Why video AI can't tell left from right—and how to fix it

Video-LLMs struggle with directional motion despite sophisticated architecture—most guess randomly on "did it move left or right" questions. The researchers traced where this breaks down: motion information exists in the model's hidden states but fails to bind to the correct answer. They created MoDirect, a diagnostic dataset, and DeltaDirect, a method that predicts 2D motion vectors from frame differences. The fix improves real-world motion direction accuracy by 22 points without requiring real-world training data, while maintaining performance on standard benchmarks.