UMM-Reflection trains BAGEL to diagnose and revise its own images
机器之心 reports that researchers from Nanyang Technological University, Shanghai Jiao Tong University, and the University of Tokyo developed UMM-Reflection, which uses full-trajectory reinforcement learning to train BAGEL for multi-round diagnosis and image revision without an external scorer at inference time. The reported GenEval score rose from 0.72 with reflection SFT to 0.84 with reinforcement learning, while the repair rate for initially incorrect images increased from 20.6% to 64.9%. The team released the code, model weights, and training data, although multi-round inference can require four generated images.
The reported results show that full-trajectory reinforcement learning can make a unified multimodal model more reliable at iterative image self-correction while removing external critics from inference, with code, weights, and data released for reproducibility.
机器之心 2026-11-10 23:26 北京
同一个模型写出诊断、再按诊断重画,推理时不再需要任何外部打分器。
统一多模态模型既能看图,也能画图,那么它能不能像人一样,画完之后自己检查、发现问题、再动手改?
来自南洋理工大学、上海交通大学和东京大学的研究者给出了一个肯定的回答。他们提出 UMM-Reflection,在开源统一模型 BAGEL 上用强化学习训练完整的「反思—修改」多轮轨迹:同一个模型写出诊断,再按诊断重画,推理时不再需要任何外部打分器。