LLM-powered reasoning in agent-based modeling
06:00 · July 9, 2026 · arXiv cs.AI RSS
Agent-based modeling (ABM) has the capability to model millions of individuals and their interactions, which is useful for policy making. However, ABMs have traditionally relied on static prior, which prevents the models from adapting to real-time changes. Our research provides a novel approach to addressing this information gap. Large language models (LLMs) offer new opportunities to predict human decision-making. Here, we introduce a scalable Hybrid Agent-based and Language-driven Epidemic (HALE) modeling framework that leverages LLMs to predict human decision-making in an ABM simulation. As a proof-of-concept, we use HALE to simulate COVID-19 and its effects in Salt Lake County, UT.
Summary
Traditional agent-based models excel at representing interactions among millions of individuals yet have long depended on fixed historical priors for behavior, mobility, and contact patterns. This static foundation prevents them from reflecting how people adjust their activities in response to evolving conditions such as rising case counts or changing public-health guidance. The Hybrid Agent-based and Language-driven Epidemic (HALE) framework addresses this limitation by inserting large language models into the simulation loop to generate dynamic predictions of deliberate human choices.
Within HALE, an LLM receives current epidemic indicators and demographic context for each agent and returns updated probabilities for activities such as leaving home, attending work or school, or engaging in leisure. These predictions replace or augment the original activity schedules derived from the 2017 National Household Travel Survey, allowing the underlying temporal contact network to evolve during the run. Because only the decision layer is off-loaded to the language model, the approach remains computationally tractable for populations on the order of 1.1 million while still preserving the detailed spatial and temporal resolution of conventional ABM.
As a proof-of-concept, the authors applied HALE to a COVID-19 scenario in Salt Lake County, Utah, covering September 2020 through February 2022. The hybrid model reproduced the timing and magnitude of the observed epidemic peak more closely than a baseline ABM that relied solely on static priors; the latter systematically overestimated total infections. The framework also revealed spatial and demographic heterogeneity in behavioral responses that pure ABM could not capture, demonstrating how LLM-driven reasoning can supply the missing feedback between perceived risk and individual mobility.
Why it matters
This research is highly relevant for Dutch AI researchers and policy-makers, as it offers a novel methodology for dynamic policy simulation and epidemiological modeling. Dutch institutions can adapt this LLM-powered ABM framework to improve local public health strategies, urban planning, and socio-economic simulations.
