Reproduce “Critic-CoT: Boosting the reasoning abilities of large language model via Chain-of-thoughts Critic” 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
Xin Zheng, Jie Lou, Boxi Cao, Xueru Wen, Yuqiu Ji, Hongyu Lin, Yaojie Lu, Xianpei Han and 2 more
The Critic-CoT framework from the Chinese Information Processing Laboratory and Xiaohongshu Inc. trains Large Language Models (LLMs) for deliberate, step-by-step self-critique and refinement inspired by System-2 thinking. This approach boosts LLM reasoning abilities, with the 'Critic As Filter' strategy achieving 95.4% accuracy on GSM8K and 68.4% on MATH500, and generalizing effectively to out-of-domain tasks like code generation and multi-hop reasoning.
Key insights
- Explicitly training an LLM for critique and refinement capabilities directly improves its fundamental task-solving performance, indicating a mutual reinforcement between these abilities.
- A System-2-like step-wise Chain-of-Thought (CoT) critique is crucial for precise error identification within multi-step reasoning, outperforming simpler, instance-level feedback.
- High-quality training data for critique and refinement can be automatically constructed via weak supervision, significantly reducing the reliance on costly human annotation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.