DeepSeek-R1: pure reinforcement learning (RL), no supervised fine-tuning (SFT), no chain-of-thought…