Pass IndexThe State of AISign in

RLHF

Training a model on human preferences: people rank two answers, a reward model learns the pattern, and the model is tuned to score well against it.

how it works · the vocabulary

reinforcement learning from human feedback

It is what makes a model helpful and polite rather than merely fluent, and it is also where refusal behaviour and house style come from. Expensive, because it needs human judgement at scale.

Nearby

AgentAgent memoryAlignmentAutoregressiveBM25ChunkingCold startComputer use