From Risk Perception to Behavior Large Language Models-Based Simulation of Pandemic Prevention Behaviors

Individual prevention behaviors are a primary line of defense during the early stages of novel infectious disease outbreaks, yet their adoption is heterogeneous and difficult to forecast-especially when empirical data are scarce and epidemic-policy contexts evolve rapidly. To address this gap, we develop an LLM-based prevention-behavior simulation framework that couples (i) a static module for behavior-intensity prediction under a specified external context and (ii) a dynamic module that updates residents' perceived risk over time and propagates these updates into behavior evolution. The model is implemented via structured prompt engineering in a first-person perspective and is evaluated against two rounds of survey data from Beijing residents (R1: December 2020; R2: August 2021) under progressively realistic data-availability settings: zero-shot, few-shot, and cross-context transfer. Using Kolmogorov-Smirnov tests to compare simulated and observed behavior distributions (p > 0.001 as the validity criterion), the framework demonstrates robust performance and improves with limited reference examples; reported predictive accuracy increases from 72.7% (zero-shot) to 81.8% (few-shot), and remains high at 77.8% under transfer to novel contexts. We further apply the framework to simulate behavior changes during China's December 2022 policy relaxation and to stress-test behavioral responses across 120 systematically varied epidemic conditions (R0, CFR, and control-measure tiers). Results indicate broad behavioral loosening under relaxation but a distinctive counter-trend increase in drain-related disinfection, highlighting how low-cost, low-friction behaviors may persist or intensify even when external constraints recede-raising a potential environmental tradeoff.