“The final reinforcement learning stage for DeepSeek R1 mixes verifiable domain prompts with standard RLHF preference tuning to improve helpfulness and harmlessness.”