New Benchmark Finds Video AI Models Fail to Track Off-Screen Events

A new benchmark called WRBench, published in June 2026, tested 23 video AI models across nearly 10,000 clips to evaluate whether they can accurately represent what happens in a scene when the camera looks away. The study found that current video generation systems consistently fail at this task, resetting off-screen objects to their original positions rather than reflecting logical changes. Notably, scaling models to larger sizes made the problem worse, not better — bigger models produced more visually convincing frames but were less accurate about off-screen continuity. Researchers attribute this to a fundamental architectural gap: video models are trained to render visible content convincingly but lack any persistent internal representation of world state beyond the camera's current view. Four independent research groups published related findings in the same month, all converging on the conclusion that this off-screen tracking failure is a structural limitation with significant implications for AI systems like robots and autonomous vehicles.