EpiEvolve: Self-Evolving Agents for Streaming Pandemic Forecasting under Regime Shifts

This paper introduces EpiEvolve, a self-evolving agent that wraps a frozen LLM forecaster for streaming pandemic forecasting under regime shifts. EpiEvolve adapts through hierarchical episodic memory, regime-conditioned retrieval, and outcome-informed lesson distillation, without updating model parameters. On weekly COVID-19 hospitalization forecasting across five variant regimes, EpiEvolve reaches 0.629 average accuracy (vs. 0.561 for the static backbone and 0.325 for the CDC ensemble) and cuts recovery lag after regime shifts from 5 to 2 weeks.