CReTooL: Enhancing Complex Reasoning through Diversified Tool Learning

Abstract

Large Language Models (LLMs) [1] have exhibited significant reasoning capabilities across various domains and modalities. Tool learning has been proven to enhance LLMs' performance in complex reasoning tasks [2]. This approach leverages external tools such as the web search API [3] for knowledge retrieval and Python-based calculators for computations [4]. However, current research faces three challenges: (1) Lack of diverse and specialized tool design for complex reasoning tasks. (2) Insufficient adaptive mechanisms for leveraging diverse tools. (3) Absence of effective self-evaluation mechanisms for the final answer. To mitigate these challenges, we propose CReTooL (Complex Reasoning through diversified Tool Learning), designed to improve LLMs' complex reasoning by adaptively leveraging diverse tools. CReTooL comprises four components: diversified tool construction, adaptive tool selection, answer generation and self-evaluation module. Extensive experiments on benchmarks demonstrate that the proposed framework CReTooL achieves excellent performance in complex reasoning, with strong tool adaptive planning and leveraging capabilities.

DOI: https://doi.org/10.24135/ICONIP11