OOD Detection

OOD detection asks whether an input falls outside the data a model learned from, helping prevent confident errors when conditions or classes change. Common approaches score softmax or energy outputs, compare feature distances with Mahalanobis or k-nearest-neighbor methods, model density with generative models, or learn boundaries using synthetic outliers; pretrained vision-language models and test-time adaptation are increasingly used as well. Methods are tested on near- and far-OOD image benchmarks such as CIFAR and ImageNet, and in settings including medical imaging, language models, graphs, and autonomous driving. A central challenge is that strong benchmark scores may not transfer across domains: subtle shifts can evade detectors, while rejecting unknowns can also discard valid in-distribution inputs or fail to predict downstream harm.