# Synthetic Prompting: Generating Chain-of-Thought Demonstrations for Large Language Models

**Authors:** Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, Weizhu Chen  
**arXiv:** [2302.00618](https://arxiv.org/abs/2302.00618)  
**Date:** 2023-02-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.00618v1/

## Key points

- The paper introduces a method to improve the quality of prompts used in large language models for various reasoning tasks.
- Current methods rely on handcrafted examples which are expensive and time-consuming to create.
- The proposed method leverages a few handcrafted examples to prompt the model to generate more examples by itself and selects effective demonstrations to elicit better reasoning.
- The approach involves alternating between a backward process and forward process to generate new examples.
- The authors evaluate their method on numerical, symbolic, and algorithmic reasoning tasks and demonstrate its superiority over existing prompting techniques.
- This paper presents an innovative approach that can significantly reduce costs associated with generating high-quality prompts while improving model performance in various reasoning tasks.

## Summary

The paper titled "Synthetic Prompting: Generating Chain-of-Thought Demonstrations for Large Language Models" by Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen introduces a novel method to improve the quality of prompts used in large language models for various reasoning tasks. The current methods rely on handcrafted examples to guide the models through step-by-step demonstrations. However, creating many such examples is expensive and time-consuming. The proposed method leverages a few handcrafted examples to prompt the model to generate more examples by itself and selects effective demonstrations to elicit better reasoning. The approach involves alternating between a backward process and forward process to generate new examples. In the backward process, a question matching a sampled reasoning chain is generated so that it is solvable and clear. In contrast, the forward process produces a more detailed reasoning chain for the question, improving example quality. The authors evaluate their method on numerical, symbolic, and algorithmic reasoning tasks and demonstrate its superiority over existing prompting techniques. Overall, this paper presents an innovative approach that can significantly reduce costs associated with generating high-quality prompts while improving model performance in various reasoning tasks.