arXiv:2611.10552v2 Announce Type: replace Abstract: Multimodal Large Language Models (MLLMs) inherit strong textual reasoning capabilities from their language backbones and text pretraining, but often fail to apply the same level of competency to visual inputs.
