Researchers from Nanyang Technological University, Shanghai Jiao Tong University, and the University of Tokyo have proposed the UMM-Reflection method, built on the open-source unified multimodal model BAGEL. Through "reflection SFT cold start + full-trajectory reinforcement learning," it trains complete multi-turn "reflect-revise" trajectories, enabling the same model to autonomously perform image inspection, diagnosis, revision, and stopping decisions—with no external scorer needed at inference time. Experiments show that the method raises the repair success rate of initially erroneous images from 20.6% to 64.9% and the GenEval metric from 0.72 to 0.84, with the spatial position sub-score jumping from 0.47 to 0.89. It also shows notable gains on benchmarks such as WISE that were unseen during training, far outperforming applying reinforcement learning directly to the generator. The study confirms that the model itself possesses the ability to judge whether an image is correct or incorrect, and that the role of reinforcement learning is to teach the model to select effective revision paths that can truly correct erroneous images into correct results. The relevant code, models, and training data have all been open-sourced.