolayer MoS₂, MoSe₂, and WS₂. In biology it predicted swarming phenotypes of engineered E. coli, and in computer science it discovered a superior inference‑time scaling architecture for HealthBench. A double‑blind study with 30 experts showed that Co‑Scientist’s reliability modules reduce hallucination and plagiarism while improving research safety.
By Samuel Schmidgall, Xiaokai Zhu, Marian Shaw, Lin Yang, Valentin Li\'{e}vin, Jingyun Yang, Yuchen Zhuang, Tim Strother, Alex Bijamov, Min Woo Sun, Anil Palepu, Justin Chen, David Steiner, Jacqueline Shreibati, Wei-Hung Weng, Yilin Zhao, Xingjian Hu, Nicholas Zahn, Sadhya Garg, Julia Kirby, Yuxiang Gan, Jiaoli Li, Divy Thakkar, Shekoofeh Azizi, David Racz, Juraj Gottweis, Vivek Natarajan, Chenglin Wu, Tal Danino, Keran Rong, Haozhe Wang, Benoit Schillings, Yong Cheng, Quoc V. Le, Tao Tu
arXiv:2608. 27260v1 Announce Type: new Abstract: LLM agents increasingly rely on generated interaction data to learn how to interact with external environments.
By Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, Weiwen Liu
Naive Prompt Optimization (NPO) is a lightweight, single‑lineage method that iteratively refines prompts using a teacher model’s rollout feedback. It matches or surpasses the performance of more complex optimizers like GEPA while requiring fewer rollouts, and its advantage grows with stronger teacher models. In interactive games, NPO remains competitive, and prompts optimized by NPO transfer well to other student models within the same family.
By Yuan Chang, Xiaoqi Chen
The paper introduces HarnessLens, a budget‑aware framework that evolves language‑model agent harnesses by jointly exploring task spaces and user‑configurable components. It derives candidate modifications from execution trajectories and selectively verifies them on behavior‑relevant tasks using an attributable‑evidence gate, thereby avoiding wasteful rollouts on unrelated behaviors. Experiments on three harnesses and four benchmarks show that HarnessLens improves held‑out performance by 7.6‑13.6% while using less evaluation budget than existing methods.
By Jinghan Xu, Yikai Zhang, Aili Chen, Weiyuan Li, Jiaqing Liang, Deqing Yang
AgentJudgeBench is a new benchmark that evaluates the reliability of large language model (LLM) judges on agentic tool‑calling tasks involving workflow directed acyclic graphs (DAGs). It contains 3,808 instances across six DAG topologies and three difficulty tiers, tested with five generators (3B–70B open‑weight models and GPT‑5.4) and six judges (20B to frontier scale) under both paired‑with‑and‑without‑ground‑truth conditions. The study finds that judge alignment degrades with task difficulty, ground‑truth exposure can sometimes hurt alignment, and structured evaluation rubrics provide modest improvements, revealing a structural ceiling that model capacity alone cannot surpass.
By Abhigya Verma, Amit Kumar Saha, Seganrasan Subramanian, Sai Harshitha Aluru
DuMateBench is a new benchmark for autonomous agents that uses real user sessions from a large production platform, preserving interaction history, configurations, and workspace state. It contains 200 tasks across 8 scenarios and 17 capability categories, many requiring coordination of multiple capabilities. The benchmark tests agents in Docker containers with real-world complexities—Insufficient, Unstable, and Noisy—and evaluates performance with a hybrid deterministic and LLM-as-Judge protocol, revealing significant gaps in task completion across various agent frameworks and LLMs.
By Zechun Niu, Yukun Zhao, Jiaxin Zhang, Xu Shen, Jinhua Si, Han Tian, Can Xu, Yunfan Song, Jiaxin Mao, Yansong Gao, Yuchen Li, Jianmin Wu, Lingyong Yan, Shuaiqiang Wang, Dawei Yin
WikiSkill is a framework that separates raw execution experience, accumulated knowledge, and executable skills, continuously consolidating experience into a persistent knowledge base (wiki). By co‑evolving agent skills with this wiki, the method consistently outperforms state‑of‑the‑art skill‑evolution techniques across diverse benchmarks and models. The study shows that larger models benefit more from evolved skills, smaller models can outperform larger ones when equipped with skills, and that skills transfer effectively across model families, with the wiki’s persistent knowledge being critical for success.
By Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu
PILOT introduces a live self‑improvement framework for long‑horizon agents, enabling a supervisor to redirect or abort an active worker during execution and to distil learned procedures into reusable skills. Unlike existing single‑agent or subagent architectures, PILOT’s supervisor‑worker harness supports real‑time steering and self‑evolution. Experiments on two frozen backbones across three benchmarks show PILOT outperforming other harnesses, achieving significant gains in performance metrics and efficiency.
By Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Wen Da, Zhaokai Luo, Mu Chuan, Yao Hu, Wenjie Li, Chengyue Jiang
The paper argues that large language models need adaptive reasoning rather than fixed reasoning budgets. It shows that over‑reasoning leads to high computational cost without accuracy gains, while under‑reasoning results in incorrect or incomplete solutions. The authors evaluate these failure modes on MATH‑500 and the GAIA benchmark, highlighting the need for dynamic reasoning allocation in agentic AI systems.
By Md Jueal Mia, M. Hadi Amini
The paper presents a modeling and simulation framework to study reinforcement learning (RL) control of connected and automated vehicle (CAV) platoon joining maneuvers in mixed traffic. Using SUMO and agent-based modeling, it evaluates Deep Q-Network (DQN), Double DQN (DDQN), and Proximal Policy Optimization (PPO) algorithms, finding that PPO achieves a 98 % joining success rate with less than 1 % collision rate by incorporating risk penalties. The study also shows a trade‑off between safety, joining effectiveness, and decision efficiency, and demonstrates that an external safety controller can prevent collisions but may reduce joining efficiency.
By Biao Yin, Abderrahmane Kasmi, Nadir Farhi
AgentFold is a multi‑agent framework that treats protein‑folding model design as a closed‑loop search over executable code variants. Starting from the ESMFold codebase, the agents generate hypotheses, modify and debug code, evaluate model variants, and store both successes and failures in structured memory, guided by an MCTS‑style policy that allocates GPU resources. In an engineering‑scale experiment, AgentFold explored about 80 variants using 5,000 GPU‑hours and 170 million LLM tokens, improving the best lDDT score by 7.5% over independent Codex proposals and outperforming a random‑search baseline, while also uncovering empirical design patterns such as the benefits of early, soft, learnable priors.
By Mingquan Liu, Jiangyu Chen, Hanqun Cao, Xujun Zhang, Pengsen Ma, Xiangru Tang, Shuting Jin, Zhuo Yang, Tianfan Fu, Fang Wu, Xiangxiang Zeng