This paper introduces Causal-JEPA (C-JEPA), a world model that learns by hiding entire objects in its memory and forcing itself to predict them from other objects.
This paper shows how to give AI a steady “mental map” of the world that keeps updating even when the camera looks away.