MedCalc-Pro Enables LLM Agents for Complex Medical Calculations.

Current benchmarks for evaluating Large Language Models (LLMs) in medical calculations often simplify real-world clinical scenarios, typically assuming a single calculator per patient case and explicit tool specification. However, actual clinical practice frequently demands multiple calculators, nested calculations, and the ability to handle ambiguous queries. To address this, the MedCalc-Pro benchmark has been developed, featuring 2,268 real-world clinical cases across 77 medical calculators and 14 clinical departments. It covers three progressively challenging task settings: single-calculator, multi-calculator, and nested-calculator scenarios. Alongside the benchmark, a generalizable agent framework is proposed. This framework enhances LLM performance in complex clinical settings by supporting multi-tool selection and nested-tool calling, while also mitigating parameter error propagation through structured validation and evidence review. Systematic comparisons against open-source, closed-source, and medical-specialized LLMs demonstrate that this new framework achieves the best performance across all three task settings.