Epidemics as Dialogue: How Language Models Teach Epidemiologists to Listen to Human Behavior ✛
Context
Epidemics are coupled behavioral-biological systems: human behavior determines where and between whom contacts occur, while infection risk and restrictions in turn change that behavior. Yet behavior remains one of the least observed components of the epidemic system. Contact surveys capture population mixing but are conducted rarely and with delay; mobility data provide an indirect signal but do not reflect the age structure of contacts. Existing models either rely on already realized behavior or preset response functions to risk, limiting their ability to anticipate reactions to novel conditions. This work sits alongside studies that use large language models to reproduce human judgments and behavior, as well as attempts to link genes and diseases or predict the health of entire countries, but differs by embedding the language model directly into the dynamic feedback loop of the epidemic.
Methods
The authors proposed a layer called GABLE (Generative Adaptive Behavioral Layer for Epidemics), which at each simulation step feeds the language model current policy conditions, epidemic indicators, and pre-pandemic contact diaries. The language model decides which recorded contacts a person would keep under those conditions, and these decisions are aggregated into an age-specific contact matrix. The matrix drives a stochastic age-structured model of SARS-CoV-2 transmission that simulates hospitalization dynamics. The updated epidemic state then returns to the behavioral layer, closing the feedback loop. Three language models were used to generate behavior: GPT-4o mini, Gemini 2.5 Flash, and Grok 3 mini. The approach was evaluated on COVID-19 data from France from March 2020 to July 2021, comparing against mobility-based matrices and a static pre-pandemic matrix.
Results
All three language models reproduced the overall temporal dynamics of contacts: during the first lockdown, the average number of contacts dropped from 12.67 to 3.15–3.74 per person per day, close to the 3.44 value for mobility-based matrices. Correlation with mobility ranged from 0.60 to 0.85. Age structure differed across models: for example, during the second lockdown, Gemini 2.5 Flash preserved more adult contacts but fewer adolescent contacts than mobility. In retrospective reconstruction, all configurations reproduced the cumulative hospitalization waves with an error of 2.3–3.6%, but language models reduced the imbalance between adults and elderly compared to the static matrix. In short-term forecasting over 4 weeks, GABLE with GPT-4o mini achieved a mean weighted interval score of 1298 versus 2443 for mobility and 3047–4148 for statistical methods, with 95% interval coverage of 96.2%. When evaluating alternative policies, a two-week delay in lockdown increased contacts by 46% and hospitalizations by 45.8%, while maximal measures reduced them by 17.2% and 14.0%, respectively.
Implications
This work differs from neighboring studies in that it does not merely use language models to reproduce surveys or predict treatment effects, but embeds them in a closed dynamic system where behavior and epidemic influence each other. This approach removes the need to preset response functions for each factor and allows evaluating the consequences of measures not yet introduced. The resulting age-specific contact matrices carry epidemiologically meaningful information even after calibrating overall transmission intensity, confirming the importance of mixing structure rather than just the total number of contacts.
Future development
The method can be extended to other pathogens and behavior types, such as testing, vaccination, or healthcare seeking. It is promising to explore ensembles of language models and explicitly account for uncertainty in behavioral responses. An important task remains validation on truly novel epidemics not represented in the training data of language models, to rule out leakage of historical information.
Impact
The approach could influence epidemiological modeling, public health decision support, and computational social science, where modeling adaptive behavior in response to changing conditions is needed.
Next steps
Next steps include validation in other countries and pathogens, studying differences between language models, and developing methods to handle genuinely novel epidemic situations.
Key open problems
This work directly addresses a fundamental problem in epidemiology: how to predict human behavior in conditions that have not yet occurred. Traditional models either ignore adaptation or prescribe it with rigid rules, limiting their predictive power. This approach offers a flexible mechanism that learns from context, but the question remains how well it generalizes beyond historical data.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors combined large language models with a mechanistic model of the COVID-19 epidemic in France. Based on descriptions of policies and the epidemiological situation, the language model generated age-specific contact matrices that drove infection transmission. This approach reproduced real waves of hospitalizations, improved short-term forecasts, and allowed assessing the consequences of alternative policy scenarios before they were implemented.
Where its strength is
The work embeds the language model directly into the dynamic feedback loop of the epidemic, enabling the simulation of adaptive behavior without pre-specified risk-response functions.
Where it could go next
- It might be interesting to compare GABLE's behavioral layer with game-theoretic models where individuals make decisions based on personal cost-benefit calculations, which would allow comparing aggregated LLM responses with rational strategies.based on: Epidemics as a Game: How Personal Calculations Reshape the Wave of Infections arXiv:2604.04305
- Closely related is work on a digital twin of an entire country's health: it uses big data to predict individual trajectories. Combining such a twin with a behavioral layer could yield a more detailed picture of the link between behavior and outcomes.based on: A Digital Twin for an Entire Country's Health arXiv:2608.16273v1
Worth trying
- Compare the contact matrices generated by the LLM with results from contact surveys, if available for the same period.
- Test how the choice of language model affects the age structure of contacts and the resulting epidemic trajectories.
- Pilot the approach in other countries or for other epidemics to assess its transferability.
Why this may matter
This approach could reduce reliance on scarce and delayed behavioral data, opening the way to more flexible and timely epidemic forecasts.
Nearby works in our archive
- A Digital Twin for an Entire Country's Healthsimilarity 0.62
- Why Not Every Contact Leads to Illnesssimilarity 0.57
- A Virus Hits Like a Punch: Why Outbreaks Don't Follow the Averagesimilarity 0.57
- Epidemics Breathe in Tandemsimilarity 0.54
- Epidemics as a Game: How Personal Calculations Reshape the Wave of Infectionssimilarity 0.53
The closest work in the archive sits at 0.62 by meaning — the usual level for our corpus.
During the first lockdown in France, the average number of daily contacts dropped roughly fourfold: from 12.67 to 3.15–3.74. That is like each person, instead of a full bus, talking only to three fellow passengers.
The idea of a model that predicts crowd behavior in response to changing conditions evokes the psychohistory from Isaac Asimov's Foundation, where mathematical models predicted the fates of civilizations. Here, though, the scale is humbler: not a galaxy, but one country and one epidemic.
Key numbers
- drop in contacts during first lockdown: from 12.67 to 3.15–3.74 per person per day
- mean weighted interval score for GABLE: 1298 versus 2443 for mobility
- 95% forecast interval coverage: 96.2%
- increase in hospitalizations with a 2-week lockdown delay: 45.8%
- reduction in contacts under maximal measures: 17.2%