behavior-driven epidemic modeling

This study addresses key challenges in epidemic decision-making, including hidden disease burdens, imperfect surveillance signals distorted by policy interventions, and intervention effects mediated by human behavior. It pioneers the systematic integration of the world model paradigm into computational epidemiology, framing epidemics as controlled partially observable dynamical systems. By jointly learning latent dynamics, endogenous observation mechanisms, and behavioral feedback loops, the approach enables counterfactual simulation and sequential decision planning under uncertainty. Evaluated across three case studies, the method effectively mitigates issues such as behavior-induced surveillance false positives and signal lags, demonstrating its necessity and superiority for policy evaluation and intervention analysis.

Traditional epidemiological models often yield biased predictions because they neglect heterogeneity in individual risk perception and behavioral responses. This study proposes a unified transmission framework that explicitly incorporates behavioral heterogeneity by categorizing the population into risk-neutral and risk-averse groups, each exhibiting distinct contact behaviors. For the first time in epidemiology, a Bayesian mixture approach is introduced to characterize how such behavioral diversity shapes epidemic dynamics. Built upon an extended SIR structure, the model integrates simulation and empirical analysis, demonstrating superior performance over conventional approaches in parameter recovery, trajectory fitting, and forecasting accuracy. It effectively mitigates common issues such as underestimation of infection peaks and artificial elongation of epidemic curves, thereby enhancing both behavioral realism and predictive reliability.

This study investigates the decision-making capabilities of generative AI agents acting as policymakers in complex social systems, with a focus on recurrent policy choices in epidemic control. The AI agent assumes the role of a mayor within an SEIR epidemic simulation environment, adjusting business restriction policies weekly based on evolving outbreak dynamics. A dynamic memory mechanism with time-decay weighting integrates historical information to inform decisions. Crucially, a theory-guided prompting strategy—incorporating concise epidemiological principles—is introduced to enhance policy reasoning. Experimental results demonstrate that this approach enables the AI agent to exhibit human-like responsiveness and significantly improves both the quality and stability of its policy decisions. The effectiveness of theory-guided prompting in shaping coherent policy behavior is validated across both single-agent and multi-agent evaluation settings.

Existing epidemiological models often neglect human behavioral responses and undetected infections—particularly asymptomatic cases—leading to biased characterizations of COVID-19 transmission dynamics. To address this, we propose a novel Bayesian stochastic differential equation (SDE) model that jointly integrates: (1) multi-source-driven population behavior dynamics—including both policy interventions and spontaneous behavioral adaptations; (2) data-generation uncertainty arising from undetected infections; and (3) coupled inversion of case, hospitalization, and mortality data. Leveraging Bayesian inference and rigorous uncertainty quantification, our framework enables joint estimation of the true infection trajectory and time-varying, nonlinear transmission rates. Empirical validation on Montreal and Miami datasets demonstrates substantially improved forecasting accuracy and robust separation of behavioral feedback effects, quantitatively revealing their critical role in suppressing transmission.

This work proposes a novel approach to epidemic modeling by framing disease transmission as an iterative program synthesis problem, addressing the limitations of traditional models that rely on fixed structures and require extensive manual intervention to adapt to evolving pathogens, changing interventions, or shifting scenario assumptions. Central to this framework is an explicit epidemiological flow graph serving as an intermediate representation, which enables modular verification and interpretable parameter learning. By integrating agent-driven program synthesis, mechanism-based model compilation, and parameter optimization constrained by both physical and epidemiological principles, the method accurately captures complex transmission dynamics across diverse scenarios, generates counterfactual predictions grounded in epidemiological logic, and significantly accelerates convergence to high-quality models.

This study addresses the modeling of individual self-reporting decision-making during infectious disease outbreaks to support precision public health interventions. To this end, we develop a spatially explicit agent-based simulation framework that integrates real census data and, for the first time, deeply couples large language model–generated individual decisions with fine-grained geographic and social structures. The framework incorporates contextual factors such as household influence and information framing. Our approach systematically reveals income and education level as key drivers of disparities in influenza-like illness reporting rates and successfully captures behavioral heterogeneity across both social and geographic dimensions in simulations of San Francisco and Atlanta. This work establishes a novel paradigm for high-resolution behavioral epidemiological modeling.

Quantifying the impact of human behavior on disease transmission during pandemics remains highly challenging. This work proposes the Epi-LLM framework, which for the first time integrates large language models (LLMs) into an agent-based SEIR epidemic model, leveraging data from pandemic-related behavioral game experiments and generalized linear models to simulate agents’ dynamic reasoning and adaptive behaviors within contact networks. The study reveals that architectural variance among LLMs significantly affects the validity of behavioral simulations and demonstrates the necessity of explicitly parameterizing attitudes to capture cultural differences. Experimental results show that all four LLM architectures effectively reduce infection peaks, achieving quarantine compliance rates of 58–65%. Perceived health severity emerges as the strongest behavioral predictor, and the model’s pseudo-R² aligns closely with empirical findings from human experiments.

This work proposes a scientific machine learning approach to construct an efficient surrogate model for agent-based epidemic simulations, which are traditionally computationally prohibitive for real-time decision-making. By embedding a mechanistic SEIR model with a neural network–parameterized contact rate into a universal differential equation (UDE) framework, the method integrates multi-stage shooting and an observer-based prediction error method (PEM) for dynamical system identification. This formulation ensures solution positivity and mass conservation while enhancing numerical stability and interpretability. Evaluated on the ExaEpi scenario, the PEM-UDE approach reduces mean squared error by 77% and 20% compared to single-stage and multi-stage UDE baselines, respectively, and achieves 90-day forecasts in just 20–35 seconds—accelerating simulation by approximately 10⁴-fold and enabling real-time “what-if” analyses on standard laptop hardware.

This work addresses the challenge of manually optimizing non-pharmaceutical intervention (NPI) strategies due to the vast combinatorial space of possible measures. To overcome this, the authors propose ADIOS, a system that automatically searches for highly effective NPI policies with minimal societal disruption. ADIOS integrates agent-based epidemic simulation with grammar-guided genetic programming (GGGP) and introduces a domain-specific language (DSL) tailored for NPIs. This DSL structures the policy space using a context-free grammar and incorporates semantic constraints to eliminate infeasible strategies, substantially improving search efficiency. Experiments on GEMS, a high-resolution microsimulation platform for epidemics in Germany, demonstrate that ADIOS efficiently discovers near-optimal intervention strategies applicable to complex real-world scenarios.