AI Glossary Term

What does “Reinforcement Learning from Human Feedback (RLHF)” mean?

Definition

A training method that uses human preference signals to shape model behavior.

Browse the full AI GlossaryExplore all free AI guides

How Reinforcement Learning from Human Feedback (RLHF) works

Diagram for Reinforcement Learning from Human Feedback (RLHF), steps in order: Model writes several answers; then People rank them; then Reward model learns the preferences; then RL tunes the model toward preferred answers.Model writes several answersPeople rank themReward model learns the preferencesRL tunes the model toward preferredanswers

Example

Raters compare two chatbot replies to the same question and pick the more helpful and safer one. Thousands of such comparisons train a reward model, which is then used to nudge the chatbot toward replies people prefer.

Why it matters

RLHF is a main method for making chatbots more helpful and less harmful. The outcome reflects the values and judgment of the raters, which is itself a limitation.

Common questions

Why use human feedback?

Qualities like helpfulness are hard to write as a formula, but people can compare answers.

What are the limits?

Raters can be inconsistent or biased, and models may learn to please raters rather than be accurate.

Alternatives?

Methods such as DPO skip the separate reward model, and some approaches use AI-generated feedback.

Related terms

Reinforcement LearningTraining by reward signals where an agent learns actions that maximize long-term return.Reward ModelA model that scores outputs based on preference signals, often used in RLHF pipelines.DPO (Direct Preference Optimization)A training method that fine-tunes models directly on preference pairs without needing a separate reward model.Continual LearningTraining approaches that let a model keep learning from new data without forgetting prior knowledge.Few-Shot LearningLearning or adapting behavior from only a small number of examples.Model DriftPerformance degradation over time as real-world conditions diverge from training assumptions.

Learn more in our free guides

Deep LearningReinforcement LearningMachine Learning BasicsSupervised Learning

Test yourself

Deep Learning QuizReinforcement Learning QuizMachine Learning Basics Quiz

See also

RetrievalRed TeamingRecommendation SystemRobustnessRecallSafety FilterRAG (Retrieval-Augmented Generation)Scaling Law