MENTIS Adds Beliefs and Intent to LLM World Models, Tests 8

World models in AI mostly track physics: where objects are, what changes when an action fires. A new paper called Mental World Modeling from Hao Fei and Yiran Zhao argues the reason those models keep mispredicting what a human will do next is that they leave out the mind of the human doing it. Not the scene, the agent inside the scene.

Their proposal is a framework named MENTIS, described as training-free and fully inspectable, that breaks decision prediction into five stages: parse the state, generate the target observation, decompose the action, run a coupled physical-and-mental transition, then evaluate branches. State itself is split into a physical component (entities, attributes, relations, environmental conditions) and a mental component (beliefs, attention, goals, intentions, emotions, preferences, relations, norms, and, evocatively, atmosphere). The claim is that this decomposition is inspectable in a way an end-to-end policy is not.

Why this is interesting if you are building agents rather than benchmarking them: the false-belief and visibility-and-affordance scenes the paper leans on are the exact failure modes that plague partial-observability tasks, where a model watches the same scene as the user and still reaches the wrong action because it assumed the user knew what it knew. The authors tested 8 modern LLM-based world models on a manually constructed dataset of 448 process-annotated situated decisions across text, image, and sounding-video stories, in settings like roadwork sites, neighborhoods, parks, and offices. Reportedly, F1 increases at every rung when the mental channel is added, removing either the mental or physical channel degrades every model, and the authors call transition simulation the largest single bottleneck. They also note that direct answering does not suffice, even with more compute, which is a quiet swipe at the default 'just scale it' response.

The honest caveat is that what the reporting doesn't give you is which eight models were tested, the absolute F1 numbers separating conditions, or how MENTIS compares on latency and token cost against a direct-answer baseline. Menti-Bench is also manually curated at 448 items, so take the ranking as directional, not settled.

If the finding holds up, the beneficiaries are the agent-framework crowd, anyone building planners on top of GPT-class models who has been patching around theory-of-mind gaps with prompt tricks and now has a factored decomposition they can steal and probe.