Harness-Zero: Harness Distillation via Agent-as-Harness

ID: 3c0948c4
原文链接: https://arxiv.org/abs/2609.24974
PDF: https://arxiv.org/pdf/2609.24974v1
作者: Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song
日期: 2026-09-21
更新: 2026-09-21
分类: agents
来源类型: paper
标签: arxiv, agent-harness, distillation, fine-tuning, paper
质量评分: 4/5
抓取时间: 2026-09-23T04:27:36+00:00

中文导读

Agent harness（围绕冻结底座模型的提示词控制流工具记忆与上下文管理）能大幅提升性能，但收益绑定在部署时的特定 harness 上，而最优 harness 随领域实例和模型变化论文研究 harness 蒸馏：把领域/实例优化后的 harness 当作训练期指导，将其诱导的行为迁移进模型权重，让单一固定 harness 也能保留收益Harness-Zero 用 agent-as-harness 实现：harnessing agent 在目标 harness 的动作空间里先修正学生模型的输出，把 harness 指导转化为训练示范，微调后行为被内化，部署时可以移除专用 harness横跨知识工作工具使用科学三个领域，基座模型 macro 平均任务成功率从 23.3% 提到 44.3%，甚至超过挂着该 harness 时的 41.7%；28 个 harness 诱导行为模式平均恢复 82.3%

为什么值得关注

Harness-Zero：把专用 agent harness 蒸馏进权重，部署时不带 harness 反超带着的，任务成功率 23.3% 升至 44.3%

关键信息

论文标题：Harness-Zero: Harness Distillation via Agent-as-Harness
作者：Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song
arXiv：https://arxiv.org/abs/2609.24974
发布时间：2026-09-21
arXiv 分类：cs.AI, cs.CL, cs.NE
评论/页数：N/A
关联标签：arxiv, agent-harness, distillation, fine-tuning, paper

English Abstract

Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instances, and models, a general-purpose agent must either settle for a suboptimal shared harness or route among an ever-growing set of specialized ones. We therefore study agent harness distillation: using a domain- or instance-optimized harness as training-time guidance and transferring the behaviors it induces into model weights, so that its gains survive under a single fixed target harness. The challenge is that the two harnesses differ in action space and available information, so guidance from the optimized harness cannot serve directly as supervision for the target one. We introduce Harness-Zero, which enables harness distillation through agent-as-harness. Guided by the optimized harness, a harnessing agent corrects student responses before execution in the target harness's action space, turning harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behavior into the model, so the specialized harness can be removed at deployment. Our experiments spanning knowledge work, tool use, and science domains show that: (1) For frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness. (2) With the specialized harness removed at deployment, Harness-Zero improves the base model's macro-average task success from 23.3% to 44.3%, even exceeding the 41.7% it reaches with that harness still attached. (3) Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains.

English Summary

Agent harnesses (prompts, control flow, tooling, memory, and context management around a frozen backbone) improve performance but tie gains to the deployed harness, and the best harness varies by domain, instance, and model. The authors study harness distillation: use a domain- or instance-optimized harness as training-time guidance and transfer its induced behaviors into model weights so gains survive under one fixed target harness. Harness-Zero enables this via agent-as-harness: a harnessing agent corrects student responses before execution in the target harness's action space, turning harness guidance into training demonstrations; fine-tuning on those trajectories internalizes the behavior so the specialized harness can be removed at deployment....

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。