Unified multimodal models can both view and render images, which in principle allows them to repair their own generations: identify what is wrong in an image, revise it, inspect the result, and diagnose again. A new paper posted on Hugging Face, Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning, tackles this self-repair loop directly.
The authors note a core difficulty: whether a revision actually helps is only known after the image is rendered, so the reflection text and the image generation must be learned jointly across the entire loop. Supervised fine-tuning on reflection trajectories offers a cold start but does not discover high-success repair paths, while naive reinforcement learning that optimizes only the renderer or a single head leaves much of the potential gain unused.
Their proposed method, UMM-Reflection, applies reinforcement learning to complete reflection trajectories within one unified model. Sibling trajectories share a single initial image, so a group-relative advantage can compare different reflection strategies, and one trajectory-level advantage updates both the reflection tokens and the flow-based revisions. This design, the authors say, avoids the combinatorial blow-up of assigning credit round by round.
Unlike single-round editing or pipelines relying on an external critic, credit in UMM-Reflection flows across rounds and to both roles of the same model, and no verifier is required at inference time.
On BAGEL, the method improves GenEval by 12.05 points over SFT. The authors report that the gains transfer to benchmarks not used in training: WISE (+10.97), OneIG-Bench (+3.48), and T2I-CompBench++ (+4.63).