# On the Advance of Making Language Models Better Reasoners

**Authors:** Yifei Li, Zeqi Lin, Shizhuo Zhang, Qiang Fu, Bei Chen, Jian-Guang Lou, Weizhu Chen  
**arXiv:** [2206.02336](https://arxiv.org/abs/2206.02336)  
**Date:** 2022-06-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.02336v1/

## Key points

- Large language models like GPT-3 and PaLM have shown impressive performance in few-shot learning but struggle with reasoning tasks like the arithmetic benchmark GSM8K.
- Recent advances have focused on guiding language models to generate a chain of reasoning steps before producing the final answer, which has improved problem-solving rates for GSM8K from 17.9% to 58.1%.
- The DiVeRSe approach proposed by Yifei Li and colleagues aims to further enhance reasoning capabilities by exploring different prompts to increase diversity in reasoning paths, introducing a verifier that can distinguish good answers from bad ones for better weighted voting, and verifying each single step's correctness individually instead of verifying all steps as a whole.
- Extensive experiments using the latest language model code-davinci-002 show that DiVeRSe outperforms PaLM with 540B parameters on six out of eight reasoning benchmarks (e.g., GSM8K improved from 74.4% to 83.2%).
- The proposed approach demonstrates significant progress towards making language models better reasoners and could have implications for various applications requiring advanced reasoning abilities.

## Summary

In recent years, large language models such as GPT-3 and PaLM have demonstrated impressive performance in few-shot learning. However, they still face challenges when it comes to reasoning tasks like the arithmetic benchmark GSM8K. To address this issue, recent advances have focused on guiding language models to generate a chain of reasoning steps before producing the final answer. This approach has successfully boosted the problem-solving rate for GSM8K from 17.9% to 58.1%. In this paper titled "On the Advance of Making Language Models Better Reasoners," Yifei Li and colleagues propose a new approach called DiVeRSe (Diverse Verifier on Reasoning Step) to further enhance reasoning capabilities. The DiVeRSe approach consists of three key components: exploring different prompts to increase diversity in reasoning paths; introducing a verifier that can distinguish good answers from bad ones for better weighted voting; and verifying each single step's correctness individually instead of verifying all steps as a whole. The authors conducted extensive experiments using the latest language model code-davinci-002 and found that DiVeRSe outperforms PaLM with 540B parameters on six out of eight reasoning benchmarks (e.g., GSM8K improved from 74.4% to 83.2%). The proposed approach demonstrates significant progress towards making language models better reasoners and could have implications for various applications requiring advanced reasoning abilities.