statistical-test

USAD proposes two new statistics, Variance Discrepancy and Perturbation-based Covariance Discrepancy, to capture global and local uncertainty patterns of adversarial examples, achieving superior detection performance over baseline methods.

UnpredictaBench is a benchmark for evaluating how well large language models can sample from target distributions, including statistical and natural-language random processes. Experiments show that current models struggle to capture true underlying distributions, with no model exceeding 40% on the KS@100 metric.