EASYTOOL: Enhancing LLM-based Agents with Concise Tool Instruction

EASYTOOL: Enhancing LLM-based Agents with Concise Tool Instruction

27 Mar 2024 | Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Kan Ren, Dongsheng Li, Deqing Yang
EASYTOOL is a framework designed to enhance the tool utilization capabilities of large language models (LLMs) by transforming diverse and lengthy tool documentation into concise and unified tool instructions. The framework addresses the issues of inconsistency, redundancy, and incompleteness in existing tool documentation, which can hinder LLMs' ability to effectively use tools. EASYTOOL purifies essential information from various sources, creating standardized tool descriptions and functionalities. It also provides detailed guidelines for tool usage, including parameter demonstrations, to improve the accuracy and efficiency of tool invocation. Extensive experiments on multiple datasets demonstrate that EASYTOOL significantly reduces token consumption and improves the performance of LLM-based agents in real-world scenarios. The framework is applicable to both open-source and closed LLMs, showing its broad applicability and effectiveness in enhancing tool utilization.EASYTOOL is a framework designed to enhance the tool utilization capabilities of large language models (LLMs) by transforming diverse and lengthy tool documentation into concise and unified tool instructions. The framework addresses the issues of inconsistency, redundancy, and incompleteness in existing tool documentation, which can hinder LLMs' ability to effectively use tools. EASYTOOL purifies essential information from various sources, creating standardized tool descriptions and functionalities. It also provides detailed guidelines for tool usage, including parameter demonstrations, to improve the accuracy and efficiency of tool invocation. Extensive experiments on multiple datasets demonstrate that EASYTOOL significantly reduces token consumption and improves the performance of LLM-based agents in real-world scenarios. The framework is applicable to both open-source and closed LLMs, showing its broad applicability and effectiveness in enhancing tool utilization.
Reach us at info@study.space