Mental World Modeling

Researchers propose Mental World Modeling (MWM) to predict human decisions by tracking both physical scenes and agents’ hidden mental states.

MWM extends world models to represent beliefs, intentions, desires, and social constraints, not just visible environment dynamics.

They also introduce MENTIS, a training-free baseline that parses scenarios, generates target-specific observations, decomposes actions, and evaluates branches.

Across curated decision scenarios in text, images, and video, modeling mental state improved prediction of human choices.

The work suggests world-model research may need to shift from simulating scenes to simulating minds for better human-centered planning and prediction.