aclanthology.orgSource publication:
Loading content…
Epidemic interventions are textual policies, yet a plain LLM lacks the epidemic dynamics to project their consequences, the surveillance signals to judge severity, and the institutional constraints on admissible actions. EpiWorld grounds an LLM policy agent in an action-conditioned epidemiological world model, screens candidate interventions through counterfactual rollouts, and distils simulated outcomes into reusable lessons. On retrospective COVID-19 and influenza data, the full configuration cuts 8-week cumulative hospitalisations by 59.3% versus the reference historical policy on real US Omicron, and by about 16% on average across six LLM backbones.
An epidemic policy agent can now trial an intervention inside learned epidemic dynamics before committing to it. Statistical and neural forecasters estimated what was likely to happen, causal-inference methods estimated retrospective intervention effects, and reinforcement learning or optimal control depended on predefined dynamics or simulation environments. Over a 12-week autoregressive rollout, EpiWorld beats all baselines on out-of-distribution Peak-MAE across the three datasets, by roughly 6%, 4% and 8% on COVID-19 SEIR, Influenza SEIR and real US Omicron against the closest competitor, EARTH + Policy.
The system couples a regional world model that predicts epidemic evolution with a language-model policy actor that acts under public-health protocols. The world model combines delay-aware temporal encoding, graph-structured regional coupling, residual next-state prediction and policy-conditioned latent modulation, so an intervention reshapes every latent channel rather than entering as an ordinary covariate. The actor draws on three skill tiers, with protocol constraints fixing the admissible action envelope, surveillance skills returning quantitative summaries such as Rt and severity ratios, and adaptive lessons adjusting surveillance parameters; on the COVID-19 SEIR state-OOD split the Peak-MAE advantage over EARTH + Policy holds at paired Wilcoxon p = 0.039 across 50 held-out replicates.
The closed-loop gain comes mainly from the language model choosing when and where to spend a comparable intervention budget, with after-action lessons adding a smaller increment. Replacing the actor with a fixed severity-proportional rule that calls no language model recovers only 59% of the base LLM's reduction at nearly matched mean intensity (0.485 against about 0.50); the base LLM already accounts for most of the gap, and skill evolution adds a further 2.8 points. Re-executing every method in the mechanistic simulator that generated the COVID-19 SEIR state-OOD corpus preserves the ordering exactly, each landing within 0.97 to 1.16 times its in-model value; the margin over the strongest baseline, EpidRLearn, is 73% on independent ground truth versus 58% in-model.
The evaluation covers retrospective COVID-19 and influenza data: weekly state-level surveillance for the real US, plus COVID-19 and influenza SEIR corpora calibrated to real mobility, with policy rollouts over an 8-week horizon. The system is positioned as a decision-support layer whose proposals and counterfactual rollouts assist public-health authorities, while final decisions and accountability remain with human operators.
Adaptive lessons are calibrated to the infection regime they are distilled from and do not transfer across regimes: a bank mined under Omicron, applied to a first wave in a fully susceptible population (R0 about 1.7 to 2.9), degrades outcomes by 12.1%. Transfer across world models is also not guaranteed: a bank mined under the real-US world model, applied to the synthetic SEIR world model on five held-out OOD states, degrades pooled performance by 12.1%. On states whose baseline error is already low (FL, NJ, CA) the multiplicative correction over-applies and per-state MAE regresses. The five-member ensemble ranks forecast reliability well (Spearman ρ = 0.83) but is under-dispersed, with nominal 95% intervals covering about 68% of outcomes, so it is not a calibrated interval.