We propose SRPO, a reflection-aware RL method that significantly improves multimodal LLM reasoning by explicitly teaching self-reflection, outperforming state-of-the-art models on multiple benchmarks.