Reinforcement Learning from Human Feedback is a training phase used on LLMs after supervised fine-tuning to further improve LLM responses. It was one of the key innovations behind ChatGPT.