RLHF (Reinforcement Learning from Human Feedback) is an LLM training technique that, after pre-training, adjusts the model based on comparisons made by humans between several answers, so that it becomes more helpful, honest and safe.
The classic method has three steps: supervised fine-tuning on examples written by humans, training a reward model that learns to predict human preferences, then optimizing the LLM to maximize that score, usually with the PPO algorithm. Variants such as DPO learn directly from preferences, without a separate reward model.
RLHF played a key role in taking LLMs from research to the general public, starting with ChatGPT. Anthropic proposed Constitutional AI, where part of the feedback is produced by a model based on written principles (known as RLAIF).
Why it matters when hiring
RLHF concerns a handful of teams that train or adapt models themselves: AI labs, applied research teams. Profiles who have practiced it at scale are very rare. Do not confuse them with a candidate who did a simple fine-tuning; ask how they built the preference data and evaluated the result.