Identifies inputs that differ from the training data distribution. Enhances model safety and reliability in real-world applications.