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

2021-05-31 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

EpiEvolve is a self-evolving agent designed for streaming pandemic forecasting, specifically addressing the challenges of real-world operational environments with delayed labels and disease regime shifts. It wraps a pre-trained LLM forecaster, such as Qwen3-14B-Base, keeping its weights fixed while adapting through a hierarchical episodic memory. This memory stores past forecast outcomes, reflects on delayed ground truth, retrieves cases relevant to the current epidemiological regime, and distills recurring errors into strategic rules. Evaluated on weekly COVID-19 hospitalization trend forecasting across five variant regimes, EpiEvolve achieved 0.629 average accuracy, significantly surpassing its static backbone (0.561) and the external CDC ensemble (0.325). Crucially, it reduced recovery lag after regime shifts from 5 weeks to 2 weeks, demonstrating robust adaptation without costly gradient updates.

Key takeaway

For Machine Learning Engineers deploying LLM-based forecasters in dynamic environments, you should prioritize memory-based adaptation over continuous parameter fine-tuning. EpiEvolve demonstrates that a frozen LLM backbone, augmented with hierarchical episodic memory and strategic rule distillation, significantly improves accuracy and reduces recovery lag after regime shifts. This approach offers a robust and cost-effective alternative to retraining, ensuring your models remain performant as data distributions evolve.

Key insights

LLM forecasters can adapt to streaming data and regime shifts via memory-based self-evolution, not parameter updates.

Principles

• Memory-based adaptation outperforms static models in streaming contexts.
• Hierarchical episodic memory enhances regime-aware retrieval.
• Distilling errors into rules improves post-shift recovery.

Method

EpiEvolve uses a frozen LLM backbone, updates hierarchical episodic memory with reflections on delayed labels, retrieves regime-relevant cases, and distills errors into strategic rules for context.

In practice

• Implement hierarchical memory for LLM agents in streaming tasks.
• Use reflection and rule distillation to adapt to concept drift.
• Prioritize memory updates over parameter fine-tuning for deployment.

Topics

• Pandemic Forecasting
• LLM Agents
• Concept Drift Adaptation
• Episodic Memory
• Streaming Machine Learning
• COVID-19

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• EpiEvolve: Self-Evolving Agents for Streaming Pandemic Forecasting under Regime Shifts — EpiEvolve improves streaming pandemic forecasting by enabling LLMs to self-evolve through episodic memory, reflection, and strategic rule distillation.
• EVIL: Evolving Interpretable Algorithms for Zero-Shot Inference on Event Sequences and Time Series with LLMs — LLM-guided evolution can discover compact, interpretable algorithms for zero-shot dynamical systems inference, outperforming complex neural networks.
• Harnessing Generalist Agents for Contextualized Time Series — TimeClaw's agentic harness framework enables LLMs to perform robust, contextualized time series reasoning by providing native temporal runtime support.
• Learning, Fast and Slow: Towards LLMs That Adapt Continually [R] — Combining slow parameter updates with fast context optimization enhances LLM adaptation and mitigates forgetting.
• Introduction to quantitative finance Part 26: Whether to use forecasting methods, or to tell an LLM… — Forecasting in quantitative finance faces challenges with LLMs, necessitating memory-enhanced reward shaping for consistency.
• The Epi-LLM Framework: probing LLM behavioral priors through epidemiological agent-based models — The Epi-LLM framework uses LLMs in agent-based models to simulate epidemic behavior, showing LLMs can reduce infection peaks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.