Closed-loop Auto Research for Molecular Property Prediction: Discovering and Certifying Generalizable Improvements

The article presents advancements in Closed-loop Auto Research for molecular property prediction, employing language-model agents to dynamically edit model code and representations while acquiring external evidence. The study evaluates three axes of Auto Research—features, models, and external evidence—across 36 endpoints in three benchmark suites, achieving positive held-out gains with varying configurations, notably a gain of 0.042 on the test set for certain data configurations. This work highlights the potential for generalizable improvements in model performance through adaptive workflows, emphasizing the importance of separating discovery from certification in closed-loop systems, a concept applicable across various domains in AI.