GRPO
GRPO is a reinforcement learning algorithm that improves an AI model by comparing rewards among responses to the same input.
GRPO, or Group Relative Policy Optimization, is a reinforcement learning algorithm that generates multiple responses to the same prompt and compares their rewards within a group to improve a model's policy. DeepSeek introduced it in its 2024 DeepSeekMath paper. Unlike PPO, it does not train a separate value model, instead estimating each response's advantage relative to the group's average reward.
It is used in reinforcement learning to strengthen the mathematical and reasoning abilities of language models and was applied in training DeepSeek-R1.
This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.