Epidemic Modeling with Generative Agents: Methodology, Prompt Sensitivity, and LLM Sensitivity

Abstract

Generative agent-based modeling (GABM) replaces the rule-based decision-making of classical agent-based models with agent decisions powered by large language models (LLMs), enabling researchers to simulate human behavior that resists clean mathematical specification. This dissertation establishes the methodology, characterizes its sensitivity to prompt design, and characterizes its sensitivity to the underlying LLM. Three studies — conducted within a shared epidemic GABM framework in which generative agents decide each day whether to stay home or go out — support the central claim: GABM is a viable simulation method, but the prompt and the LLM are both parameters of any GABM finding. Agents endogenously self-isolate as community case counts rise and quarantine when they feel ill. Collectively, this produces multi-wave dynamics followed by an endemic period. Heterogeneous behavior emerges from Big Five personality traits, age, and gender. The prompt is a parameter of GABM, but only along certain dimensions. Synonymous rewording and persona names leave epidemic trajectories statistically unchanged; minor wording variations and contextual shifts both alter outcomes. LLM choice is also a parameter. Across 21 LLM configurations from OpenAI, Anthropic, and Google, most LLMs agree on the direction of persona effects but disagree several-fold on magnitude. Standard model attributes — size, version, knowledge cutoff, release date — do not predict where an LLM's collective agent behavior lands. Mentioning a trait in the response text moves that trait's effect on the decision beyond the trait alone. Together the three studies argue that GABM is most defensible for qualitative claims about agent behavior and weakest for the quantitative claims modelers most often want to make. LLM sensitivity testing should become standard practice in GABM publications, alongside prompt sensitivity. The methodology is real; its parameters include the prompt and the LLM.