Epi-LLM framework uses LLM agents to simulate realistic epidemic behavior

Researchers Petra Ferenz, Ava Keeling, and colleagues introduced Epi-LLM, a framework that combines agent-based modeling, real-life epigames, and large language models (LLMs) to simulate how human behavior affects epidemic dynamics. They created a synthetic society of LLM-driven agents that reason and adapt over a contact network during a simulated outbreak. Comparing against an SEIR no-intervention baseline and human data from the AUIB epigame study, they found that LLM agents across four different architectures reduced peak active infections. Quarantine compliance peaked at 58-65% on day six of the 15-day simulation. A binomial generalized linear model revealed that perceived health severity was the strongest predictor of quarantine behavior (β = 0.33, p = 0.002), yielding a pseudo-R² of 0.055—close to the 0.072 observed in human trials.

Crucially, the choice of LLM architecture significantly influences epidemic outcomes: low-variance architectures provide better internal validity for testing behavioral rules, while high-variance models may better capture real-world decision-making diversity. Geographic labels alone did not induce culturally differentiated behavior—explicit attitudinal parameterization was required. This proof-of-principle work lays the foundation for using Epi-LLM as a scalable, risk-free simulation environment for pandemic preparedness, allowing researchers to probe behavioral priors and test interventions without real-world consequences.