# Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves

**Authors:** Yihe Deng, Weitong Zhang, Zixiang Chen, Quanquan Gu  
**arXiv:** [2311.04205](https://arxiv.org/abs/2311.04205)  
**Date:** 2023-11-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.04205v1/

## Key points

- Authors address the issue of misunderstandings between humans and Large Language Models (LLMs)
- Proposed method called "Rephrase and Respond" (RaR) allows LLMs to rephrase and expand questions posed by humans
- RaR serves as a simple yet effective prompting method for improving LLM performance
- Two-step variant of RaR introduced, where a rephrasing LLM first rephrases the question before passing it to a different responding LLM
- Experiments demonstrate significant enhancement in performance across different tasks using RaR
- Comprehensive comparison between RaR and Chain-of-Thought (CoT) methods shows complementarity and potential for combined use
- Paper contributes to enhancing LLM performance efficiently and highlights fair evaluation practices
- Data and codes available on GitHub for further exploration.

## Summary

In the paper titled "Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves," authors Yihe Deng, Weitong Zhang, Zixiang Chen, and Quanquan Gu address the issue of misunderstandings that can arise between humans and Large Language Models (LLMs). These misunderstandings can lead to LLMs interpreting questions in unexpected ways and providing incorrect responses. To address this gap, the authors propose a method called "Rephrase and Respond" (RaR), which allows LLMs to rephrase and expand questions posed by humans and provide responses in a single prompt. This approach serves as a simple yet effective prompting method for improving LLM performance. Additionally, the authors introduce a two-step variant of RaR where a rephrasing LLM first rephrases the question before passing both the original and rephrased questions to a different responding LLM. This enables the effective utilization of rephrased questions generated by one LLM with another. The experiments conducted by the authors demonstrate that their methods significantly enhance the performance of various models across different tasks. Furthermore, they provide a comprehensive comparison between RaR and the popular Chain-of-Thought (CoT) methods both theoretically and empirically. The results show that RaR is complementary to CoT and can be combined with it to achieve even better performance. Overall, this paper contributes to enhancing LLM performance efficiently while also shedding light on fair evaluation practices for assessing LLM capabilities. The authors make their data and codes available on GitHub for further exploration.