跳到正文
arXiv cs.CL· Yuanhao Zeng, Shaocheng Lan, Ao Lu, Kuang Yan, Heng Lin, Xinang Xu, Yuan Luo, Yueting Wang, Yexin Li, Ruo Shan, Kan Ren·· 23 小时前AI 评分55

分辨率假说:智能体进化的关键在于识别改进

The Resolution Hypothesis: Agent Evolution's Key Is Identifying Improvements

AI 导读

论文提出“分辨率假说”,认为智能体进化停滞的原因在于带噪声的评估无法可靠识别出更好的候选产物,而非方法提不出更好的产物。作者在 7 种进化方法和 6 个模型上做了超过 1000 万次 LLM 调用验证:在候选层面,选择机制区分有益与有害产物的能力并不优于随机,误差会逐轮累积;在方法层面,几个百分点的排名差距小于同一方法两次独立运行的差异,已发表的排名常依赖运气。

正文 · 原文

Authors:Yuanhao Zeng, Shaocheng Lan, Ao Lu, Kuang Yan, Heng Lin, Xinang Xu, Yuan Luo, Yueting Wang, Yexin Li, Ruo Shan, Kan Ren

View PDF HTML (experimental)

Abstract:Agent evolution follows a simple loop: a proposer writes candidate artifacts (prompts, skills, memories) and selection accepts those that score higher. This loop has shown promise, yet practitioners face a puzzle: methods trade wins across papers, and evolution often fails to bootstrap on certain tasks. We trace this puzzle to the Resolution Hypothesis: agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably identify them. Across seven evolution methods and six models (over 10M LLM calls), we find support for the hypothesis at both levels where evaluation scores are used. At the candidate level (choosing artifacts), helpful candidate artifacts exist, but selection is no better than random at telling them from harmful ones, so errors accumulate across rounds. At the method level (ranking methods), typical margins of a few percentage points are smaller than the difference between two independent runs of the same method, so published rankings often rest on luck. Together, these findings indicate that agent evolution is often bottlenecked by reliably identifying improvements. This calls for a shift in how we develop and evaluate self-improving agents, from designing ever more candidate improvement systems to making true improvements detectable and accumulable.
Comments: 38 pages, 12 figures, 13 tables. Project page: this https URL
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as: arXiv:2611.23068 [cs.AI]
  (or arXiv:2611.23068v1 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2611.23068

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Yuanhao Zeng [view email]
[v1] Tue, 17 Nov 2026 16:57:03 UTC (4,314 KB)

来源:arXiv cs.CL · arxiv.org