MLLMs Produce 5–19x Shorter Outputs on Image Text Input, Suppressing Reasoning Not Perception | HACKOBAR_