Core Concepts
Meta Prompting is a novel technique that focuses on the structure and syntax of information, rather than just the content, to enhance the reasoning and problem-solving capabilities of large language models.
Abstract
The paper introduces Meta Prompting (MP), a comprehensive study of an innovative technique that reshapes the utilization of language models (LMs) and AI systems in problem-solving and data interaction. Grounded in type theory and category theory, Meta Prompting emphasizes the structure and syntax of information over traditional content-centric methods.
The key highlights of the paper are:
Formal definitions of Meta Prompting and its distinction from few-shot prompting. Meta Prompting is defined as a functor that maps tasks to structured prompts, capturing the reasoning structure of problems.
Exploration of Meta Prompting's effectiveness in various AI applications, with a focus on complex reasoning tasks. Meta Prompting can effectively deconstruct intricate problems into simpler sub-problems, enhancing token efficiency and enabling more equitable problem-solving comparisons.
Introduction of Meta Prompting for prompting tasks, allowing LLMs to self-generate new prompts in a recursive, metaprogramming-like manner. This Recursive Meta Prompting (RMP) showcases the system's ability to dynamically generate and refine prompts, making it highly adaptable and responsive to task complexities.
Empirical experiments demonstrating the superior performance of meta-prompted LLMs on MATH and GSM8K benchmarks, as well as their ability to solve the Game of 24 tasks with a 100% success rate, highlighting the transformative impact of Meta Prompting on AI problem-solving.
The paper emphasizes the importance of structural and syntactical elements in enhancing the reasoning and problem-solving capabilities of large language models, going beyond traditional content-driven approaches.
Stats
"Meta Prompting significantly reduces the number of tokens required compared to few-shot prompting."
"The zero-shot meta-prompted Qwen-72B model achieved a PASS@1 accuracy of 46.3% on the MATH dataset, outperforming open-source models and proprietary models like GPT-4."
"The zero-shot meta-prompted Qwen-72B model achieved an accuracy of 83.5% on the GSM8K benchmark, surpassing the best results from both few-shot prompting approaches and fine-tuned counterparts."
"The MP-CR Agent achieved a 100% success rate in solving all 1362 samples of the Game of 24 tasks."
Quotes
"Meta Prompting extends beyond existing methods by abstracting and generalizing key principles for enhanced cognitive processing."
"The functorial nature of Meta Prompting allows for this advanced capability, where LLMs can not only solve problems but also generate the structures to solve them."
"Meta Prompting stands out for its token efficiency and its ability to provide a fairer, more unbiased approach to problem-solving compared to few-shot examples."