# Contrastive Chain-of-Thought Prompting

**Authors:** Yew Ken Chia, Guizhen Chen, Luu Anh Tuan, Soujanya Poria, Lidong Bing  
**arXiv:** [2311.09277](https://arxiv.org/abs/2311.09277)  
**Date:** 2023-11-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.09277v1/

## Key points

- The paper explores the concept of chain of thought in enhancing language model reasoning.
- Previous studies have shown minimal impact of invalid demonstrations on reasoning performance.
- Conventional chain of thought lacks guidance to avoid mistakes, potentially leading to more errors.
- The authors propose a new approach called contrastive chain of thought.
- Contrastive chain of thought provides both valid and invalid reasoning demonstrations to guide language models while reducing mistakes.
- An automatic method for constructing contrastive demonstrations is introduced to improve generalization.
- Experimental results demonstrate that contrastive chain of thought enhances reasoning capabilities compared to traditional approaches.
- Code implementation is provided for further exploration and research.

## Summary

, , , , 
The paper titled "Contrastive Chain-of-Thought Prompting" explores the concept of chain of thought in enhancing language model reasoning. Previous studies have shown that even invalid demonstrations have minimal impact on reasoning performance, which is surprising considering the importance of logically sound reasoning. Additionally, the conventional chain of thought does not provide guidance to language models on avoiding mistakes, potentially leading to more errors. To address these limitations, the authors propose a new approach called contrastive chain of thought. Inspired by how humans learn from both positive and negative examples, this approach provides both valid and invalid reasoning demonstrations to guide language models in step-by-step reasoning while reducing mistakes. To improve generalization, an automatic method for constructing contrastive demonstrations is introduced. Experimental results on reasoning benchmarks demonstrate that contrastive chain of thought can serve as a general enhancement of chain-of-thought prompting. The authors also provide their code implementation for further exploration and research. In summary, this paper introduces a novel approach to enhance language model reasoning through contrastive chain-of-thought prompting. By incorporating both valid and invalid demonstrations, this method improves reasoning capabilities and reduces errors compared to traditional approaches.