Reordering a model's next-token predictions by likelihood or quality rather than accepting the top-1 choice.