A freshly pretrained language model is a brilliant autocomplete — it predicts likely next words, but it has no sense of being helpful. RLHF (reinforcement learning from human feedback) is the process that bridges that gap.
Three steps
RLHF works in three stages. First, supervised fine-tuning: show the model examples of good responses to imitate. Second, train a reward model on human preferences — people rank pairs of answers, and the reward model learns to predict which humans will prefer. Third, optimize the model against that reward, nudging it toward answers people like.
What it really teaches
The key insight: for open-ended questions there's no single "correct" answer, only a distribution of good ones and a sense of what's better. Imitation alone can't capture that. RLHF lets the model learn from a preference signal — better vs. worse — rather than a single target.
RLHF didn't make models smarter. It made them aligned with what people actually want — which is a different, and crucial, thing.
The evolution
Modern pipelines increasingly replace the full RL loop with simpler methods like DPO, or use RL with verifiable rewards for reasoning. But the core idea RLHF introduced — train on human preference, not just imitation — is now foundational to every capable assistant. It's the step that turned raw models into the tools people use every day.
