Author | Chen Wen

Source | Insight New Research Society

Over the past year, the term “World Model” has swiftly transitioned from academic jargon to a cornerstone concept in the AI and robotics sectors. Concurrently, VLA (Vision-Language-Action), the predominant technical approach for embodied AI, has repeatedly taken center stage in these discussions.

In the last 18 months, over $10 billion in funding has poured into world model and robotics AI firms, betting on their potential to “comprehend the physical world”.

Yann LeCun, founder of Advanced Machine Intelligence Labs (AMI Labs), optimistically predicted that “within three to five years, world models will supplant LLMs as the dominant AI paradigm”. Meanwhile, Jim Fan, NVIDIA’s head of robotics, sparked intense debate over technical route transitions by declaring that “VLA is dead”.

Despite extensive debate, the industry remains divided on “what precisely a world model is”. Definitions and methodologies vary widely, encompassing renderers, simulators, planners, video generation, latent space prediction, and more. Each faction stakes its claim, as if discussing the same concept across different dimensions.

This conceptual disarray starkly contrasts with the urgent expectations for real-world implementation timelines.

To grasp world models, one must first distinguish them from large language models (LLMs).

The core mechanism of LLMs is “predicting the next token”—given preceding words, they forecast the likelihood of the next word. An LLM knows that “a glass falling to the ground will shatter” because this sentence has appeared countless times in its training data, not because it understands elastic modulus, stress propagation, or impact energy.

World models aim to bridge this gap. Instead of predicting the next word, they predict the next state—how an object’s position in space will change or what chain reactions an action will trigger.

As Wang Zhongyuan, dean of the Beijing Academy of Artificial Intelligence (BAAI), noted, the paradigm shift in AI is moving from “predicting the next token” to “predicting the next physical state”. World models, as the next-generation foundational models for the real physical world, represent a significant leap in AI paradigms by focusing on “predicting the next physical state”.

However, “world model” is not a clearly defined technical concept at present. The work being done by different teams varies far more than their names suggest. Li Feifei and the World Labs team openly state that “world model” is one of the most important yet overused terms in AI today.

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

Beyond these four categories, BAAI is also exploring a fifth possibility: full-modal representation fusion based on a unified latent space. This approach compresses modalities like text, images, and videos into the same semantic space for native training, with plans to incorporate more physical world modalities in the future. Wang Zhongyuan judges that full-modal latent space modeling may be the true breakthrough path for world models.

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

By introducing classical structures from reinforcement learning, Li Feifei divides the current complex landscape of generative models, physical simulation systems, and embodied AI methods into three functional categories:

Renderer: Outputs pixel-based visuals for human viewing, with visual fidelity as the core metric.

Simulator: Outputs environment states that align with objective laws. Li Feifei particularly notes that simulators receive the least attention but are the most critical, serving as the bridge between rendering and planning.

Planner: Outputs action commands for intelligent agents.

Currently, these three directions are beginning to merge. Li Feifei judges that “when their boundaries disappear, they will collectively reshape something grander: the relationship between machine intelligence and the physical world it inhabits”, culminating in a unified world model capable of rendering photorealistic views, generating physically accurate structures, and planning action sequences.

If the bottleneck for LLMs is computational power, then for world models, the primary bottleneck is data—the first “mountain” they must climb for real-world implementation.

At Cisco’s AI Summit in February, Li Feifei stated in her speech that the core bottleneck holding back AI for the physical world compared to language models lies in the data signal-to-noise ratio. Textual data has clear semantics and is easy to acquire, whereas pixel and voxel data in the physical world are noisy, and high-quality 3D and 4D data are extremely scarce.

A stark contrast: While training data for LLMs in the digital world has reached the scale of hundreds of trillions of tokens, training data for visual-language-action models in the physical world often amounts to just one-ten-thousandth of that. This lack of real-world data directly leads to weak model capabilities.

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

Beyond data, the second challenge for world models is that generating realistic visuals does not equate to understanding physical laws.

A video generation model can create a scene of pigs flying in the sky because it learned this pattern from science fiction films, but it does not understand the physical common sense that “pigs cannot fly”.

Moreover, current models have significant shortcomings in two core capabilities: causal reasoning and predicting complex dynamic systems. Many of their physical scene deductions fall short of practical standards.

Given these issues, what does the future hold for world model implementation? To answer this, we must first clarify the relationship between world models and VLA.

VLA (Vision-Language-Action) is currently the mainstream technical route for embodied AI, unifying vision, language, and actions into an end-to-end large model that takes images and instructions as input and directly outputs action sequences. Over the past two years, VLA was once seen as the “standard answer” for embodied AI. When Google DeepMind’s RT-2 paper was released, analysts moved the commercialization timeline for embodied AI forward by three years based on its findings.

However, after two years of development, VLA’s shortcomings have gradually emerged: Robots can recognize objects but do not understand that “pushing a cup will cause it to fall”; they can follow instructions but cannot predict “how much force is needed to twist a bottle cap”. Engineers comment that the physics VLA learns is a “pseudo-physics” based on superficial correlations.

As a result, the industry has begun to debate how VLA and world models should coexist.

Guo Yandong, founder and CEO of AiSquared, offered his perspective at the 2026 BAAI Conference: World models are not a competing route to VLA but a core component within the VLA framework.

Guo redefined VLA as a general term for end-to-end model architectures that fuse multiple modalities and are driven by big data. Under this definition, world models and VLA are not fundamentally different, nor are they substitutes for each other.

In layman’s terms, world models are responsible for understanding the world, while VLA is responsible for acting on it. The two are not opposing forces but a naturally unified whole. The division of labor between VLA and world models resembles that of the “cerebral cortex” and “cerebellum”: the cerebral cortex handles understanding and planning, while the cerebellum handles prediction and correction.

Huang Tiejun, chairman of BAAI, holds a similar view: VLA and world models are not contradictory. While VLA is a practical choice for enterprises, the goal of world models is to create a general-purpose brain. A powerful world model should serve as the “subconscious” and “intuitive module” for VLA.

In practical implementations, this integrated approach is already evident. At CVPR 2026, XPENG Motors showcased a world model technology roadmap for the first time, adopting a dual-pillar architecture of “VLA + world model”. VLA learns driving logic from massive real-world driving data, while the world model focuses on proactive judgment and multi-step reasoning for traffic scenarios.

Specifically for robotics, VLA is more mature than world models. The reason is that current robotics deployments in industrial scenarios, where tasks are well-defined and action types are limited, allow companies to collect large amounts of data in advance and train models to achieve near-100% success rates. In contrast, the main advantage of world models lies in cross-scenario and multi-task generalization, making them more suitable for open environments like homes. However, as everyone knows, home scenarios are still far from mature commercialization.

Overall, both routes have achieved some commercial implementation at scale in the short term or in specific scenarios, but the actions and tasks they can truly perform remain limited.

From the above analysis, it is clear what issues world models and VLA face, making the industry’s next competitive focus equally clear: shifting from “capable of prediction” to “capable of action”.

At the 2026 BAAI Conference, StarOrigin Intelligence unveiled ω-EVA, the world’s first embodied interaction world model, which for the first time achieved a closed-loop action decision-making process for world models in robotics.

The model establishes a decision-making closed loop of “preview, verify, act”. Before executing a command, the robot first predicts the environmental changes caused by the action and then optimizes the plan based on the reasoning results.

The release of StarOrigin’s ω-EVA reveals an important trend: World models must not remain offline “thinkers” but must become real-time “decision-makers”. More fundamentally, world models must evolve from one-time prediction and action generation to continuous perception, imagination, correction, and self-updating through real interactions.

From a global technical perspective, the action-driven route is emerging as an important direction. It skips unnecessary pixel generation steps and concentrates all computational resources on “understanding physical interactions” and “generating optimal actions”. This approach is closer to the essence of biological intelligence. When humans act, they do not need to render high-definition 3D movies in their minds but directly react based on an intuitive understanding of the physical world.

So, how long until world models truly enter production and implementation?

Wang Zhongyuan offers a cautiously optimistic judgment: “At least for the next three to five years, world models will continue to evolve and iterate. Research is unpredictable—it might get stuck on a difficult point for three to five years without breakthroughs, but a technological explosion could also suddenly occur”.

This judgment is echoed by multiple sources. BAAI predicts that “several more years are needed”, with the next three to five years marking a period of continuous evolution and iteration for world models.

In the short term, world models are more likely to play a role in production processes as data engines, training tools, and environment construction tools rather than being massively deployed on real machines for real-time inference. The mid-term competitive focus will shift to state maintenance, physical consistency, and cross-shot continuity. In the long term, they are expected to further integrate into robotics, gaming, digital twins, and agent task closed loops.

From a commercialization perspective, the industry is moving from “multimodal generation” to “interactive workflows”. NVIDIA released Cosmos 3, an open-source full-modal physical AI model that integrates three core capabilities: visual reasoning, world generation, and action prediction. Alibaba released the Qwen-Robot series, an embodied intelligence large model that includes three modules: a VLA operation model, a VLN navigation model, and a world model.

These signs indicate that the industry is transitioning from technical exploration to product validation.

Looking back, large language models have enabled machines to talk about the world; the emergence of world models allows machines to understand, imagine, reason, and interact with the world. Currently, this transition from the digital world to the physical world has only just begun, and the next three to five years will be a critical window to determine who can reach the destination first.