Oxford and NUS Propose Next-Generation World Models: Mental World Modeling Emerges

Imagine this scenario: Person A puts a cup into a cabinet, and Person B does not see it.

Question: Where will B look next for the cup?

If the world model only knows physical facts, it will incorrectly predict B's next decision: "The cup is in the cabinet now, so I should go look in the cabinet." (Above)

But a model that also understands the human mind can correctly predict: "B doesn't know the cup was moved, so believes it's still on the table and will likely return to the table to look for it." (below)

The world's next state is also determined by mental variables.

This reveals a key issue: a world model that only tracks objects, locations, and motion may produce predictions that are physically plausible but behaviorally incorrect.

Recently, a research team from the University of Oxford and the National University of Singapore introduced a novel framework called Mental World Modeling (MWM) in their latest study, “Mental World Modeling.” This work integrates these mental variables directly into the world state, rather than treating them merely as post-hoc explanations.

• Paper Title: Mental World Modeling
• Paper link: https://arxiv.org/abs/2607.27201
• Project homepage: https://mental-world.github.io/
• Open-source code: https://github.com/mental-world/Mentis

As of press time, MWM has ranked #1 on Hugging Face's Daily Papers.

Hugging Face paper: https://huggingface.co/papers/2607.27201

Most existing world models primarily focus on the physical aspects of the world:

• What objects and agents are present?
• Where are they?
• How will the visible scenarios evolve?

In these models, humans are often merely objects that move and perform actions, with their internal states not truly incorporated into the world model. However, in the real world, people are not simple moving objects.

This is far from sufficient for human-centered intelligence.

Why? Because human behavioral decisions are not determined solely by external environments such as objects, spatial structures, and physical laws, but rather emerge from the interaction between external environments and internal psychological-social factors.

For example, a service robot must determine whether a user is confused, losing patience, or seeking help indirectly; a medical assistant must consider the patient’s cognition, risk perception, fear, and level of trust; and a collaborative agent must recognize that certain actions, while physically possible, may be inappropriate due to social norms, role relationships, or the context of interpersonal interaction.

These scenarios require intelligent systems to continuously track agents’ mental and social states, including what they know, believe, attend to, desire, intend to do, feel, and consider socially acceptable. Even if two scenarios are physically identical, differences in people’s beliefs, goals, emotions, relationships, or social responsibilities can lead to entirely different behaviors.

Even if a world model can accurately reconstruct physical scenes, it may still fail to correctly predict human behavior.

Cognitive science has investigated these abilities through theories such as mental models, theory of mind (ToM), the BDI agent model (Belief–Desire–Intention), and embodied cognition.

However, most current artificial intelligence research either builds physical world models lacking mental states or reduces mental reasoning to isolated theory-of-mind question-answering tasks.

Neither of these perspectives is sufficient to describe a truly complete world—because the next state of the world is not determined solely by physical factors, but emerges from the joint evolution of physical and mental states.

How can mental variables be genuinely incorporated into the state space of a world model?

This study attempts to propose a formal framework for mental world modeling, aimed at constructing world models that simultaneously represent physical and mental dynamics.

The goal of mental world modeling is not to simulate an individual's private subjective consciousness and experiences, but to construct an external, approximate, task-relevant world simulator. The global state of this simulator includes both physical variables and mental variables.

Two entities, three operational modules, and one physical-mind coupling state. The mental world model, functioning as a global simulator, first generates the information that the target agent can perceive and infer, then enables the agent to take actions based on this cognition.

First, mental world modeling maintains a coupled physical–mental world state, simultaneously representing the physical environment and the agent's latent mental state.

Second, a target agent can only observe a partial, first-person rendering of the global state. For each target agent, MWM generates a localized observation perspective—namely, the information that this agent can actually see, hear, know, and infer;

Subsequently, the target agent takes an action based on this observation, while the world model simulates how this action will simultaneously alter:

1. Objective scenarios in the physical world;
2. Mental-social configuration in the mental-social world.

To make this framework verifiable, the team implemented MENTIS, a modular reference system that is training-free and fully observable, forcing LLM-based systems to reason in the manner of a mental world model.

MENTIS process.

The system first converts the input scenario into a structured current state.

Generate observations for the target pseudo-agent

Parse each candidate option as a behavioral branch, parallelly simulate the subsequent changes in physical and mental states, and finally evaluate different future outcomes to select the final action.

Given a contextual scenario, a target agent, and a set of candidate behaviors, MENTIS decomposes the decision prediction process into a series of distinct stages: state parsing → goal observation generation → behavior decomposition → physical-cognitive coupled state transition simulation → branch-level evaluation → final decision.

• First, it parses the scenario into physical–mental states with a typed structure;
• Then, generate the partial observation that the target agent can obtain from that state;
• Next, break down each candidate behavior into its corresponding physical-level impact and mental-level impact;
• Then, simulate the next state of the physical–cognitive coupling for each possible behavioral branch;
• Finally, score each branch along three dimensions—physical plausibility, mental consistency, and social normativity—and select the final action using deterministic rules.

In other words, MENTIS no longer allows the model to jump directly from scenario to behavior prediction; instead, it requires the model to explicitly represent: the current state of the world, how the target agent perceives the world, how different actions alter physical and mental states, and which behavior is most reasonable both physically and socially.

MENTIS requires no training, so its results reflect the reasoning capability inherent in the model architecture itself, rather than parameters learned through fitting. Each stage in the system produces intermediate outputs that can be machine-verified, enabling states, observations, and simulated futures to be recorded, compared with human annotations, or directly replaced with ground-truth labels for analysis.

Researchers designed a series of experimental groups to specifically test potential failure points of the framework, and the results revealed consistent patterns.

First, explicit mental world modeling is necessary for predicting human decisions.

Necessity阶梯 (Final behavior F1, 448 records). (a) F1 scores of eight world models across different necessity stages (thin lines represent individual model results, thick line represents the average result). (b) S6 ablation results across eight models (white dots indicate values for each model).

Among the eight large language model (LLM)-based world models tested, the average F1 score for direct answers was 63.3; the full MWM configuration achieved the best performance at 87.9. Removing the mental channel caused a performance drop of approximately 12.1 points across all models, removing the physical channel resulted in a drop of 16.5 points, and separating the prediction of state transitions led to a decline of 6.4 points.

F1 scores of four experimental setups across different scenario categories

Moreover, the greatest performance improvements occur in interpersonal interaction scenarios, where decisions rely primarily on hidden mental variables. For GPT-5.6-SOL, the full MWM increases the final action F1 score from 66.5 with direct responses to 92.9, an improvement of 26.4 points. This result comes from a controlled test in the paper but aligns with the study’s motivation: when beliefs, roles, and relationships govern the next step, physical variables alone cannot account for actions.

So, what is the real bottleneck?

In addition to demonstrating its necessity, the paper conducted a more targeted diagnostic: replacing a specific intermediate step with correct information (Oracle intervention) to measure how much the final prediction improves, thereby further attributing the differences between the model and humans.

Oracle intervention experiment (gpt-5.6-sol). Final F1 score after replacing model predictions with ground truth labels in one or more stages; bar labels indicate performance gains relative to the full prediction version S6.

The experimental results show that for gpt-5.6-sol, the full MWM is 90.7; using ground-truth state transitions improves it to 94.2, yielding a +3.5 performance gain—the largest among individual interventions. Simply by intervening on state transitions, the model can recover 45% of the 7.8-point gap to human performance.

The results indicate that the current largest remaining bottleneck is state transition simulation—the model's insufficient ability to predict how the physical-mind coupled world changes. This provides a clear direction for future improvements to the MWM system.

The application value of MWM should not be measured by "how many social domains it can list." After all, virtually all systems designed for humans are, in a broad sense, inherently "social."

A more accurate and meaningful criterion is:

MWM has true value when the validity of an action depends on variables that cannot be inferred solely from the physical scenario but determine how humans perceive, choose, accept, resist, or learn.

From the perspective of decision theory, the value of mental state information can be defined as:

The expected utility gap between the optimal intervention chosen when the physical–mental coupling state is accessible and the optimal intervention chosen based solely on the physical state.

The gap is significant when beliefs, goals, attention, trust, obligations, emotional states, or social norms alter what constitutes beneficial behavior; it is small when the next action is almost entirely determined by physical feasibility.

For example, a mechanical gripper controller does not require MWM to perform the action of "reaching for a cup"; however, a home service robot deciding whether to "hand the cup to its owner" must understand the underlying mental and social context—whether this action constitutes assistance, disturbance, rudeness, risk, or miscommunication.

This work proposes a formal framework for mental world modeling, aimed at constructing world models capable of simultaneously representing physical and mental dynamics. The core contributions can be summarized in three points:

Propose mental world modeling: a formal mathematical framework that integrates physical dynamics with mental dynamics under a unified perspective on world modeling, enabling AI systems to not only reason about how the environment changes, but also understand how agents perceive, interpret, and mentally respond to these changes.

Introducing MENTIS: A training-free benchmark implementation that realizes and operationalizes the MWM framework through modules including state parsing, observation generation, behavior decomposition, physics-mind coupled state transitions, and branch-level value evaluation.

Conduct systematic empirical research, yielding two core conclusions from the experimental results:

Explicit mental world modeling is necessary for predicting human decisions, especially in social interaction and interpersonal relationship scenarios;

The current key bottleneck in the MWM system is insufficient capability for state transition simulation, which also points to directions for future improvements.

This article is from the WeChat public account "Machine Heart".