Title: Investigating LLM Behavioral Priors via Epidemiological Agent-Based Models: The Epi-LLM Framework

Abstract

Quantifying the impact of human behavior on infectious disease dynamics during epidemics presents a significant challenge. To address this, we present the Epi-LLM framework, a novel system that integrates large language models (LLMs), real-life epigames, and agent-based modeling. In this setup, a synthetic population of agents utilizes LLMs to reason and adapt dynamically within an outbreak contact network.

We evaluated the behavior of synthetic agents against two benchmarks: a no-intervention SEIR baseline and data from human participants in the AUIB epigame study. Our results indicate that LLM agents, across four distinct architectures, successfully reduced peak active infections. Specifically, quarantine compliance among these agents reached between 58% and 65% on the sixth day of the 15-day simulation.

Statistical analysis using a binomial generalised linear model identified perceived health severity as the most robust predictor of quarantine behavior ($\beta = 0.33, p = 0.002$). This model achieved a pseudo-$R^2$ of 0.055, a figure closely mirroring the 0.072 value observed in the human trial. Furthermore, our findings suggest that LLM architecture significantly influences epidemic dynamics. Low-variance architectures provide higher internal validity for testing specific behavioral rules, whereas high-variance models may more accurately reflect the complexity of real-world decision-making.

Notably, we found that geographic labels alone are insufficient to generate culturally differentiated behavior; explicit attitudinal parameterisation is necessary. As a proof-of-principle study, this work establishes the foundation for utilizing the Epi-LLM framework as a scalable, risk-free simulation environment to advance pandemic preparedness research.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC