MemoBench: World Models Fail Object Permanence Test
机器之心 · wechat · 2026-07-06
Researchers from Harvard, MIT, Google, CMU, and IBM have proposed MemoBench—the first "Visible-Disappeared-Reappeared" (V-D-R) world modeling evaluation benchmark for dynamic environments, accepted by ECCV 2026. Using 360 high-quality ground truth videos, it systematically tests whether 10 mainstream video world generation models can remember an object's identity, deduce its state changes while out of view, and accurately restore it upon reappearance.
Results show that no model scored above 0.6 (out of 1) in "object reappearance score," indicating that current video generation models, despite producing coherent visuals, almost entirely fail to correctly restore the state changes that should have occurred during occlusion when the object reappears. The research highlights a significant gap between "generating realistic visuals" and "truly understanding the world," providing a quantifiable diagnostic metric (object permanence capability) for next-generation world models.
Related event: MemoBench Reveals Top Video Models Fail at Object Permanence(2 posts)→
More from Multimodal
- A fine-tuned Krea 2 raw model produced a rainy-night driving scene — darlens13 · 2026-07-27
- Users ask whether Video2X can load custom OpenModelDB models — Used-Profit2355 · 2026-07-27
- A builder wants AI to reverse-engineer viral video effects into ComfyUI workflows — stale2000 · 2026-07-27
- Midjourney V8.2 adds personalization and shows off stylized image outputs — Mr_AllenT · 2026-07-27
- Midjourney’s image variety draws a Krea 2 comparison and asks how to reproduce it — diffusion_throwaway · 2026-07-27
- AI short film sets a 1985 dystopia to music and leans into cinema — ProfessorKey98 · 2026-07-27