← Concept Index

Training & adaptation

Alignment and RLHF

Also called: reinforcement learning from human feedback, instruction tuning

DEFINITION

After pretraining, models are shaped to be helpful and to refuse harmful requests using human (and AI) feedback. RLHF — reinforcement learning from human feedback — is one common method.

WHY IT MATTERS

Alignment is why a raw text-predictor behaves like a helpful assistant, but it's a tendency, not a guarantee. It can be over-cautious, inconsistent, or bypassed, which is why safeguards can't live in the model alone.

COMMONLY CONFUSED WITH

A hard rule engine. Alignment nudges probabilities; it isn't a filter that provably blocks every bad output.