Reinforcement Learning from Human Feedback (RLHF) is a training technique that uses human preference judgements as a reward signal to align a language model's outputs with human values and intentions. Human annotators compare pairs of model responses and label which is better; those labels train a Reward Model that can score any response; and a reinforcement learning algorithm-typically Proximal Policy Optimisation-then updates the language model to produce outputs the reward model rates highly.

Think of it like coaching a new employee through peer reviews rather than a rulebook. Instead of enumerating every policy they must follow, you have colleagues rate their work samples, identify patterns in what earns praise, and give the employee feedback based on those patterns. Over time the employee learns the implicit standards the organisation values.

RLHF became widely known as the key technique behind InstructGPT and later ChatGPT. The process runs in three stages: (1) supervised fine-tuning on demonstration data so the model can follow instructions at all; (2) reward model training on pairwise human ratings of model outputs; (3) RL fine-tuning where the policy is updated to maximise the reward model score while a KL-divergence penalty prevents it from deviating too far from the supervised baseline. The result is a model that is markedly more helpful and less harmful than a raw pretrained LLM.

RLHF sits at the intersection of Reinforcement Learning, Fine-tuning, and Reward Model design; its failure modes include reward hacking and sycophancy, and it directly informs Responsible AI practices around alignment and safety. Variants such as DPO (Direct Preference Optimisation) and Constitutional AI attempt to achieve similar alignment goals with less complexity.

You will encounter RLHF whenever you read about how modern chat assistants like Claude, ChatGPT, or Gemini are trained to be helpful, harmless, and honest rather than merely statistically fluent.