New video test reveals AI models struggle with timing details
Original: TempCloze: Can Video-LLMs Identify the Missing Middle?
Source: arxiv.org ↗
Who: Wenqi Pei and six collaborators from academic institutions posted this paper on arXiv, presenting work accepted as a Findings paper at EMNLP 2026, a major conference on natural language processing.
What's new: The researchers created a new test called TempCloze to check whether video AI models truly understand the order of events in videos, rather than guessing from language clues. The test shows a video's beginning and ending clips, then asks the model to pick which of four middle clips actually happened between them.
How it works: Earlier tests let AI models cheat by using wording in questions or common language patterns to pick answers without really watching the video. TempCloze removes that advantage by using only visual content. The test includes 1,521 videos from seven sources, mainly long continuous shots and first-person footage. The wrong answer options are deliberately tricky in three ways: some show a plausible but wrong event, some show the right kind of event at the wrong moment, and some test whether events unfold in the correct sequence. All options share similar scenes and objects, forcing the model to rely on timing rather than appearance.
The numbers: The team evaluated 31 different video AI models, including 10 proprietary ones and 21 open-source ones. They found that models fail most often at alignment, meaning they struggle to determine when events should occur, even when they can recognize what is happening and how individual moments progress. They also ran deeper analyses on four representative models, looking at how changing the order of answer choices, showing more or less of the video, and varying frame density affected performance.
Why it matters: This work exposes a key blind spot in current video AI systems. Being able to name objects and actions in a video is not the same as understanding the timeline of events. The finding suggests that progress in video understanding may require new training approaches focused specifically on temporal alignment, rather than just more data on visual recognition.
Caveats: The test focuses on long-take and egocentric videos, so results may not generalize to heavily edited footage with cuts and scene changes. The paper does not report specific accuracy numbers in the abstract, making it hard to gauge how large the alignment gap is across model categories.