arxivcs.CV2026-07-10
Model Guides You How to Draw: Adaptive Visual Gating for Unified Multimodal Reasoning
Wenxi Gao, Guanxi Lu, Didi Zhu, Hao Mark Chen, Quan Deng, Zhican Wang, et al.
Unified multimodal models (UMMs) with interleaved reasoning, which generate both textual and visual steps as part of intermediate reasoning traces, have demonstrated great potential for visual mathematical reasoning tasks. However, we identify a key insight in this paradigm: gene…