“DeepSeek uses the Group Relative Policy Optimization (GRPO) algorithm for reinforcement learning, which is a variant of PPO with a different value approximation method.”