A new framework called Mental World Modeling argues that AI systems cannot reliably predict human actions by tracking physical scenes alone. It adds hypotheses about beliefs, attention, goals, intentions, emotions, social norms, and relationships while explicitly treating those hidden states as uncertain rather than as measurements of consciousness.
The researchers built MENTIS, a training-free reference pipeline that parses a scene, renders an agent’s partial perspective, separates possible actions into physical and mental components, and simulates their effects. It scores each branch for physical plausibility, mental consistency, and social appropriateness. Machine-readable intermediate outputs are intended to make errors traceable.
Evaluation used Menti-Bench, which contains 448 decision scenes across text, images, and audio-video. Across eight tested language models, direct answers averaged an F1 score of 63.3, repeated-answer self-consistency reached 77.9, and the full pipeline reached 87.9; humans scored 98.5 under the same protocol. GPT-4.1 with the framework scored 84.9, slightly above GPT-5.6-Sol using direct answers with self-consistency at 83.6.
Removing either the mental or physical channel reduced scores by double digits. About 80% of the remaining gap was attributed to intermediate prediction errors, especially state transitions. The results indicate that structured mental modeling can improve behavior prediction, but accurately simulating how physical and inferred mental states change together remains the central bottleneck.