BAGEL

Researchers from NTU, Shanghai Jiao Tong University and the University of Tokyo propose UMM-Reflection, which uses reinforcement learning to train full multi-turn reflect-then-revise trajectories on the open unified model BAGEL, letting a single model write its own diagnosis and redraw accordingly with no external scorer at inference. Repair success on initially flawed images rises from 20.6% to 64.9%, GenEval improves from 0.72 to 0.84, the spatial-position sub-score jumps from 0.47 to 0.89, and the unseen WISE benchmark gains 11 points. Code, models and training data are open-sourced.