In 2026, "world model" remains one of the least agreed-upon terms in AI.
A system that generates a continuous video gets called a world model. So does a digital environment that changes in real time as a user moves a mouse or presses keys. The same label is also used for systems that predict future states in latent space and for policies that output robot actions directly. All of them process information about the world, but their limits are far apart. A video generator may produce strikingly realistic frames while breaking basic physical rules. A robot may complete a grasp once, yet only under the specific objects, camera views and motion paths it saw in a lab. A demo alone rarely shows whether a system has learned visual correlation, physical structure, or the relation between action and outcome.
That lack of clarity makes technical judgment harder. Image quality, geometric accuracy, prediction ability and task success are often compared as if they measured the same thing. Teams may appear to be racing in one category while actually answering different questions. The discussion has therefore shifted back to a simpler one: what capabilities does a model need before it can move from content generation to understanding and changing the world?
Three ways of defining the term
Fei-Fei Li and World Labs divide world models by function into renderers, simulators and planners, which output pixels, world states and actions, respectively.
Yann LeCun takes a different view. He argues that a system should learn predictable structure in an abstract latent space and retain the information that matters for understanding, reasoning and planning.
Zhu Jun of Tsinghua University offers a third framing. When Motus was released in December last year, he publicly laid out the broader idea of a "general world model." In March this year, he described it as a foundation that connects the digital world and the physical world. A few months later, Zhu and his team expanded that framework in the paper General World Models from First-Principles, defining its core capabilities from first principles and presenting a five-level evolutionary roadmap.
In that paper, the core abilities of a general world model are grouped into understanding, imagination and action. A model needs to form a judgment about the current world from historical observations, project the futures that may follow from different choices, take action, and then use new observations to revise itself. Under that framing, video generation, real-time interaction, latent-state prediction and robot control can be seen as different stages on one capability path.
That also turns the broad question of how far world models are from general intelligence into narrower tests: can the system maintain a coherent world, predict the result of intervention, learn from real feedback, and gradually form goals while organizing more complex action?
First principles, explained with riding a bicycle
The article uses a familiar example: learning to ride a bicycle. At the start, the body wobbles. A rider sees the handlebar tilt, feels the center of gravity shift, adjusts direction, and then keeps correcting based on new feedback. With practice, the brain starts to predict what a given action will lead to.
That, in its simplest form, is what a world model means. In 1943, Kenneth Craik proposed that humans build "small-scale models" of reality in the mind and use them to reason about the outcomes of different actions. The POMDP framework in reinforcement learning describes a similar loop: an agent receives partial observations, estimates the world state, takes an action, and updates its judgment after seeing the next observation.
Zhu’s team condenses that logic into three linked capabilities for a general world model:
- Understanding: integrating vision, language, sound, touch, or robot sensor input into an internal judgment of the current world;
- Imagination: projecting multiple possible futures from the current state, especially counterfactual questions such as what happens if action A is taken instead of action B;
- Action: turning prediction into intervention in a digital or physical environment, and using the resulting observations to test and revise the model.
Together they form a closed loop. Understanding infers the current state. Imagination projects possible futures. Action changes the world and produces evidence for the next round of understanding.
That is also why video generation is not the whole story. A video model can answer what the next frames may look like. To become part of a general world model, it also needs to answer what changes if a condition is altered, an object is moved, or an action is applied. That is where causality, counterfactual reasoning and executable decision-making begin to matter.
A five-level roadmap from world generation to world organization
To describe how world models may evolve, Zhu’s team proposes five levels: L1 world generation, L2 interactive worlds, L3 acting in the world, then L4 autonomous world agents and L5 world organizers.
The article uses a glass pushed to the edge of a table as an example. An L1 model can generate a continuous sequence in which the glass slides, hits the floor and shatters. An L2 model can keep the world evolving after a user changes the viewing angle or the direction of the push. At L3, the model must judge whether the glass is about to fall, predict whether reaching for it will be fast enough, and output a robotic grasp that can actually be executed. Only after the action is taken can it find out whether its estimate of weight, friction and timing was right.
At L4, the system no longer waits for a human command. It needs to detect risk, observe proactively, gather missing information and revise strategy after failure. L5 goes one step further into multi-agent coordination: who should grab the glass, who should avoid an obstacle, who should allocate tools, and how the task should be rearranged after the environment changes.
The point of the roadmap is to turn "world model" from a loose label into a ladder of testable questions:
- L1 asks whether generated trajectories are coherent;
- L2 asks whether the world can keep responding over time;
- L3 asks whether the model can change the physical environment;
- L4 asks whether the system can explore on its own and keep learning;
- L5 asks whether it can organize multi-agent collaboration in open environments.
By the paper’s own assessment, existing systems have reached into the first three levels. L4 and L5 remain open. The main gaps include causal and physical grounding, persistent memory, online learning, efficient deployment and safety control. Evaluation also has to compare prediction with actual outcomes and test transfer across unfamiliar tasks, environments and embodiments.
From video to embodiment: one hypothesis tested twice
The difficulty of world models is not just algorithmic. It is also about data.
Internet video is massive. It records a wide range of objects, people, scenes and motion, letting models learn spatial structure, object permanence, human behavior and the way events usually unfold. But video rarely records the actions, forces and intentions that caused those changes. Robot trajectories connect observation, action and result more directly, yet they are much more expensive to collect and often tied to specific hardware and tasks.
Zhu’s team places those data sources into a pyramid that moves from observation toward action. At the base sits internet-scale video, used to acquire world knowledge and dynamic priors. Above that come domain video, egocentric human video, and human demonstrations with recorded actions. At the top is real robot interaction data. As one moves upward, the data becomes scarcer and more expensive, but the mapping among action, task and embodiment grows clearer.
That framing explains why Shengshu is pushing both video generation and embodied intelligence at the same time. Video provides breadth of world knowledge. Robot data grounds action.
On the digital side, the Vidu series continues to explore generation and interaction in the visual world. Video is not the end point of a world model, but it is a meaningful starting point because it lets a system learn how objects, scenes, motion and events evolve across large-scale spatiotemporal changes. Vidu S1 pushes this further from one-shot generation to real-time response, allowing user input to keep altering what comes next and testing whether the model can preserve state continuity during interaction.
On the physical side, Motus and Motubrain connect scene understanding, state prediction and robotic action in one chain. The model needs to perceive the environment, predict the outcome of intervention, turn its judgment into an executable action, and then face the verdict of the real result.
These product lines become two tests of the same hypothesis. The digital environment tests whether the model can understand and imagine a world. The physical environment tests whether that knowledge can support action and improve through feedback.
MoT and the problem of multimodal coordination
If those two lines of validation are to share underlying capabilities, images, video, language and robot actions must compute around the same world state. Traditional modular systems pass information from stage to stage and can lose geometric, temporal and uncertainty information at the interfaces. Full parameter sharing across all modalities can create training conflict in the opposite direction, letting abundant video data overwhelm scarce robot signals.
To address that, Zhu’s team proposes MoT, or Mixture-of-Transformers. Different modalities keep their own expert parameters, while a shared attention mechanism exchanges context across them. Vision tells the model what is happening in the environment. Language provides goals and constraints. Actions intervene in the environment. Together they update the model’s judgment of the world.
MoT does not solve everything by itself. Physical law, cross-embodiment transfer, real-time inference and safety control still depend on data, training methods and system engineering. The direction, however, is clear: a general world model needs a world state that can be read and written by different modalities and kept up to date over time.
Motus2 and a closed loop that tries to improve itself
If Motus and Motubrain pushed Shengshu’s world-model work into L3 by letting robots generate actions from environment understanding and future prediction, Motus2 adds result evaluation and policy feedback in an attempt to give that action loop a self-improving mechanism.
Motus2 does not separate action generation, future simulation and outcome evaluation into three unrelated systems. It uses one shared-parameter video-action model with three control interfaces:
- Policy / World-Action Model: proposes executable candidate actions;
- Simulator / Action-Conditioned World Model: predicts the visual consequences those actions may cause;
- Evaluator / Value Model: judges which result is closer to the task objective and feeds that judgment back into selection and improvement.
The article describes this as letting the robot rehearse internally before it moves. Instead of seeing an object and outputting one action immediately, the system first proposes several candidates, plays out their possible visual outcomes and task progress inside the model, and then picks the better step. After execution, success, failure and suboptimal results do not just become discarded logs; they serve as evidence for the next round of dynamics modeling and value learning.
Action, prediction, evaluation, feedback and renewed action are folded into one decision-and-learning loop. That is the self-evolving closed loop Motus2 is meant to demonstrate.
On the data side, Motus2 follows a layered training path, starting from monocular egocentric video, extending to synchronized binocular first-person data, and then adapting to the robot body with robot data. Its training uses about 130,000 hours of human egocentric recordings and more than 100 hours of robot trajectories and human-robot alignment data.
The logic is straightforward: first accumulate a common experience of how humans interact with the world, then use robot data to teach the system how this body turns understanding into action.
On the execution side, Motus2 also uses an independent lightweight tactile expert to refine short actions right before execution and to predict post-contact force feedback. Vision is good at locating objects and tracking broad changes in the world. Touch covers the hardest "last inch" when contact actually happens, giving the model another layer of immediate correction for fine operations such as pressing, bumping, grasping and twisting.
Even so, Motus2 is still in the stage of extending from L3 toward L4. Current validation remains focused on specific robots and manipulation tasks. Autonomous goal formation, persistent memory, generalization in open environments, long-horizon online learning and safety control are still unresolved.
Its main advance, as presented in the article, is the integration of policy generation, future simulation, value evaluation and tactile feedback into one system, giving robots a way to preview consequences before acting and revise strategy after acting.
One Chinese research path from probabilistic learning to world intelligence
The general world model framework also extends a longer research line built by Zhu Jun and Tsinghua TSAIL. The article notes that Zhu has long worked on Bayesian methods, probabilistic machine learning, generative models and reinforcement learning, areas that deal with uncertainty, data distributions, future generation and action decisions.
TSAIL has also trained and influenced a group of researchers who later joined leading global teams. Song Yang worked with Zhu on Bayesian learning during his undergraduate years, and Song Jiaming also took part in related work as an undergraduate. They later made notable contributions in score-based generative models and DDIM, respectively.
From probabilistic modeling to generation and decision-making, this academic thread reflects two entry points into AGI. One starts from language and keeps expanding reasoning, tool use and task execution. The other starts from vision and dynamic environments, then moves into real-time interaction and embodied action.
Placed in China’s AGI industry map, the article treats Zhipu and Shengshu as examples of these two routes. Zhipu starts from language models and keeps strengthening reasoning, coding and agent capabilities. Shengshu starts from video generation and extends toward real-time interaction, embodied action and policy optimization. The two routes are also beginning to overlap. Language agents need visual feedback and environment models. World models need language to express goals, rules and plans. Different entry points may still converge on joint modeling of language, world state and action.
From demo quality to the ability to survive feedback
The article closes by arguing that competition in world models will move away from the immediate impact of demos and toward a system’s ability to withstand feedback. Visuals can create a convincing first impression. Geometry and physics determine whether a world can be reused. Action is what brings a model’s judgment into reality. Whether a prediction is correct can only be confirmed after intervention. Whether a task succeeds must also be tested in unfamiliar scenes, over long runs and under changing conditions.
Zhu’s team does not claim to settle the argument over world models. What the paper offers instead is a set of testable questions: what exactly has the model understood, how far can it imagine, and can it revise itself after acting?
Over the past decade, AI’s biggest gains have come from learning patterns in language and data. The next stage, as the article frames it, asks models to face environments that keep changing. Only when a model can keep forming judgments in such environments, bear the consequences of action and continue learning will a world model move from a crowded technical phrase to a foundation for general intelligence.
This article was sourced from the WeChat public account Machine Heart (ID: almosthuman2014) and written by Yang Wen.