Reproduce “Outcome-Grounded Advantage Reshaping for Fine-Grained Credit Assignment in Mathematical Reasoning” 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
Ziheng Li, Liu Kang, Feng Xiao, Luxi Xing, Qingyi Si, Zhuoran Li, Weikang Gong, Deqing Yang and 2 more
Outcome-grounded Advantage Reshaping (OAR) refines credit assignment in Group Relative Policy Optimization for mathematical reasoning, redistributing sequence-level advantages based on each token's influence on the final answer distribution. This approach improved performance on benchmarks like AIME Pass@1 (e.g., from 9.4 to 11.8 on Qwen2.5-7B) and enhanced training stability by preventing premature entropy collapse.
Key insights
- A token's direct influence on the model's final answer distribution provides a more causally aligned signal for credit assignment than intrinsic uncertainty measures like entropy.
- Controlled, bi-level reallocation of advantages with conservation of total advantage mass is crucial for maintaining stable and efficient training in critic-free RL settings.
- Gradient-based approximations can efficiently estimate token importance, offering a scalable solution for fine-grained credit assignment in large-scale online training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.