Reproduce “RM-OPD: On-Policy Distillation for Generative Reward Models” 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
Qinglin Su, Licheng Pan, Zhen Fang, Jiakang Wang, Qingnan Ren, Yunsheng Lu, Yan Gao, Zhixuan Chu
RM-OPD trains domain-specific reward-model teachers, then uses them to guide one student on its own critiques and verdicts, token by token. Domain contribution reweighting balances updates across domains, while optional criteria context distillation supplies teachers with domain-specific evaluation criteria. On 16 subsets, its Qwen2.5-7B-Instruct model reaches 75.41% macro accuracy versus 71.45% for matched MixRL-GRPO, under tested 7B conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.