Reproduce “A Survey on Feedback-based Multi-step Reasoning for Large Language Models on Mathematics” 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
Ting-Ruen Wei, Haowei Liu, Xuyang Wu, Yi Fang
This survey systematically organizes feedback-based strategies for improving Large Language Models' (LLMs) mathematical reasoning, categorizing methods by feedback granularity (step-level vs. outcome-level) and implementation (training-based vs. training-free). It provides a comprehensive taxonomy of approaches, identifies key challenges like reward hacking and efficiency, and outlines future research directions in this domain.
Key insights
- Feedback mechanisms for LLMs in mathematical reasoning are categorized into step-level (process rewards) and outcome-level (outcome rewards), with both training-based and training-free approaches.
- Mathematics provides a robust domain for designing and validating feedback due to its verifiable logical structure, unlike more subjective domains.
- There is a clear evolution from basic prompting (Chain-of-Thought) to sophisticated feedback integration, including trained reward models and external tool utilization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.