Agentic Tool Use in Large Language Models

A new survey paper organizes the fragmented research on how large language models use external tools, proposing three distinct paradigms to clarify how these autonomous agents retrieve information, compute, and act in the real world [1]. The paper, authored by Jinchao Hu and posted to the arXiv preprint server, argues that existing studies on agentic tool use are scattered across tasks, tool types, and training settings, lacking a unified view of how methods differ and evolve [1][2]. Large language models, or LLMs, are neural networks trained on vast amounts of text for tasks such as generation and translation, and they serve as the foundation for modern chatbots [3]. Their deployment as autonomous agents depends on reliable tools for information retrieval, computation, and external action [2]. To address the fragmentation, the survey organizes the literature into three paradigms: prompting as plug-and-play, supervised tool learning, and reward-driven tool policy learning [1][2]. It analyzes the methods, strengths, and failure modes of each approach and reviews the evaluation landscape [2]. The work aims to provide a structured evolutionary view of agentic tool use, moving beyond isolated benchmarks [1]. LLMs such as Meta’s Llama family, which ranges from 1 billion to 2 trillion parameters, exemplify the rapid scaling of these models [5]. Instruction fine-tuned versions of Llama have been released alongside foundation models starting with Llama 2, and the latest iteration, Llama 4, arrived in April 2025 [5]. The survey’s framework is intended to help researchers situate new tool-use methods within a broader trajectory, rather than treating each as a standalone contribution [2]. The paper was submitted to arXiv on April 1, 2026, and revised on June 29, 2026 [1]. arXiv, an open-access repository of electronic preprints founded in 1991, now receives about 24,000 articles per month and hosts over two million papers [9]. The repository is not peer-reviewed, but it has become the primary distribution channel for research in computer science and physics [9]. The survey’s abstract and bibliographic tools are accessible through arXiv’s abstract page, which includes experimental community features developed under the arXivLabs framework [1][8].