Group relative policy optimisation
A reinforcement learning method from DeepSeek that scores each answer against a group of answers to the same question.
Also: GRPO
GRPO was introduced in DeepSeek's DeepSeekMath paper in 2024 and used to train DeepSeek-R1. For each prompt the model writes several answers, each is rewarded, and every answer is judged by how much better or worse it did than the group's average. This removes the need for a separate value model that older methods such as PPO required, saving memory and compute. Its simplicity and DeepSeek's open reporting made GRPO a popular choice for open reasoning-model training in 2025.