Homepage: https://meta-prompting.github.io.
Official implementation of paper "Meta Prompting for AGI Systems" (https://arxiv.org/abs/2311.11482).
Meta Prompting is an advanced prompting technique that focuses on the structural and syntactical aspects of problems, prioritizing the general format and pattern over specific content details. It aims to construct a more abstract, structured approach to interacting with large language models (LLMs), emphasizing the structure and syntax of information. This technique is particularly effective in contexts where the underlying pattern or framework of a problem is crucial for understanding or solving it.
A Meta Prompt is an example-agnostic structured prompt designed to capture the reasoning structure of a specific category of tasks. It provides a scaffold that outlines the general approach to a problem, enabling LLMs to fill in specific details as needed. This approach allows for more efficient and targeted use of LLM capabilities by focusing on the "how" of problem-solving rather than the "what".
-
Syntax-Oriented: Meta Prompting prioritizes the form and structure over the content, using syntax as a guiding template for the expected response or solution.
-
Abstract-Example-Based: It employs abstracted examples as frameworks, illustrating the structure of problems and solutions without focusing on specific content.
-
Type Theory Inspiration: Drawing from type theory, Meta Prompting emphasizes the categorization of components in a prompt, such as problem statements, solution steps, or conclusions. It focuses on their logical arrangement and interrelationships, ensuring a coherent and structured approach to problem-solving.
-
Adaptability: Meta Prompting is versatile, applicable across various domains, and capable of providing structured responses to a wide range of problems.
-
Guidance for Detailed Exploration: It provides a clear roadmap for problem-solving, focusing on structural patterns to aid in navigating complex topics.
We evaluated our models on the MATH and GSM8K datasets:
- MATH: A competition-level math word problems benchmark with 5000 test problems.
- GSM8K: The most widely used math word problem dataset with 1319 test grade school math problems.
Models were evaluated using the vLLM framework for Qwen-14B and Qwen-72B base language models with Meta Prompt. The evaluation was performed with a rule-based evaluator and SymPy to align model responses with ground-truth solutions.
Our evaluation demonstrates superior performance of the zero-shot meta-prompted Qwen-72B base language model on both MATH and GSM8K datasets, showcasing Meta Prompting's efficacy in mathematical problem-solving.
| Model | FT-Dataset | Tool Usage | Eval Method | MATH (%) |
|---|---|---|---|---|
| Proprietary Models | ||||
| Claude-2 | - | No | CoT | 32.5 |
| Minerva-540B | Arxiv+Web | No | CoT | 33.6 |
| PaLM-2 | - | No | CoT | 34.3 |
| GPT-4 (2023-0314) | - | No | CoT | 42.5 |
| Open-source Models | ||||
| Qwen-14B (base) | - | No | CoT | 24.8 |
| Qwen-14B (base) | - | No | MP | 28.9 |
| Llama-2-70B (base) | - | No | CoT | 13.5 |
| Qwen-72B (base) | - | No | CoT | 35.2 |
| Qwen-72B-MetaMathQA | MetaMathQA | No | CoT | 41.7 |
| Qwen-72B (base) | - | No | MP | 46.3 |
| Model | FT-Dataset | Tool Usage | Eval Method | GSM8K (%) |
|---|---|---|---|---|
| Llama-2-70B (base) | - | No | CoT | 13.5 |
| Qwen-14B (base) | - | No | CoT | 61.3 |
| Qwen-14B (base) | - | No | MP | 64.8 |
| WizardMath-70B | WizardMath | No | CoT | 81.6 |
| MetaMath-70B | MetaMathQA | No | CoT | 82.3 |
| Qwen-72B (base) | - | No | CoT | 78.9 |
| Qwen-72B (base) | - | No | MP | 83.5 |
| Method | LLM Sessions (per sample) | Generate/Prompt tokens | Cost | Success Rate |
|---|---|---|---|---|
| IO (best of 100) | 100 | 1.8k / 1.0k | $0.13 | 33% |
| CoT (best of 100) | 100 | 6.7k / 2.2k | $0.47 | 49% |
| ToT (breadth=5) | 61.72 | 5.5k / 1.4k | $0.74 | 74% |
| MP | 100% |
"Meta Prompting for Complex Reasoning" (see /.prompts/cr-agent* ) is a specialized adaptation of the general Meta Prompting approach, specifically tailored for tackling intricate and multifaceted problems, particularly in fields requiring in-depth analytical and logical reasoning. This version emphasizes not just the structure and syntax of the problem-solving process, but also delves deeply into the content, ensuring a thorough and comprehensive approach to each problem.
-
Complex Problem Decomposition: The technique begins with a complex problem or question, which is then broken down into smaller, more manageable sub-problems or questions. This decomposition is crucial for tackling complex issues in a systematic and methodical way.
-
Detailed Preliminary Content: Before addressing the main problem, the AI provides extensive preliminary content, including foundational concepts, relevant theories, and useful hints. This step ensures that all necessary background information is covered to understand and solve the problem.
-
Step-by-Step Problem Solving:
-
Intermediate Questions: The AI formulates a series of intermediate questions, each targeting a specific aspect of the complex problem. These questions guide the problem-solving process in a structured manner.
-
Answer Sketches and Code Execution: For each question, the AI develops a detailed answer sketch, which is then tested and refined through code execution. This process not only verifies the accuracy of the answer but also deepens the understanding of the problem.
-
Detailed Answers: Based on the code execution results, the AI provides comprehensive and detailed answers for each intermediate question, gradually building towards the solution of the original complex problem.
-
-
Final Solution Presentation:
-
Solution Synthesis: After addressing all intermediate questions, the AI synthesizes the findings into a complete solution for the original complex problem.
-
Code for Final Solution: The final solution is further verified or solved using coding, ensuring accuracy and precision.
-
Formatted Final Answer: The solution is presented in a clear, concise, and formally correct format, often using LaTeX for mathematical precision and enclosed within
\boxed{}for emphasis.
-
See ./prompts/cr-agent-assistant-v0.1.md for a minimalist implementation based on OpenAI Assistant API as a System Message.
- please visit https://chat.openai.com/g/g-L3a4ZCIHx-cr-agent-v0-1 for an online demo.
See ./prompts/cr-agent-xml-assistant-v0.1.xml for a minimalist XML-style implementation based on OpenAI Assistant API.
- please visit https://chat.openai.com/g/g-4ir4la2Z6-cr-agent-xml-v0-1 for an online demo.
See ./prompts/cr-agent-xml-assistant-v0.2.xml for a minimalist XML-style implementation based on OpenAI Assistant API.
- please visit https://chat.openai.com/g/g-cJV031wLP-cr-agent-xml-v0-2 for an online demo.
The prompt of CR Agent XML v0.2 is autonomously generated by CR Agent v0.1 (this process can be seen as metaprogramming).
In the realm of advanced machine learning and AI systems, the task of automatically generating structured prompts, termed Meta Prompting for Prompting Tasks (MP-PT) or simply Meta Prompting in this specialized case (Reynolds & McDonell, 2021; Hou et al., 2022), emerges as a critical component. This process entails utilizing language models to interpret input strings as instructions and consequently generate prompts that guide further tasks. We formalize this concept within the framework of General Meta Prompting with special tasks called prompting tasks, detailing its categorical and functorial properties.
- please visit https://chat.openai.com/g/g-o54JV8zr7-mp-pt for an online demo.
See ./prompts/mp-pt-reasoning-v0.1.tex for a minimalist latex-style implementation based on OpenAI Assistant API.
See ./prompts/mp-icpd-v0.1.md for a minimalist implementation based on OpenAI Assistant API.
- please visit https://chat.openai.com/g/g-9d0iBPnzR-mp-icpd for an online demo.
<|User|> [Input Document]: <your_system_prompt_itself>
\textbf{Output Prompt: [to be generated using the same latex format as this prompt]}
The generated is shown in ./prompts/mp-icpd-v0.2.md.
-
Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pp. 1–7, 2021.
-
Yutai Hou, Hongyuan Dong, Xinghao Wang, Bohan Li, and Wanxiang Che. Metaprompting: Learning to learn better prompts. arXiv preprint arXiv:2209.11486, 2022.
Please cite the paper and star this repo if you use Meta Prompting (MP) and find it interesting/useful, thanks! Feel free to contact zhangyif21@tsinghua.edu.cn or open an issue if you have any questions.
@article{zhang2023meta,
title={Meta Prompting for AGI Systems},
author={Zhang, Yifan and Yuan, Yang and Yao, Andrew Chi-Chih},
journal={arXiv preprint arXiv:2311.11482},
year={2023}
}