Silent Hazards of Token Reduction in Vision-Language Models: The Hidden Impact on Consistency

Open in new window