Improving LLM-based Wargame Engagement Scenario Generation via DEVS-based Combat Modeling and Simulation

With the emergence of the Mosaic Warfare paradigm and advances in artificial intelligence (AI) technologies, modern battlefield environments have become increasingly complex due to the proliferation of diverse intelligence, surveillance, reconnaissance (ISR), and weapon systems. This complexity significantly increases the cognitive and temporal burden on military commanders in achieving timely situational awareness and making effective decisions. In this paper, we propose an LLM-based course-ofaction (COA) generation and refinement framework with a DEVS-based engagement-level simulation model, TDGSim. Given an operational scenario, the large language model (LLM) autonomously generates multiple candidate COAs, which are evaluated through simulation. Based on the evaluation results, the COAs are either refined iteratively or selected according to our proposed improvement and selection strategies. Our experimental results show that iterative, simulation-informed self-refinement of COAs by the LLM outperforms simple evaluation-based COA selection methods. Through this work, we demonstrate the feasibility of an LLM-enabled, automated military decision-making process (MDMP) support system for future defense modeling and simulation (M&S) environments.