Grokking refers to a phenomenon in deep learning where a neural network initially memorizes training data and only much later—sometimes after thousands of additional training steps—suddenly generalizes to new, unseen data. It is distinct from standard learning, where test accuracy improves smoothly as the model learns.