HarnessEvo：揭示 LLM 代理优化收益的局部性与预算分配陷阱
原标题：Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents
AI 摘要
本文提出 HarnessEvo 方法，将 LLM 代理的文本脚手架分解为角色、任务策略、工具/格式规则和反思/控制四个槽位，并采用留一进/留一出协议进行信用分配。实验表明，在 ALFWorld 基准上，优化收益完全集中在反思/控制槽位，而均匀分配预算会导致次可加性陷阱，使所有槽位冻结。作者建议先进行信用分配并集中预算，而非均匀拆分。
正文节选
Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents Abstract A growing body of work makes a frozen large language model (LLM) into a better agent by evolving its harness—the textual scaffolding (persona, strategy, format rules, control heuristics) wrapped around the model. The dominant recipe, exemplified by reflective prompt evolution, treats the harness as a single flat string and optimizes it as a whole. We ask a sharper ques