# On Meta-Prompting

**Authors:** Adrian de Wynter, Xun Wang, Qilong Gu, Si-Qing Chen  
**arXiv:** [2312.06562](https://arxiv.org/abs/2312.06562)  
**Date:** 2023-12-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.06562v2/

## Key points

- Study evaluates effectiveness of meta-prompting in improving text generation tasks, focusing on Ideation and Creativity
- Utilization of GPT-4 through Azure Open AI API for data generation and analysis
- Theoretical framework based on category theory allows effective generalization and description of meta-prompting approaches
- Strong preference (p < 0.01) for meta-generated prompts and outputs over hardcoded baseline prompts shown in results, indicating effectiveness of meta-prompting
- Approach involves replacing strings with task descriptions and context to guide model in generating relevant prompts, enhancing output quality

## Summary

, , , , 

In this study, we evaluate the effectiveness of meta-prompting in improving text generation tasks, specifically focusing on Ideation and Creativity. We utilize GPT-4 through the Azure Open AI API for data generation and analysis. Our theoretical framework based on category theory allows us to effectively generalize and describe meta-prompting approaches. Through 5 experiments, we compare the outputs generated using meta-prompts against baseline prompts across 300 samples from WikiText-103 and TOEFL11 corpus for Ideation, and TOEFL11 and DailyMail/CNN corpus for Creativity. The results show a strong preference (p < 0.01) for meta-generated prompts and outputs over hardcoded baseline prompts, indicating the effectiveness of meta-prompting in producing desirable outcomes. Our approach involves replacing strings with task descriptions and context to guide the model in generating relevant prompts. This method proves successful in enhancing text generation tasks by providing clear instructions for improved output quality. In the Ideation task, we explore ways to enhance existing text by suggesting improvements such as making it more concise or adding details to support claims. For Creativity, we focus on generating natural transitions between passages or completing passages seamlessly. Overall, our study demonstrates that  is a more effective strategy than basic prompting in generating creative and ideative text outputs. The full details of our experiments and code are available for reproduction on GitHub at https://github.com/adewynter/metaprompting.