The AI Handbook Open in the app →

Reinforcement Learning from Human Feedback

Training & Tuning · Advanced · 5 min read

What is it?

RLHF is a training method that fine-tunes a model using human preferences — people rank the model's answers, and the model learns to produce responses people prefer.

Explain like I'm 5

It's like coaching. The model gives answers, humans say 'this one's better,' and the model gradually learns to give the kind of answers people like.

Why was it created?

A base language model predicts likely text, not helpful or safe text. RLHF was developed to align models with what humans actually want — helpful, honest, and harmless responses.

Where is it used?

  • Chat assistants
  • Making models follow instructions
  • Reducing harmful or off-topic output
  • Tuning tone and helpfulness

Why should developers care?

RLHF is the step that turns a raw next-word predictor into a usable assistant, so it underlies the chat models most developers build on.

How does it work?

Humans rank multiple model responses to the same prompt. Those rankings train a 'reward model' that scores answers. The language model is then optimized with reinforcement learning to maximize that reward, nudging it toward preferred responses.

Real-world example

Given a question, the model drafts several answers; annotators rank them; the model learns to favor the clear, helpful style people ranked highest.

Common use cases

  • Instruction following
  • Safety and refusal behavior
  • Aligning tone and formatting
  • Preference-based fine-tuning

Advantages

  • Aligns output with human preferences
  • Improves helpfulness and safety
  • Captures fuzzy goals hard to write as rules
  • Works on top of a pretrained model

Disadvantages

  • Expensive — needs human labelers
  • Reward model can be gamed
  • Can over-refuse or become bland
  • Hard to reproduce consistently

When should you use it?

When you need a model to behave like a helpful, aligned assistant rather than just predict text.

When should you avoid it?

When a task is fully specified by rules or when simpler prompting/fine-tuning already meets the goal.

Alternatives

Supervised fine-tuning aloneDirect Preference Optimization (DPO)Constitutional / AI-feedback methods (RLAIF)

Related terms

Fine TuningTrainingLarge Language ModelHallucination

Interview questions

Beginner

  • What kind of feedback does RLHF use?
  • Why isn't a base language model already a good assistant?

Intermediate

  • What is the role of the reward model in RLHF?
  • How can a reward model be 'gamed'?

Senior

  • How does DPO simplify the RLHF pipeline?
  • What failure modes appear when over-optimizing a reward model?

Common misconceptions

  • "RLHF teaches the model new facts" — it mainly shapes behavior and style, not knowledge.
  • "More RLHF is always better" — over-optimizing can make models evasive or bland.

Fun facts

  • InstructGPT showed a small RLHF-tuned model could be preferred over a much larger untuned one.
  • The reward model is itself a trained neural network.

Timeline

  • 2017 — Deep RL from human preferences demonstrated
  • 2022 — InstructGPT popularizes RLHF for chat models

Learning resources

Quick summary

RLHF fine-tunes a model on human preference rankings via a reward model, turning a raw text predictor into a helpful, aligned assistant.

Cheat sheet

  • Humans rank answers → reward model
  • Model optimized to maximize reward
  • Turns base model into an assistant
  • Costly; reward can be gamed

If you remember only one thing

RLHF coaches a model with human preferences so it produces answers people actually want, not just statistically likely text.

Further reading