VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions Paper • 2603.23495 • Published Mar 24 • 3