Conformal regression provides prediction intervals with finite-sample marginal coverage, but the width of those intervals depends on how uncertainty is allocated across inputs. The new CLAPS method tackles this by explicitly combining two sources of uncertainty: aleatoric, the irreducible noise in the data, and epistemic, the model's own uncertainty.

CLAPS uses a learned heteroscedastic scale for aleatoric uncertainty and a last-layer Laplace approximation for epistemic uncertainty. A distinctive step is that the learned variance also weights the training features used to construct the posterior, yielding a noise-weighted support correction in the learned representation. This correction contracts as weighted feature information increases and becomes most influential under localized support imbalance.

In experiments on real regression benchmarks, CLAPS achieved near-nominal marginal coverage and modest efficiency improvements over a matched aleatoric baseline. The work is published in Transactions on Machine Learning Research.