New AI Finds the Key Moment in a Long Video

Have you ever tried to find one specific moment in a long video? Maybe the goal of a game or a face in a crowd? Computers struggle with this too. Even the newest AI video tools often pick the wrong clip when asked a question about something really long. Why? Because they find clips that are broadly related to the question, but miss the tiny detail that proves the correct answer over a similar-sounding wrong one.

Researchers came up with a clever fix called PACE, which works like a detective. First, it quickly watches the entire video and writes down a short description of each small clip. Then, when you give it a multiple-choice question, it uses those answer choices as "clues" and goes back to find the exact clip that supports the right answer. It's the difference between looking for "a car" in a parking lot and looking for "a red sedan with a dent on the left bumper."

In tests, PACE beat previous methods on several video question-answering challenges. On one benchmark called MMR-V, it got 42.6% accuracy, a clear improvement over older agentic systems that could find relevant footage but couldn't pick the correct answer. It also recovered 66.9% of the important clues that human experts had identified, showing it truly finds the right evidence, not just a good guess.

This matters because AI that understands long videos could save people real time. Imagine asking an AI to find every scene in hours of security video where a certain person appears, or having it pull the key play from a game that lasted three hours. It's not perfect—it still works best with multiple-choice questions and can be fooled by poor video quality—but it's a big step toward AI that actually watches TV, sports, and surveillance footage for us.