PaperlayerView exact revision on arXiv

Source-linked AI summary

Integrating adaptive human behavior into epidemic models with large language models

Yicheng Mao, Haoyang Li, Rob Deardon, Hongru Du

arXiv:2608.29535v1physics.soc-phcs.AI

TL;DR

Epidemic models have limited direct evidence of adaptive, age-specific mixing, while mobility is an indirect signal. GABLE uses LLMs to generate age-structured contacts within a mechanistic model; in France, it improved short-term forecasts over reference models and produced distinct trajectories for alternative policy scenarios.

  • Problem

    Epidemic models lack timely, direct evidence of adaptive age-specific mixing: contact surveys are unavailable for many populations and retrospective, while mobility is indirect and does not uniquely determine contacts.

    Source: Introduction · p. 2

  • Method

    GABLE uses an LLM behavioral layer to translate epidemic and policy context into age-structured contact matrices that iteratively drive a mechanistic epidemic model.

    Source: Results · p. 4, GABLE framework · p. 14

  • Results

    GABLE generated epidemiologically informative temporal and age-specific mixing changes, improved short-term forecasts over mobility-driven and statistical references, and produced distinct trajectories under alternative future policies.

    Source: Introduction · p. 4

  • Takeaways & Limitations

    GABLE offers a decision-support interface for evaluating candidate interventions through the behavioral responses they are expected to induce.

    Source: Discussion · p. 13

    Counterfactual outcomes cannot be validated directly because alternative policy scenarios were not observed.

    Source: Validation under the factual policy… · p. 42

Abstract

from arXiv · show

Infectious disease transmission is shaped by patterns of human interaction, which adapt as epidemic conditions change. Capturing these context-dependent behaviors remains a fundamental challenge for epidemic models. Here, we recast this challenge by using large language models (LLMs) to represent adaptive human behavior within mechanistic epidemic models. We operationalize this idea through Generative Adaptive Behavioral Layer for Epidemics (GABLE), which adapts LLMs to infer behavioral responses to epidemic and policy conditions and translates them into age-structured contact matrices coupled to a mechanistic epidemic model. Applied to COVID-19 in France, GABLE reproduced responses in population mixing and age-specific contact structures that remained epidemiologically informative. In short-term forecasting, LLM-generated contact matrices outperformed mobility-driven matrices derived from real-world mobility data, with the largest gains at longer horizons. GABLE also extends beyond forecasting to prospective policy evaluation by projecting behavioral and epidemic responses to candidate interventions before implementation. When supplied with subsequently implemented policies, GABLE reproduced epidemic trajectories and generated distinct responses to alternative policy timing and composition. By leveraging LLMs as a flexible behavioral layer, GABLE provides a framework for coupling context-sensitive behavioral generation with epidemic dynamics.

Introduction

The introduction frames epidemics as co-evolving with adaptive human behavior, which observed-behavior data and predefined response rules limit the ability to represent prospectively. It presents GABLE as an LLM-based behavioral layer coupled to mechanistic transmission models for forecasting and policy scenarios.

Introduction

  • Contact surveys are scarce, resource-intensive, and retrospective, while mobility is an indirect signal that does not uniquely determine interpersonal contact or age-specific mixing.Both approaches rely on realized behavior, limiting anticipation beyond observed conditions or under interventions not yet implemented.Source: Introduction · p. 2
  • Existing endogenous approaches enable prospective simulation through predefined population response functions (14,15,16,17) or individual decision frameworks (18, 19, 20, 21, 22), but require behavioral mechanisms to be specified in advance.Real-world responses reflect interacting policy, risk, economic, caregiving, setting, fatigue, and compliance influences (23,24).Source: Introduction · p. 2
  • LLMs may jointly interpret heterogeneous policy, epidemic, demographic, and activity context, motivating their use to generate structured population responses for transmission models.Prior work reports LLM reproduction of population-level patterns and epidemic policy responses despite systematic biases (26, 27, 28, 29, 30,27,31,32,33,34,35,36,37,38,39).Source: Introduction · p. 2
  • GABLE translates policy, epidemic, demographic, and contact-diary context into age-structured contacts that govern transmission, with epidemic outcomes feeding back to update later behavior.The framework also supports historical reconstruction, rolling forecasts under fixed policies, and prospective comparisons of alternative policies.Source: Introduction · p. 3, Figure 1 · p. 3
  • In COVID-19 in France, GABLE generated epidemiologically informative temporal and age-specific mixing, improved short-term forecasts versus mobility-driven and statistical references, and produced distinct trajectories under alternative future policy scenarios.Source: Introduction · p. 4

Results

GABLE uses an iterative feedback loop in which LLMs generate age-structured contacts from epidemic and policy conditions, and those contacts drive epidemic projections. Results span adaptive contact reconstruction, hospitalization modeling, short-term forecasting, and prospective policy evaluation.

LLMs generate context-dependent and age-structured contact adaptation

  • During the first lockdown, daily contacts fell from 12.67 pre-pandemic to 3.15–3.74 across LLMs, close to 3.44 in mobility-driven matrices.Contact trajectories broadly agreed with mobility-driven matrices, most strongly for GPT-4o mini (r=0.85), followed by Gemini 2.5 Flash (r=0.74) and Grok 3 mini (r=0.60); contact intensity declined with policy stringency across LLMs (correlations −0.67 to −0.76).Source: LLMs generate context-dependent and… · pp. 4–5, Figure 2 · p. 6
  • During the second lockdown, adolescent contacts fell 62.1%–80.1% across LLMs versus 53.0% in mobility-driven matrices, while adult contacts fell 51.8%–70.3% versus 61.2%.Adolescents had the largest maximum reductions in every LLM configuration and seniors the smallest; the models also differed in age allocation and assortative mixing.Source: LLMs generate context-dependent and… · p. 5, Figure 2 · p. 6

LLM-generated behavioral structure informs age-specific epidemic dynamics

  • After calibration, all five contact configurations reproduced national hospitalization-wave timing and scale, with median cumulative totals 2.3%–3.6% above observed.No representation consistently outperformed the others on held-out age-specific admissions; GPT-4o mini and Grok 3 mini reduced the static matrix’s adult–senior imbalance, though mobility-driven matrices remained a strong retrospective benchmark.Source: LLM-generated behavioral structure… · pp. 6, 8, Figure 3 · p. 7

LLM-based behavioral adaptation improves prospective epidemic forecasts

  • GABLE achieved the lowest overall mean WIS at 1298, versus 2443 for the mobility-driven configuration and 3047–4148 for statistical references.At one week, GABLE’s WIS was 782 versus 781 for automatic ARIMA; at weeks two through four it outperformed every reference, with relative WIS of 0.46–0.54 versus the mobility-driven configuration. Its 95% prediction-interval coverage was 96.2%, versus 71.2%–82.7% for reference methods, and it had the lowest WIS in three of four age groups.Source: LLM-based behavioral adaptation… · pp. 8, 10, Figure 4 · p. 9
  • Forecasts spanning the second lockdown had the largest errors: GABLE’s projected-to-retrospective contact ratios were 1.51 and 1.38, and it beat the mobility-driven configuration at 11 of 13 origins.Errors clustered in forecasts that carried pre-lockdown policy conditions forward after restrictions tightened; under limited policy change, projected contact intensity differed from retrospective estimates by less than 3.0% through two weeks.Source: LLM-based behavioral adaptation… · pp. 8, 10, Figure 4 · p. 9

GABLE enables prospective evaluation of alternative policy strategies

  • Under the factual policy sequence, mean absolute relative error rose from 2.8% at one week to 21.7% at four weeks, with 96% coverage of 95% prediction intervals.At the two policy decision points, four-week cumulative admission biases were −9.7% and −2.9%.Source: GABLE enables prospective evaluation… · p. 10
  • At the lockdown branch, retaining late-October measures produced 41.4% more contacts and 45.8% more cumulative admissions than factual scenarios over four weeks.At the relaxation branch, maintaining pre-relaxation measures reduced contacts by 10.8% and admissions by 9.6%; maximum measures reduced contacts and admissions by 17.2% and 14.0% at the lockdown branch and by 27.0% and 22.7% at relaxation.Source: GABLE enables prospective evaluation… · p. 10, Figure 5 · p. 11
  • At relaxation, school closure and a one-week delay had the same mean stringency index of 74.5 but reduced contacts by 15.9% versus 2.9% and admissions by 11.4% versus 5.3%.School closure reduced adolescent contacts by 45.3% but senior contacts by 2.4%; across ten non-factual scenarios, contact-intensity and cumulative-admission changes were strongly associated (r=0.95).Source: Figure 5 · p. 11, GABLE enables prospective evaluation… · p. 12

Discussion

GABLE uses LLMs to translate changing epidemic and policy conditions into structured behavioral responses coupled to mechanistic epidemic dynamics. The framework supports prospective policy evaluation while highlighting uncertainty in behavioral representations and limits on validation and generalization.

Discussion

  • GABLE generates adaptive responses without contemporaneous behavioral measurements or separate response functions for contextual factors, and feeds behavior–transmission feedback into evolving epidemic dynamics.Epidemic conditions shape behavior, behavior modifies transmission, and the resulting trajectory informs subsequent adaptation.Source: Discussion · p. 12
  • Age-structured contact matrices translate contextual cues into transmission-relevant behavior; LLMs differed in age-specific structure, affecting hospitalization dynamics even after aggregate transmission intensity was calibrated.The models shared broad temporal adaptation patterns, while their distinct age-specific structures showed that contact patterns carry information beyond overall contact intensity.Source: Discussion · p. 12
  • With subsequently implemented French policies supplied and future hospitalization outcomes withheld, GABLE reproduced subsequent epidemic trajectories with useful prospective accuracy.This factual-policy evaluation provides an empirical reference for counterfactual analysis, but cannot directly validate effects of unobserved alternative policies.Source: Discussion · p. 13
  • GABLE could support comparison of intervention timing, targeted behaviors, and affected population groups, and could extend beyond COVID-19 contact behavior to other pathogens and responses.Potential extensions include modeling testing, vaccination, or healthcare seeking.Source: Discussion · p. 13
  • Contact matrices lacked direct validation against longitudinal realized-contact measurements, LLMs showed model-specific behavioral tendencies (40), and pretraining influence could not be excluded despite sensitivity tests.The tests found similar contact structure and epidemiological performance across prompt conditions, but evaluation in genuinely novel epidemic settings remains necessary.Source: Discussion · p. 13
  • Alternative-policy outcomes remain conditional projections, and generalizability beyond COVID-19 in France across pathogens, populations, and policy environments remains unestablished.Factual-policy validation supports GABLE’s capability but does not establish the accuracy of any individual counterfactual effect.Source: Discussion · p. 13

Materials and Methods

GABLE couples LLM-generated, context-sensitive contact matrices to an age-structured stochastic epidemic model. The methods describe its data, calibration, forecast and policy scenarios, and evaluation measures.

Data sources

  • The study used French pre-pandemic contact diaries, government policy records, epidemic surveillance, age-stratified admissions, population estimates, and vaccination records.Admissions summed across age groups were used for calibration; age-stratified admissions were withheld for evaluation. The four age classes were children 0–10, adolescents 11–18, adults 19–64, and seniors 65 years or older.Source: Data sources · p. 14

GABLE framework

  • GABLE updates contact matrices from epidemic and policy context, then feeds them into an age-structured stochastic transmission model.Behavioral contexts and matrices update weekly; epidemiological transitions are simulated daily with each weekly matrix held fixed. The layer evaluates baseline contact diaries, adjusts retention decisions for calendar and household constraints, and aggregates weighted contacts into age-structured matrices.Source: GABLE framework · pp. 14–15

Model configurations for comparative evaluation

  • Five configurations shared the same downstream epidemic model and differed in their contact matrices: static pre-pandemic, mobility-driven, and three LLM-generated variants.The LLM variants used GPT-4o mini, Gemini 2.5 Flash, or Grok 3 mini with the same baseline diaries and weekly context.Source: Model configurations for comparative… · p. 16

Model calibration and inference

  • Calibration estimated the per-contact transmission rate before lockdown, then fitted a time-varying common multiplier while preserving relative age mixing.A simulation-based Poisson objective used daily admissions summed across age groups; age-stratified admissions were excluded from fitting.Source: Model calibration and inference · p. 16

Real-time forecasting

  • Real-time forecasts covered four weeks from 13 rolling origins, with GPT-4o mini generating contacts and projected admissions recursively updating future epidemic context.Future policy was held at origin-week levels, and no post-origin admissions informed future behavioral context. The mobility-driven matrix was held at its origin-week value because later mobility observations were unavailable.Source: Real-time forecasting · pp. 16–17

Policy-conditioned validation and counterfactual projections

  • Policy-conditioned validation supplied subsequently implemented policies without future admissions; counterfactual projections compared six policy scenarios at pre-lockdown and pre-relaxation branch points.All prospective settings recursively updated epidemic context using projected admissions. Counterfactuals included the factual sequence, no change, one- or two-week delays, maximum containment, and school closure.Source: Policy-conditioned validation and… · pp. 17–18

Evaluation

  • Evaluation assessed contact intensity and structure, retrospective admissions and age-specific outcomes, forecast accuracy, prospective contact generation, and policy-conditioned performance.Forecasts used 13 origins and four horizons; metrics included weighted interval score, mean absolute error, mean absolute relative error, cumulative bias, and 95% prediction-interval coverage. Counterfactual comparisons summarized four-week contact and admission outcomes relative to factual-policy projections.Source: Evaluation · pp. 18–19

Supplementary Materials for Integrating adaptive human behavior into epidemic models with large

The supplementary PDF contains supporting methods, text, figures, and tables.

This PDF file includes:

Materials and Methods · Data sources

The data sources supplied contact diaries, weekly policy and epidemic inputs for behavioral generation, plus age-specific hospital admissions for model calibration and evaluation.

Data sources

  • 650 French contact-survey participants remained after weekday, non-holiday diaries with more than 20 professional contacts were excluded; each contributed one diary.The panel comprised 211 children, 76 adolescents, 211 adults and 152 seniors; reported household size set the household floor for contact-matrix construction.Source: Data sources · p. 27
  • Participants’ contact diaries recorded their age, day and holiday status, and each contact’s age and setting across six settings.The survey was accessed through the SOCRATES data tool; settings were home, work, school, transport, leisure and other.Source: Data sources · p. 27
  • Weekly French policy conditions came from seven Oxford tracker indicators, using each indicator’s modal level across the week.Indicators C1–C7 covered school and workplace closures, public events, gatherings, public transport, stay-at-home requirements and internal movement.Source: Data sources · p. 27
  • Retrospective epidemic prompts used weekly national cases and deaths, withholding unreliable early case counts and showing deaths alone when needed.Each prompt included the current and previous week and cumulative totals.Source: Data sources · p. 27
  • Age-specific daily hospital admissions from SI-VIC were delay-corrected, imputed for missing days, and used to calibrate and score the transmission model.Admissions summed across age groups formed the calibration target.Source: Data sources · p. 27
  • Hospital admissions were withheld from the LLM during retrospective contact reconstruction but informed prospective analyses before each target week.For prospective forecasting, policy-conditioned and counterfactual analyses, the model extended the epidemic input using its own preceding projections.Source: Data sources · p. 27

LLM prompts and queries

GABLE prompts an LLM with a participant’s age and pre-epidemic contact diary alongside weekly policy and epidemic context. It returns contact-specific decisions that are converted into a binary retention vector, with prompt inputs adapted to retrospective, forecasting, and policy-conditioned settings.

LLM prompts and queries

  • A shared system message combines a fixed task instruction with the week’s policies, policy duration, epidemic state, and school-calendar status.The task describes the disease in general terms and defines six contact settings; policy duration conveys how long the current measures have remained unchanged, while holidays are distinguished from policy-driven school closures.Source: LLM prompts and queries · pp. 28–29
  • For each participant-week, the LLM evaluates the full diary in one query and decides which contacts remain, weighing individual circumstances and restrictions rather than assuming perfect compliance.The prompt asks it to consider both reasons to give up contacts, such as restrictions or epidemic severity, and reasons to keep them, such as essential needs or fatigue with prolonged measures.Source: LLM prompts and queries · pp. 29–30
  • The user message supplies the participant’s age and numbered diary contacts, and the model returns a JSON list that defines the binary contact-retention vector.Queries used JSON mode at temperature 0.3; out-of-range contact numbers were discarded and replies were cached for matrix rebuilding without new queries.Source: LLM prompts and queries · p. 29
  • Prompt inputs vary by setting: retrospective prompts use observed policies and same-week cases and deaths, while forecasts freeze origin policies and use admissions through the previous week.Policy-conditioned prompts use target-week policies and the forecasting epidemic block; counterfactual prompts render scenario policies and recompute their duration counter.Source: LLM prompts and queries · pp. 29, 32

Contact-matrix construction

GABLE reconstructs baseline and weekly LLM contact matrices with the same procedure, changing only contact weights. The resulting matrices incorporate reciprocity correction and behavioral adjustments, and support age-specific and aggregate contact-intensity measures.

Contact-matrix construction

  • The same procedure converts weighted diary contacts into a 4 × 4 contact-rate matrix for both the pre-pandemic baseline and weekly LLM matrices.Only the weight assigned to each reported contact changes between matrix types.Source: Contact-matrix construction · p. 32
  • Contacts are assigned to age classes from exact ages when available, otherwise from reported age-range midpoints, then averaged by participant and contact age.Each reported contact contributes according to its weight; rows index participant age and columns contact age.Source: Contact-matrix construction · p. 32
  • Population-size-based reciprocity correction addresses asymmetric reported contacts caused by different sampling rates across age classes.The corrected matrix retains participant age in rows and contact age in columns; C_ij is the contact rate from participant group i to group j.Source: Contact-matrix construction · pp. 32–33
  • For heatmaps and structural comparisons, the display matrix transposes the transmission matrix so rows index contact age and columns participant age.Source: Contact-matrix construction · p. 33
  • Baseline weights are 1 for every diary contact, while weekly LLM weights begin with the binary retention vector and are then deterministically adjusted.The baseline is defined by the survey and the reconstruction equations.Source: Contact-matrix construction · p. 33
  • School-only contacts are zeroed during academic holidays, and a household-contact floor preserves co-resident mixing even under strict stay-at-home orders.The floor is based on the average household size in France, 2.18 (53), and is skipped when a participant reports no household contact.Source: Contact-matrix construction · p. 33
  • Age-specific contact intensity is the mean daily contact count after reciprocity correction, and aggregate intensity is its population-weighted average.Source: Contact-matrix construction · pp. 33–34

Transmission model

The transmission model couples behavioral-layer contact matrices to age-, strain-, and vaccination-stratified epidemic dynamics. It represents disease progression and intervention-related behavior changes while using daily hospital admissions as its calibration target.

Transmission model

  • The transmission engine is an age-stratified stochastic two-strain model based on (48, 12), with vaccination strata and daily hospital admissions as its calibration target.It tracks susceptible, latent, infectious, waiting, hospitalized, and recovered states; simulations start with ten presymptomatic adults on 16 January 2020 and use one-day τ-leaping, with ensemble trajectories summarized by their median and 2.5th and 97.5th percentiles.Source: Transmission model · pp. 34–35
  • Behavioral-layer contact matrices determine transmission, with age- and strain-specific susceptibility and infectiousness and a fitted factor scaling overall contact levels.Epidemiological inputs and model structure remain fixed across matrix configurations; vaccination reduces susceptibility and the onward transmission contribution of vaccinated infectious cases.Source: Transmission model · p. 34
  • Severe symptoms reduce contacts by 75 percent, while testing reduces contacts by 90 percent for 50 percent of infectious individuals after the first lockdown exit.The testing-identified fraction is zero through the first wave and first lockdown, and presymptomatic cases retain their contacts.Source: Transmission model · p. 35
  • Disease progression uses age- and strain-specific symptom probabilities, while Alpha and vaccination alter transmission, hospitalization risk, and progression to symptomatic or severe disease.Alpha is introduced in winter 2020–2021 at the prevalence observed in variant surveillance (54); vaccine effects on disease given infection are applied separately from protection against infection to avoid double counting.Source: Transmission model · p. 35

Model inference

The model was calibrated with a shared transmission rate and window-specific correction factors using a simulation-based Poisson objective. Retrospective and forecasting analyses used defined fitting windows to align calibration and evaluation periods.

Transmission rate

  • The fitted per-contact transmission rate was β = 0.098, shared across retrospective configurations using the pre-pandemic contact matrix before lockdown.It was selected by maximizing the simulation-based Poisson objective over β values from 0.060 to 0.120 in steps of 0.002, with 50 stochastic runs per grid point.Source: Model inference · p. 35

Correcting factors

  • With β fixed, one correction factor α was fitted separately for each window, beginning with the lockdown-matrix window.The search used golden-section optimization on [0.15, 3.0] to width 0.3, followed by quadratic refinement on a grid of width 0.2 in steps of 0.02; boundary optima triggered interval widening. The coarse stage used 10 stochastic runs and refinement used 50.Source: Model inference · p. 36

Simulation-based calibration objective

  • Candidate parameters were compared using a simulation-based Poisson objective on ensemble-median daily hospital admissions summed across age groups.Repeated simulations produced the daily ensemble median to reduce the influence of highly variable stochastic runs; the objective guided parameter selection and was not treated as a full probabilistic likelihood.Source: Model inference · p. 36

Fitting windows

  • Retrospective calibration used a shared partition: the lockdown window ran from 16 March to 10 May 2020, followed by fifteen consecutive windows through 4 July 2021.The windows tracked lockdown exits, holidays, and later lockdown periods; their boundaries matched the published mobility-driven matrix results, avoiding differences due to window choice. Forecasting instead used each week from the week beginning 11 May 2020 through the forecast origin as its own window.Source: Model inference · p. 36

Real-time forecasting

Real-time forecasts used rolling origins across the autumn 2020 second wave, projecting four weeks ahead while restricting generated contact matrices and epidemic inputs to information available at each origin. The evaluation compared LLM-generated and mobility-driven matrices under real-time constraints and summarized probabilistic accuracy with the weighted interval score.

Real-time forecasting

  • Each rolling origin from 14 September to 7 December 2020 produced a four-week forecast from 100 stochastic runs with daily age-resolved quantile trajectories.The origins covered the autumn second wave.Source: Real-time forecasting · p. 36
  • For each future week, the LLM regenerated contact matrices without post-origin observations, using a frozen policy, the published school calendar, and epidemic inputs lagged by one week.Reported admissions were used through the origin; within the horizon, the model’s own median predictions supplied the epidemic block, and no week received its own admissions.Source: Real-time forecasting · pp. 36–37
  • Correcting factors were fitted through each origin and frozen forward at the mean of the last K=4 weekly factors; the lockdown period retained its 16 March to 10 May 2020 factor.Weeks up to the origin kept their own fitted weekly values. For comparison, mobility-driven matrices used published values through the origin and were frozen thereafter.Source: Real-time forecasting · p. 37
  • Probabilistic forecast accuracy was summarized by the weighted interval score (WIS), using the predictive median and K=11 central prediction intervals at the standard set of 23 quantile levels.The nominal interval levels ranged from 0.02 to 0.90.Source: Real-time forecasting · p. 37

Statistical forecast baselines

The section defines three non-epidemiological statistical forecasting references and specifies their real-time fitting and scoring setup.

Statistical forecast baselines

  • The references are automatic ARIMA (49), automatic exponential smoothing (50), and Theta (51), run through statsforecast (55) in a non-seasonal setting.They provide relative-score references without epidemiological structure.Source: Statistical forecast baselines · p. 37
  • Each model uses weekly admissions from 11 May 2020 to each forecast origin to produce four-week forecasts on the log(1 + x) scale.For age-specific evaluation, each model is fitted independently to each age group. Central prediction intervals are converted to 23 quantile levels, clipped to be non-negative and non-decreasing; the relative weighted interval score reference set also includes the mobility-driven matrix configuration.Source: Statistical forecast baselines · p. 37

Counterfactual policy experiments

The section describes four-week counterfactual projections from selected policy branch points, comparing alternative policy sequences and their behavioral and epidemic responses.

Analytical design

  • Each counterfactual projected four weeks from a fixed origin, sharing pre-branch conditions and then recursively using its own prior admission projections without post-branch observed admissions.Scenarios shared the epidemic state, contact matrices, transmission parameters, vaccination and variant inputs, frozen correcting factor, random seeds, and observed school calendar; their initial differences were policy sequences and duration counters.Source: Counterfactual policy experiments · pp. 38–39

Counterfactual outcome summaries

  • Policy intensity used a nine-indicator Oxford stringency index; scenario responses were summarized through four-week contacts and admissions, with behavioral–epidemic association tested across non-factual scenarios.C1–C7 followed scenario levels while C8 and H1 retained observed values; the association used Pearson correlation between percentage changes in aggregate contact intensity and cumulative admissions.Source: Counterfactual policy experiments · pp. 38–40

Selection of branch points

  • The lockdown branch was the week beginning 26 October 2020, and the relaxation branch the week beginning 23 November 2020, selected for the largest positive and most negative estimated policy-change effects.Positive effects indicate that the factual policy change reduced projected admissions relative to holding origin-week policies fixed; negative effects indicate an increase.Source: Counterfactual policy experiments · pp. 38–39

Scenario specification

  • Six scenarios covered factual policies, no change, one- and two-week delays, maximum containment measures, and school closure.For school closure, the school-closing indicator was set to its maximum while the other six containment indicators followed the factual sequence.Source: Counterfactual policy experiments · p. 39

Supplementary Text

The supplementary analyses compare age-structured contact patterns across LLMs, evaluate projections using the factual policy sequence, and test how prompt disclosure and historical identity cues affect generated contacts and epidemiological performance.

Structural comparison of LLM and mobility-driven contact matrices

  • GPT-4o mini had adult–senior slope b=0.72 (r=0.96) and median intensity ratios of 0.75 for adults and 0.67 for seniors, while child/adolescent slope was b=0.99 (r=0.90).Its age-specific ratios were 0.92 for children and 1.01 for adolescents, indicating that lower contact intensity was concentrated among adults and seniors.Source: Structural comparison of LLM and… · p. 41
  • Gemini 2.5 Flash had adult–senior slope b=1.14 (r=0.94) and young-related slope b=0.78 (r=0.87), with median intensity ratios of 1.21 for adults and 0.89 for seniors.The reduction among younger groups was concentrated among adolescents: their ratio was 0.84, compared with 0.96 among children.Source: Structural comparison of LLM and… · p. 41
  • Grok 3 mini had lower contact intensity across all ages, with adult–senior slope b=0.86 (r=0.90), young-related slope b=0.75 (r=0.82), and age-specific ratios below one.Ratios were 0.82 for children, 0.79 for adolescents, 0.89 for adults, and 0.71 for seniors.Source: Structural comparison of LLM and… · p. 41
  • Within-group fractions were higher for GPT-4o mini among children and adolescents, lower for Gemini 2.5 Flash and Grok 3 mini in those groups, and more similar across models among adults and seniors.All time-varying configurations retained comparatively strong adult assortativity and weaker senior assortativity; the reported structural differences contextualize age-specific hospitalization biases.Source: Structural comparison of LLM and… · pp. 41–42

Validation under the factual policy scenario

  • Factual-policy projections reproduced the autumn 2020 wave’s principal turning points, while mean absolute relative error rose from 2.8% at one week to 21.7% at four weeks and mean 95% interval coverage was 96%.Signed errors reached a mean of −12.2% at four weeks, indicating that the framework tended to overstate the policy-driven reduction in admissions; alternative policy outcomes themselves cannot be directly validated.Source: Validation under the factual policy… · p. 42

Sensitivity to historical information leakage

  • To probe historical-information leakage, the study compared its masked prompt with deeper masking of geography and numeric epidemic quantities and full disclosure of COVID-19, SARS-CoV-2, and target dates.The behavioral task and downstream pipeline were held fixed across conditions, and all analyses used GPT-4o mini.Source: Sensitivity to historical information… · p. 43
  • Removing country identity and numerical epidemic counts did not materially alter contact structure or epidemiological performance, while full disclosure did not systematically improve on the masked baseline.Together, the perturbations indicate that the main behavioral and epidemiological findings were insensitive to substantial changes in prompt historical identifiability.Source: Sensitivity to historical information… · p. 49
  • GPT-4o mini was evaluated on weekly aggregate daily contacts per person under masked-baseline, deeper-masking, and full-disclosure information conditions, with three national lockdowns shown as shaded intervals.Source: Figure S4 · p. 53
  • Figure S4 presents age-structured contact matrices for the week beginning 9 November 2020 under baseline, deeper masking, and full disclosure information conditions.Source: Figure S4 · p. 53
  • Retrospective cumulative hospitalization bias was +2.9% under baseline, +2.4% under deeper masking, and +2.2% under full disclosure; fitted correcting-factor medians were close to 1.5 in all three conditions.The distributions of fitted correcting factors were highly similar across the three conditions.Source: Sensitivity to historical information… · p. 48, Figure S4 · p. 53
  • Rolling four-week mean weighted interval scores were 1298 under baseline, 1380 under deeper masking, and 1241 under full disclosure, with WIS ratios of 1.06 and 0.96 relative to baseline.95% prediction-interval coverage was 96.2% under baseline and deeper masking and 94.2% under full disclosure.Source: Sensitivity to historical information… · p. 49, Figure S4 · p. 53

Related papers

This paper cites

Loading 2608.29535v1…