Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training
| Source: arXiv AI
Tags: Harness-S, Reinforcement Learning, AppWorld, SWE-Gym, DBBench, Agent Harnesses
Harness-S trains a single agent policy by rescuing failed rollouts with specialized sibling harnesses, and the resulting policy beats policy-only training on AppWorld, SWE-Gym and DBBench even when run under the original harness.
Details
Agent results depend heavily on the harness wrapped around the model, and different harness variants fix different failures. Harness-S treats this as a resource for training rather than searching for one best harness. It builds a fixed pool of sibling harnesses along distinct repair directions. When every attempt under the base harness fails, an Adaptive Harness Router picks Top-k siblings using behavior profiles and voting across several LLMs, and the fresh rollouts replace the failed group in reinforcement learning. The authors report that the trained policy substantially outperforms policy-only training on AppWorld, SWE-Gym and DBBench when returned to the base harness, and that reusing the router for Rank-1 inference adds gains at the same rollout count. The abstract gives no exact numbers.