PLAY2PROMPT: Zero-shot Tool Instruction Optimization for LLM Agents via Tool Play
Keywords: large language models, zero-shot tool use, prompt optimization
TL;DR: We propose PLAY2PROMPT, a framework to optimize tool documentation and generate usage examples by enabling LLMs to interact with tools, which empowers LLMs to utilize tools more effectively in zero-shot settings.
Abstract: Large language models (LLMs) are increasingly integrated with external tools to complete user requests. Many real-world applications require LLMs to use specialized tools in a zero-shot setting. To achieve this, current methods primarily rely on prompting LLMs with tool-specific information, yet tool documentation is often underspecified or noisy, limiting effectiveness. Manual improvements are inefficient and impractical, as they require domain expertise to rewrite documentation and test on carefully curated held-out datasets to evaluate performance gains. Automatic prompt engineering techniques are not applicable either, because they require labeled examples, which is unavailable in the zero-shot setting. In this work, we introduce PLAY2PROMPT, an automated framework that iteratively refines tool documentation and generates usage examples. PLAY2PROMPT enables LLMs to explore tool input-output behaviors, allowing us to effectively search the space of possible tool descriptions and examples. The generated examples not only guide LLM inference but also serve as validation data to ensure more effective tool use. Extensive experiments on real-world tasks demonstrate significant improvements in zero-shot tool performance across both open- and closed-source models.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13080
Loading