A small AI watches video like a detective — and beats a model 10x its size

Most AI that "understands" video does it the dumb way: it watches every frame, start to finish, whether your question needs all of it or none of it. Ask "what color was the car at minute 12?" and the model still chews through the whole movie. The longer the video, the more it costs — even for a trivial question.

OmniAgent flips this. Instead of watching everything, it acts like a detective working a case. It looks at a piece of the video, thinks about what it found, then decides what to check next — sound, a later scene, a specific moment. Each useful clue gets written down as plain text into a running notebook. So its thinking depends on how hard your question is, not on how long the video runs. A two-hour video and a two-minute clip cost the same if the question is equally simple.

Two training tricks make this work. First, the team had the model practice many possible investigation paths and kept only the good ones to learn from. Second, they rewarded it most for the turns where it actually discovered something pivotal, rather than spreading credit evenly across every step.

The payoff: give it more thinking turns and it gets better — proof the detective approach earns its keep. On the long-video test LVBench, the 7-billion-parameter agent scored 50.5% against a rival ten times larger that hit 47.3%.

That's one benchmark and a narrow margin; "state-of-the-art" here means only among open-source models, not closed commercial ones.

Why you should care: Searching a long lecture or security recording for one answer shouldn't require a giant model grinding through every second. OmniAgent shows a small model that investigates on demand can beat brute force.