Type: url
Original: https://arxiv.org/pdf/2603.22862
Status: generated
Summary
“The Evolution of Tool Use in LLM Agents: From Single-Tool Call to Multi-Tool Orchestration” (Xu et al., Harbin Institute of Technology) is a survey arguing that the central problem of LLM tool use has shifted from whether a model can make one correct call to orchestrating many tools over long trajectories with intermediate state, execution feedback and constraints on safety, cost and verifiability. It organizes the literature along six dimensions — inference-time planning and execution, training and trajectory construction, safety and control, efficiency, capability completeness in open environments, and benchmark design — and reviews applications in software engineering, enterprise workflows, GUIs and mobile systems. It closes with open problems centred on evolvable tool governance, verifiable closed-loop execution and adaptive orchestration.