Overview
This zhichai.net forum post provides a Feynman-style deep-dive commentary (in Chinese) on a new paper by researchers from Google DeepMind and Harvard University: *Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search* (Martinson, Brenner, Plomecka, Williams, Reich, & Shamsi; arXiv:2605.16238).
The Problem
Infectious disease forecasting is hard because it couples virus evolution, human behavior, healthcare capacity, seasonality, and immune dynamics across multiple timescales. The CDC's forecasting hub aggregates models from dozens of human expert teams, each requiring weeks to months of labor from epidemiologists, data scientists, software engineers, and statisticians. This hand-crafted approach does not scale to novel pathogens, county-level granularity, or compressed time windows.
The Method: LLM + Tree Search
The core insight is that scientific forecasting model development can be framed as a code optimization problem. The system combines:
- LLM: proposes code improvements based on epidemiological theory, statistical methods, or novel ideas
- Tree search: systematically explores the code space, evaluates candidates (e.g., via weighted interval score), and retains only performance-improving variants
- Labor bottleneck broken: weeks of expert modeling compressed to hours or days; new pathogens could get initial models within days
- Transparency: generated code is fully readable and verifiable, unlike black-box neural networks
- Generality: the same LLM-guided tree search approach has produced breakthroughs in single-cell RNA-seq analysis (40 new methods), geospatial modeling (U-Net + Transformer hybrids), and zebrafish neural activity prediction
- Martinson, S., et al. (2026). Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search. arXiv:2605.16238
- Aygün, E., et al. (2025). An AI system to help scientists write expert-level empirical software. arXiv:2509.06503
- CDC FluSight Forecasting: https://www.cdc.gov/flu/weekly/flusight/
The result is an iterative evolution loop: the LLM suggests changes, the system generates and evaluates code, and feedback guides the next proposal.
Prospective Evaluation
Unlike retrospective evaluations that leak future information, this work was fully prospective: during the 2025-2026 US respiratory season, the system ran blind to future data, submitted real-time forecasts for influenza, COVID-19, and RSV, and was scored against actual outcomes. Its autonomously discovered, methodologically diverse ensembles matched or exceeded the CDC human-curated hub ensemble out-of-sample, including in RSV's data-scarce cold-start setting.
Guarding Against Reward Hacking
The authors identified two safeguards via controlled retrospective ablations:
1. Log-scale distance metrics — forcing improvements to be meaningful in orders of magnitude, not trivial metric exploits 2. Automated judge-in-the-loop — checking that generated code remains structurally consistent with scientific theory (e.g., SIR model assumptions), like automated peer review
Why It Matters
Limitations
The paper acknowledges dependence on input data quality, predefined structural constraints for the judge, substantial compute (300-2000 tree nodes per search), and continued need for human experts in initial configuration and objective definition.