Reinforcement learning from human feedback: tuning a model toward outputs people rate highly.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
RLHF closes the loop with human judgment: people compare pairs of model outputs, a reward model learns to predict those preferences, and the language model is then optimized against that learned reward. It is the technique that turned raw base models into helpful, conversational assistants.
It is also the main practical alignment tool in production today, with a known caveat baked into its design: it optimizes for what evaluators approve of, which is not the same thing as what is true or what is good. Models can learn to be agreeable rather than accurate.