Having "burned" tens of billions of dollars, there is still no unified definition for the concept of a world model.

Over the past year, the "World Model" has rapidly expanded from an academic jargon to a key term in the AI and robotics industry. Meanwhile, VLA (Vision-Language-Action), as the mainstream technical route for embodied intelligence, has also been repeatedly pushed to the forefront of discussions.

In the past 18 months, over 100 billion US dollars in capital has flowed into world model and robotics AI companies, betting on the disruptive potential of world models to "understand the physical world."

Representing the optimists, Yann LeCun, founder of Advanced Machine Intelligence Labs (AMI Labs), has publicly predicted that "within three to five years, world models will replace LLMs as the mainstream AI paradigm." Meanwhile, Jim Fan, NVIDIA's top robotics executive, even sparked intense debates on the replacement of technical routes with his statement that "VLA is dead."

There have been numerous arguments, but the industry still has not reached a consensus on "what a world model is." Renderers, simulators, planners, video generation, latent space prediction... A wide variety of definitions and paths have staked their own claims, as if a group of people are talking about the same thing across different dimensions.

This conceptual chaos forms a stark contrast with the urgent expectations for implementation timelines.

"The Next Token" and "The Next Physical State"

To understand world models, we first need to grasp their essential difference from large language models.

The core mechanism of large language models is "predicting the next token." Given the preceding words, it predicts the probability of the next word appearing. It knows that "a glass will shatter when it falls to the ground" only because this sentence has appeared countless times in the training data, not because it understands elastic modulus, stress transmission, and impact energy.

The starting point of world models is to fill this gap. Instead of predicting the next word, they predict the next state—how the position of an object in space will change, and what kind of chain reaction an action will trigger.

As Wang Zhongyuan, President of the Beijing Academy of Artificial Intelligence (BAAI), put it, the paradigm iteration of artificial intelligence is moving from "predicting the next token" to "predicting the next physical state." As the next-generation foundational model oriented toward the real physical world, the world model takes "predicting the next physical state" as its core, representing another important paradigm leap for artificial intelligence.

However, "world model" is not a technically well-defined concept at present. The work done by different teams varies far more than the name implies. Li Feifei and the World Labs team bluntly stated that "world model" is one of the most important yet most overused terms in today's AI field.

In response to the current generalization and misuse of the world model concept in the industry, Wang Zhongyuan divides existing technical routes into four categories: the first category is language-centric world models, including VLM, VLA, etc.; the second category is pixel-centric world models, such as video generation models that learn videos or images in the visual space; the third category is 3D structure-centric world models, including 3D reconstruction and related spatial models; the fourth category is visual representation-centric world models, such as the JEPA series of models.

Beyond these four paths, BAAI is also exploring a fifth possibility: full-modal representation fusion based on a unified latent space, which natively trains text, images, videos and other modalities by compressing them into the same semantic space, and will further incorporate more physical world modalities in the future. Wang Zhongyuan judges that full-modal latent space modeling may be the real breakthrough path for world models.

If Wang Zhongyuan's classification starts from the technical implementation path, then Li Feifei and the World Labs team provide a clearer framework from the functional dimension.

By introducing the classic structure in reinforcement learning, Li Feifei functionally divides the current complex generative models, physical simulation systems, and embodied intelligence methods into three categories:

Renderer: Output pixel frames for human eyes to view, with visual fidelity as the core metric.

Simulator: Output environmental states that conform to objective laws. Li Feifei specifically pointed out that simulators receive the least attention but are the most critical, serving as the bridge connecting rendering and planning.

Planner: Output action commands for agents.

At present, these three directions are beginning to merge with each other. Li Feifei judges that "when their boundaries disappear, they will jointly reshape something grander: the relationship between machine intelligence and the physical world it inhabits." The end point is a unified world model that can render photo-realistic views, generate physically accurate structures, and plan action sequences.

Implementation Still Needs to Surmount Two "Big Mountains"

If the bottleneck of large language models is computing power, then the bottleneck of world models is first and foremost data—the first "big mountain" that needs to be crossed for their implementation.

At Cisco's AI Summit in February this year, Li Feifei stated in her speech that the development of physical-world AI lags behind that of language models, and the core bottleneck lies in the signal-to-noise ratio of data. Text data has clear semantics and is easy to obtain, while the pixel and voxel data of the physical world is full of noise, and high-quality data in 3D and 4D dimensions is extremely scarce.

An intuitive comparison: the current training data volume of large language models in the digital world has reached the level of hundreds of trillions of tokens, while the training data volume of vision-language-action models in the physical world is often only one ten-thousandth of that. The lack of real data directly leads to weak model capabilities.

Wang Zhongyuan also admitted that the current bottlenecks of world models are mainly reflected in the scarcity of real physical data, the lack of convergence in technical routes, and the imperfect evaluation system. Data collection in the physical world is not only expensive, but also lacks samples of extreme working conditions.

Apart from data, the second challenge for world models is that generating seemingly realistic images does not equal understanding physical laws.

Video generation models can create scenes where a group of pigs fly in the sky. The reason is that the model has learned this pattern from sci-fi movies, but it does not understand the physical common sense that "pigs cannot fly."

In addition, current models still have major shortcomings in two core capabilities: causal reasoning and prediction of complex dynamic systems. Many of their inferences about physical scenarios have not reached practical standards.

Given so many problems, what exactly is the implementation prospect of world models? To answer this question, we first need to clarify the relationship between world models and VLA.

VLA (Vision-Language-Action) is the current mainstream technical route for embodied intelligence, which unifies vision, language, and action in an end-to-end large model, outputting action sequences directly by inputting images and instructions. Over the past two years, VLA was once regarded as the "standard answer" for embodied intelligence. At that time, when Google DeepMind's RT-2 paper was just released, analysts moved the commercialization timeline of embodied intelligence three years forward based on the findings of the paper.

However, after VLA has been in operation for two years, its shortcomings have gradually emerged: robots can recognize objects, but do not understand that "pushing a cup will make it fall"; they can understand instructions, but cannot predict "how much force is needed to unscrew a bottle cap." Engineers commented that the physics VLA has learned is a kind of "pseudo-physics" based on surface correlations.

As a result, the industry began to debate: what kind of relationship should VLA and world models have with each other?

Guo Yandong, Founder and CEO of Zhih Square, gave his understanding at the 2026 BAAI Conference: the world model is not a competitive route for VLA, but a core component of the VLA system.

Guo Yandong redefined VLA as the general term for end-to-end model architectures that integrate multiple modalities and are driven by big data. Under this definition, there is no essential difference between world models and VLA, let alone a substitution relationship.

In layman's terms, the world model is responsible for understanding the world, and VLA is responsible for acting on the world. The two are not opposites, but a naturally unified whole. The division of labor between VLA and the world model is similar to the relationship between the "cerebral cortex" and the "cerebellum": the cerebral cortex is responsible for understanding and planning, and the cerebellum is responsible for prediction and correction.

Huang Tiejun, Chairman of BAAI, holds a similar view: VLA and world models are not contradictory. Adopting VLA by enterprises is a realistic choice, while the goal of world models is to create a general brain. A powerful world model should be the "subconscious" and "intuition module" of VLA.

In actual implementation, this integration idea has already been reflected. The world model technology map first demonstrated by XPeng Motors at CVPR 2026 adopts the "VLA + World Model" dual-pillar architecture: VLA relies on massive real driving data to learn driving logic, while the world model focuses on forward-looking judgment and multi-step deduction of traffic scenarios.

Specifically for robot implementation, VLA is more mature than world models. The reason is that in current industrial scenarios where robots are widely deployed, tasks are clear and the types of actions are limited. Enterprises can collect a large amount of data in advance to train the model to a near-100% success rate. The main advantage of world models lies in cross-scenario and multi-task generalization, making them more suitable for open environments such as homes. However, as we all know, home scenarios are still far from mature commercialization.

In general, these two routes have achieved certain scales of commercial implementation in the short term or in specific scenarios, but the actions and tasks they can actually perform are still relatively limited.

Action Closed-Loop Becomes the Competitive Focus of the Next Stage

It is not difficult to see from the above analysis that the problems faced by world models and VLA are actually very clear, and the competitive focus of the industry in the next stage is also very clear: shifting from "being able to predict" to "being able to act."

Xingyuan Intelligence released the world's first embodied interactive world model "ω-EVA" at the 2026 BAAI Conference, which for the first time realized the closed-loop of robot action decision-making based on the world model.

This model sets up a decision-making closed-loop process of "rehearsal, verification, action." Before the robot executes an instruction, it first predicts the environmental changes brought about by the action, and then optimizes the plan based on the deduction results.

The release of Xingyuan Intelligence's "ω-EVA" reveals an important trend: world models cannot just be offline "thinkers," but must become real-time "decision-makers." More profoundly, world models need to move from one-time prediction and action generation to continuous perception, imagination, correction, and self-update from real interactions.

From the perspective of global technical routes, the action-driven route is becoming an important direction. It directly skips unnecessary pixel generation steps and concentrates all computing resources on "understanding physical interactions" and "generating optimal actions." This route is closer to the essence of biological intelligence: when humans act, they do not need to render a high-definition 3D movie in their minds, but directly generate reactions based on their intuitive understanding of the physical world.

So, how long will it take for world models to truly enter production and implementation?

Wang Zhongyuan gave a cautiously optimistic judgment: "At least in the next three to five years, world models will be in a stage of continuous evolution and iteration. Scientific exploration is unpredictable—it may get stuck at a difficult point without breakthroughs for three to five years, but it may also suddenly usher in a technological explosion."

This judgment has been echoed by many parties. BAAI predicts that "it will take at least several more years," and the next three to five years will be the stage for the continuous evolution and iteration of world models.

In the short term, world models are more likely to play a role in production links first, acting as data engines, training tools, and environment construction tools, rather than being deployed in large numbers on real machines for real-time reasoning. In the medium term, the competition focus will shift to state retention, physical consistency, and cross-shot continuity. In the long term, it is expected to further connect to the closed loops of robots, games, digital twins, and Agent tasks.

From the perspective of commercial signals, the industry is moving from "multimodal generation" to "interactive workflows." NVIDIA released the open-source full-modal physical AI model Cosmos 3, which integrates three core capabilities: visual reasoning, world generation, and action prediction. Alibaba released the Qwen-Robot series of embodied intelligence large models, which includes three major modules: VLA operation model, VLN movement model, and world model.

These signs indicate that the industry is moving from technical exploration to product verification.

Conclusion

Looking back now, large language models enable machines to talk about the world; the emergence of world models allows machines to understand, imagine, reason, and interact with the world. At present, this leap from the digital world to the physical world has just begun, and the next three to five years will be a critical window to determine who can be the first to reach the other side.

* Images are from the Internet; please contact for removal in case of infringement.

This article is from the WeChat public account "Insight New Research Community" (ID: DJXYS-0309), author: Chen Wen, published with authorization by 36Kr.