Simulating population compliance with pandemic interventions using large language models.

Area of Science:

• Computational epidemiology
• Agent-based modeling
• Artificial intelligence in public health

Background:

• Effective pandemic response relies on modeling population compliance with non-pharmaceutical interventions (NPIs).
• Existing epidemic models often use fixed behavioral scenarios, not emergent processes.
• Modeling individualized behavioral responses to dynamic NPIs and disease risk is crucial.

Purpose of the Study:

• To evaluate the capability of large language model (LLM)-based agents in generating individualized behavioral responses to time-varying NPIs and disease risk.
• To assess LLM agents' performance in simulating mobility changes in response to pandemic conditions.

Main Methods:

• Instantiated demographically representative LLM agents in three U.S. cities (Boston, Denver, San Antonio).
• Conditioned agents on evolving COVID-19 outbreak conditions and policies without fitting to observed mobility data.
• Benchmarked LLM-generated mobility changes against cellphone-derived foot-traffic records across various venue types.

Main Results:

• LLM agents generated zero-shot mobility changes across restaurants, retail, and entertainment venues.
• Simulations captured average mobility trends across cities and venue types.
• LLM simulations exhibited overly narrow within-city variation, with distinct biases among different LLMs.

Conclusions:

• LLM agents present a promising framework for modeling adherence to NPIs during pandemics.
• Ensemble approaches enhance the robustness and performance of LLM-based simulations.
• Further fine-tuning and empirical validation are necessary for LLM agents to support public policy analysis.