SRUM: Fine-Grained Self-Rewarding for Unified Multimodal Models | Read Paper on Bytez