RLVR and Verifiable Rewards

RLVR stands for Reinforcement Learning with Verifiable Rewards: the signal comes from a rule-verifier (did the test pass, is the answer correct) rather than a trained model. Agent tasks are often verifiable – thus RLVR is the main line. A special case of the general scheme: the reward model can be deterministic code when correctness is definable by a rule. Asymmetry of verification and generation (“easier to verify than to do”) is the fundamental reason RL can outperform any demonstrator.

Related: [Reward for Process vs Reward for Outcome], RLHF, [Data and Environment are More Important than Algorithm]