Self-Supervised Alignment with Mutual Information Learning to Follow Principles without Preference Labels