Contrastive Learning from AI Revisions (CLAIR): A Novel Approach to Address Underspecification in AI Model Alignment with Anchored Preference Optimization (APO)