To Call or Not to Call: Framework for Assessing and Optimizing LLM Tool Calling Decisions

arXiv·medium signal

From Max Planck Institute: evaluates web search tool-use decisions along necessity, utility, and affordability, combining normative (true need inferred from optimal allocation) and descriptive (model's self-perceived need from observed behaviors) perspectives. Shows that while tool calling helps in aggregate, optimal calling achieves much better performance than always-call with significantly fewer calls. Directly applicable to reducing latency and cost in production agent systems.