Small Vision-Language Models are Smart Compressors for Long Video Understanding Paper โข 2604.08120 โข Published Apr 9 โข 20