Mental World Modeling
The paper argues world models miss human action when they track scenes but not beliefs.
Hao Fei and Yiran Zhao define Mental World Modeling as a framework that keeps physical state and hidden mental state together. Their baseline, MENTIS, breaks the task into observable steps including state parsing, target-specific observation, action decomposition, transition modeling, and value evaluation. In tests across text, image, and sounding-video decision scenarios, eight LLM-based world models performed better when mental state was modeled explicitly. HF Daily Papers' note
Hao Fei and Yiran Zhao define Mental World Modeling as a framework that keeps physical state and hidden mental state together. Their baseline, MENTIS, breaks the task into observable steps including state parsing, target-specific observation, action decomposition, transition modeling, and value evaluation. In tests across text, image, and sounding-video decision scenarios, eight LLM-based world models performed better when mental state was modeled explicitly. HF Daily Papers' note
score 4