Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery

Details

Google Research published a paper in Nature today describing ERA (Empirical Research Assistance), an AI system built on Gemini that writes and optimizes scientific code. Given a problem and a success metric, ERA searches scientific literature, writes code, explores thousands of solution paths via tree search, and evaluates results — iterating until it finds high-quality implementations. The Nature benchmark results are notable: ERA achieves expert-level performance across genomics, public health, satellite imagery analysis, neuroscience prediction, time-series forecasting, and mathematics. Google claims it could democratize access to expert-level computational modeling for scientists who lack deep software engineering backgrounds. ERA is not just a benchmark result — it's already in production use. Google says it has run ERA on eight research manuscripts, including newly released papers on epidemiology, CO2 mapping, snow runoff, solar energy design, cosmology, and neuroscience. Five of those papers are being released today. The tool is also the foundation of Computational Discovery, a new product launching through a trusted tester program in Google Labs. The caveat is that source material is limited to Google's own assessment of expert-level performance. Independent replication on ERA's benchmark claims hasn't yet occurred. Still, the Nature publication adds credibility, and the breadth of scientific domains tested is wider than most comparable systems.