# Automatic Chain of Thought Prompting in Large Language Models

**Authors:** Zhuosheng Zhang, Aston Zhang, Mu Li, Alex Smola  
**arXiv:** [2210.03493](https://arxiv.org/abs/2210.03493)  
**Date:** 2022-10-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.03493v1/

## Key points

- Authors Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola explore chain-of-thought (CoT) prompting in large language models (LLMs)
- Two major paradigms within CoT prompting: simple prompt for step-by-step thinking vs. manual demonstrations with reasoning chains
- Manual demonstrations typically outperform simple prompts due to task-specific nature
- Auto-CoT method proposed to automate generation of reasoning chains in LLMs using "Let's think step by step" prompt
- Emphasis on diversity in constructing demonstrations to address errors in automatically generated chains
- Experiments show Auto-CoT matching or surpassing traditional CoT methods on ten public benchmark reasoning tasks using GPT-3
- Code repository available at https://github.com/amazon-research/auto-cot for further exploration and implementation

## Summary

In their paper titled "Automatic Chain of Thought Prompting in Large Language Models," authors Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola explore the concept of chain-of-thought (CoT) prompting in large language models (LLMs). They highlight two major paradigms within CoT prompting: one that uses a simple prompt like "Let's think step by step" to encourage step-by-step thinking before answering a question, and another that involves manual demonstrations composed of questions and reasoning chains leading to answers. The latter paradigm typically outperforms the former due to the task-specific nature of manually crafted demonstrations. The authors propose an innovative approach to streamline the process of generating reasoning chains for demonstrations in LLMs through their method called Auto-CoT. By leveraging the "Let's think step by step" prompt, they aim to automate the generation of these chains one by one, eliminating the need for manual demonstration design. However, they note that automatically generated chains may contain errors. To address this challenge, they emphasize the importance of diversity in constructing demonstrations. Through experiments on ten public benchmark reasoning tasks using GPT-3, Auto-CoT consistently matches or surpasses the performance of traditional CoT prompting methods that rely on manual demonstration designs. The authors provide access to their code repository at https://github.com/amazon-research/auto-cot for further exploration and implementation. Overall, this research sheds light on an efficient and effective approach to enhancing complex reasoning capabilities in LLMs through automated chain-of-thought prompting techniques.