Definition
Structural transfer is when a model trained on one task picks up the structure of another related task with little or no additional supervision — the cleanest evidence that the model learned something generalizable rather than memorized. It’s how a lot of capability emerges in pretraining.