The paper introduces Meta-Prompting, a scaffolding technique designed to enhance the functionality of language models. This approach transforms a single Language Model into a multi-faceted conductor, capable of managing and integrating multiple independent queries. It uses high-level instructions to break down complex tasks into smaller, more manageable subtasks. These subtasks are then handled by expert instances of the same language model, each operating under specific instructions. The paper shows that Meta-Prompting significantly enhances the performance of GPT-4 across a wide array of tasks, surpassing standard, expert, and multipersona prompting.
Publication date: 23 Jan 2024
Project Page: https://github.com/suzgunmirac/meta-prompting
Paper: https://arxiv.org/pdf/2401.12954
