RLHF trains a reward model on human preferences between answers, then tunes the AI to score higher — turning raw prediction into something worth talking to.
