What is the name of the alignment method that replaces many human preference labels with model self-critique and revision guided by a written set of principles?
Show answer
Constitutional AI
Constitutional AI uses a fixed list of principles (a "constitution") to have the model critique and revise its own responses, generating preference data with far less human labeling. The phase that optimizes against this machine-generated feedback is often called RLAIF, RL from AI Feedback.
Also asked as:
- Which alignment technique uses a model's own self-critique and a set of rules to generate preference data, reducing the need for human labeling?