HuggingFace Daily Papers（社区热门论文）· HuggingFace Daily Papers（社区热门论文）·· 2026-06-04精选AI 评分70
RHO：利用过往轨迹优化LLM智能体工具链的自监督方法
AI 导读
Retrospective Harness Optimization (RHO) 是一种自监督方法，仅利用过往轨迹优化LLM智能体的工具链（技能、工具和工作流程集合）。RHO从历史任务中选取多样化的困难任务核心集，并行重新执行；智能体通过自我验证和自我一致性分析回放，生成候选工具链更新，并依据自身成对自我偏好选择最有效更新。在软件工程、技术工作和知识工作三个领域评估中，单轮优化将SWE-Bench Pro通过率从59%提升至78%，无需外部评分。分析表明RHO有效针对先前失败模式，优化后的工具链改变智能体行为模式，在长周期会话中维持更高准确率。
推荐理由
不靠人工标注就能让 Agent 自我提升，单轮直接把 SWE-Bench Pro 通过率从 59% 拉到 78%，做自主 Agent 优化的同学应该仔细读一下。
正文
Abstract:AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.
| Comments: | Accepted to EMNLP 2026 (Findings). Code: this https URL ; Project website: this https URL |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2606.05922 [cs.AI] |
| (or arXiv:2606.05922v3 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2606.05922 arXiv-issued DOI via DataCite |
Submission history
From: Wenbo Pan [view email]
[v1]
Thu, 4 Jun 2026 09:26:00 UTC (525 KB)
[v2]
Wed, 10 Jun 2026 05:04:53 UTC (513 KB)
[v3]
Sat, 29 Aug 2026 06:46:46 UTC (515 KB)
来源：HuggingFace Daily Papers（社区热门论文） · arxiv.org