(
Reinforcement
Learning from
Human
Feedback) A machine learning technique used in developing AI chatbots and agents. Regular "reinforcement learning" (RL) comprises automated methods with predefined goals. In contrast, the human feedback in RLHF means that AI engineers and annotation specialists use their own judgement to rank results and make modifications.
After a model is pretrained, both RL and RLHF refines the model by feeding prompts and responding to the answers. However, RLHF is used to eliminate destructive, insulting and vulgar answers. It is also used to add interactive responses such as "that's a very good question."
Human fine tuning is an essential stage in developing models that deliver answers in language and idioms people relate to. While pretraining large models takes months, RL and RLHF take weeks. Reinforcement learning is especially important in developing AI agents because agents can take action on their own rather than only generate textual responses. See
AI training vs. inference and
AI agent.