# Large Language Models Cannot Self-Correct Reasoning Yet

**Authors:** Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou  
**arXiv:** [2310.01798](https://arxiv.org/abs/2310.01798)  
**Date:** 2023-10-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.01798v1/

## Key points

- Large Language Models (LLMs) have gained attention for text generation capabilities
- Concerns about accuracy and appropriateness of generated content
- Self-correction as a potential solution
- Study examines role and efficacy of self-correction in LLMs with focus on reasoning
- Intrinsic self-correction: LLMs attempt to correct initial responses without external feedback
- Research findings indicate LLMs struggle to self-correct without external feedback and may perform worse after attempting self-correction
- Multi-agent debate as another approach for LLMs to self-correct through critique and debate via multiple model calls
- Further research and practical applications suggested in the field
- Insights into limitations and potential of self-correction in LLMs for improving generated content

## Summary

Large Language Models (LLMs) have gained significant attention for their text generation capabilities. However, concerns remain about the accuracy and appropriateness of their generated content. To address these issues, the concept of self-correction has emerged as a potential solution. This paper critically examines the role and efficacy of self-correction in LLMs with a focus on reasoning. The study explores intrinsic self-correction where LLMs attempt to correct their initial responses without external feedback. The research findings indicate that LLMs struggle to self-correct without external feedback and may even perform worse after attempting self-correction. Additionally, it discusses multi-agent debate as another approach for LLMs to self-correct by allowing models to critique and debate through multiple model calls. The paper suggests further research and practical applications in this field and provides insights into the limitations and potential of self-correction in LLMs for improving their generated content.