Conformal Certification Certifies 16.7% of Aggressive Proposals in Offline MBO

16.7% of candidates from an aggressive proposal pool receive a statistical certificate of meeting a target threshold, with empirical coverage hitting 0.990 against a nominal 0.90. Standard conformal prediction that ignores covariate shift? It collapses to 0.416 coverage.

That gap is the point of a new paper proposing Conformal Candidate Certification (CCC), a post-hoc wrapper for offline model-based optimization (MBO).

Why Offline MBO Needs Statistical Certificates

Offline MBO uses a surrogate model trained on a fixed historical dataset to propose candidates. The optimizer aggressively searches out-of-distribution designs where the surrogate is least reliable. Existing methods give no per-candidate statistical guarantee that a design actually meets the target metric.

CCC adds a calibrated one-sided lower bound to each candidate. Only candidates whose bound exceeds the target get advanced. This turns an unbounded optimization gamble into a decision with a known false-positive rate.

Entropy-Regularized Surrogates Unlock Weighted Conformal Prediction

The trick: entropy-regularized surrogate maximization naturally produces a Gibbs-tilted proposal distribution. That same entropy term supplies importance weights for weighted conformal prediction, eliminating the need for a separate density-ratio estimation step.

In a controlled synthetic study, the authors show that CCC certifies 16.7% of an aggressive proposal pool with empirical coverage 0.990 at nominal 0.90. Without the importance-weight correction, standard conformal prediction - which assumes the test data matches the training distribution - yields only 0.416 coverage.

What This Changes

CCC turns offline MBO from a black-box optimization loop into one where you can reject or accept candidates with a statistical guarantee. For any domain where proposed designs must meet a hard threshold - drug molecule properties, material strength, device performance - this means you can deploy aggressively optimized candidates with a per-design lower bound, not just a surrogate's best guess.

The method is post-hoc: it wraps any existing surrogate and optimizer without retraining. For engineers running offline optimization pipelines, that's the difference between shipping a candidate you trust and hoping the surrogate wasn't lying.

Source: Conformal Candidate Certification for Offline Model-Based Optimization
Domain: arxiv.org