Google’s Empirical Research Assistance uses LLMs to automate scientific discovery

A Google team led by John Platt built Empirical Research Assistance (ERA), a system that automates parts of scientific discovery by framing problems as "scoreable tasks." The core idea is to define a score function for a problem and then search for code that maximizes that score. While formulating the score remains a human task, ERA automates the subsequent search process.

ERA operates by maintaining a tree of past experiments (notebooks) and using an optimistic selection rule—Upper Confidence Bound—to choose which branches to mutate next. At each iteration, the system proposes around ten mutations per promising branch, enabling different branches to share history and learn from one another. Platt likened ERA to "having a hyper-eager grad student who doesn’t sleep."

The system’s performance improved markedly between Gemini 2.0 and 2.5, transitioning from non-functional to highly effective. ERA has already produced at least ten peer-reviewed papers, including climate-related work on contrail reduction.

Platt emphasized that ERA functions as a "power tool" that can lead to misleading results if not used carefully. He advised grounding claims in descriptive modeling and cautioned against overfitting to metrics, citing a Kaggle competition where winners exploited a half-pixel labeling error to win $15,000 without solving the underlying problem.

In climate work, ERA helped model both the heat-trapping and sunlight-reflecting effects of contrails, enabling the team to quantify warming prevented by rerouting flights. Platt described climate as a moving attractor of weather outcomes, making modeling inherently difficult as the attractor itself shifts due to human activity.