Reproduce “MetaTool: Facilitating Large Language Models to Master Tools with Meta-task Augmentation” on OpenResearch

About this paper

Xiaohan Wang, Dian Li, Yilin Zhao, Sinbadliu, Hui Wang

This paper introduces MetaTool, a method that enhances Large Language Models' (LLMs) ability to master tools by training them on self-supervised 'meta-tasks' designed to impart a task-agnostic understanding of tool mechanics. The approach enables open-source LLMs like LLaMA3-8B to achieve tool-use performance competitive with, and in some cases exceeding, state-of-the-art proprietary models like ChatGPT and GPT-4, demonstrating strong generalization to unseen tools and complex scenarios.

Key insights

• Effective and generalizable tool use requires LLMs to develop a task-agnostic understanding of tool mechanics, encompassing causality, preconditions, and functional boundaries, beyond just learning specific task solutions.
• Self-supervised 'meta-tasks' (e.g., predicting effects, preconditions, or reachable states) can generate scalable training data that imparts this fundamental tool knowledge.
• Integrating meta-task data through instruction tuning alongside conventional solution data allows LLMs to learn both the foundational properties of tools and their practical application.

Summary by alphaXiv.