RLHF

Reinforcement Learning from Human Feedback

Training method that aligns LLM behavior to human preference ratings.

When you'd see it: Discussions of how modern chat models are trained to be helpful and safe. RLHF is a key step after pretraining.

Why it matters: RLHF tunes a model toward human preferences by training it on people's ratings of its outputs. It is much of why chat models feel helpful and aligned rather than just predicting text.