DocsChisel Optimizes LLM Agent Tool Documentation for Better Performance.

Large language models increasingly rely on external tools, making the quality of tool documentation crucial for agent performance. Current approaches often treat documentation as a static input, or attempt generic optimization without considering how information fields impact agents across different settings. A new study reveals that the effectiveness of documentation fields varies significantly based on task domain, LLM backbone, and agent paradigm, indicating no single documentation style works universally. To address this, researchers developed DocsChisel, an adaptive framework. DocsChisel works by analyzing an LLM agent's failed execution traces to pinpoint documentation-related issues. It then iteratively refines the tool documentation by adding, removing, or modifying specific information fields for each tool. Evaluations show DocsChisel dramatically improves LLM agent task success rates, achieving a 95.89% increase over original documentation and an average 75.15% improvement compared to state-of-the-art baselines like EasyTool and DRAFT. This optimization is achieved with minimal overhead in terms of time and tokens.