Glossary

RL with verifiable rewards (RLVR)

Beginner

Training an AI by letting it try problems with answers that can be checked, and rewarding the right ones.

Novice

Reinforcement learning where the reward comes from a checker (is the math answer right, do the tests pass) instead of a trained reward model.

Expert

The RLHF objective with the reward model replaced by a verification function. It resists reward hacking better than neural rewards, but works only where answers can be checked, and mostly sharpens what the base model can already sample.