Reinforcement learning from human feedback (RLHF)

A preference-tuning method in which human judgments of which answers are better are used to further train a model toward preferred behavior. It is a major reason two versions of the same base model can behave differently, and it is the kind of tuning that produced the divergence in the Maverick episode.

A specific fine-tuning technique in which human reviewers rank or rate model outputs, and the model is further trained to produce outputs more like the highly rated ones; a major factor in the shift from raw base-model behavior to a consistently helpful, instruction-following deployed assistant.

A weight-layer method for tuning a model in which human ratings of its outputs are used to nudge the model toward the rated-good behaviors. A common place the alignment tax appears.