Research
Study finds video world model prediction error does not reliably measure physical surprise
Researchers tested V-JEPA2 and other video world models on IntPhys2 and Kubric, finding that impossible events can yield lower prediction error than possible ones. A fixed error threshold also wrongly flags many possible events as impossible.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
When Does Latent Prediction Error Support Physical Surprise Claims in Video World Models?