AI · 1h ago
AI Video Models Fail to Track Off-Screen Events, New Benchmark Shows
A new benchmark called WRBench tests whether video AI systems can track objects and changes that occur off-screen. Across 23 models, none reliably tracked off-screen events, and larger models performed worse. The findings highlight a fundamental architectural gap in current video generation models.
Meridian48 take
The scaling failure is a red flag for AI labs betting on larger models to achieve world understanding—fluency and true modeling are not the same thing.
video-aiworld-models