Agentic Calibration of Grey-Box Simulation Models: An LLM-Driven Alternative

This paper treats a large language model as the optimizer for calibrating grey-box simulation models. Plausibility and clinical constraints are supplied as plain-language instructions in the system prompt, avoiding dedicated constraint-modeling machinery at the modeller-facing interface. On an anal cancer simulation model, the agentic method achieves substantially lower best error than Nelder–Mead and Bayesian Optimization in the unconstrained setting while using fewer model evaluations. Under clinical constraints, it reaches comparable error levels, with both the agentic method and Bayesian Optimization outperforming Nelder–Mead. The trade-off is higher inference time per iteration.