A phenomenon in machine learning where a model generalizes well on unseen data after overfitting its training data.