Providing feedback at each individual token in a generated sequence, rather than a single score for the whole output.