In 2026, “world models” is one of the most contested terms in the AI community.
A video that can be continuously generated is called a world model; a digital environment that changes in real time with mouse and keyboard inputs is also called a world model; systems that predict future states in latent space, as well as policies that directly output robot actions, use the same name.
These models all process information about the world, yet their capabilities vary widely. The visuals generated in videos can be highly realistic but still violate physical laws; robots may successfully perform a single grasp but only be familiar with objects, camera angles, and motion trajectories within a lab setting. It is difficult to determine whether a system has learned visual correlations, physical structures, or the relationships between actions and outcomes—just by looking at demos.
Conceptual confusion increases the difficulty of technical evaluation. Image quality, geometric accuracy, predictive capability, and task success rate are being compared interchangeably; different teams appear to be competing in the same space, yet they may actually be answering entirely different questions.
Therefore, research on world models is returning to a fundamental question: what capabilities must a model possess to move from generating content to understanding and altering the world?
Li Fei-Fei and World Labs classify world models by function into renderers, simulators, and planners, which output pixels, world states, and actions, respectively.
Link: https://www.worldlabs.ai/blog/taxonomy-of-world-models
LeCun advocates learning predictable world structures in an abstract latent space, enabling models to retain information useful for understanding, reasoning, and planning.
Link: https://ai.meta.com/research/vjepa/
Zhu Jun from Tsinghua University offered a third perspective. As early as December last year, when Motus was released, he publicly articulated the overall concept of a “General World Model.” In March of this year, he further positioned it as the foundation connecting the digital and physical worlds. Several months later, Zhu Jun and his team refined this framework in their paper, “General World Models from First-Principles,” defining core capabilities from first principles and presenting a five-level evolutionary roadmap.
Paper link: https://www.shengshu.com/assets/gwm-principles-and-roadmap.pdf
Technical report: https://www.shengshu.com/zh/general-world-model/
The paper summarizes the core capabilities of a general world model as understanding, imagining, and acting. The model must form a judgment about the current world from historical observations, simulate potential futures resulting from different choices, take actions, and then update itself using new observations.
Within this framework, video generation, real-time interaction, latent space prediction, and robotic control can be viewed as different stages along the same capability continuum.
How far are world models from general intelligence has been transformed into more specific questions: Can they maintain a coherent world? Can they predict the consequences of interventions? Can they learn from real-world feedback? And can they gradually form goals and organize more complex actions?
First principles of the world model
Many people have experience learning to ride a bicycle.
At first, your body will wobble. When you see the handlebars tilt to one side, you’ll feel your center of gravity shift and instinctively adjust your direction, then make further corrections based on the new feedback. With more practice, your brain gradually learns to anticipate what result a particular action will produce.
This is precisely the most fundamental meaning of a world model.
In 1943, Kenneth Craik proposed that humans form small-scale models of reality in their minds to simulate the outcomes of different actions. POMDP in reinforcement learning further describes this process: an agent receives partial observations, estimates the state of the world, takes actions, and updates its judgments based on new observations.
Zhu Jun’s team summarized this logic as three core capabilities of a general world model, which are interconnected and form a continuously updating loop:
Understanding: Integrating information from visual, linguistic, auditory, tactile, or robotic sensor inputs to form an internal assessment of the current world;
Imagination: Projecting multiple possible futures from the current state, especially considering “What if I take action A instead of action B?”
Action: Turn predictions into interventions in digital or physical environments, and use new observations after the action to test and refine the model.
The three core capabilities of a general world model form a closed loop: understanding infers the current state, imagination predicts possible futures, and action changes the world, providing new evidence for the next round of understanding.
This is also why video generation does not equate to the entirety of a world model.
The video model can answer the following
Come draw
The surface may look one way, but to move toward a general world model, it must answer questions like: “If I change this condition, move this object, or apply this action, how would the world be different?” Only then does it touch upon causality, counterfactuals, and actionable decision-making.
Understand, imagine, act—how do they form a closed loop?
To describe how world models evolve, Zhu Jun’s team further proposes a five-level roadmap: starting from L1 generating worlds, L2 interactive worlds, and L3 acting within worlds, gradually progressing toward L4 autonomous world agents and L5 world organizers.
It’s easier to understand with the analogy of a glass pushed to the edge of a table.
The L1 model can generate a continuous sequence of images showing a cup sliding, colliding, and shattering; the L2 model can continue evolving this world after the user changes the viewpoint or alters the direction of the push; at L3, the model must determine whether the cup will fall, predict whether a reaching motion can succeed in time, and output a tangible grasping action that the robot can execute. Only after taking action will it know whether its judgments about the cup’s weight, friction, and timing of the grasp were correct.
Above that, L4 no longer merely waits for instructions—it must be able to detect risks, proactively observe, supplement information, and adjust its strategy after failure. L5 goes further by handling multi-agent collaboration, such as determining who grabs the cup, who avoids obstacles, who manages tools, and how to re-prioritize tasks after environmental changes.
The significance of this approach is that it transforms the “world model” from a noun into a set of progressively testable questions.
L1: Check whether the generated trajectory is coherent.
L2: Assess whether the world can respond continuously.
L3: Determine whether the model can alter the physical environment.
L4: Evaluate whether the system can actively explore and continuously learn.
L5: Assess whether it can orchestrate multi-agent collaboration in an open environment.
Formal modeling objectives and learned capabilities of the General World Model (GWM) five-level roadmap.
According to the paper’s assessment, existing systems have reached the first three levels; L4 and L5 remain open problems, with key gaps including causal and physical grounding, persistent memory, online learning, efficient deployment, and security control. Evaluations must also compare predictions with ground-truth results to assess the model’s ability to transfer across unfamiliar tasks, environments, and ontologies.
From video to embodiment: Two validations of the same hypothesis
The challenges of world models are not only in the algorithms.
Internet videos are vast and capture a wide range of objects, people, scenes, and motions, allowing models to learn spatial structures, object permanence, human behavior, and how events typically unfold. However, videos rarely synchronously record the actions, forces, and intentions that cause these changes. Robot trajectories can link observations, actions, and outcomes, but they are more expensive to collect and are often tied to specific hardware and tasks.
The approach proposed by Zhu Jun’s team is to place these two types of data into a data pyramid that moves from observation to action.
The data pyramid of the General World Model (GWM) and its relationship to the five-level capability roadmap.
At the base are internet-scale videos used to acquire world knowledge and dynamic priors; above them are domain-specific videos, first-person human videos, and human demonstrations with action recordings; at the top are real robot interaction data. The higher up the hierarchy, the scarcer and more costly the data becomes, but the clearer the correspondence between actions, tasks, and the robot’s physical form.
This explains why Shengshu Technology is advancing both video generation and embodied intelligence: videos provide breadth of world knowledge, while robot data enables action grounding.
On the digital side,
View
The series continues to explore the generation and interaction with the visual world. Video is not the endpoint of world models, but a crucial starting point—it enables models to learn how objects, scenes, motion, and events evolve across large-scale spatiotemporal variations.
View
S1 further advances from one-time generation to real-time responsiveness, allowing user input to continuously alter subsequent content and testing whether the model can maintain state continuity during interaction.
On the physical side, Motus and Motubrain integrate environmental understanding, state prediction, and robotic actions into a single pipeline. The model must perceive the environment, predict the consequences of interventions, translate its judgments into executable actions, and then be evaluated against real-world outcomes.
Implement existing representative systems from L1 to L3.
The two product lines thus provide two validations of the same assumption about the world model: the digital environment tests whether the model can understand and imagine the world, while the physical environment tests whether this knowledge can support action and be refined through feedback.
For two types of verification to share underlying capabilities, images, videos, language, and robotic actions must collaboratively compute around a unified world state. Traditional modular systems pass information sequentially, often losing geometric, temporal, and uncertainty information at interfaces. Fully sharing parameters across all modalities may also cause training conflicts, overwhelming scarce robotic signals with vast amounts of video data.
The MoT (Mixture-of-Transformers) architecture proposed by Zhu Jun’s team seeks to strike a balance between the two. Each modality retains its own expert parameters while exchanging context through a shared attention mechanism. Vision informs the model what is happening in the environment, language provides goals and constraints, and actions enable intervention in the environment—these three types of information collectively update the model’s understanding of the world.
MoT provides the foundation for unified computation, but physical laws, cross-ontology transfer, real-time reasoning, and security control still rely on data, training, and systems engineering to be resolved. Its target is clear: a general world model requires a world state that can be jointly read and written by different modalities and continuously updated.
Motus2:
Let the bot act before doing anything
Try it out in your own world first.
If Motus and Motubrain advanced the exploration of Shushu’s world model to L3, enabling robots to generate actions based on environmental understanding and future prediction, then the recently introduced Motus2 further incorporates outcome evaluation and strategic feedback, aiming to give the action loop self-improving capabilities.
Motus2 does not implement action generation, future simulation, and outcome evaluation as three separate systems. Instead, it uses a single video-action model with shared parameters, exposing three control interfaces:
Policy (World-Action Model): Proposes candidate actions that can be executed;
Simulator (Action-Conditioned World Model): Predicts the potential outcomes of these actions in the visual world;
Evaluator / Value Model: Determines which outcome is closer to the task objective and uses this judgment to select and improve.
Think of it as a true “mind first, then skill” approach.
Robots no longer simply react to objects by immediately performing an action; instead, they first generate several candidate options, simulate the resulting scenarios and task progress within their models, and then select the optimal next step. After execution, successful, failed, or suboptimal outcomes are no longer discarded as data waste—they become valuable evidence for the next round of dynamics modeling and value learning.
Action, prediction, evaluation, feedback, and re-action form a single decision and learning loop, which is the closed-loop self-evolution emphasized by Motus2.
On the data side, Motus2 employs a hierarchical training pipeline that progresses from monocular first-person video to synchronized binocular first-person data, and finally achieves embodied adaptation through robotic data. The training utilizes approximately 130,000 hours of human first-person recordings, along with over 100 hours of robotic trajectories and human-robot alignment data.
This approach first allows the model to accumulate shared experiences about how humans interact with the world, then enables it to learn through robotic data how this body translates understanding into action.
On the execution side, Motus2 refines upcoming short actions and predicts force feedback after contact through an independent lightweight haptic expert. While vision excels at identifying where objects are and how the world is generally changing, haptics fill in the hardest part—the “final centimeter”—of contact, giving the model an additional layer of real-time correction for delicate maneuvers such as pressing, touching, grasping, and rotating, beyond what it can see.
Motus2 is still in the phase of extending from L3 to L4. Current validation primarily focuses on specific robots and operational tasks, while autonomous goal formation, persistent memory, generalization in open environments, long-term online learning, and safety control remain to be explored.
Its key advancement lies in integrating strategy generation, future simulation, value evaluation, and haptic feedback into a single system, enabling robots to simulate outcomes before acting and refine their strategies using results after acting, providing a testable pathway toward L4 autonomous world agents.
From Probabilistic Learning to World Intelligence
A Chinese technological pathway
The framework of the General World Model continues the long-standing research direction of Zhu Jun and Tsinghua’s TSAIL.
Professor Zhu Jun has long researched Bayesian methods, probabilistic machine learning, generative models, and reinforcement learning, addressing uncertainty, data distributions, future generation, and action decision-making.
TSAIL has also trained and influenced a group of researchers who joined leading global teams. During his undergraduate studies, Song Yang collaborated with Zhu Jun on Bayesian learning research, and Song Jiaming also contributed to related work during his undergraduate years; both later made significant contributions in the fields of score-based generative models and DDIM.
From probabilistic modeling to generation and decision-making, this academic lineage reflects two entry points in AGI development: one begins with language, gradually enhancing reasoning, tool use, and task execution capabilities; the other starts with vision and the dynamic world, progressing toward real-time interaction and embodied action.
Within China’s AGI industry landscape, Zhipu and Shengshu represent two distinct pathways. Zhipu begins with language models and continuously enhances its reasoning, coding, and agent capabilities; Shengshu starts with video generation and further expands into real-time interaction, embodied action, and strategic optimization.
The two approaches are increasingly intersecting. The language agent requires visual feedback and an environmental model, while the world model also needs linguistic expressions of goals, rules, and plans—ultimately, all pathways converge toward a unified modeling of language, world states, and actions.
Conclusion
The competition among world models will ultimately shift from the impact of demos to the system’s ability to withstand feedback.
The interface creates a trustworthy first impression; geometry and physics determine whether this world can be reused. Actions bring the model’s judgments into reality. Whether a prediction holds true can only be confirmed after intervention occurs; whether a task succeeds must be tested in unfamiliar environments, over long-term operation, and amid continuous change.
Zhu Jun’s team did not provide a final answer to the debate over world models, but they offered a set of questions for further investigation: What does the model understand? How far can it imagine? And can it correct itself after taking action?
Over the past decade, the most significant advances in AI have come from learning patterns in language and data. The next stage requires models to operate in continuously changing environments. When models can continuously form judgments, take actions, bear the consequences, and keep learning in such environments, world models will evolve from a crowded technical term into a foundation for general intelligence.
This article is from the WeChat public account “Machine Heart” (ID: almosthuman2014), authored by Yang Wen.
