Providing training feedback at the level of refinement steps in an iterative process, rather than at individual token positions.