Abstract for: Integrating Simulation Modeling with Large Language Models to Evaluate Spatially-Aware Strategies for Pandemic Response

This study explores the integration of Large Language Models (LLMs) with simulation engines, such as both system dynamics (SD) and agent-based models (ABM), for enhancing public health intervention outcomes during disease outbreaks. It aims to refine spatially-disaggregated simulations with LLM-derived local knowledge and develop a framework for interoperability between LLMs and simulation engines at specific time steps. Initial experiments demonstrate that LLMs can improve models by adjusting spatial preferences and behaviors based on simulation data and local insights, leading to a deeper understanding of infectious disease spread. The LLM-ABM integration faces several challenges, including ensuring consistent behavior among agents in ABMs and the computational demands of frequent LLM queries. For the LLM-SD integration, we developed an XMILE Stella wrapper to dynamically adjust parameters based on the LLM output. This revealed further issues, such as efficiently incorporating varied spatial information within LLM constraints and scaling the model. The study underscores LLMs' potential to craft bespoke public health strategies using their spatial knowledge and reasoning capabilities, while also highlighting the current limitations.