What is it?
Preference learning uses comparisons or feedback to teach a model which responses are preferred.
RLHF means Reinforcement Learning from Human Feedback. It is one preference-alignment approach, not the only possible method.
04 · Training & Model Adaptation · Preference Learning / RLHF
What is Preference Learning?
Preference learning trains a model using feedback about which outputs are more desirable.
Think of it like...
A writing coach comparing two answers. The coach says, “This answer is better than that one.” Over many comparisons, the learner begins to understand which style is preferred.
How it works
1 · Prompt
Give the task
A model receives a question or instruction.
2 · Candidates
Produce responses
Candidate outputs can be compared.
3 · Feedback
Choose a preference
People or another process supplies a preference signal.
4 · Training
Shape behavior
The signal is used in further training; implementations differ.
Real-world examples
Helpful answers
Prefer direct, useful responses.
Safety behavior
Prefer responses that follow policy boundaries.
Style
Prefer concise or clearly structured responses.
What it is NOT
RLHF ≠ Reinforcement Learning in Every Form
RLHF uses human feedback; reinforcement learning is the broader family.
Preference Learning ≠ Fine-tuning in General
Preference learning specifically uses preference or feedback signals.
Human Preference ≠ Objective Truth
People can disagree, and preference data can contain bias.
Related concepts
Pre-training
↓
Instruction Tuning
↓
Preference Learning
Remember this
Preference learning uses feedback about better and worse outputs to shape model behavior.
Video