RLAIF: The Model as Preference Labeller

How AI feedback replaces human preference labels, what the published comparisons actually claim, and the biases a single labeller model concentrates rather than averages out.