Reproduce “Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training” 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

Chenlu Ye, Zhou Yu, Ziji Zhang, Hao Chen, Narayanan Sadagopan, Jing Huang, Tong Zhang, Anurag Beniwal

PROF uses process reward models (PRMs) not as direct optimization targets but to filter responses after checking answer correctness. It ranks correct and incorrect rollouts and selects a balanced training set. For Qwen2.5-Math-7B-base, PROF-BOTH’s accuracy across five math benchmarks is 51.7%, versus 49.9% for GRPO and 47.3% for direct PRM/ORM blending. On AIME2024, it scores 17.5%, below GRPO’s 20.6%.

Summary by alphaXiv.

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