Conformal Certification guarantees reliable offline model optimization candidates

Offline model-based optimization (MBO) is widely used to design molecules, proteins, and engineering structures by optimizing a surrogate model over a fixed historical dataset. The core challenge is that the surrogate’s predictions are least reliable for the very candidates the optimizer pushes hardest—those far from the training distribution. Existing validation techniques cannot provide per-candidate statistical guarantees that a design meets a performance target, leaving practitioners blind to whether a top-ranked candidate is genuinely good or just a lucky extrapolation.

Seungjin Choi’s new paper, presented at ICML 2026 Workshop, proposes Conformal Candidate Certification (CCC) to solve this. CCC is a lightweight post-hoc wrapper: for each candidate it outputs a calibrated one-sided lower bound on the true objective. Only candidates whose bound exceeds a user-defined threshold are advanced. The key technical insight is that entropy-regularized surrogate maximization naturally induces a Gibbs-tilted proposal distribution, which allows the same surrogate to supply importance weights for weighted conformal prediction—eliminating the need for a separate density-ratio estimation step. In controlled experiments on a synthetic test function, CCC certified 16.7% of aggressive proposals with empirical coverage of 0.990 at a nominal 0.90 level. In contrast, standard conformal prediction that ignores covariate shift collapsed to 0.416 coverage, illustrating the dangers of naïve application. The method promises to make MBO safer for high-stakes scientific discovery and engineering design where overconfident models can lead to costly failures.