Epi-LLM Framework Probes Epidemic Behavioral Priors

A preprint posted to arXiv as 2606.02867 on June 1, 2026, titled "The Epi-LLM Framework: probing LLM behavioral priors through epidemiological agent-based models," introduces a method that combines agent-based modeling, data from real-life "epigames," and large language models, the authors report. In the framework, a synthetic society of LLM-driven agents reasons and adapts over an outbreak contact network, according to the paper.

The authors report that LLM agents across four model architectures reduced peak active infections relative to a no-intervention SEIR baseline, a standard compartmental epidemic model. Quarantine compliance among the agents peaked at 58 to 65 percent on day six of a 15-day simulation, the paper states. A binomial generalized linear model identified perceived health severity as the strongest predictor of quarantine behavior, yielding a pseudo R-squared of 0.055, which the authors compare with 0.072 observed in human-trial data from an earlier epigame study.

The paper characterizes the work as a proof-of-principle for scalable, risk-free simulation that could support pandemic preparedness, the authors report. They note that the choice of model family affected results, with lower-variance LLMs producing more internally consistent experiments and higher-variance models potentially capturing more of the heterogeneity seen in real human behavior.

Researchers have been testing whether LLMs can stand in for human participants in social-science and public-health simulations, where running real-world behavioral experiments during an outbreak is impractical or unethical. Approaches like this are typically judged on how well agent behavior matches human data and how sensitive results are to model choice and prompting. The relatively low pseudo R-squared values, in both the model and the human comparison, suggest behavioral prediction here explains only a small share of variance, a caution against over-reading the agreement.

As a preprint, the work has not yet completed peer review, and its findings should be read as preliminary rather than validated for policy use. Signals to watch include peer-reviewed publication, replication across other outbreak settings and model families, and independent comparison of LLM-agent behavior against larger human datasets.