LLM-powered reasoning in agent-based modeling

Traditional agent-based models (ABMs) for epidemic simulation often rely on static parameters, failing to account for how human behavior changes in response to real-time epidemic developments. This paper asks whether Large Language Models (LLMs) can be integrated into ABMs to predict and simulate these dynamic behavioral shifts, thereby creating more accurate population-scale digital twins.

The authors developed the Hybrid Agent-based and Language-driven Epidemic (HALE) framework. This system combines a standard ABM—which handles disease transmission dynamics—with an LLM component that acts as a feedback loop. The LLM monitors the simulated epidemic status and adjusts the mobility patterns of agents based on their demographic and geographic characteristics. To maintain scalability, the authors grouped the 1.1 million agents in Salt Lake County into smaller subpopulations based on location, age, race, and sex, allowing the LLM to provide reasoning for these groups rather than for every individual separately.

The HALE framework successfully captured the observed epidemic peak and total size of the COVID-19 outbreak in Salt Lake County from September 2020 to February 2022. In contrast, traditional ABM simulations that lacked the LLM-driven behavioral feedback loop significantly overestimated the epidemic's impact. The study demonstrates that LLMs can effectively translate perceived threat levels into actionable changes in agent mobility, providing a more realistic simulation of human responses to public health crises.

This research provides a scalable, computationally efficient method for incorporating generative AI into complex social simulations. By allowing models to adapt to real-time changes in human behavior, HALE offers policymakers a more robust tool for evaluating the potential impact of public health interventions, such as lockdowns or social distancing mandates, in a way that reflects the inherent uncertainty of human decision-making.