VecUQ-OT: Aggregating Uncertainty Measures via Multivariate Ranks

Decision-making under uncertainty typically requires a single scalar score, forcing practitioners to commit to one uncertainty measure. However, different measures often capture complementary failure modes, and relying on a single measure may be insufficient for downstream tasks. We propose \emph{VecUQ-OT}, a label-free procedure that aggregates multiple scalar uncertainty measures into a single ranking score via entropy-regularized optimal transport. Rather than selecting a single measure, our approach combines signals of (possibly) different natures through non-additive fusion based on \emph{multivariate ranks}. The construction is motivated by the theory of multivariate ranks via exact optimal transport and realized through a tractable entropic approximation that generalizes to unseen inputs without retraining. Experiments across synthetic, image, and text domains demonstrate that \emph{VecUQ-OT} provides stable performance across tasks, remains reliable when individual measures fail, and outperforms the natural additive fusion alternative.