# PACE: Improving Prompt with Actor-Critic Editing for Large Language Model

**Authors:** Yihong Dong, Kangcheng Luo, Xue Jiang, Zhi Jin, Ge Li  
**arXiv:** [2308.10088](https://arxiv.org/abs/2308.10088)  
**Date:** 2023-08-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.10088v1/

## Key points

- The paper introduces a method called Prompt with Actor-Critic Editing (PACE) to enhance the performance of large language models (LLMs) by automatically editing prompts.
- LLMs are affected by the quality of human-written prompts, and PACE addresses this issue by leveraging the actor-critic algorithm from reinforcement learning.
- PACE treats LLMs as both actors and critics, refining prompts based on feedback from both actors performing the prompt and critics criticizing the response.
- This process allows LLMs to align the prompt more effectively with a specific task by incorporating real responses and thinking from LLMs.
- Extensive experiments on 24 instruction induction tasks and 21 big-bench tasks demonstrate that PACE significantly improves the relative performance of medium/low-quality human-written prompts by up to 98%.
- PACE achieves comparable performance to high-quality prompts and exhibits notable efficacy in prompt generation.
- Overall, PACE offers an automated approach for enhancing LLM performance by refining prompts, reducing the need for manual effort in improving prompt quality.

## Summary

The paper titled "PACE: Improving Prompt with Actor-Critic Editing for Large Language Model" introduces a method called Prompt with Actor-Critic Editing (PACE) to enhance the performance of large language models (LLMs) by automatically editing prompts. LLMs have shown great potential in various tasks when conditioned on prompts, but the quality of human-written prompts can significantly affect their performance. To address this issue, PACE leverages the actor-critic algorithm from reinforcement learning and treats LLMs as both actors and critics. It considers prompts as a type of policy and refines them based on feedback from both actors performing the prompt and critics criticizing the response. This process allows LLMs to align the prompt more effectively with a specific task by incorporating real responses and thinking from LLMs. The researchers conducted extensive experiments on 24 instruction induction tasks and 21 big-bench tasks. The results demonstrate that PACE significantly improves the relative performance of medium/low-quality human-written prompts by up to 98%, achieving comparable performance to high-quality prompts. Additionally, PACE also exhibits notable efficacy in prompt generation. Overall, PACE offers an automated approach for enhancing LLM performance by refining prompts, reducing the need for manual effort in improving prompt quality.