How vision-language models are learning to follow long and live video streams: they remember what they have seen, decide for themselves when to respond, and hold real-time audio-visual conversations.
Photo: Umut Tuluoglu
Researchers at Sejong University and NAVER Cloud pared 24,416 video test questions to 11,033 that survived shortcut checks; on those, even the top-scoring model got fewer than half right.