Overview

EGO-Prompt can be used to find better prompts for domain-specific tasks.

Read Paper Here

Setup

conda create -n ego_prompt python=3.11
source activate ego_prompt
pip install python-dotenv tqdm textgrad numpy scikit-learn matplotlib pandas

🔄 Update Checklist

• Update Evaluation Code
• Update instructions on how to use your own datasets
• Merge GPT-5 calling code

Results summary

This is the summary of the results shown in our demos (3 runs).

Method	Swissmetro	TrafficSafe	Pandemic
Organized Prompt (Mean ± Std)	0.4075 ± 0.0168	0.2269 ± 0.0364	0.3477 ± 0.0183
Ego-Prompt (Mean ± Std)	0.4915 ± 0.0330	0.3156 ± 0.0123	0.3656 ± 0.0447
Organized Prompt (Best)	0.4238	0.2682	0.3662
Ego-Prompt (Best)	0.5262	0.3291	0.3948

Citation

@inproceedings{zhao2025how,
  author = {Yang Zhao, Pu Wang, Hao Frank Yang},
  title = {How to Auto-optimize Prompts for Domain Tasks? Adaptive Prompting and Reasoning through Evolutionary Domain Knowledge Adaptation},
  booktitle={The Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
  year = {2025},
}

Contact yzhao229@jh.edu or open an issue if you have any questions.

Acknowledgments

The code is based on the following repositories:

• TextGrad