HALE framework combines LLMs with large-scale epidemic simulations

Researchers have unveiled HALE (Hybrid Agent-based and Language-driven Epidemic), a framework that combines classic agent-based modeling with large language models. The system lets an LLM predict the decisions of millions of simulated people in real time, adapting to real-world changes. Using COVID-19 in Salt Lake County as an example, the scientists showed that the hybrid approach can become a tool for more accurate forecasting in policymaking.

The problem: static models

Traditional agent-based models (ABMs) can simulate the behavior and interactions of millions of people, which is invaluable for policy development. But they rely on static assumptions: agent behavior rules are fixed in advance and do not change.

During an epidemic, people's behavior shifts radically. They start wearing masks, avoid crowds and change their work schedules. A static model cannot account for these shifts in real time; it is locked into its initial scenario and cannot learn from new data.

How HALE solves the problem

HALE inserts an LLM into the agent-based simulation loop. Instead of hard-coding behavior rules, the system uses a neural network to predict each agent's decisions dynamically. At every simulation step, the LLM analyzes the current context (the number of infected people in a district, the level of public fear, vaccine availability) and predicts the person's decision: whether they will go to a store, whether they will keep their distance, whether they will agree to be vaccinated.

• The framework uses an LLM to forecast behavior at every simulation step
• HALE was tested on COVID-19 in Salt Lake County, Utah
• The system adapts to real data instead of following a prescribed scenario
• It scales to millions of agents

Why this matters for policy

Governments use epidemiological models to make decisions about quarantine, vaccination and intervention. If a model is inaccurate, a policy may cause unnecessary harm or prove ineffective.

HALE improves accuracy through adaptivity. The system does not simply follow a plan; it changes as new data and the population's reactions emerge. Using COVID-19 in Salt Lake County, the researchers showed that HALE captures the real dynamics of infection and behavioral shifts better than static models.

This is especially important for rapid response: with a more accurate forecast, policymakers can adjust measures before an infection wave spreads.

What this means

The hybrid approach suggests that LLMs can be more than an analysis tool: they can be an active component of the simulations themselves. If HALE works for epidemics, it could be applied to other crises: economic shocks, social conflicts, natural disasters. That would open up a new class of models that are live, adaptive and driven by neural networks.