GRPO

A reinforcement learning algorithm that estimates a baseline from a group of sampled outputs for the same prompt rather than training a separate value model, reducing training cost. Popularized by DeepSeek's use of it to train reasoning models.