We show that dynamics-aligned representations improve zero-shot generalization for contextual world models.