Theory-Informed Generative Agents for Human Mobility Modeling

Human mobility follows robust population-level regularities, yet individual behavior remains highly heterogeneous and context-dependent. The advances of generative agents (large language model (LLM)–driven, persona-conditioned computational agents) offer a promising approach to modeling rich, individualized behavior, but they often lack theoretical grounding and do not readily scale to population-level simulation due to the computational cost of generating decisions for large numbers of agents over long horizons. To reconcile mechanistic understanding with flexible, human-like decision modeling, we introduce a theory-informed mobility agent (TIMA) framework that integrates the physical constraints of mobility theory with the semantic reasoning of LLMs. Mechanistically, the LLM parameterizes the decision tendencies within the physical mobility model process to capture both the scaling laws of aggregate flows and the heterogeneity of individual preferences. Unlike prior LLM agents that are typically evaluated on small, bespoke samples, TIMA enables population-scale simulation while reconstructing key mobility patterns and diverse activity portfolios across multiple U.S. cities in a data-light manner. Beyond pattern reconstruction, we leverage TIMA to investigate mobility-mediated social phenomena. Specifically, we examine how individual choices shape experienced segregation and drive behavioral adaptations in response to external shocks. Using the COVID-19 pandemic as a case study, we demonstrate that the framework naturally captures shifts in segregation and reductions in activity as a function of evolving disease risk and policy constraints. Ultimately, this work illustrates a paradigm for integrating physical theory with generative AI, providing a scalable and flexible foundation for the next generation of policy-relevant behavioral simulation.