Antagonistic Evolution for LLM Tool Use

Yuefeng Huang, Xu Huang, Yi Yang, Hong Xie, Defu Lian

Keywords: Large Language Model, Tool Learning, Adversarial Evolution

Abstract: Tool use has emerged as a pivotal mechanism for enhancing Large Language Models (LLMs), allowing them to interact with external tools to solve complex tasks and access knowledge beyond their static pre-trained parameters. However, most existing studies rely on advanced LLMs to improve tool-use capabilities via data synthesis, often resulting in suboptimal data quality or mismatched task difficulty, thereby limiting model performance. To address these limitations, we propose a novel antagonistic evolution framework for tool-use tasks, involving a query-generation model and a tool-use model updated in an adversarial manner. The query-generation model is optimized to produce increasingly challenging and high-quality queries, which the tool-use model then learns to solve. This adversarial process is iteratively executed, enabling both models to co-evolve and progressively enhance the tool-use capabilities. Experiments on three comprehensive tool-use benchmarks demonstrate evolving performance improvements, validating the effectiveness of our approach.

Primary Area: generative models

Submission Number: 4709