Memorization vs. Generalization in Large Models

Memorisation and generalisation are often spoken of as opposites, yet in large models they can occupy the same training run in sequence. First the network drives training error down, sometimes to perfect classification or exact reconstruction of the examples it was given. At that moment it is natural to say it has “learned the data.” Held-out inputs may still be handled poorly. Only later—sometimes much later—does performance on those unseen cases climb. That late transition is what the grokking literature names and tries to explain.

The editorial pressure of the topic is simple: a perfect training score does not certify that a rule has been found. It certifies that the training set has been fit. Rule-like behaviour has to be demonstrated on inputs that do not reward mere storage. When those two milestones separate in time, experimenters get a slow-motion view of a change that is usually compressed into a brief phase of ordinary training.

One line of work approaches the delay through an effective theory of representation learning. The claim is not that a full microscopic account of every weight is in hand, but that useful intermediate descriptions—how representations reorganise, how structured encodings become available—can explain why generalisation can arrive long after memorisation. Early representations may be enough to fit the training set. Later ones, if they form, may support behaviour that transfers. The paper title itself signals caution: towards understanding, not a closed theory.

Separately, comparing what a model does on one collection of inputs versus another is a statistical problem. Two-sample testing asks whether two samples appear to share a distribution. Projected Wasserstein distance is one construction for that task: projections reduce the comparison, and a Wasserstein metric judges how mass would need to move to match the measures. In this lesson’s setting, the two samples might be activations, errors, or outputs on training-like inputs versus held-out or rule-structured ones. A test that still separates them is evidence that the model’s behaviour has not collapsed onto a single regime. A test that no longer separates them is weaker evidence than a mechanistic proof; it only says the chosen probe no longer sees a difference.

Taken together, the two sources push in complementary directions. One names and theorises a training-time phenomenon in which generalisation lags memorisation and ties that lag to representation learning. The other supplies a careful distance-based way to test whether two clouds of model-related observations differ. Neither source, on the facts given here, hands down a single schedule, architecture list, or universal threshold at which “true” learning begins. What they do support is a disciplined habit: treat train fit and held-out rule-following as distinct events; watch representations rather than loss alone when the events separate; and when comparing regimes, say plainly what distributional test was used and what it can and cannot show.

For anyone training or auditing large models, the practical stance is unhurried. Log training and validation trajectories far past the point of training saturation. Probe with inputs that encode the putative rule rather than near-duplicates of training rows. When claiming that a model has moved from memorisation toward generalisation, bring evidence about representations or about distributional agreement on held-out behaviour—not only a green training metric. Where the evidence is thin, say so. Grokking is a named pattern in some training runs, not a guarantee that every late uptick is deep understanding.