Learning the differences between consecutive states rather than full states, reducing compression complexity for time-evolving data.