GRPO
Policy optimisation that scores a group of answers relative to each other.
Aperçu
Group Relative Policy Optimization drops the value model that PPO needs by sampling a group of answers to the same question and using their relative scores as the baseline. It solves the problem that critic models are large, expensive and another thing to train. It exists because removing the critic makes reinforcement learning on reasoning tasks practical at scale, which is how it became central to reasoning-model training.
Graphe des technologies
Comment GRPO se rattache au reste de la pile.
Utilise et dépend de
- Reinforcement Learning concept Algorithms GRPO based on Reinforcement Learning High ✓ source