Mental World Modeling: Extending World Models to Beliefs, Intentions, and Social Norms
Original title:Mental World Modeling
AI Summary
This paper proposes Mental World Modeling (MWM), a framework that treats beliefs, desires, intentions, feelings, and perceived social permissibility as first-class components of a world model. MWM couples physical and mental state, generates target-specific partial observations, and simulates how candidate actions update both. The authors instantiate it in MENTIS, a training-free and inspectable baseline with explicit stages for parsing, observation generation, action decomposition, transition modeling, and branch evaluation. Experiments on a manually constructed, quality-controlled dataset spanning text, images, and sounding-video stories use eight modern LLM-based world models and report that explicit mental-state modeling is important for predicting human decisions.
Why it's worth reading
The work reframes world modeling around what different agents know, want, and intend, giving researchers a concrete framework and inspectable baseline for evaluating social reasoning, collaboration, and embodied planning.
Deep Read
What Happened
Original facts: The paper introduces Mental World Modeling (MWM), which makes mental variables core state components of a world model instead of post-hoc explanations of physical predictions. It also presents MENTIS, a training-free and fully inspectable baseline.
Core Tech
Original facts: MWM maintains a coupled physical-mental world state, renders a target-specific partial observation, and simulates how candidate actions jointly update physical and mental components. MENTIS decomposes this into state parsing, target-observation generation, action decomposition, coupled transitions, and branch-level value evaluation.
Key Evidence & Numbers
Original facts: The evaluation uses a manually constructed, quality-controlled dataset covering text, image, and sounding-video stories. Eight modern LLM-based world models are evaluated. The abstract reports that explicit mental-state modeling is essential for predicting human decisions and describes additional analyses of current bottlenecks.
Why It Matters
Analysis: Physical scene prediction alone cannot capture actions driven by asymmetric information, false beliefs, intentions, or social norms. MWM offers a unified formulation in which an agent’s observable environment and its internal representation of that environment are simulated together.
Practical Impact
Analysis: The framework could support multi-agent collaboration, robot interaction, social simulation, educational assistants, and narrative decision evaluation. MENTIS’s inspectable stages may help separate failures in parsing, observation modeling, mental transitions, and value estimation.
Limitations & Uncertainty
Original facts: The abstract does not report dataset size, exact task metrics, the identities of the eight models, or improvement magnitudes, and it does not provide the full error analysis. Analysis: Manually authored scenarios may limit distributional representativeness, while persistent mental-state tracking in real environments remains substantially harder than story-based evaluation. Unverified inference: MENTIS being training-free does not establish lower inference cost or better out-of-distribution generalization than trained alternatives.
Original Sources
- arXiv:2607.27201
- Source: Hugging Face Papers
- Publication date: 2026-08-03