Neural Posterior Estimation (NPE) trains an amortized conditional density estimator on simulated data, but its log-likelihood objective can push flexible models such as normalizing flows to assign extreme density to training points. This overconfidence does not transfer to held-out data. Early stopping and other regularizers help only partially, because they are stuck on an optimization path where the posterior mean is underfit while the posterior variance is overfit.

The proposed method, Cross-Fitting NPE (CF-NPE), avoids this trade-off by using sample splitting. With K-fold cross-fitting, it first fits K separate posterior-mean regressors, then fits a conditional density estimator to the residuals from out-of-fold predictions. This cleanly separates estimation of the conditional mean from estimation of higher moments, preventing the variance from being distorted by in-sample overfitting.

In experiments on synthetic data and in a cosmological application, CF-NPE achieved higher held-out log-likelihood than standard NPE and typically improved calibration. The authors note the method is most useful in high-dimensional, simulation-limited settings, where the cost of extra simulations is high and the benefits of a well-calibrated posterior are greatest.