What is RL from Verifiable Rewards (RLVR), and why does it produce more reliable training signal than a learned reward model?

RL from Verifiable Rewards (RLVR) utilizes tasks where correctness can be determined by a checker or unit test rather than a human-labeled reward model. Because the reward is binary (correct or incorrect), the model cannot ‘game’ the reward signal in the way it might with a learned proxy model. This creates a stable, non-hackable training signal, though it is restricted to domains where ground truth is explicitly available, such as formal proofs or programming.