Google’s ERA (Empirical Research Assistance) uses Gemini and tree-search to write and optimize scientific code, achieving expert-level performance across six benchmark domains per a new Nature paper.
Key Takeaways
Nature paper benchmarks ERA across genomics, public health, satellite imagery, neuroscience, time-series forecasting, and mathematics – expert-level results on all.
ERA uses tree search to explore thousands of code solutions, iteratively optimizing against a defined scientific goal given a problem statement.
Applied results include ERA forecasts ranking near the top of CDC leaderboards for flu, COVID-19, and RSV hospital admission predictions.
ERA-built models outperformed California’s official B120 seasonal runoff forecast and a commercial retail consensus estimate.
Computational Discovery, built with ERA and AlphaEvolve, is now rolling out via Gemini for Science at labs.google/science.
