Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu, Rong Shan, Shijian Wang, Hao Wang (Shanghai Jiao Tong University) · Sibling Harness Training

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Harness-S trains an agent policy using a fixed pool of sibling harnesses that rescue failed rollouts, with an adaptive router choosing among them. The trained policy is reported to substantially outperform policy-only training when run under the original base harness, across AppWorld, SWE-Gym and DBBench. The authors also report behavioural analyses and held-out cases supporting transfer of harness-elicited behaviours into the policy, and additional gains from reusing the router at inference. The page gives no specific figures.