Post
1383
The beauty in GRPO is the fact that it doesn’t care if the rewards are rule-based or learned, the hack: let the data self-normalize— trajectories in a batch compete against their mean, no value model, no extra params, just clean, efficient RL that cuts memory usage by 50%, while maintaining SOTA performance. btw it was introduced 9months prior to R1: arxiv.org/pdf/2402.03300
