Reproduce “Reward-Adaptive Co-Evolution of Policy and Rubric” 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

Yizhu Chen, Jianguang Jiang, Jiameng Huang, Chuyu Fang, Bojun Zhang, Zhenyu Jiao, Xing Yuan, Tong Lin

RACE adapts reinforcement-learning training by routing response groups according to reward spread and diversity, then switching between policy and rubric training as reward conditions change. It masks low-diversity collapsed groups and caches varied groups with indistinguishable rewards for rubric training. On eight tasks it averages 73.0, versus EvoLM’s 71.3, and scores higher on each, in the tested shared-backbone, frozen-judge setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.