Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents
Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents
Title: Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents 标题： 提示词框架优化的价值究竟在哪里？自进化大模型智能体中的局部增益与预算分配陷阱
Abstract: A growing body of work improves frozen large language models (LLMs) as agents by evolving their harness: the textual scaffolding around the model, including persona, strategy, format rules, and control heuristics. Existing reflective prompt-evolution methods usually optimize this harness as one flat string. We instead ask where the optimization value actually resides.
摘要： 越来越多的研究致力于通过进化“提示词框架”（Harness）来提升冻结大语言模型（LLM）作为智能体的表现。提示词框架是指围绕模型构建的文本脚手架，包括角色设定、策略、格式规则和控制启发式逻辑。现有的反射式提示词进化方法通常将整个框架视为一个扁平的字符串进行优化。而我们则探讨：优化的价值究竟真正存在于何处？
We introduce HARNESSEVO, which decomposes the harness into four separately evolvable slots: role, task-strategy, tool/format-rules, and reflection/control. Using the same reflective optimizer under an iso-budget setting, we pair this decomposition with leave-one-in and leave-one-out attribution to measure the contribution of each slot.
我们引入了 HARNESSEVO，它将提示词框架分解为四个可独立进化的槽位：角色（role）、任务策略（task-strategy）、工具/格式规则（tool/format-rules）以及反射/控制（reflection/control）。在等量预算（iso-budget）设置下，我们使用相同的反射式优化器，并结合“留一法”（leave-one-in/out）归因分析来衡量每个槽位的贡献。
On ALFWorld with a frozen 7B backbone, HARNESSEVO does not significantly improve the overall binary success rate over either the stock harness or flat-string evolution: 0.657 versus 0.642 and 0.642, respectively. However, the slot-level analysis reveals that nearly all useful optimization value is localized in the reflection/control slot, which achieves a leave-one-in gain of +0.119. The other slots are individually null.
在基于冻结 7B 参数模型的 ALFWorld 任务中，HARNESSEVO 的整体二元成功率相比原始框架或扁平字符串进化方法并无显著提升（分别为 0.657 对比 0.642 和 0.642）。然而，槽位级分析显示，几乎所有有效的优化价值都集中在“反射/控制”槽位中，该槽位实现了 +0.119 的“留一法”增益。其他槽位单独来看则几乎没有贡献。
We further show that uniform budget splitting is harmful: allocating 64 rollouts across four slots leaves only 16 per slot, below the optimizer’s effective search floor, causing every slot to freeze at its empty seed. Concentrating the budget on the high-credit control slot recovers the lost gain, reaching 0.761 with half the split budget.
我们进一步证明，均匀分配预算是有害的：将 64 次运行（rollouts）分配给四个槽位，每个槽位仅分得 16 次，这低于优化器的有效搜索阈值，导致每个槽位都停滞在初始的空状态。将预算集中在贡献度高的“控制”槽位上可以挽回损失的增益，仅用一半的预算就达到了 0.761 的成功率。
The effect is task-contingent. On WebShop, all slots freeze empty and all methods tie, indicating a genuine absence of recurrent, verbalizable control failures rather than budget starvation. Overall, our results suggest that harness value is localized, uniform budget splitting can be actively harmful, and credit assignment should precede structured agent-evolution.
这种效应取决于具体任务。在 WebShop 任务中，所有槽位均停滞在空状态，且所有方法表现持平，这表明该任务中确实不存在可重复、可言语化的控制缺陷，而非预算不足所致。总的来说，我们的研究结果表明，提示词框架的价值是局部化的，均匀分配预算可能适得其反，且在进行结构化智能体进化之前，应优先进行贡献度归因分析。