RLHF (Reinforcement Learning from Human Feedback) is the model’s upbringing stage: humans rank several model answers, a reward model learns those preferences, and the main model is trained to maximize the reward. The result: more helpful, accurate and polite.
This is the trick that separated ChatGPT from a raw text model; lighter variants (DPO) are standard today. The known critique: the model can drift toward pleasing humans rather than being correct — so high-quality human data stays the bottleneck.
In short: pre-training gives understanding, RLHF gives manners and honesty. Modern model training is the balance of these stages.
