Preprint filters false positives from AI materials searches

Researchers report in a preprint that their CISE method returned only candidates that passed high-fidelity evaluation across three AI-driven materials science search tasks. Baseline methods returned more candidates but included false positives. CISE constructs statistical intervals around estimated rewards, uses conservative scores to guide subsequent searches, and returns a candidate only when every required property interval falls within its feasible region. The approach targets searches where evaluating every candidate accurately is prohibitively expensive. Its theoretical coverage results depend on explicit assumptions of independence and covariate shift, and its experiments produced a smaller shortlist.