智能体
x.com ·
@omarsar0
Salesforce 论文：harness 与模型协同演化，模仿专家轨迹反而拖累弱模型
Salesforce 在七个企业 agent 任务上先演化出一个适配弱模型的 harness，再用更强专家在同一 harness 下的完整轨迹微调该弱模型，结果 Qwen3-Coder 和 Gemma 4 在全部七项任务上性能下降 4 至 30 分；而在未演化的 harness 下同样的微调却有帮助，说明结果取决于模型与 harness 的匹配。分析认为模仿虽传递知识并提升 scaffold 使用率，但弱模型会照搬专家的规划策略却无力执行，且不再匹配围绕自身规划风格演化的 harness。作者提出的修复方式是让 meta-level agent 定位弱模型 rollout 中的失败轮次，仅请专家重写该轮，以保留原有规划风格并叠加 harness 演化与权重更新的收益。
AI推荐理由 该帖转述了 Salesforce 一篇关于 harness 与模型协同演化的论文，给出了具体实验设置（7 个企业 agent 任务、Qwen3-Coder 与 Gemma 4）和量化结果（性能下降 4 至 30 分），并指出失败原因在于弱模型模仿专家规划策略却缺乏执行能力，以及提出的修复思路（只重写失败轮次而非复制整条轨迹）。这些是输入中可核验的具体新信息，对 agent harness 工程与微调实践有参考价值。局限在于：内容为二手转述，未提供论文原文细节，实验规模、任务定义和统计显著性均未说明，且互动量仅代表传播信号，不能替代质量判断。
78
浏览 1,580
点赞 10
评论 6
转发 1
正文
Nice paper from Salesforce on co-evolving harnesses and models.
Harness engineering is a hot topic right now. So this is a great read.
(bookmark it)
Salesforce evolved a harness with a weak model across seven enterprise agent tasks, then trained that model on a stronger expert's full trajectories under the same harness.
Performance dropped on all seven tasks, by 4 to 30 points across Qwen3-Coder and Gemma 4.
The same fine-tuning helps under the unevolved harness. So the harness is what changes the outcome.
Their analysis points at model-harness fit.
Imitation transfers knowledge and increases scaffold usage, but the weaker model adopts the expert's planning strategy without the competence to execute it, and it no longer matches a harness that was evolved around its own native planning style.
The fix is to stop copying whole trajectories.
A meta-level agent finds the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. That keeps the model's planning style intact and combines the gains from harness evolution and weight updates.
Paper: [academy.dair.ai/papers/co-evol…](https://academy.dair.ai/papers/co-evolving-harnesses-and-models-on-policy-correction-helps-weaker-models-catch-2609.09134)
