AlphaGRPO: Unlocking Self-Reflective Multimodal Generation in UMMs via Decompositional Verifiable Reward

Researchers introduced AlphaGRPO, a reinforcement-learning framework for unified multimodal models.

It applies GRPO with decomposed, verifiable rewards to improve text-to-image generation and self-correction.

Experiments show gains on multimodal benchmarks and editing tasks, even without direct editing-task training.