Research Papers 论文研究 6h ago Updated 2h ago 更新于 2小时前 48

Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents Harness优化价值存在于何处?自演化LLM智能体中的局部收益与预算分配陷阱

HARNESSEVO decomposes agent harnesses into four evolvable slots (role, task-strategy, tool/format-rules, reflection/control) instead of optimizing as a flat string On ALFWorld with a frozen 7B backbone, overall success rate (0.657) did not significantly exceed stock harness or flat-string evolution (both 0.642) Nearly all optimization value is localized in the reflection/control slot, which achieved a leave-one-in gain of +0.119; other slots showed null contributions Uniform budget splitting acr 提出HARNESSEVO框架,将LLM Agent的harness分解为角色、任务策略、工具/格式规则、反思/控制四个独立槽位进行演化 在ALFWorld(7B backbone)上发现优化价值几乎完全集中在reflection/control槽位(leave-one-in增益+0.119),其他三个槽位单独贡献为零 均匀预算分配存在"陷阱":64次rollout均分至4个槽位后每个仅16次,低于优化器有效搜索下限,导致所有槽位冻结于空种子 将预算集中于高价值control槽位可恢复性能至0.761,且仅需一半预算 效果具有任务依赖性:WebShop上所有方法打平,表明该任务不存在可重复的言语

65
Hot 热度
74
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • HARNESSEVO decomposes agent harnesses into four evolvable slots (role, task-strategy, tool/format-rules, reflection/control) instead of optimizing as a flat string
  • On ALFWorld with a frozen 7B backbone, overall success rate (0.657) did not significantly exceed stock harness or flat-string evolution (both 0.642)
  • Nearly all optimization value is localized in the reflection/control slot, which achieved a leave-one-in gain of +0.119; other slots showed null contributions
  • Uniform budget splitting across four slots is actively harmful: 16 rollouts per slot fell below the optimizer's effective search floor, causing all slots to freeze at empty seeds
  • Concentrating budget on the high-credit control slot recovered lost gains, reaching 0.761 with half the split budget; WebShop showed no verbalizable control failures, confirming task-contingency

Why It Matters

This work challenges the common assumption that structured prompt decomposition inherently improves agent performance, revealing that optimization value is highly localized rather than distributed. For AI practitioners building self-evolving agents, the findings directly impact how computational budgets should be allocated during harness evolution, warning against naive uniform splitting that can degrade performance below baseline.

Technical Details

  • HARNESSEVO introduces a four-slot harness decomposition: role, task-strategy, tool/format-rules, and reflection/control, each independently evolvable under a reflective optimizer
  • Attribution is measured using leave-one-in and leave-one-out methods to quantify each slot's marginal contribution to agent performance
  • Experiments use a frozen 7B LLM backbone evaluated on ALFWorld (interactive navigation task) and WebShop (e-commerce browsing task)
  • The iso-budget setting ensures fair comparison: the same total number of rollouts is allocated across conditions, revealing that uniform splitting (16 per slot) falls below the optimizer's effective search floor
  • Budget concentration on the reflection/control slot alone achieved 0.761 success on ALFWorld, surpassing all flat and uniformly-split variants

Industry Insight

  • Credit assignment should precede structured agent evolution; blindly decomposing prompts without identifying high-impact slots wastes compute and may harm performance
  • Practitioners should adopt adaptive budget allocation strategies that concentrate optimization effort on slots demonstrating measurable gains rather than distributing evenly
  • The task-contingent nature of harness value (present on ALFWorld, absent on WebShop) suggests that harness optimization should be evaluated per-domain, and negative results on one benchmark should not be generalized to all agent applications

TL;DR

  • 提出HARNESSEVO框架,将LLM Agent的harness分解为角色、任务策略、工具/格式规则、反思/控制四个独立槽位进行演化
  • 在ALFWorld(7B backbone)上发现优化价值几乎完全集中在reflection/control槽位(leave-one-in增益+0.119),其他三个槽位单独贡献为零
  • 均匀预算分配存在"陷阱":64次rollout均分至4个槽位后每个仅16次,低于优化器有效搜索下限,导致所有槽位冻结于空种子
  • 将预算集中于高价值control槽位可恢复性能至0.761,且仅需一半预算
  • 效果具有任务依赖性:WebShop上所有方法打平,表明该任务不存在可重复的言语化控制失败

为什么值得看

本文揭示了当前self-evolving Agent研究中一个被忽视的关键问题:harness优化的价值分布高度不均匀,盲目采用均匀预算分配可能适得其反。对Agent系统开发者而言,这提供了从"整体prompt演化"转向"信用分配驱动的结构化演化"的方法论依据。

技术解析

  • HARNESSEVO框架:将harness(文本脚手架,含persona、策略、格式规则、控制启发式)分解为四个独立槽位:role、task-strategy、tool/format-rules、reflection/control,支持分别演化
  • 归因方法:采用leave-one-in和leave-one-out策略,在相同iso-budget条件下测量每个槽位的独立贡献
  • 实验设置:冻结7B backbone,在ALFWorld和WebShop两个benchmark上测试,使用64次rollout预算
  • 关键发现:uniform budget splitting导致每个槽位仅获16次rollout,低于optimizer的有效搜索floor,所有槽位冻结;集中预算至control槽位后达到0.761成功率
  • 任务差异性:WebShop上所有槽位均冻结为空,所有方法表现持平,说明该任务不存在recurrent verbalizable control failures,而非预算饥饿所致

行业启示

  • 信用分配优先于结构化演化:在分解harness进行优化前,应先通过归因分析识别高价值槽位,避免资源浪费在无效维度上
  • 警惕均匀分配陷阱:在预算受限的self-evolving Agent系统中,均匀分配优化预算可能比集中分配更有害,应优先保障高credit槽位的搜索深度
  • 任务适配性考量:harness优化的收益高度依赖任务特性,在缺乏可言语化控制失败的任务上,结构化演化可能无益,需先评估任务本质再决定优化策略

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Alignment 对齐 Evaluation 评测