Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning
UMM-Reflection uses interleaved RL so unified models repair their own image generations.
UMM-Reflection applies reinforcement learning to complete reflection trajectories inside one unified multimodal model, so the same model diagnoses image errors and revises them. Sibling trajectories share an initial image, and one trajectory-level advantage updates both reflection tokens and flow-based revisions without an external verifier at inference. On BAGEL it improves GenEval by 12.05 points over supervised fine-tuning. Gains transfer to WISE (+10.97), OneIG-Bench (+3.48), and T2I-CompBench++ (+4.63), none of which were used in training.
- RL updates both reflection tokens and flow-based image revisions jointly.
- No external critic or verifier is required at inference.
- On BAGEL, GenEval rises 12.05 points over supervised fine-tuning.
- Gains transfer to WISE, OneIG-Bench, and T2I-CompBench++.
- AI models
- BAGELUMM-Reflection
From the original article
1 min read at arxiv.orgUnified multimodal models can both look at and render images, so in principle they can repair their own generations: diagnose what an image gets wrong, revise it, observe the result, and diagnose again. Whether a revision helps is known only after it is rendered, so the reflection text and the image generation must be learned jointly, over the whole loop.
By Yijia Fan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.