Video AI can describe a scene. It still struggles to notice what changed.
Researchers introduced DELTAVID, a training framework that feeds video multimodal large language models a specific kind of problem: find the difference between two nearly identical clips. The task forces the model to pinpoint where a change happened, when it happened, and collect spatial evidence — rather than relying on broad scene-level pattern matching. To make the approach trainable at scale, the team built DELTAVID-10K and DELTAVID-Bench, datasets that embed controllable, labeled differences into real video footage for both training and evaluation.
The gains carry beyond the spot-the-difference task itself. Models trained with DELTAVID improved across eight established video understanding benchmarks — MMVU, MLVU, Video-MME, VideoHolmes, VideoMMMU, LVBench, TempCompass, and LongVideoBench — suggesting that fine-grained local reasoning is a transferable skill, not a narrow trick. That is the more interesting finding: a proxy task designed for one failure mode appears to strengthen general spatiotemporal reasoning.
Current video models are already fluent at summarizing clips; the gap is localization and evidence. DELTAVID is a research result, not a shipping product, but the benchmark sweep makes it harder to dismiss as incremental — eight is a lot of leaderboards to move at once.