Adaptive Visual Token Reduction for Accelerated Image Understanding

Large Vision-Language Models achieve strong VQA performance, but processing high-resolution, information-rich images requires substantial computation, motivating visual token reduction.