All terms
Training
Group Relative Policy Optimization
A training method that scores several answers per prompt against each other instead of using a separate scoring model.
Definition
Group Relative Policy Optimization is a reinforcement-learning technique (learning by trial and reward) that generates several answers to a prompt and judges each one against the group's average. This relative comparison replaces the separate scoring model used by methods like PPO, cutting memory and computation. Popularized by DeepSeek-R1, it is widely used to train reasoning models, often with checkable rewards such as whether the math answer is correct.