22 August 2026

World models miss how humans think, leading to wrong action predictions

• Existing world models, including Sora and Genie, simulate physical scenes but ignore mental states like beliefs, desires, and social norms that drive human behavior.
• Researchers created Mental World Modeling, a framework that tracks both what happens physically and what people think, want, and intend during interactions.
• A test system called MENTIS using this framework reached 87.9 percent accuracy on predicting the right action, compared to 63.3 percent for models answering directly.
• The biggest accuracy gains appear in scenes involving multiple people, where mental modeling improved predictions by 26.4 points versus only 14.0 points for object-focused scenes.
• Humans solved the same test at 98.5 percent accuracy, with most remaining errors traced to difficulty simulating how physical and mental states change together.

Reported by The Decoder