Reproduce “Harness-Aware Distillation for Small Language Model Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Moonseok Choi, Taehong Moon, Giung Nam, Juho Lee
HAD distills an agent to use a fixed harness by comparing the same teacher’s actions with and without harness information at student-visited states. It adds a preference for harness-conditioned actions and drops preference pairs contradicted by harness records. On unseen ALFWorld tasks, a Qwen3 1.7B student achieves 63.43% success, versus 47.01% for OPD, in tested settings with the harness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.