Google 研究：Harness-Zero 在训练后移除专用 harness，任务成功率反升
Super interesting paper from Google and colleagues. It studies where it's possible to distill an ag...
Google 这篇论文把 agent harness 蒸馏进模型，移除专用 harness 后成功率反而从 23.3% 涨到 44.3%，比带 harness 还高，做 agent 的都该看看。
Google 与合作者的论文提出 Harness-Zero，研究能否把 agent harness 蒸馏进模型本身。移除专用 harness 后，宏观任务成功率从 23.3% 升至 44.3%，高于基础模型带 harness 时的 41.7%。方法是在训练时使用优化过的 harness，由 harnessing agent 在部署动作空间中纠正学生模型的响应，纠正后的运行轨迹作为训练演示。在知识工作、工具使用和科学领域的 28 种 harness 诱导行为上，平均恢复 82.3%。对于使用相同进化 harness 的前沿模型，agent-as-harness 形式也优于 code-as-harness 形式。
Super interesting paper from Google and colleagues. It studies where it's possible to distill an ag...
Super interesting paper from Google and colleagues. It studies where it's possible to distill an agent harness. With the specialized harness removed, macro task success goes from 23.3% to 44.3%. That is higher than the 41.7% the base model reaches with the harness attached. Harness-Zero uses the optimized harness only during training. The optimized harness and the deployment harness have different action spaces, so a harnessing agent guided by the optimized harness corrects the student's responses in the deployment action space before they run. Those corrected runs become the training demonstrations. Across 28 harness-induced behaviors in knowledge work, tool use and science, 82.3% are recovered on average. For frontier models using the same evolved harness, the agent-as-harness form also beats the code-as-harness form. It remains to be seen how robust the approach is, but it's very interesting to see potential in harness distillation. Paper: arxiv.org/abs/2609.24974 Chat with Paper: academy.dair.ai/papers/harness… 💬 6 🔄 5 ❤️ 16 👀 1917 📊 10 ⚡
