Reinforcement Learning from Human Feedback (RLHF) is a training method used to align language models with human preferences. After pre-training, humans rank or rate model outputs, those rankings train a reward model, and the language model is then optimized to produce responses the reward model scores highly. RLHF is what turned raw next-token predictors into helpful, harmless assistants, and it underlies the instruction-following behavior of ChatGPT, Claude, and Gemini. Related and successor techniques include RLAIF (feedback generated by an AI instead of humans), Constitutional AI, and DPO (Direct Preference Optimization), which achieves similar alignment without a separate reinforcement-learning loop.
Frequently Asked Questions
What is RLHF?
RLHF (Reinforcement Learning from Human Feedback) aligns a language model with human preferences by training a reward model from human rankings and then optimizing the model against it.
Why is RLHF important?
RLHF is the key step that makes base models into helpful assistants; it shapes tone, safety, and instruction-following in products like ChatGPT and Claude.