Understanding and Enhancing Chain-of-Thought Prompting with Coherent Reasoning and Sensitivity Analysis

Introduction: This paper explores the effectiveness of Chain-of-Thought (CoT) prompting in large language models (LLMs), focusing on the impact of considering the entire reasoning chain during training and inference. It contrasts this "Coherent CoT" approach with the more traditional "Stepwise ICL" method, where each reasoning step is treated in isolation. The authors also investigate the sensitivity of CoT models to errors in different reasoning steps and propose a novel prompting technique to improve performance.

Coherent vs. Stepwise CoT

The paper argues that training LLMs with Coherent CoT, where the model considers the entire reasoning chain, leads to better performance than Stepwise ICL, which treats each step independently. This is attributed to the model's ability to self-correct by considering previous predictions in subsequent steps.

Sensitivity Analysis

The research reveals that Coherent CoT models are more sensitive to errors in intermediate reasoning steps within demonstration examples than to errors in the final conclusions. This suggests that the accuracy of intermediate steps is crucial for overall performance.

Enhanced Prompting Technique

Based on the sensitivity analysis, the authors propose a new prompting method that incorporates both correct and incorrect reasoning paths in demonstration examples. This approach aims to improve the accuracy of intermediate steps and, consequently, boost overall CoT performance.

Conclusion

By considering the entire reasoning chain during training (Coherent CoT), LLMs can achieve better performance due to their ability to self-correct. Furthermore, the sensitivity of these models to errors in intermediate reasoning steps highlights the importance of accurate demonstrations. The proposed prompting technique, incorporating both correct and incorrect reasoning paths, offers a promising approach to enhance CoT effectiveness.

Source(s):

• [Wei, J., et al. (2024). Understanding and Enhancing Chain-of-Thought Prompting with Coherent Reasoning and Sensitivity Analysis. arXiv preprint arXiv:2410.16540v1.]