Agentic Tool Use in Large Language Models

2026-04-01

AI Summary

This work addresses the limitations of large language models as autonomous agents in real-world applications due to unreliable tool use, a challenge exacerbated by the fragmented nature of existing research across tasks, tool types, and training methodologies. The paper introduces, for the first time, a structured evolutionary perspective on tool use, systematically unifying three dominant paradigms: plug-and-play prompting, supervised tool learning, and reward-driven strategies. It synthesizes their methodological foundations, strengths, and failure modes, integrating techniques from prompt engineering, supervised fine-tuning, and reinforcement learning. By encompassing diverse tool categories—including information retrieval, computation, and external action—the study establishes a cohesive taxonomy and evaluation framework, offering a clear roadmap and highlighting critical open problems for future research on agent-based tool utilization.

Abstract

Large language models are increasingly being deployed as autonomous agents yet their real world effectiveness depends on reliable tools for information retrieval, computation and external action. Existing studies remain fragmented across tasks, tool types, and training settings, lacking a unified view of how tool-use methods differ and evolve. This paper organizes the literature into three paradigms: prompting as plug-and-play, supervised tool learning and reward-driven tool policy learning, analyzes their methods, strengths and failure modes, reviews the evaluation landscape and highlights key challenges, aiming to address this fragmentation and provide a more structured evolutionary view of agentic tool use.

Problem

• agentic tool use
• large language models
• tool-use paradigms
• fragmentation
• evolutionary view

Innovation

• agentic tool use
• large language models
• tool learning paradigms
• reward-driven learning
• evaluation framework