RLAIF (Reinforcement Learning from AI Feedback)

The stage of Constitutional AI where an AI (rather than a human) provides feedback on which response better follows the constitution. The feedback is used to train a reward model, which is then used for RL optimization.