Information Theoretic Guarantees For Policy Alignment In Large Language Models