What Is a World Model? The Four Major Schools and How to Evaluate Them
The term "world model" is used inconsistently across AI research, with different communities defining it in fundamentally different ways. This article traces the concept's eighty-year history—from psychologist Kenneth Craik's 1943 idea of a small-scale model of reality held in the brain, through its roots in cognitive science, to its modern revival driven by deep learning, generative modeling, and reinforcement learning. It organizes today's work into four schools—video generation, abstract latent-space prediction, 3D spatial understanding, and reinforcement learning planning—explaining the core claim, strengths, and limitations of each. Finally, it presents six dimensions for evaluating a world model: predictive fidelity, causal consistency, long-horizon stability, decision usability, abstraction and interpretability, and data and compute efficiency, along with the trade-offs among them.
1. A Widely Used Term Without a Widely Shared Definition
If you pick up a handful of AI survey papers published after 2024, you will most likely run into a candid admission: the term “world model” is used in a highly inconsistent way, the literature is deeply fragmented, and different research communities define it in their own terms.
The root of the problem is that the label “world model” carries enormous expectations while offering far too many paths to deliver on them. One camp holds that a model which can generate sufficiently realistic video has effectively built a “world simulator,” because it can imagine the many scenarios of the physical world. Another camp argues the opposite: this path is fundamentally unworkable, and genuine world understanding should take place in an abstract space, used to predict the causal relationships between things. A third insists that the real world is three-dimensional, and staying at the level of two-dimensional pixels falls far short. Meanwhile, in the reinforcement learning community the criteria become strictly practical: if it helps an agent plan, it qualifies as a world model.
The same word is thus burdened with four definitions, four methodologies, and four sets of evaluation criteria. The first step to understanding world models is to accept that it is not a single-point concept but a family of methods with family resemblances. The goal of this article is to lay out that map: where it came from, which schools it has split into, the approach and representative ideas of each, and how we should actually measure it.
2. An Eight-Decade Conceptual History: From Psychology to Reinforcement Learning
World models are not a term invented from nothing in 2024. They have a conceptual history spanning more than eighty years, moving through psychology, symbolic AI, and reinforcement learning. Each stage added a new layer of meaning to the word.
The Origin in Psychology: A Small-Scale Model in the Brain
The earliest intellectual source can be traced to 1943. In his book The Nature of Explanation, the British psychologist Kenneth Craik wrote a passage that has been quoted ever since: if an organism carries in its head a “small-scale model” of external reality and of its own possible actions, it can try out various alternatives, conclude which is the best, react to future situations before they occur, and use knowledge of the past to deal with the present and the future.
Craik was a brilliant but short-lived thinker; he died in 1945 in a cycling accident at the age of only 31. Yet his metaphor of “a miniature model of external reality inside the brain” eventually settled into one of the core assumptions of cognitive science. The flourishing “mental model” theories of cognitive psychology in the 1970s and 1980s essentially all start from Craik’s insight.
It is worth noting that when Craik proposed this concept, the term “artificial intelligence” did not yet exist—that term was not formally coined by John McCarthy until 1956, and the noun “world model” was nowhere in sight. But Craik in effect set an initial goal for later AI: to give machines an internal model akin to the human brain, so that they can predict the consequences of their own actions. Today almost every paper discussing world models cites this intellectual origin in its introduction.
From Metaphor to Engineering Goal
In the era of symbolic and connectionist AI, the “internal model” gradually shifted from a philosophical metaphor into a computable object. What has truly brought world models back into the spotlight in recent years is the maturation of deep learning, generative modeling, and reinforcement learning: for the first time, they made “building a predictable, imaginable model of the environment inside a machine” an engineering-feasible task. And so, watered by different technical traditions, the same ancient metaphor has grown into stylistically distinct branches.
3. The Four Major Contemporary Schools and Their Methodologies
The most effective way to understand the current disagreements is to group the main research orientations into four schools. It is worth stressing that these four are not strictly mutually exclusive, and real research often borrows across them.
School 1: The Video Generation School—Generative Prediction as World Simulation
The core claim of this school is that if a model can generate sufficiently realistic, temporally coherent video, it has to some degree “simulated” the world and can therefore be regarded as a world simulator.
Its strength is high visual fidelity: it can directly present complex scenes, object interactions, and lighting changes, and is easy for humans to observe and verify. Its limitations are equally clear: it largely stays at the pixel and two-dimensional level, cannot guarantee the long-range consistency of physical laws, and struggles to explicitly express causal relationships and three-dimensional structure. When sequences are stretched very long, problems such as error accumulation, loss of object permanence, and inverted causality quickly surface.
School 2: The Abstract Latent-Space Prediction School—Predicting Causality in an Abstract Space
In contrast to the video generation school, this one argues that predicting directly in pixel space is both expensive and fragile, and that genuine intelligence should operate in a highly abstract latent space. The model needs to strip away irrelevant detail and focus on capturing the causal relationships between states, so as to predict at a lower cost and in a more essential way.
Its philosophical grounding traces back to the belief that “intelligence requires understanding how the world works.” The strengths of latent-space prediction are high data efficiency and great potential for interpretability; the difficulty lies in defining “what the correct abstraction is”—if the granularity and structure of the abstraction are poorly designed, the model may learn representations that are irrelevant to, or even harmful for, decision-making.
School 3: The 3D Spatial Understanding School—The World Is Three-Dimensional
This school emphasizes a simple stance: the real world is fundamentally three-dimensional, and if a model stays only at the level of two-dimensional video, it will lose a large amount of geometric and spatial information. A world model should therefore include explicit modeling of three-dimensional structure, changes in viewpoint, and spatial relationships.
Its approach is to reconstruct and understand spatial geometry, enabling the model to infer a consistent 3D scene representation from multi-view, multi-modal observations. Its value lies in stronger geometric consistency and spatial reasoning ability, at the cost of a heavier dependence on multi-view data, reconstruction accuracy, and representation design.
School 4: The Reinforcement Learning Planning School—If It Works for Planning, It Is a Good Model
In the reinforcement learning tradition, the criteria are the most pragmatic: whether a world model qualifies depends on whether it can be used to plan within “imagination” and thereby improve decision-making performance in the real environment. This school typically has the agent simulate future trajectories inside a learned model of the environment’s dynamics, performing look-ahead search and policy optimization, so that it can learn efficiently even when samples are limited.
Its biggest advantage is that it is directly tied to decision-making and markedly improves data efficiency; its main challenge is that model bias gets amplified by the planning process—once the internal model misestimates the true dynamics, the “beautiful future” it imagines may be badly distorted.
4. How to Evaluate a World Model
Precisely because the schools are so numerous, discussing which is better or worse apart from evaluation criteria is meaningless. The following dimensions can serve as a general evaluation framework:
- Predictive fidelity: how close the generated or predicted result is to real observations, covering both the visual level and the state level.
- Causal consistency: whether the model has learned correlation or transferable causality, and whether it stays stable under distribution shift.
- Long-horizon stability: whether error remains controllable as the sequence lengthens, and whether object permanence and physical laws are maintained.
- Decision usability: whether predictions can be effectively used by downstream planning or control modules, thereby improving performance on real tasks.
- Abstraction level and interpretability: whether the internal representation sits at the pixel, feature, or semantic level, and whether it is easy for humans to understand and intervene in.
- Data and compute efficiency: the number of samples and amount of compute required to reach the same level of prediction or decision-making.
It is worth noting that there is tension among these dimensions: pursuing visual fidelity often sacrifices abstraction and causality, while pursuing extreme efficiency may damage long-horizon stability. Evaluating a world model is essentially about how it trades off these metrics for a given task.
5. Conclusion: Ask “Which School?”
To recap, the world model is a conceptual family with an eighty-year history that spans multiple disciplines and has now split into several methodologies. It has both Craik’s psychological origin and the contemporary engineering divergence into four schools—video generation, abstract latent-space prediction, 3D spatial understanding, and reinforcement learning planning—along with a whole set of multidimensional evaluation criteria.
So the next time you see the word “world model,” the most useful thing to do is to ask one question: which school do you mean, and what is its evaluation standard? Only when these two points are clarified can a discussion of world models truly become substantive, rather than getting stuck in the interchangeable use of terminology.
