Grokking

Grokking is a phenomenon in machine learning in which a neural network first memorises its training data — reaching perfect training accuracy while test accuracy sits near chance — and then, after a long plateau, abruptly begins to generalise, with test accuracy jumping toward 100% far later in training than standard expectations would predict.^[1] The name, borrowed from Robert A. Heinlein's coined verb meaning deep understanding, was introduced by OpenAI researchers in a 2022 paper on small algorithmic datasets.

History and discovery

The phenomenon was documented by Alec Power, Yuri Burda, Harrison Edwards, David Gross, and Ilya Sutskever in Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets (2022). Training simple transformer networks on small algorithmic tasks such as modular arithmetic, they observed a striking sequence: training loss fell quickly to zero, validation loss worsened as the network overfit, and then — often after hundreds of epochs of apparent stagnation — validation loss collapsed and the model generalised to unseen inputs.^[1]

Grokking was surprising because classical statistical wisdom holds that once a model has fitted the training set and validation performance has bottomed out, further training only deepens overfitting. Grokking showed that, given enough training, a path from memorisation to generalisation can still exist. Later work gave the effect a more neutral name, delayed generalisation, and connected it to the broader bias-variance trade-off as networks move from memorising examples to discovering compact underlying rules.^[3]

Key concepts

Several findings characterise the effect:

• Memorisation precedes rule learning. Networks typically encode the training examples first; generalising circuits emerge later, when weight decay or the optimiser's implicit bias gradually prunes the memorising solution in favour of a simpler one.^[2]
• Weight norm growth. Studies using an effective theory of representation drift showed that grokking correlates with the growth of weight norms, connecting the delayed transition to how representations evolve during training rather than to a single sudden event.^[2]
• Accelerating grokking. The Grokfast method (2024) amplified slow-varying gradient components during training, accelerating the onset of grokking by more than 50× with only a few lines of code added to standard optimisers, evidence that the delayed phase is driven by low-frequency gradient signals.^[4]
• Relevance to large models. Researchers have hypothesised that grokking-like transitions may occur during large language model training — for example when models suddenly acquire a new capability — though direct evidence remains an active research question.^[3]

Why it matters for AI

Grokking matters for three practical reasons. First, it shows that checkpoint selection based only on early validation curves can be misleading — a model that looks fully overfit may still be on a path to generalisation. Second, it informs interpretability: grokking transitions are one of the cleanest settings in which researchers can watch a network shift from memorisation to algorithmic circuits, making it a favourite probe for mechanistic interpretability. Third, it touches efficiency: understanding what triggers the delayed transition could shorten training runs by bringing the generalisation phase forward.

>See Also

• Overfitting
• Bias-Variance Tradeoff
• Regularisation
• Mechanistic Interpretability
• Transformer Architecture
• Continual Learning