reward design

AlphaGRPO teaches multimodal models to self-correct

AlphaGRPO adds verifiable reward signals to multimodal models so they can reason, refine outputs, and improve generation without cold-start training.