Kangqi Li

GLIMPSE: Do Large Vision-Language Models Truly Think With Videos or Just Glimpse at Them?

Existing video benchmarks often resemble image-based benchmarks, with question types like “What actions does the person perform throughout the video?” or “What color is the woman’s dress in the video?” For these, models can often answer by scanning just a few key frames, without deep temporal reason