Abstract
Simulation-based inference (SBI) methods often produce posterior approximations whose credible sets are poorly calibrated, underestimating uncertainty around the true parameters. We introduce CP4SBI, a conformal calibration framework that recalibrates credible sets to achieve local Bayesian coverage. We propose two variants – one based on regression trees and another on cumulative distribution functions – that provide finite-sample coverage guarantees under several scoring functions, including highest posterior density, symmetric, and quantile-based scores. Experiments on standard SBI benchmarks show that our approach substantially improves the uncertainty quantification of neural posterior estimators built with normalizing flows and score-diffusion models.