Loading video...

Video Failed to Load

Go Home

⚡️ Introducing video world models that know when they don't know! 📽️ Action-conditioned video models can hallucinate a lot, generating videos that look plausible but are incorrect! How can we rigorously quantify their uncertainty, and know when to trust generations?

12,038 views • 11 months ago •via X (Twitter)

4 Comments

Anirudha Majumdar's profile picture
Anirudha Majumdar11 months ago

❌ Hallucinations hinder trustworthy video generation, e.g., for scalable evaluation of generalist robot policies and visual planning! 🚀 Using proper scoring rules, our approach generates dense confidence predictions at the subpatch level that are physically interpretable and aligned with observations.

Anirudha Majumdar's profile picture
Anirudha Majumdar11 months ago

Check out the paper and the project website to find out more! 📃 Paper: 🌐 Website:

Decoded Institute's profile picture
Decoded Institute11 months ago

This is the missing layer: not just generating, but calibrating what the model doesn’t know. Without uncertainty, ‘plausible’ becomes a liability.

Himanshu Kumar's profile picture
Himanshu Kumar11 months ago

Uncertainty quantification in action-conditioned models requires evaluating calibration metrics; the ECE score is 0.12.

Related Videos