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

Researchers introduce EpiEvolve, a self-evolving AI agent that improves pandemic forecasting by adapting to changing disease patterns in real-time streaming scenarios. The system achieves 12% higher accuracy than static models and reduces recovery time after major shifts from 5 weeks to 2 weeks by leveraging episodic memory and strategic rule learning.

EpiEvolve addresses a fundamental gap between how pandemic forecasting models are typically developed and how they operate in practice. Traditional machine learning approaches train models offline and deploy them as static systems, but real-world epidemiology involves continuous data streams, delayed information arrival, and shifting disease dynamics as new variants emerge. This research demonstrates that adaptive architectures can significantly outperform conventional approaches by treating forecasting as an evolving problem rather than a fixed one.

The underlying innovation combines several complementary mechanisms: episodic memory stores historical predictions and outcomes, reflection on delayed labels identifies recurring error patterns, and regime-aware retrieval finds analogous past situations to inform current decisions. By keeping the core LLM weights frozen while building context dynamically, EpiEvolve maintains stability while enabling rapid adaptation. The 12.2% accuracy improvement over the static baseline and substantial reduction in post-shift recovery lag suggest this approach has practical merit for operational settings where response speed and accuracy directly impact public health outcomes.

The broader implications extend beyond epidemiology to any forecasting domain requiring adaptation under distribution shift. Healthcare systems, supply chain networks, and financial institutions face similar challenges of streaming data with delayed feedback and regime changes. EpiEvolve's architectural pattern—combining fixed pretrained models with dynamic episodic reasoning—offers a generalizable framework for building more robust forecasting systems. The comparison to the CDC ensemble baseline provides important grounding in real-world performance standards, suggesting the approach could have immediate deployment value for organizations managing pandemic response.

• →EpiEvolve achieves 62.9% accuracy versus 56.1% for static models and 32.5% for CDC ensemble through self-adaptation mechanisms
• →Hierarchical episodic memory combined with reflection on delayed labels enables the system to extract and apply insights from past prediction failures
• →Recovery lag after disease regime shifts decreases from 5 weeks to 2 weeks, indicating substantially faster adaptation to new variants
• →The frozen-weight architecture maintains model stability while allowing context-based adaptation, providing a reusable design pattern for streaming forecasting
• →Regime-aware retrieval mechanisms identify relevant historical cases, demonstrating that past pandemic dynamics can inform future predictions across different variants