正文 · AI 翻译
Westlake World Model Center Established
On October 31, the Westlake World Model Center was established at the Yungu Campus. The name may puzzle many people: they know what a “large language model” is, but what is a “world model”?
Simply put, it is another buzzword in artificial intelligence following “large language models.”
In the past few years, the rapid development of large language models has amazed us with their vast reading and ability to talk about almost anything. But mastery of language is not the same as understanding the world. World models aim to give artificial intelligence the wisdom to investigate things and gain knowledge, and the insight to discern broader patterns from small details: to explore the laws of the physical world, predict how the real world operates, and move from “observing the world” to forming a “worldview.”
From late 2025 through 2026, AI leaders including Fei-Fei Li and Yann LeCun have entered world-model startups one after another. Just a few days ago, Google DeepMind released Genie4, combining a world model with a reasoning model.
The establishment of the Westlake World Model Center signals that Westlake University will focus on developing this field. In fact, its Autonomous Intelligence Laboratory, AutoLab, began researching world models 3 years ago. It has since incubated a startup and worked with companies to put the technology into practice.
The center’s establishment therefore attracted not only a group of AI researchers but also nearly 90 heads of industry research and development departments and investors. The “Frontiers of Physical World Models Forum,” held after the founding ceremony, was also a focus of their attention.
So what exactly is a world model? What changes might it bring? And why, as something new, does it draw so much discussion? There is also the story of the work begun 3 years ahead of time.
01 World Models
From “what the world is like” to “what will happen in the world”
As early as 1990, Richard Sutton, one of the principal founders of reinforcement learning, offered a definition of a world model: given a current state and an action, predict the next state. This requires the model to answer not just “What is the world like?” but “What will happen in the world?”
Today, getting AI to generate a video is not difficult. What is harder is whether, when a cup slides toward the edge of a table, a model can predict what will happen. Yu Kaicheng likens past ways of training AI to making it “memorize a large number of exam questions.” World models, by contrast, seek to make accurate predictions by “understanding” the laws of the physical world.
Yann LeCun, known as the “father of convolutional neural networks” and a Turing Award winner, put it more directly: “By simply predicting text, AI can never attain human-level intelligence. We need models that can understand physical reality.”
Large language models are essentially probabilistic predictors of textual symbols, while world models aim to “predict the next state of the real world.” Their goal is to give AI a grasp of time, space, causality and physical common sense.
Perhaps the early prototype of a world model most familiar to the public is the 2024 video generation model Sora. Its most iconic image shows a wooden sailing ship riding waves inside a coffee cup, with strikingly realistic wave motion and texture. Sora was described at the time as a “general-purpose simulator of the physical world.”
Soon afterward, autonomous driving became the first field to put world models to practical use. These models can generate highly realistic driving videos and even actively create “edge cases,” such as severe weather and sudden accidents, to train autonomous driving models. This addresses a weakness in autonomous driving training: in the real world, one cannot cause car crashes just to train a model.
But the question is: does the ability to generate realistic scenes mean that a model truly understands the world?
A simulated rainy scene for autonomous driving training (demonstration of early work by Yu Kaicheng’s team)
02 “Self-Questioning”
— To help machines better understand the world
Yu Kaicheng, born in Shandong in 1993, presents a contrast: he describes himself as a “sunny, cheerful and enthusiastic young man,” but in his technical work he is a little “rebellious.”
He took the college entrance examination in 2012 and scored high enough to enter Tsinghua University or Peking University, but ultimately chose the University of Hong Kong. After graduating from HKU with a first-class honours Bachelor of Engineering degree, Yu went directly into a doctoral program at the Swiss Federal Institute of Technology Lausanne, specializing in automated machine learning. After earning his doctorate, he worked on related research at Intel’s Intelligent Systems Lab, Abacus.AI and Alibaba’s DAMO Academy.
At DAMO Academy, Yu mainly researched autonomous driving. In the process, he gradually recognized the limits of traditional training methods. Traditional autonomous driving technology essentially relies on “rote memorization” and depends heavily on training datasets: a model can handle scenarios it has seen in its “question bank,” but is prone to errors when rare, unexpected situations arise.
In 2023, he left Alibaba to join Westlake University, where he established the Autonomous Intelligence Laboratory, AutoLab, and began researching world models. World models, a hot topic this year, were still a relatively niche concept at the time.
At the time, there were 2 main approaches to the difficulty of training for rare, long-tail autonomous driving scenarios. The first was to generate synthetic data, creating examples of rare scenarios and feeding them to models for training. The second was to teach models to generalize from the ground up, giving them the ability to reason through unfamiliar scenarios. The latter, of course, was harder.
Soon after arriving at Westlake University, Yu and his team first explored the relatively easier route: generating videos of extreme scenarios to train autonomous driving models. This approach quickly produced results. In 2024, they won the world championship in the ECCV autonomous driving track, scoring far higher than the runner-up.
During an extensive deployment with a leading domestic automaker, Yu and his team began to question the foundation of this approach. If the model could already generate realistic data, that suggested it understood the world to some extent. What, then, was the point of continuing to render pixels?
At the end of 2024, AutoLab shifted to a new technical approach and proposed a “conceptual world model”: instead of generating pixels, it would represent physical concepts in an abstract mathematical latent space. Yu said the model needed to act like a person, breaking complex scenes into a finite set of concepts and then recombining them to “understand” new scenes that had never appeared in full in its training data.
In other words, the “conceptual world model” breaks complex physical environments, object properties, spatial relationships and the consequences of actions into reusable physical concepts, such as spatial relationships, material properties and contact states. When faced with an entirely new scene, the model does not need to learn the whole scene anew; instead, it recombines existing concepts in latent space to reason through it.
As the research progressed into 2025, the team began observing in experiments that, once mechanisms for breaking down and linking concepts were established, the model’s success rate on complex tasks improved significantly over traditional imitation learning and reinforcement learning baselines while maintaining relatively low prediction costs.
This is a technical approach to an implicit world model. AutoLab has also become one of the few teams in China to complete preliminary validation of it, and these research advances laid the groundwork for the subsequent startup.
03 Beyond Autonomous Driving
— When AI moves through the real world
In early 2026, Westlake Digital Intelligence Awomo was formally established. The company positions itself as a developer of foundational large models for the physical world. Taking large language models as a benchmark, it aims to build a native large physical model, rather than simply deliver customized solutions to automakers.
The “conceptual world model” developed by AutoLab also received its formal name in 2026: Awomo. Awomo ranked first on all 4 internationally recognized robotics benchmark datasets, further validating this approach’s ability to generalize across environmental disturbances, fine manipulation and long-horizon tasks.
To date, the startup has raised more than 100 million yuan in total through seed and angel financing rounds. Its investors include Innoangel Fund, Oriental Fortune Capital, Zhengxuan Investment, Tianqi Capital, Westlake Innovation Investment Fund and Jinma Investment, among other hard-technology investment institutions.
Autonomous driving is only one application of world models. A world model can also serve as a robot’s “physical brain.” When facing complex disturbances, a robot could quickly reason about object properties and the consequences of actions in latent space, and accurately perform grasping, assembly and long-horizon collaborative tasks.
An even bigger change could come in how data is used for training. Traditional AI training requires vast quantities of training data, which entails extremely high computing costs. But Yu envisions that, with a general-purpose physical foundation model, downstream developers or small businesses might no longer need to collect millions of hours of data. They might be able to customize a model for a specific field quickly with just 20 to 30 hours of post-training.
In the long run, Yu firmly believes that world models will ultimately serve ordinary people’s needs for personalized creation. Without much of the burden of training on massive amounts of data, users could draw on a model’s ability to “build worlds” to create their own private robots or personalized intelligent assistants, much as they might set up a workshop in a private garage.
All these applications can be understood as actions AI attempts to take when facing the world, and every action leaves a trajectory. That is also why Yu is researching the ancient art of calligraphy. Chinese characters are pictographic and have an inherent sense of structure and space. The development of Chinese characters from oracle bone script to the present is, in essence, a process of compressing entities in the real world into 2-dimensional symbols.
“Oracle bone script was the earliest latent space,” Yu said.
The process of turning an intention or concept into brushstrokes also shares an underlying mathematical logic with the action control and trajectory planning used in robotics. Thousands of years later, we are still trying to abstract the laws of this world, only with different tools.
Appendix: Introduction to the Westlake World Model Center
The Westlake World Model Center was formally established in October 2026 as a university-level, interdisciplinary research platform at Westlake University for physical intelligence. Professor Yu Kaicheng serves as its director, Academician Jin Yaochu as its chief scientist, and several principal investigators jointly lead research in their respective areas.
As artificial intelligence moves from understanding language toward interacting with the physical world, the center focuses on world models and physical intelligence. It is committed to overcoming key obstacles to intelligent systems’ ability to understand physical laws, predict environmental changes and make decisions through interaction, and to building an internationally leading academic center and engine for innovation in physical intelligence.
With world models as its core research focus, the center explores intelligent systems that combine perception, cognition, prediction and action, advancing AI’s ability to understand the world, anticipate the future and act autonomously in real environments. Drawing on Westlake University’s research base and interdisciplinary strengths, the center will bring together researchers from multiple fields to foster coordinated progress in foundational theory, key technologies and applied research. It aims to develop general-purpose world models that can understand physical laws, predict the consequences of actions and continue learning from real-world feedback, supporting China’s independent innovation and frontier research in the next generation of artificial intelligence.