meta-prompting/meta-prompting

What it solves

Meta Prompting addresses the limitations of few-shot prompting, which relies on providing specific, content-rich examples to guide a model. Instead of teaching a model what to think via examples, Meta Prompting provides a structural template for how to think, allowing LLMs to handle entire categories of tasks more efficiently and robustly without needing task-specific examples.

How it works

Meta Prompting uses an example-agnostic scaffold—a high-level blueprint—that defines the formal procedure, syntax, and compositionality of problem-solving.

• Theoretical Foundation: The framework uses category theory to model prompting as a "functor," ensuring that complex tasks can be systematically decomposed into modular prompt structures.
• Recursive Meta Prompting (RMP): This is an automated self-improvement loop where the LLM generates and refines its own prompts, modeled mathematically as a "monad" to ensure the refinement process is stable and consistent.
• Multi-Modal Extension: Using type theory, the framework creates typed "slots" for different data streams (like images or audio), allowing the model to synthesize information across modalities into a structured output.

Who it’s for

This project is for AI researchers and prompt engineers looking to improve the reasoning capabilities of base LLMs without the need for extensive fine-tuning or large sets of few-shot examples.

Highlights

• Example-Agnostic: Works using a single structural template rather than multiple concrete examples.
• High Performance: Enabled a base Qwen-72B model to outperform fine-tuned models and early GPT-4 versions on MATH and GSM8K benchmarks.
• Token Efficiency: Significantly reduces token usage and cost compared to iterative methods like Tree-of-Thought (ToT).
• Formal Rigor: Provides a mathematical foundation for prompt engineering using category theory and type theory.