Reward model
Beginner
A second AI trained to give answers a score, so the first AI knows what “good” looks like.
Novice
A neural network trained to score a model’s responses, for example by how likely a solution is to be correct or how much people would prefer it. An outcome reward model gives one score per answer.
Expert
A learned scalar scorer, trained on correctness labels (outcome reward model, ORM) or human preference pairs. It extends rewards beyond checkable tasks but becomes a target for exploitation under strong optimization.