Tool learning with language models: a comprehensive survey of methods, pipelines, and benchmarks

Abstract Tool learning has emerged as a key capability for enhancing the reasoning and decision-making abilities of large language models (LLMs) by enabling them to interface with external tools such as application programming interfaces (APIs), search engines, and calculators. This survey provides a systematic overview of the tool learning paradigm, focusing on how LLMs can decompose complex tasks, select appropriate tools, invoke them correctly, and generate coherent responses. We summarize a unified four-stage framework comprising task planning, tool selection, task execution, and response generation, which captures the core processes underlying tool-augmented language modeling. We further analyze major learning methodologies, including tuning-free methods, supervised fine-tuning methods, and reinforcement learning, and discuss how they contribute to different stages of the tool-use pipeline. The survey also reviews recent benchmarks designed to evaluate tool-use competence across both general and domain-specific scenarios. Finally, we highlight open challenges in safety, efficiency, and generalization, and outline promising directions for future research. This work aims to serve as a conceptual roadmap and practical reference for researchers and developers working on tool-augmented artificial intelligence (AI) systems.