CAST is a critique‑aware training framework that transforms sparse task outcomes into action‑level supervision for both critique learning and policy optimization. By analyzing agent trajectories, CAST synthesizes structured rationales that explain action validity under partial observability, enabling the creation of richer training data. Fine‑tuned Qwen3‑family models trained with CAST show significant reliability gains, outperforming GPT‑OSS‑120B by over 10% on Retail tasks and improving Telehealth performance by 9% in an out‑of‑domain setting.
By Amir Saeidi, Zehua Zhang, Rishitosh Singh, Naman Ahuja, Vivek Gupta, Ali Payani, Gaowen Liu, Jayanth Srinivasa, Chitta Baral
The paper investigates how language‑model judges can make version‑dependent errors when evaluating upgraded agents. Using 35 public coding‑agent submissions, two customer‑service agents, and over a thousand expert‑labeled trajectories, the authors show that fixed judges often reject task‑conditioned error invariance and can incorrectly approve failed patches, especially as agent capability increases. Paired audits of current outputs reduce interval width only marginally, and the study concludes that independent human patch review is still necessary.
By Jiapeng Li
The paper introduces POrchestra, a proactive agentic orchestration framework that lets sub‑agents request harness adjustments during execution, enhancing adaptability to dynamic environments. It also proposes On‑Policy Orchestration Distillation (OPOD), which uses hindsight feedback from sub‑agents’ requests to provide targeted improvement signals for training the orchestration policy. Experiments on three long‑horizon benchmarks show that POrchestra boosts task success by about 9.1% over the best baseline, and with OPOD the improvement rises to 10.0%, while also improving cost‑performance trade‑offs.
By Xueyang Feng, Chaozhuo Li, Xiaohe Bo, Jiakai Tang, Chenxi Li, Qiwei Ye, Xu Chen
The paper investigates how altering the harness—specifically the way a coding agent manages context and tool outputs—affects performance when the underlying model and task remain unchanged. Two harness configurations were compared on three coding benchmarks: a control that preserves the full conversation in order, and a treatment that mechanically shortens older tool results to keep the context tight. Across all benchmarks, the treatment increased the mean per‑task fail‑to‑pass fraction and, in some cases, the number of complete solutions, demonstrating that the harness itself can significantly influence a frozen model’s effectiveness.
By Sydney Lewis
The paper investigates the use of test‑time search, specifically Monte Carlo tree search (MCTS), to improve agent performance in terminal‑based computer‑use environments. It finds that conventional MCTS does not significantly boost task success due to limited critic information and poor intermediate state evaluation. The authors propose a new strategy called Probe‑and‑Resume, where a critic is invoked only when the agent believes it has finished, and it is given read, write, and execute access to a forked copy of the environment to probe hidden state. This approach raises success rates by 8.4 percentage points over standard MCTS and nearly matches a more expensive Codex CLI agent.
By Jiakai Xu, Tiancheng Ge, Tujie Guan, Yusheng Li, Rohan Timmaraju, Eugene Wu, Kostis Kaffes
The paper introduces VERSE, a Verified Self‑Evolving optimizer that enhances LLM agent harnesses by allowing the optimizer to test edits, replay failures, and perturb steps while tracking fixes and regressions. VERSE builds its own tools for failure analysis, verification, training audits, and workflow control, and uses this feedback to revise the harness’s prompts, skills, tools, hooks, and notes without changing model weights. In experiments across five executors and multiple languages, VERSE improves all evaluated harness optimizers, achieving higher accuracy on held‑out and out‑of‑distribution tasks compared to the strongest baselines.
By Zekai Wang, Yingqiang Ge, Zekun Wang, Hai Wang, Yuhui Xu, Joshua Frandsen, Shancong Fu, Ashia C. Wilson, Chandan K. Reddy
The paper introduces the Agent-Editing World Model (AEWM), a new approach that models how reasoning and actions influence future task progress instead of simulating tool responses. AEWM includes an Action Judge that classifies decisions as Critical, Exploratory, or Noisy, and a State Revision mechanism that edits noisy reasoning–action continuations from the same observed history. The integrated system, EditAct, directly updates the underlying state during real execution, leading to significant performance gains across multiple benchmarks and agent backbones.
By Shuang Sun, Guoxin Chen, Fanzhe Meng, Jia Deng, Huatong Song, Jinhao Jiang, Wayne Xin Zhao, Hongteng Xu, Ji-Rong Wen
arXiv:2610.07335v1 Announce Type: cross
Abstract: Improving the reliability of large language model (LLM) agents in long-horizon decision-making remains a key challenge. When deployed as autonomous a...
By Heewon Park, Somin Im, Minhae Kwon
arXiv:2601. 22758v2 Announce Type: replace Abstract: Large language model agents repeatedly encounter related tasks, yet systems that learn from trajectories commit every lesson to one predefined artifact form.
By Libin Qiu, Zhirong Gao, Junfu Chen, Yuhang Ye, Liangyu Li, Weizhi Huang, Xiaobo Xue, Wenkai Qiu, Shuo Tang
SWE‑Duel proposes a self‑scaling benchmark for coding agents, where tasks and oracles are generated at evaluation time to prevent data leakage and adapt difficulty to participant performance. In the arena, two agents take turns as Red and Blue: Red proposes a feature, injects a subtle bug, and must repair it blind, while Blue reviews the pull request without knowing the bug and must pass all tests. A tournament of 11 model‑harness participants on 15 repositories across five languages showed GPT‑5.5/Codex CLI as the top solver and Claude Opus 4.8/mini‑swe‑agent as the top task author, with newer model versions producing harder tasks for older ones.
By Wei Minn, Zhensu Sun, Lwin Khin Shar, David Lo
A ReAct-style agent accumulates observations, intermediate results, and failed attempts in a single context. As this context grows, stale and off-path information consumes its limited capacity and obs...
The paper introduces Belief-Calibrated Optimization (BCO), a method that records and updates a persistent in‑context document representing an agent’s belief about how the environment responds to edits. By continually revising this world model as new candidates are evaluated, BCO improves the performance of frozen LLM agents across five benchmarks, outperforming a control lacking the world model. An offline ablation shows that the document’s content provides reusable, accurate predictions of environmental responses, beyond mere form.
By Yuhan Chen, Zhihua Tian, Mahavir Dabas, Charith Peris, Rahul Gupta, Ming Jin, Feiyang Kang, Siyuan Zhang, Nan Wang, Ruoxi Jia