Coordinated Pandemic Control with Large Language Model Agents as Policymaking Assistants

TL;DR

The paper tackles fragmentation and reactive responses in pandemic policymaking by introducing a decentralized framework where each administrative region is represented by an autonomous LLM agent that reasons over local epidemiological dynamics while coordinating with others through structured communication. A closed-loop simulation combines region-specific SEIQRD dynamics with temporal inflow reallocation of inter-regional mobility, enabling proactive, cross-region policy optimization. Empirical results on US COVID-19 data show substantial reductions in infections and deaths at both state and aggregate levels, with up to $63.7\%$ fewer infections and $40.1\%$ fewer deaths at the state level, and $39.0\%$ and $27.0\%$ reductions across states, respectively; planning horizon and policy design (TIR vs SIS/TIS) influence performance and equity. The framework demonstrates scalability to 20 states, maintains robustness across LLM backbones, and offers a general paradigm for coordinating complex, interdependent policy decisions beyond pandemics, while acknowledging limitations in model realism and real-world compliance.

Abstract

Effective pandemic control requires timely and coordinated policymaking across administrative regions that are intrinsically interdependent. However, human-driven responses are often fragmented and reactive, with policies formulated in isolation and adjusted only after outbreaks escalate, undermining proactive intervention and global pandemic mitigation. To address this challenge, here we propose a large language model (LLM) multi-agent policymaking framework that supports coordinated and proactive pandemic control across regions. Within our framework, each administrative region is assigned an LLM agent as an AI policymaking assistant. The agent reasons over region-specific epidemiological dynamics while communicating with other agents to account for cross-regional interdependencies. By integrating real-world data, a pandemic evolution simulator, and structured inter-agent communication, our framework enables agents to jointly explore counterfactual intervention scenarios and synthesize coordinated policy decisions through a closed-loop simulation process. We validate the proposed framework using state-level COVID-19 data from the United States between April and December 2020, together with real-world mobility records and observed policy interventions. Compared with real-world pandemic outcomes, our approach reduces cumulative infections and deaths by up to 63.7% and 40.1%, respectively, at the individual state level, and by 39.0% and 27.0%, respectively, when aggregated across states. These results demonstrate that LLM multi-agent systems can enable more effective pandemic control with coordinated policymaking...

Figures (6)

• Figure 1: Comparison between traditional pandemic policymaking and our proposed LLM-agent-enabled coordinated policymaking. a, Traditional pandemic policymaking under fragmented coordination. Regional authorities operate largely in isolation, relying on delayed or incomplete information and independent analyses. Limited cross-region communication and asynchronous interventions lead to misaligned policies and inefficient pandemic control. b, LLM multi-agent system for coordinated pandemic policymaking. Each administrative region is represented by an LLM agent that exchanges information with others, enabling collaborative reasoning and coordinated policy decisions across regions while accounting for inter-state mobility. c, Architecture of the proposed LLM-agent system for pandemic policymaking. Each regional agent integrates epidemiological modeling, real-world data inputs (case records, mobility flows, demographics, and healthcare resources), and structured prompts to reason about policy actions. Inter-regional mobility is modeled through state-level mobility flows, inducing cross-region transmission dynamics. Agents communicate, evaluate outcomes, and generate coordinated policy decisions, which are instantiated as intervention strategies (e.g., temporal inflow reallocation, spatial inflow suppression, and targeted inbound screening) and fed back into the pandemic simulation loop.
• Figure 2: Effectiveness of the LLM multi-agent policymaking framework in mitigating pandemic spread. The cumulative confirmed infections (a) and cumulative deaths (b) vary across the time (from May 2020 to December 2020). Results are compared across different intervention strategies, including the LLM agent–based policymaking, expert-experience-based policymaking, random policymaking, and the observed ground-truth policymaking. c,d, The overall reduction in cumulative infections and deaths achieved by the LLM agent–based policy differ across five states, where numerical annotations on each bar denote the percentage reduction in cumulative infections and mortality relative to the ground truth. Furthermore, e-h, LLM agent-based policy stabilizes the temporal dynamics of effective reproduction number $R_t$ and reduces its mean value.
• Figure 3: State-level intervention illustration and interpretation.a, The real-world policy is compared with the mobility reallocation policy generated by the LLM agent in MS (Mississippi), together with the corresponding daily incidence rate. Each reallocation cycle is highlighted in the background using alternating light blue and yellow shading. b,c, The specific mobility adjustments during the second reallocation cycle and the reasoning process of LLM agents are illustrated. Moreover, d, LLM agent policy are classified into three types (strict-first, balanced, and relaxed-first). Policy-type distributions for Mississippi and Texas are presented, along with each feature’s marginal contribution (i.e., the Shapley Values) on strict-first policy.
• Figure 4: Performance comparison across multi-dimensional policies. First, a,b, the state-level performance of different intervention strategies (TIR, SIS, and TIS) in terms of total infections and deaths is illustrated. The real-world conditions are indicated by black dashed lines, with the corresponding increase or reduction percentages under different intervention strategies annotated above. Then, c,d, the system-level indicators, equity coefficient and the total reductions in infections and deaths, are compared across strategies. Each strategy is evaluated over five runs, with each point representing a single run. e, The selection probabilities of different states under SIS and TIS across the study period are presented. Moreover, f–h, TIR under planning horizons of 4, 6, 8, and 10 weeks is compared, evaluated by incidence rate (IR), death rate (DR), and total infections.
• Figure 5: Pandemic control performance in 20-state experiments.a, State-level comparisons of the average daily active case rate (ACR), daily incidence rate (IR), and daily death rate (DR) across 20 states demonstrate the generalizability of the proposed framework. b, Total cumulative infected cases under the LLM agent–based policy are compared with the ground-truth outcomes over the study period. In addition, we project the evolution of cumulative infections over the subsequent 180 days under this intervention pattern to demonstrate the long-term performance.