Generalization refers to a model's ability to perform well on previously unseen data, typically measured by it's performance on the test set.