Google researchers have built an open-source system that repeatedly edits experiment code and tests whether each change improves a defined scientific score. Empirical Research Assistance, or ERA, keeps a tree of past notebooks, chooses promising branches without abandoning exploration, and asks Gemini or another language model to propose batches of mutations.
The approach grew from an attempt to automate Kaggle-style data-science competitions. According to researcher John Platt, his team has since used it in work contributing to at least 10 papers. One example involved modeling the warming effect of airplane contrails, where ERA found a simpler predictive model incorporating confounding factors the team had not considered.
Its strength is also its central limitation. ERA optimizes whatever researchers make measurable; it does not determine whether that score captures the real scientific goal. Platt points to a contrail-detection competition where entrants exploited a half-pixel labeling error. That improved the leaderboard result without advancing the underlying problem, a familiar case of a metric becoming a bad target once optimization focuses on it.
The implementation is available on GitHub and can work with models other than Gemini, but ERA is not a Google product. Scientists still have to design valid evaluations, inspect outputs and confirm that predictive performance reflects the phenomenon they intended to study.