Model Training & Adaptation
Reinforcement learning from human feedback
RLHF
Training that uses human preference judgments to shape model behavior.
Example
People rank two answers, and those preferences help train the assistant's behavior.
Why it matters
It helps steer replies toward qualities people find helpful or appropriate.
You'll hear it
“The reviewers preferred the clearer answer, even though it was shorter.”
Can you control it?
Set by the team
The people building or running the AI choose this setup. An everyday user generally needs their help to change how this part works.
Common questions
- Do all reviewers agree about the better reply?
- No. Preferences can differ with background, expectations and the task.
- Does a preferred answer have to be factually correct?
- No. People may favor a convincing or pleasant answer that contains mistakes.
- Does every conversation immediately update the system?
- No. Feedback may be collected and used later rather than changing the AI during that exchange.