Blender-reconstruction benchmark exposes video models' spatiotemporal blind spots

Yolo Y. Tang · hf · 2026-09-15

BVB is a new benchmark that asks agents to programmatically reconstruct real-world videos in Blender. Results show current models achieve high perceptual similarity but fail to retain spatiotemporal facts, exposing a perception-vs-memory gap in video understanding.

Original post →

More from Research

Research channel →