A reward model is a model trained to predict how a human would rate a given output, providing the scalar reward signal for RLHF instead of requiring a human in every loop. It learns from datasets of human preference comparisons and outputs a score that guides the reinforcement learning policy toward generating responses that align with human preferences.