Video AI often fails to tell when an event should happen in a sequence.

Video artificial intelligence models often take shortcuts. When testing their sense of time, benchmarks use written text, so models can guess answers using word clues instead of watching the video. To fix this, researchers built TempCloze. They show a model the start and end of a video and ask it to pick the correct missing middle clip from four choices. The clips share scenes and objects to stop easy guesses based on looks. The creators tested 10 proprietary and 21 open-source models across three tasks: what happens, when it happens, and how it unfolds. They found that figuring out when events happen is the biggest bottleneck. Models understand the general action and progression, but struggle to align events correctly in time.

Why it matters

This work matters because video AI systems can be tricked by word patterns instead of truly understanding event timing. If they rely on shortcuts, they might misinterpret videos in tasks like summarization or analysis where sequence is critical.

Questions answered