Grokking: Learning That Snaps Into Place

In 2022, Power, Burda, Edwards, and colleagues at OpenAI reported something strange: a small transformer trained on modular arithmetic would reach ~100% training accuracy while test accuracy stayed near zero for thousands of epochs. Then, suddenly, test accuracy would snap to ~100%. They called the phenomenon grokking, after the Heinlein word for 'to understand fully.'

The classic curve

Accuracy
 1.0 |   Train ______________________
     |        /
     |       /         Test
 0.5 |      /          ________
     |     /          /
 0.0 |____/__________/______________ Time
     memorize      generalize
     (early)       (much later)

What is actually happening

• Early: the model memorizes training examples, complex internal representations
• Later: weight decay slowly pressures the model toward simpler solutions
• At the transition: an algorithmic circuit forms that actually generalizes
• After: performance generalizes, internal representations simplify

Why it matters beyond the toy

Grokking suggests that 'more training' can sometimes qualitatively change a model's behavior — not just improve a score but switch to a different algorithm internally. That has implications for how we evaluate safety during training.

1. 1Scores can plateau, then jump, without any architectural change
2. 2Early-stopped models can be qualitatively different from fully-trained ones
3. 3Training dynamics (not just final weights) matter for capability forecasting
4. 4Mechanistic interpretability is the tool that catches these transitions

“We show that neural networks can 'grok' algorithmic tasks, generalizing well after overfitting the training set.”

The big idea: learning is not monotonic. Grokking proves that long after training looks 'done,' the internal algorithm can still be changing.