“Multi-token prediction, which forces a model to generate an entire sequence at once without intermediate feedback, encourages a more global understanding and can improve model creativity.”