Conceptual
Login

Steerability: Why a Model Follows Some Instructions and Not Others

Training and prompt both shape behavior. Understand why some instructions hold reliably, others erode over a long session, and some are refused.