“The concept for RLVR (Reinforcement Learning from Verifiable Rewards) was inspired by a statement from John Schulman of OpenAI that major labs perform reinforcement learning on model outputs.”