arXiv cs.CL· Yongqiang Chen, Chenxi Liu, Deyu Zou, Zhenhao Chen, Tongliang Liu, Bo Han, Kun Zhang·· 13 小时前AI 评分38
ExpEvolve：为开放式发现结构化并复用经验
ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery
AI 导读
ExpEvolve 将想法级提取与控制注入进化式编码智能体，从程序中挖掘能解释评分差异的高层因素，在 11 个开放式数学任务上把最终性能和样本效率最高提升 5.0x。该方法还迁移到 4 个其他 LLM 和 MLS-Bench 的 6 个机器学习任务，并在发现进入平台期时用经验指导利用与探索。
正文 · 原文
Abstract:Evolutionary coding agents such as AlphaEvolve are one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite their success, mutations over programs are mostly low-level upon surface heuristics, without a persistent unit of control. In contrast, science compounds through ideas. In this work, we propose ExpEvolve, which injects idea-level extraction and control into evolution: it discovers and leverages the high-level factors from programs that explain scoring differences among different programs. We characterize theoretical conditions on when discovery with idea factors improves sample efficiency. We also build a practical framework to show that the extracted high-level idea factors are helpful for guiding exploitation from experience and exploration when discovery comes to a plateau. Across 11 open-ended mathematical tasks, ExpEvolve significantly improves the final performance and sample efficiency up to 5.0x. The improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench. Analysis shows that ExpEvolve can discover idea-level factors to provide better exploration and exploitation for evolutionary discovery.
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL); Machine Learning (stat.ML) |
| Cite as: | arXiv:2610.29303 [cs.LG] |
| (or arXiv:2610.29303v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29303 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yongqiang Chen [view email]
[v1]
Wed, 21 Oct 2026 07:00:23 UTC (776 KB)
来源：arXiv cs.CL · arxiv.org