Why video AI misses the crucial split-second moments

Video language models ace long-form understanding but stumble on split-second events—a quick action, a state change, a single-frame detail. Researchers created Moment-Video, a 1,000-question benchmark exposing this blind spot across 33 models. Even top performers hit only 39.6% accuracy when forced to spot, count, or reason about transient visual evidence. The problem: sparse frame sampling and temporal compression skip or blur the evidence before language reasoning can help.