Self-reasoning tokens: teaching models to think ahead