# Large Language Models As Evolution Strategies

**Authors:** Robert Tjarko Lange, Yingtao Tian, Yujin Tang  
**arXiv:** [2402.18381](https://arxiv.org/abs/2402.18381)  
**Date:** 2024-02-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.18381v1/

## Key points

- Authors: Robert Tjarko Lange, Yingtao Tian, and Yujin Tang
- Study focus: Large Language Models (LLMs) in implementing evolutionary optimization algorithms
- Novel prompting strategy for zero-shot application of LLMs to black-box optimization
- Resulting approach 'EvoLLM' outperforms traditional algorithms like random search and Gaussian Hill Climbing
- Potential of LLMs as 'plug-ins' for recombination operations
- Comparative analyses on model size, prompt strategy, and context construction
- Enhancement of EvoLLM's performance through instruction fine-tuning based on teacher algorithm information

## Summary

In their work titled "Large Language Models As Evolution Strategies," authors Robert Tjarko Lange, Yingtao Tian, and Yujin Tang explore the capabilities of large language models (LLMs) in implementing evolutionary optimization algorithms. This study delves into the zero-shot application of LLMs to black-box optimization by introducing a novel prompting strategy that involves sorting discretized population members from least to most and querying the LLM for improvements. The resulting approach, dubbed 'EvoLLM,' outperforms traditional algorithms like random search and Gaussian Hill Climbing on synthetic BBOB functions and small neuroevolution tasks. This highlights the potential of LLMs as 'plug-ins' for in-context recombination operations. The paper also includes comparative analyses on factors such as model size, prompt strategy, and context construction. Additionally, the authors demonstrate that EvoLLM's performance can be enhanced by providing teacher algorithm information through instruction fine-tuning based on previously collected teacher optimization trajectories. Overall, this study showcases the adaptability and effectiveness of LLMs in improving optimization strategies.