# Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers

**Authors:** Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, Yujiu Yang  
**arXiv:** [2309.08532](https://arxiv.org/abs/2309.08532)  
**Date:** 2023-09-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.08532v2/

## Key points

- Introduction of EvoPrompt: a framework for discrete prompt optimization
- Use of evolutionary algorithms (EAs) to automate prompt creation for large language models (LLMs)
- Combination of LLMs' language processing capabilities with EAs' optimization performance
- Iterative prompt generation process using evolutionary operators instead of gradients or parameters
- Evaluation on closed- and open-source LLMs across various datasets, showing up to 25% improvement on BIG-Bench Hard tasks
- Highlighting the effectiveness of connecting LLMs with EAs in optimizing prompts for diverse applications
- Significance of EvoPrompt in creating synergies between LLMs and EAs for further exploration and efficiency enhancement in prompt optimization

## Summary

The paper "Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers" introduces EvoPrompt - a novel framework for discrete prompt optimization. The authors propose using evolutionary algorithms (EAs) to automate the process of creating coherent and human-readable prompts for large language models (LLMs). By combining the language processing capabilities of LLMs with the efficient optimization performance of EAs, EvoPrompt aims to streamline the prompt generation process. Unlike traditional methods that rely on gradients or parameters, EvoPrompt starts with a population of prompts and iteratively generates new prompts using LLMs based on evolutionary operators. The prompts are then refined based on feedback from a development set. The authors evaluate EvoPrompt on both closed- and open-source LLMs such as GPT-3.5 and Alpaca across 31 datasets encompassing language understanding, generation tasks, and BIG-Bench Hard (BBH) tasks. Their results demonstrate that EvoPrompt outperforms human-engineered prompts and existing automatic prompt generation methods by up to 25% on BBH tasks. This highlights the effectiveness of connecting LLMs with EAs in optimizing prompts for diverse applications. Overall, the study showcases how EvoPrompt creates synergies between LLMs and EAs, paving the way for further exploration into combining large language models with conventional algorithms. The research presented in this paper was accepted at the International Conference on Learning Representations (ICLR) 2024 and offers valuable insights into enhancing the efficiency and effectiveness of prompt optimization for advanced language models.