Computer Vision Explainable AI TAB: Transformer Attention Bottlenecks enable User Intervention and Debugging in Vision-Language Models May 30, 2025