PapersOfficial

清华与 Zhipu 推出 HarnessPro,让智能体重写自身测试框架

Tsinghua and Zhipu's HarnessPro Lets Agents Rewrite Their Own Harness, Cutting Task Costs 26%

Why we picked it

清华和 Zhipu 做了个 HarnessPro,让智能体自己改写 harness,成本砍了 26%,做 agent 的可以看看怎么省 token。

清华、Zhipu 与合作者发布 HarnessPro,让智能体复盘自己的运行轨迹并重写 harness,全程不需要参考答案。在 Terminal-Bench 2 上,pass rate 提升 2 个百分点,单任务成本下降 26.1%。方法核心是递归式的 harness 自我改进,而非更换底层模型。

Original · Pandaily

Tsinghua and Zhipu's HarnessPro Lets Agents Rewrite Their Own Harness, Cutting Task Costs 26%

HarnessPro, from Tsinghua, Zhipu and partners, has agents study their own runs and rewrite their harness without reference answers, raising pass rates 2 points and cutting cost per task 26.1%.