# Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models

**Authors:** Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, Ee-Peng Lim  
**arXiv:** [2305.04091](https://arxiv.org/abs/2305.04091)  
**Date:** 2023-05-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.04091v1/

## Key points

- Large language models (LLMs) are effective in natural language processing tasks
- Multi-step reasoning tasks still pose a challenge for LLMs
- Few-shot chain-of-thought (CoT) prompting uses manually crafted step-by-step reasoning demonstrations to help LLMs generate explicit reasoning steps and improve accuracy
- Zero-shot-CoT concatenates the problem statement with "Let's think step by step" as an input prompt to LLMs, but suffers from errors such as calculation errors, missing-step errors, and semantic misunderstanding errors
- Plan and Solve (PS) Prompting is proposed to address missing-step errors in Zero-shot CoT by dividing the task into smaller subtasks and carrying them out according to a plan
- PS+ prompting extends PS prompting with more detailed instructions to further improve the quality of generated reasoning steps and address calculation errors
- The proposed zero shot prompting strategy is evaluated on ten datasets across three reasoning problems using GPT 3
- Plan and Solve Prompting consistently outperforms Zero shot CoT across all datasets by a large margin and is comparable to or exceeds Zero shot Program of Thought Prompting
- It has comparable performance with 8 shot CoT prompting on math reasoning problems
- Plan and Solve Prompting provides an effective solution for generating explicit reasoning steps in LLMs without requiring manual effort.

## Summary

Large language models (LLMs) have demonstrated impressive performance in various natural language processing tasks. However, multi-step reasoning tasks still pose a challenge for LLMs. To address this issue, few-shot chain-of-thought (CoT) prompting has been proposed, which includes manually crafted step-by-step reasoning demonstrations to help LLMs generate explicit reasoning steps and improve their accuracy. To eliminate the need for manual effort, Zero-shot-CoT was introduced, which concatenates the problem statement with "Let's think step by step" as an input prompt to LLMs. Despite its success, Zero-shot-CoT suffers from calculation errors, missing-step errors, and semantic misunderstanding errors. To address the missing-step errors in Zero-shot CoT, this paper proposes Plan and Solve (PS) Prompting. The PS prompting strategy consists of two components: devising a plan to divide the entire task into smaller subtasks and carrying out the subtasks according to the plan. To further improve the quality of generated reasoning steps and address calculation errors, PS prompting is extended with more detailed instructions to derive PS+ prompting. The proposed zero shot prompting strategy is evaluated on ten datasets across three reasoning problems using GPT 3. The experimental results show that our proposed zero shot prompting consistently outperforms Zero shot CoT across all datasets by a large margin and is comparable to or exceeds Zero shot Program of Thought Prompting. Moreover, it has comparable performance with 8 shot CoT prompting on math reasoning problems. In conclusion, Plan and Solve Prompting provides an effective solution for generating explicit reasoning steps in LLMs without requiring manual effort. It addresses several issues faced by existing strategies such as missing step errors and calculation errors while achieving superior performance compared to other state of the art methods on various datasets and reasoning problems.