A generation approach where the model generates multiple tokens in parallel or through iterative refinement, rather than one at a time.