Reproduce “LLM Critics Help Catch Bugs in Mathematics: Towards a Better Mathematical Verifier with Natural Language Feedback” 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

Bofei Gao, Zefan Cai, Runxin Xu, Peiyi Wang, Ce Zheng, Runji Lin, Keming Lu, Dayiheng Liu and 5 more

Researchers introduce MATH-Minos, a framework that enhances mathematical verifiers for Large Language Models by incorporating step-wise natural language feedback. This method improves verification accuracy by up to 1.6% on GSM8K and 0.8% on MATH datasets, and boosts PPO post-training performance by 1.8% on GSM8K and 1.5% on MATH.

Key insights

  • Providing natural language feedback, particularly with label-aware prompting, is crucial for improving verifier quality and efficiency, demonstrating that informational density of labels is more impactful than sheer quantity.
  • The two-stage training paradigm is complementary: Stage 1 builds a deep understanding of mathematical reasoning from textual explanations, while Stage 2 converts this understanding into efficient and reliable binary scoring.
  • Natural language feedback helps the verifier learn to identify specific error types (e.g., logical errors, unrelated content) more effectively than traditional binary labels, leading to a more nuanced understanding of mathematical correctness.

Summary by alphaXiv.

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