LFI

CP4SBI is a conformal calibration framework for simulation-based inference: it recalibrates posterior credible sets from neural estimators to guarantee local Bayesian coverage. Two variants, based on regression trees and on CDFs, give finite-sample guarantees and substantially improve uncertainty quantification for estimators built with normalizing flows and diffusion models.

The paper introduces a novel approach to classifying events under generalized label shift by framing classification as hypothesis testing, which leads to valid uncertainty quantification. This is demonstrated through applications in biology and astroparticle physics.