SPACR: Single-Pass Adaptive Training of Uncertainty-Aware Conformal Regressors

Soundouss MESSOUDI, Sylvain Rousseau, Sebastien Destercke

Keywords: Conformal Prediction, Calibration, Deep Learning, Uncertainty Quantification

TL;DR: We propose SPACR, which directly optimizes efficiency and validity without batch-splitting or fixed confidence levels, consistently achieving tighter intervals and superior trade-offs compared to state-of-the-art baselines.

Abstract: Conformal Prediction (CP) provides distribution-free guarantees but often yields inefficient intervals because standard models are not trained with conformal objectives. We propose the Single-Pass Adaptive Conformal Regressor (SPACR), which directly optimizes predictive efficiency and validity using a differentiable calibration proxy within its loss. Unlike prior work (e.g., DOICR), SPACR eliminates the need for batch-splitting or fixed confidence levels during training, enabling a single model to generate valid intervals for any confidence level at inference. Experiments show that SPACR consistently gives tighter intervals and superior efficiency-coverage trade-offs compared to state-of-the-art baselines.

Submission Number: 33