arXiv cs.CL· Shiding Zhu, Yudi Qi, Yajie Wang, Jiaze Li, Chao Song, Yaorui Shi, Yibo Miao, Hanqi Gao, Kai Zhang·· 12 小时前AI 评分39
EDV：面向智能体经验学习的执行-蒸馏-验证框架
Beyond Self-Reflection: An Execute-Distill-Verify Framework for Agentic Experience Learning
AI 导读
研究提出 EDV（Execute-Distill-Verify）框架，用于离线构建 LLM 智能体的经验记忆，避免反思机制把执行错误当作经验存储的“自我确认陷阱”。
正文 · 原文
Abstract:Experience learning lets large language model agents reuse lessons through external memory without updating model parameters. However, reflection can endorse execution errors and store them as future guidance, a failure mode we call the Self-Confirmation Trap. We propose EDV, an Execute-Distill-Verify framework for offline experience construction. Heterogeneous agents independently execute each task; a third-party distiller compares their trajectories and environmental feedback to extract candidate lessons; and the executors separately review each candidate, retaining only unanimously approved lessons. Evaluation uses a single solver retrieving from the frozen memory bank. Experiments on $\tau^2$-bench, Mind2Web, and MMTB show improved revisited-task performance, with additional evidence of within-domain transfer to held-out tasks on $\tau^2$-bench. On an annotated Retail subset, EDV reduces the Self-Confirmation Trap Rate from 11.0% with RB-MaTTS to 5.0%, indicating less retention of erroneous guidance.
| Comments: | 20 pages, 8 figures, 9 tables. Revised title, manuscript, and appendices |
| Subjects: | Computation and Language (cs.CL) |
| ACM classes: | I.2.7; I.2.11 |
| Cite as: | arXiv:2606.24428 [cs.CL] |
| (or arXiv:2606.24428v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.24428 arXiv-issued DOI via DataCite |
Submission history
From: Shiding Zhu [view email]
[v1]
Tue, 23 Jun 2026 11:05:05 UTC (905 KB)
[v2]
Sat, 31 Oct 2026 08:21:43 UTC (569 KB)
来源：arXiv cs.CL · arxiv.org