A new benchmark study says today's video-understanding AI agents mostly hallucinate because they look in the wrong place, not because they misread what's in frame.
Researchers built a test that separates three steps in a video-agent pipeline: finding when something happens (grounding), watching what's onscreen (visual observation), and reasoning about it. They ran a causal intervention protocol, swapping out one stage's output for a wrong answer while holding everything else fixed, then measuring how much the final answer breaks. Across three separate video-agent architectures, they ran 60,008 of these tests. Grounding errors caused roughly four times more downstream mistakes than errors in the visual-observation stage.
The finding undercuts the standard grading method for these systems, a metric called mIoU that rewards precise time-window overlap rather than pointing to the right region, when the study says picking the right region is what actually predicts reliability. It also found that a wrong answer at any stage does more damage than having no answer at all, meaning a confident but incorrect output is worse than a model admitting it can't find something. That's a case against trusting benchmark scores that look strong without checking whether they predict real downstream failures.
The paper's own audit backs this up: existing benchmark scores don't reliably predict how badly a model's answer degrades under this kind of stress test, and some fall apart once the data shifts away from what the benchmark was built on.