A new study found that AI models often know when they have the right answer but continue generating text, and a proposed method improves speed and accuracy by letting them stop early.