JHU 与 CMU 论文：4B 小模型可自动改写 Agent harness 代码并泛化到新任务
JHU 和 CMU 让 4B 小模型自己改 agent 的 harness 代码，效果反超 35B 教师模型，做 agent 开发可以试试这思路。
Johns Hopkins 与 Carnegie Mellon University 的论文提出让小模型根据失败报告自动改写 agent 的 harness 代码。harness 是决定模型看到什么内容、调用哪些工具的代码，编辑模型读取 harness 和失败信息后生成代码修改，并以新 harness 的得分作为奖励。在 21 个未见过的推理任务类型上，4B 编辑模型的平均修改得分从 0.32 升到 0.62，超过其 35B 教师。另一个在 HotpotQA 上训练的编辑模型还能在 2 个其他 QA 基准上持续改进 harness，实验显示多轮重复修改优于单次修改。
New John Hopkins and Carnegie Mellon University Paper Shows that a small model can learn to improve an agent's harness code from run results, and the skill transfers to new tasks.
A small model trained to rewrite an agent's harness code from failure reports can adapt agents to new tasks, so let it tune your harness instead of doing it by hand.
The harness is the code that decides what the model sees and which tools it calls. The editor reads the harness and what failed, then writes a code change, rewarded by how well the new harness scores.
On 21 unseen reasoning task types, a 4B editor's average edit score rose from 0.32 to 0.62, above its 35B teacher. A separate editor trained on HotpotQA kept improving harnesses on 2 other QA benchmarks.
Run it for several rounds per task, since repeated edits beat single fixes.
- elvis10-03 23:38原文
