EpiWorld grounds LLM policy agents in an epidemiological world model, cutting cumulative hospitalisation by up to 59% on retrospective COVID-19 and Influenza data

EpiWorld is a closed-loop framework that grounds an LLM policy actor in a learned action-conditioned epidemiological world model and a tiered skill library of public-health protocols, surveillance tools, and adaptive lessons accumulated through after-action analysis. Relative to a naive LLM that lacks epidemic dynamics, quantitative surveillance signals, and institutional constraints, the framework anchors policy reasoning in a world model that projects intervention consequences and in fixed protocol constraints. The abstract describes the framework components and closed-loop mechanism; no component ablations or implementation details are given.

Given a candidate intervention, the world model predicts regional epidemic evolution and enables fast counterfactual rollouts that provide feedback for policy selection and refinement; outcomes of simulated futures are distilled into reusable lessons while protocol constraints remain fixed. Combining counterfactual rollouts with lesson distillation lets the decision process improve without sacrificing interpretability or controllability. The abstract states the mechanism and design goal; no quantitative rollout speed or distillation effect is reported.

On retrospective COVID-19 and Influenza datasets, the world model achieves the best out-of-distribution Peak-MAE among all forecasting baselines. It performs best on the out-of-distribution peak-error metric relative to existing forecasting baselines. The abstract reports this comparison; specific error values, the baseline list, and data splits are not given.

The closed-loop framework reduces cumulative hospitalisation by up to 59% across datasets and by an average of about 16% across six LLM backbones, outperforming reinforcement-learning and optimal-control policy baselines. Relative to reinforcement-learning and optimal-control baselines, the framework attains lower cumulative hospitalisation on this outcome metric. The abstract reports the comparison across datasets and six LLM backbones; confidence intervals, statistical tests, and per-backbone values are not given.

The work targets public-health decision settings where intervention policies must be interpreted, justified, and revised in natural language, and it applies to retrospective COVID-19 and Influenza respiratory-disease data with regional epidemic-evolution prediction and surveillance signals. It lets an LLM policy agent select and refine interventions under fixed protocol constraints via counterfactual rollouts, distilling simulated futures into reusable lessons so the decision process improves while remaining interpretable and controllable. Beneficiaries include public-health policy researchers, epidemic modelling teams, and institutions that need an auditable policy-simulation workflow.

The abstract does not specify the world-model architecture, training data splits, baseline list, or statistical tests, nor does it give per-backbone and per-dataset values; which datasets and backbones correspond to the 59% and about 16% hospitalisation reductions, and whether uncertainty was quantified, remain to be confirmed in the full text. Evaluation is retrospective, so how surveillance delays, policy-execution deviations, and protocol updates affect closed-loop performance in prospective deployment is an open question worth watching.