Videos are flat two-dimensional recordings of a three-dimensional world, so working out where objects actually sit and how they move relative to each other over time is hard for , the systems that answer questions about images and video in words. That combination of three dimensions plus time is what the field calls 4D.
The approach trains such a model to first write down, for each moment in the clip, where the relevant objects and the camera are in space. It then reasons about how those states shift from one moment to the next and about what the footage visibly shows, treating both as evidence before giving an answer. Training used video from the wild with hand-checked state annotations plus question-answer pairs.
On two question-answering tests the model answered more accurately than the base models it was built from, and removing the state-change reasoning or the geometry-aware training step made results worse. That suggests explicit, ordered geometry earns its place as a stepping stone rather than being jumped over.