Because the memorization-vs-generalization competition account explains why grokking happens, it also tells you exactly which attention heads or neurons switched roles in a specific trained model.

False. The account explains the general competitive dynamic between memorizing and generalizing circuits and weight decay's role in tipping that competition, but it does not trace or identify the actual circuits inside any one trained model — that requires circuit-level interpretability analysis, which sits outside this mechanistic account.

Grokking phase-transition mechanistic account Interview Q&A

Grokking phase transition

The phase transition is the abrupt shift in a model’s behavior that occurs after long overfitting: the general rule takes over from memorized lookup-like structure almost all at once - even though the rule’s internal structure is amplified gradually beforehand, with the final cleanup of memorization making the switch look sudden.

Example

A model holds 100% train accuracy and near-zero test accuracy for thousands of steps, then test accuracy leaps to near 100% within a short window, mirroring the sudden representation shift.

Two circuits compete

A leading account explains grokking as competition between two circuits sharing the same weights: a 'memorizing circuit' storing training examples, and a 'generalizing circuit' implementing the true rule. Memorization wins early because it minimizes loss fastest. Once the generalizing circuit becomes efficient enough, it dominates outright — a proposed explanation for why generalization arrives abruptly, not gradually.

Setup

A model memorizes all training pairs by step 2,000 (0% test accuracy), with weight decay on. Why does test accuracy jump to 100% near step 8,000?

1
Decay shrinks every weight each step.
WhyMemorization needs large, example-specific weights, so it’s taxed hardest.
2
The generalizing circuit needs only a small, reusable rule.
WhyLower norm means less tax, so it survives and grows dominant.

Takeaway

Weight decay taxes weight norm, not correctness — the cheaper generalizing circuit eventually wins that race.

Setup

You want this account to name exactly which neurons or attention heads flipped when one specific transformer groked on a real dataset.

Expected

Since it describes circuit competition under weight decay, you might expect it to also identify which circuits in that model’s weights changed.

What breaks

It only describes the general competitive dynamic and regularization pressure — it never traces or identifies actual circuits inside any one trained model.

Lesson

Use it as a mechanism-level story, not a substitute for circuit-level interpretability of an individual model.

Grokking phase-transition mechanistic account

Grokking phase transition

Two circuits compete

Sources

A focused session on your AI Research Scientist interview

More Generalization Theory & Phenomena questions