Harness-S

TL;DR

Use alternative harnesses to rescue failed training attempts, then transfer the resulting behavior into a policy evaluated under its original harness. The authors report gains without numerical effect sizes in the abstract. (source)

Authors & Org

Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu, Rong Shan, Shijian Wang, Hao Wang, Yuan Lu, Weiwen Liu and Weinan Zhang. Affiliations are not established by the captured metadata. First submitted 2026-11-03. (source)

Method

A fixed pool of specialized sibling harnesses targets different failure modes. Rescue training calls that pool only after all initial base-harness attempts fail, replacing the failed group with fresh trajectories from the same policy for reinforcement learning. A router selects variants using behavior profiles and model voting; those profiles are refreshed as the policy changes. (source)

Results

Across AppWorld, SWE-Gym and DBBench, the authors report improvements over policy-only training when evaluation returns to the base harness. Reusing the router at inference yields further gains at the same executor rollout count. Equal rollout counts do not establish equal total routing or training cost. (source)

Significance

For Agentic Reinforcement Learning and Eval Harness Configuration, this separates a useful training scaffold from a dependency required at deployment. The evidence is the authors' abstract-level report, not an independent replication. (source)

Open Questions

The captured abstract leaves numerical gains, total compute and robustness across new harness families unspecified. Full-paper methods and code were not reviewed. (source)

Cite

• arXiv:2611.04277
• Captured abstract and metadata