Harness-S trains agents using sibling harnesses
Original title: Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training
It selectively rescues failures with sibling harnesses and transfers capabilities to the base policy.
Harness evolution shows locality, complementarity, and nonstationarity. Harness-S uses Directional Harness Evolution to create specialized siblings and a Sibling Harness Pool. Complementary Harness Rescue Training invokes the pool when base fails, replacing trajectories for RL. The policy outperforms with the router.
- Improves performance on AppWorld, SWE-Gym, DBBench
- Uses Adaptive Harness Router with Top-k variants
- Reusing router for Rank-1 inference adds gains
- Organizations
- Hugging Face
- AI models
- agent policies
From the original article
1 min read at huggingface.co
Harness evolution from a common base exhibits Locality, Complementarity, and Nonstationarity: variants repair different failures, and their utility changes as the policy learns.
By Kangning Zhang, Wenxiang Jiao, Bowen Fang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.