Pandemic Control Achieves 63.7% Improvement with Large Language Model Policymaking Assistants

Coordinated pandemic control presents a significant challenge, demanding swift and unified policymaking between interconnected administrative regions. Ziyi Shi, Xusen Guo, and Hongliang Lu, from The Hong Kong University of Science and Technology, alongside Mingxing Peng, Haotian Wang from the Guangzhou campus, and Zheng Zhu of Zhejiang University, present a novel framework utilising large language model agents to assist in this complex process. Their research proposes assigning an LLM to each region, enabling it to reason about local epidemiology while simultaneously communicating with others to address cross-regional dependencies. This innovative approach, validated using US state-level COVID-19 data, demonstrates a potential to reduce cumulative infections and deaths by up to 63.7% and 40.1% respectively, highlighting the promise of LLM-supported systems for proactive and effective pandemic mitigation.

LLM Agents for Coordinated Pandemic Policymaking

Effective pandemic control requires timely and coordinated policymaking across administratively interdependent regions. Human responses are often fragmented and reactive, with policies frequently adjusted only after outbreaks escalate, undermining proactive intervention and global pandemic mitigation. Researchers propose a large language model (LLM) multi-agent policymaking framework to support coordinated and proactive pandemic control across regions, assigning an LLM agent to each administrative region to function as an AI policymaking assistant. These agents reason over regional epidemiological data, population characteristics, and economic factors to generate policy recommendations, communicating them to neighbouring regions to initiate a negotiation process.

This process is facilitated by a shared understanding of the pandemic’s dynamics and the potential consequences of different policy options, derived from the LLM’s knowledge base. The framework aims to move beyond reactive measures towards a more anticipatory and collaborative approach to pandemic management. The research contributes a novel methodology for translating complex epidemiological data into actionable policy recommendations using LLMs, alongside a multi-agent negotiation protocol designed to foster cooperation between regions with potentially conflicting interests. The study demonstrates the feasibility of using LLMs to simulate pandemic spread and evaluate the effectiveness of different policy interventions in a realistic, interconnected environment, tested using simulated pandemic scenarios with varying parameters.

LLM Agents Model Pandemic Response Coordination

Effective pandemic control requires policymakers to implement strategies that protect public health, but coordinating responses across different regions presents a significant challenge. Jurisdictions often favour immediate containment measures over long-term, coordinated planning, which can hinder global control efforts and worsen disease spread. Regional differences in population density, movement patterns, healthcare resources, socioeconomic factors and data availability further complicate the design of interventions that are both locally relevant and globally consistent. Researchers have developed a framework utilising large language model (LLM) multi-agent systems to address these challenges, enabling agents to simulate pandemic evolution and communicate to account for regional interdependencies.

The system integrates real-world data, a pandemic evolution simulator, and structured communication between agents, allowing them to explore potential intervention scenarios and formulate coordinated policy decisions through a closed-loop simulation process. This approach overcomes the difficulties of handling large volumes of complex information under pressure and proactively intervening in the face of uncertainty. Validated using state-level COVID-19 data from the United States between April and December 2020, incorporating real-world mobility records and policy interventions, results demonstrated a reduction in cumulative infections of up to 63.7% and deaths of up to 40.1% at the individual state level, with aggregated reductions of 39.0% and 27.0% respectively. These findings suggest that LLM multi-agent systems can facilitate more effective pandemic control through coordinated policymaking, offering a potentially valuable decision-support tool for future pandemics and other complex societal issues.

LLM Framework Coordinates Pandemic Policy Across Regions

Scientists have developed a large language model (LLM) multi-policymaking framework designed to improve pandemic control through coordinated regional responses. The research team assigned an LLM to each administrative region as a policymaking assistant, enabling these agents to reason over local epidemiological data and communicate with others to account for cross-regional dependencies. This framework integrates real-world data, a pandemic evolution simulator, and structured inter-communication to explore potential intervention scenarios and synthesise coordinated policy decisions through a closed-loop simulation process. Experiments utilising state-level COVID-19 data from the United States, spanning April to December 2020, alongside real-world mobility records and observed policy interventions, yielded significant results.

The team measured a reduction in cumulative infections of up to 63.7% and a decrease in cumulative deaths of up to 40.1% at the individual state level when comparing the LLM-driven approach to actual pandemic outcomes, with aggregated reductions of 39.0% in infections and 27.0% in deaths. The framework’s ability to integrate diverse data sources and simulate intervention scenarios allows for a more nuanced understanding of pandemic dynamics than traditional, reactive approaches. The closed-loop simulation process allows for continuous refinement of policies based on evolving conditions, addressing the challenges of regional heterogeneity and uncertainty. This work presents a generalisable framework for implementing LLM agents in large-scale public policy settings, offering a promising decision-support paradigm for future pandemics and other complex societal challenges.

LLM Coordination Improves Pandemic Outcomes Significantly

This research demonstrates a novel framework employing large language models as policymaking assistants to improve pandemic control through coordinated regional interventions. By assigning an LLM to each administrative region, the system facilitates communication and considers epidemiological dynamics alongside cross-regional dependencies, enabling joint exploration of counterfactual scenarios and synthesised policy decisions. Validation using COVID-19 data from the United States showed substantial reductions in both cumulative infections and deaths compared to real-world outcomes, with improvements of up to 63.7% and 40.1% respectively at the state level. Testing with various LLM models, including GPT-3.5, Gemini-2.5, and LLaMA-3-8B, indicated stable performance across different states, suggesting the robustness of the approach. The authors acknowledge limitations inherent in the pandemic simulation and data used, noting potential discrepancies between model predictions and real-world complexities. Future work could focus on incorporating more granular data, exploring different intervention strategies, and adapting the framework to address other public health crises, potentially enhancing its predictive capabilities and real-world applicability.