Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking