Reproduce “Harness Scaling for Generalizable Agentic Reinforcement Learning” 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
Fei Bai, Huatong Song, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai, Wayne Xin Zhao, Ji-Rong Wen
The authors introduce Harness Scaling, training agents across 18 selected configurations from a 243-harness space while keeping tasks and environments fixed and assigning nearby harnesses to each task. On unseen Claude Code, Qwen3-4B-Instruct reached 37.22 average Pass@1 across three benchmarks, versus 33.38 for black-box RL trained through OpenClaw; this supports transfer only in the tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.