Tinted Frames: Question Framing Blinds Vision-Language Models

VLMs are selectively blind — they decide how much to look at an image based on question framing (open-ended vs Yes/No vs MCQ), even when the same visual reasoning is required.