AI can now scroll through video the way a researcher scrolls through sources.
A team of researchers has published VideoSearcher, a system that lets vision-language models investigate video footage by combining three capabilities in a single reasoning pass: finding the right moment in a clip, zooming in on the right region, and pulling in outside evidence through multimodal search. The goal is what the paper calls Video Deep Research — treating a video not as a closed artifact to describe but as an open-world source to interrogate. To train the system, the researchers built a custom reinforcement learning algorithm called Bi-branch Sequence Policy Optimization (BiSPO), which trains tool use and answer accuracy as separate objectives rather than forcing one loss signal to carry both jobs. They also released VideoSearch-QA, a benchmark built specifically to test this kind of open-world video grounding, because existing benchmarks leaned so heavily on text retrieval that they threw away the visual signal the task actually requires.
Most multimodal search agents today are built around images, not video, which means they lack any mechanism for reasoning across time. VideoSearcher's closed-loop design — where each tool call informs the next — is an attempt to fix that without stitching together separate specialist models. The benchmark matters too: if the field is measuring itself against text-centric tests, it has been grading on a curve.
VideoSearcher outperforms prior open-source baselines on both search-oriented and general multimodal understanding benchmarks, according to the paper — though independent replication on real-world footage is the test that will actually settle the question.