Zero-shot Synthetic Sheltering Behavior in Infectious Disease Agent-Based Models

Abstract

Agent-based models (ABMs) simulate interactions between individuals or “agents” in a virtual environment, where each agents’ decision logic has traditionally been encoded using simple rules. However, this approach relies on knowing how individuals might respond to policy interventions in specific outbreak scenarios beforehand. Large-language models (LLMs) for agent decision-making offer a promising solution due to their expansive training, though, so we explored the ability of LLMs to replicate heterogeneous human behavior of sheltering-in-place during the COVID-19 pandemic. A local 4-billion-parameter Qwen3 model and the GPT-5.6 frontier model . Bayesian logistic models were fit separately to each of the resulting synthetic datasets and to the Understanding America Survey data and the average marginal effect (AME) were compared to test whether each model recovered the association between determinants and sheltering. The average difference in AME between the Qwen sets and UAS data was small ). Nonetheless, individual-level prediction failed. A Bayesian model trained on real UAS data predicted held-out sheltering decisions at 64% accuracy compared to s. LLMs thus appear to capture population-level association structure while failing at the individual resolution ABMs require. We therefore caution directly integrating LLMs into the ABM without evaluation, though they may serve as a valuable proxy in the future with some fine-tuning.