Article title: In conversation with Westlake University's 于开丞: No matter how strong general language models become, the physical world "can't wait"
Body:

This language model received no specialized training for robots, yet it took first place in average success rate across 42 manipulation tasks, overtaking 40 public robot policies.
The world model card table has been shaken up. The question "Can language models even do the work of world models?" has been put back on the table.
More than a month later, Westlake University announced the establishment of the Westlake World Model Research Center, with President Shi Yigong personally showing his support.
This research university, known for its innovative mechanisms, amid doubts over "whether world models are still necessary," established world models as a university-level interdisciplinary research platform. The person driving this is 于开丞, assistant professor at Westlake University's School of Engineering and head of the Autonomous Intelligence Laboratory (AutoLab).
In his view, no matter how strong general language models become, the physical world "can't wait." Language models simply cannot directly replace world models.

He spent five years at EPFL in Switzerland working with Pascal Fua, a leading figure in geometric vision. In his second year of doctoral studies, he "identified" a core industry-level problem in NAS and won the Qualcomm Innovation Fellowship, when there were only four winners in Europe that year.
After returning to China, he joined Alibaba DAMO Academy and became known in the autonomous driving circle through a series of works such as BEVFusion. Two years later, due to a real-world dilemma in the autonomous driving field, he turned to world models based on research intuition.
At that time, Sora was still a year away from appearing, and few people could understand the term world model. Even DAMO Academy, known for exploring the frontier, could not take it on.
So he brought the topic back to academia to pursue it, and on Westlake University's research soil, he has worked on it for three years.
In 2024, the year when large language models were at their loudest, he used world models to beat language models in an ECCV challenge, winning the "global championship" and demonstrating the feasibility of the world model route ahead of others.
Two years later, the main subject of all the noise has finally become world models. Teams from all sides are holding technical concepts and busying themselves with fundraising.
But 于开丞, who has already worked in this direction for three years, has not chased the excitement of concepts. In his eyes, a world model is a general technology like Transformer. What he wants to do is gather "a whole round-table of people" on the domestic track to compete with AI forces in Europe and the United States.
What he wants to do has solid backing. The general world model Awomo-0.5, with 4 billion parameters, simultaneously topped six major international evaluation leaderboards, including embodied intelligence and physical world understanding, with all code and weights open-sourced. Most startups still cannot deliver their answers, but this "achieving more with less" answer has been delivered first.
Open-source link: https://awomo-westlakedi.github.io/Awomo-0.5-page/
Why did he set out three years ago when the world model track was deserted? Why is he certain that world models are 100% worth doing? In this conversation with AI Technology Review, 于开丞 gives his most fundamental answer.
The following is a transcript of the conversation, edited by AI Technology Review without changing the original meaning.

Starting point: autonomous driving money can no longer be spent
▎AI Technology Review: You started working on world models in 2023. But at that time Sora had not yet appeared, and "world model" was a term almost no one mentioned. What opportunity led you to shift your research in this direction?
于开丞: The opportunity came from a real dilemma in autonomous driving: the ceiling became lower and the cost was high.
Every time a new scenario is encountered, data has to be collected all over again and the process run through again. This is too inelegant, meaning too expensive. Since it is so expensive, a new technical path must be found. It was from this pain point that we began to pay attention to world models.
At the time, diffusion models had already been proven in image generation, so we naturally thought: since generation can be done on images, why not in physical space? So we advanced from generating scene images that conform to 3D relationships to generating scene videos that conform to 3D relationships.
Here 3D is key: 2D generation only needs to "look similar," but the physical world requires "actually similar," satisfying strict projection relationships and physical relationships.
Pushing further along this logic, we found that there is still much work worth researching. This is the starting point from which we entered world models from autonomous driving.
▎AI Technology Review: Did you already start moving into this direction at DAMO Academy?
于开丞: Yes, there was already a prototype at DAMO Academy. But later DAMO Academy, or rather the entire Alibaba Group, felt it was a bit too expensive and wanted to contract, so it no longer conducted some particularly exploratory research.
Later I thought about it and it made a lot of sense: this kind of highly exploratory research may still need to be done at universities.
▎AI Technology Review: So you left industry and returned to academia because you wanted to do world models. Why did you go to Westlake University?
于开丞: It was a coincidence and also driven by direction. At the time I had two choices, and I ultimately chose Westlake University for two reasons: first, the atmosphere here and its recognition of innovation attracted me very much, and after joining, the university indeed gave us great support; second, President Shi Yigong's spirit also moved me greatly.
It is very fortunate to do innovative things at an innovative institution. For example, the university now allows me to establish the Westlake University World Model Center in a dual industry-academia state, which is actually of breakthrough significance in the university's history. For me this is a great honor, and I also bear a heavy sense of responsibility to do this thing better.
▎AI Technology Review: When you chose such a direction full of uncertainty at that time, did you hesitate?
于开丞: At that time no one was doing world models yet, so we were venturing into uncharted territory. But our judgment was simple: do it first when no one else is doing it, and if we can make the technology leading, we may remain leading.
Now that this research direction is recognized by everyone and accepted by society, of course we are very happy, but from our perspective, this was achieved step by step, and it does not look that fantastical.
▎AI Technology Review: You once publicly mentioned that in 2024 you participated in a competition at ECCV and won the "world model world championship." Was the competition itself related to world models, or did you use world models to participate?
于开丞: We used world models to participate. This challenge was held by Huawei, with the theme "extreme scenario generation," using generative models to synthesize unseen accident scenarios in driving scenes.
We won the championship with a world model technical solution, while the other entries basically did not use world models. So at that point in time, we could be said to have seized the initiative.
▎AI Technology Review: "World models" suddenly became so popular. Did this timing come earlier or later than you expected?
于开丞: Definitely earlier than I thought. From 2023 to 2025 we were quietly doing research, and in 2026, when we felt this matter could be expanded and done "with gongs and drums," everyone had already rushed in.
This actually also represents everyone's expectations for this field, and I understand it very well, because after language models, the biggest opportunity is on the world model track.

World models are 100% valuable
▎AI Technology Review: Between 2023 and 2026, did you ever encounter times when the road ahead was unclear? Or have the urge to change research directions?
于开丞: When language models were very popular, many people would think: language models are so popular, should we switch to language models?
The temptation was great, but we still held to our original intention. The logic behind it is:
First, the question of what one is good at and not good at. Language models are not what we are best at; CV is the technical field where we can operate with ease.
Second, the question of where the value belongs. Language models are certainly very valuable, but many people are following that value, so the sense of participation certainly will not be strong; but for world models, I firmly believe they will be 100% valuable in the future, and this track needs people to work on it, and I also believe I can do it well.
Looking one level deeper, there is also the question of ecological niche. Whether then or now, startups have a chance to break out on the world model track; while large language models require burning more compute and more funds to drive, which is more the business of big companies.
But the most fundamental thinking is still: only the right thing is worth doing. If large language models were the only right thing, then no matter how late the start, it must be done. But I believe large language models are not the only thing, and world models are likewise the right thing.
▎AI Technology Review: How do you assess that world models are definitely the right thing?
于开丞: First, language models cannot walk out of the screen, but world models can.
No matter how good a language model is, it still requires a person to operate it. In the future we will go to Mars, to high mountains, deep seas, and various dangerous scenarios. Many places people are not good at going to and are not suitable to go to, so something will ultimately be needed to go in place of people. Otherwise, it can only be replaced by human labor, and that is not humane.
Second, from the technical path, world models and language models are not common, and more scientific research force needs to be invested.
Language models start from text construction, and their foundation is language and text. Language and text are abstract; they are a world that has already been compressed. They are completely different from the continuous physical world faced by world models, and technically they are not easily common. They both look like models, but in reality they differ enormously.
Moreover, world models themselves still have multiple routes: 3D is one direction, film and television is another, and connecting the routes is not easy. If it were easy, there would be only one company in the world. This also shows that this research direction of world models is full of explorable space and possibilities.
So from the underlying logic, world models are 100% valuable. Then why not do it? Then why not do it early, before others have seen this direction?

Define from the purpose, do not add modifiers randomly
▎AI Technology Review: The definitions in this track are still not unified. If you, a senior world model scholar, were to give it a definition, how would you define it?
于开丞: First, in essence, a world model is a general technology like Transformer. A consensus that everyone already has is: it can predict the next state.
Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
Title: In Conversation with Westlake University's 于开丞: No Matter How Strong General Language Models Get, the Physical World "Can't Wait"
Body: Second, definition should proceed from purpose, rather than piling up modifiers first. Discussing technology in isolation from scenarios is meaningless. Only when a technology is bound to a specific target domain does the question of whether it fits arise.
The modeling approach suited to language models is not necessarily applicable to world models; the mechanisms language models excel at, such as KV cache, may not be equally effective under world model architectures.
So my definition is: a world model is a model oriented toward events in continuous physical space, with the purpose of predicting the next state. Underneath, it adopts a Transformer architecture based on diffusion models (DiT), overlaid with a prediction mechanism for the future.
Looking further from the perspective of purpose, one category is for people to watch, such as general video generation like Seedance, which emphasizes the human viewing experience; the other category is for robots to use, which requires absolute physical realism.
There are also some very absurd definitions in the industry, such as calling ICL a world model. ICL is a capability models have by default; singling it out and talking about it is pointless.
▎AI Tech Review: So, in your view, are video generation models also world models? Previously, our interviewee mentioned that "predicting the next action" is a necessary condition for a world model, and that video generation models do not predict the next action.
于开丞: Let's start from consensus. A world model predicts the next state, and "state" can be defined very broadly—one could even define a person's psychological state as a state, which also fits the essence of a world model. It's just that we emphasize that what we do is a physical world model, and we hope it understands actions in the physical world.
But as far as the definition of a world model goes, "action" is not a necessary condition. For example: to design a home renovation, you can hire a renovation team, but you need to understand the space, understand the logic, understand my needs—that is also a kind of understanding. So a physical world model does not necessarily have to have actions.
This is a misconception that arises when world models and embodied intelligence are discussed together. In reality, embodied intelligence is an application scenario; a world model is a general technology. The two cannot be conflated.
As just mentioned, judging whether a world model is real or fake requires looking at what purpose it serves and what technical trade-offs it made.
The category meant for people to watch doesn't care about realism. For example, throwing a cup and having it fall—as long as there is the motion of going up and coming down, that's enough. Many academic papers have already proven that generated content does not obey the laws of physics, but nobody cares. As long as there are no obviously counterintuitive errors like missing hands or five fingers turning into six, it passes; as for whether the hand can pinch at a very strange angle, people watching the video won't mind.
In addition, there is some comic and anime-related content where models, just to make it look satisfying, even add counterintuitive physical effects, and the training data is also mixed with a large amount of such counterintuitive effects.
The other category needs to be used in the physical world, such as the latent-space physical world model we built: wrong is wrong, right is right, and what is pursued is the accuracy of actions and task completion.

▎AI Tech Review: Leading researchers in the industry are beginning to diverge in their technical routes. The latent space you mentioned is a route advocated by Yann LeCun; there is also 李飞飞's spatial intelligence and 英伟达's physical simulation. How do you view this division?
于开丞: Routes must be anchored to the problems each is trying to solve. The biggest reason for their division of routes is that their goals differ.
We had a conversation with Yann LeCun previously. His goal is physical understanding and physical manipulation, and he takes the model route of "compressing the world into abstract representations"; 李飞飞's team works on spatial intelligence, with product application scenarios including games and virtual assets—this has been her clearly defined target direction from the start.
英伟达 is different. Its goal is simply to sell cards. Since it sells cards, it must inevitably build its story around computing, that is, generating various solvers through computation in Omniverse and turning them into physics simulators. The value of it doing this lies in: behind physical simulation is rendering, which is large-scale pipeline unit utilization built on GPUs—and that is exactly what 英伟达 is best at.
So the divergence of routes is essentially determined by genes: different goals lead to different trade-offs, and thus different routes.

A photo of 于开丞 with Yann LeCun (photo provided by the interviewee)
▎AI Tech Review: You and Yann LeCun both belong to the latent space route. He often publicly criticizes the pixel route. Do you agree with him?
于开丞: I agree with him strongly in spirit, because I myself previously took the pixel generation route, and later personally rejected the pixel route and switched to the latent space route.
On the latent space route, our latent-space physical world model took first place on six mainstream international benchmarks, including embodied intelligence and physical world understanding. These results genuinely prove that this technical route is useful. In other words, this is not merely a matter of belief—"I believe in this track"—it has been proven effective through solid results.
But as for whether his specific technical architecture itself is worth following, I remain cautious, and I am also continuously following their progress.
▎AI Tech Review: In your view, where does the superiority of the latent space route lie?
于开丞: Latent space means compressing continuous visual information into semantic space for understanding and reasoning.
First, this route can leverage the development of language models. Language models have been trained on a massive corpus and have produced a very good "value separator." They have already separated a lot of human knowledge very well at the most fundamental logical level.
When we build world models, we must inherit this separation mechanism and make it connect hand-in-hand with language models, helping the underlying world model train better. Without language models and training only the world model, the difficulty might become 100 times, 10,000 times greater.
Second, if we regard the brain of a kitten or puppy as a world model, they can understand the world, predict the future, and also make judgments and take actions. So how do we control it and make it accurately meet our expectations? That requires language. We need to use precise language to tell it what we want it to do.
▎AI Tech Review: Regarding the accuracy issue, I have a question. For example, in the scenario of a robotic arm sorting paper clips, doesn't it need sufficiently high-definition pixels to distinguish them? Would semantic space make it unable to tell paper clips apart?
于开丞: The input to latent space can likewise be high-definition pixels; it's just that understanding and reasoning take place in semantic space. Its semantic understanding is bound to physical entities: the robotic arm can see a high-definition video about paper clips, and the semantic layer identifies it as a paper clip; but that does not mean that after identification, all pixel information is discarded and only the three characters "paper clip" are retained. This semantic concept is definitely bound to a concept of a physical phenomenon.
Let's use "cat" and "dog" as another example. Language models have already established, through large-scale corpora, a clear set of vectorized logical relationships for words: "cat" and "dog" are relatively close in semantic space because they both belong to animals; "cat" and "chopsticks" are relatively far apart.
Introducing this semantic structure into world models as guidance is of great value; otherwise, the world model would have to learn two things from scratch at the same time: visually distinguishing cats from dogs, and also learning from scratch how to write the words "cat" and "dog." That is equivalent to handling multiple tasks simultaneously, and the difficulty naturally rises sharply.
The two concepts "cat" and "dog" have long been written clearly in human textbooks; a teacher teaching a child will directly tell him "this is a cat, that is a dog." But a primitive person whom no one tells "what is what" can likewise see cats and dogs, yet no one tells him what these respectively are. The speed at which the two understand the world is completely incomparable.
▎AI Tech Review: Yann LeCun's purely latent space route has a known problem called "representation collapse." You combine it with an explicit video generation framework—will you also encounter it? How do you prevent it?
于开丞: Relatively speaking it's better, but when the training volume is not large enough, it can also occur. Therefore, on the one hand, we continue to expand the training scale; on the other hand, we introduce textual information as caption constraints to suppress representation collapse, which is equivalent to propping it up with "scaffolding."
If there are no external constraints and the model is left to learn entirely on its own, representations easily collapse. Once collapsed, the model becomes ineffective and cannot work properly; as long as the loss decreases normally, the model is fine.
▎AI Tech Review: Your technical route introduces "composable physical concepts." One of the core researchers of 李飞飞's spatial intelligence route, Stanford assistant professor 吴佳俊, also has work introducing physical priors into spatial understanding and reasoning. What is the difference between your approach and 吴佳俊's?
于开丞: Our ideas are essentially the same—both introduce physical priors into spatial understanding. I think this direction is correct, and I am also close with Jiajun; he has a strong sense of mission, and the team he leads is truly excellent: among Chinese scholars, the quality of the students he leads should be among the very best.
In addition, there is another person on the same route as us, Xu Danfei, who built a compositional world model and presented results at RSS 2026.
This is great minds thinking alike. We set out this approach earlier and are no worse than them; we will do what we should do as before. And these excellent teams choosing this direction also gives us some confidence and cross-support.

▎AI Tech Review: When the world model track became hot, it came with a lot of noise, which instead drowns out the voices of some teams that are truly doing the work. How do you view this phenomenon?
于开丞: I think it has some impact, but not much. Because in the end everyone has to hand in their exam paper. The world model Awomo-0.5 we released this time at the inauguration ceremony of the Westlake World Model Center is also a submission.
This model took first place on benchmarks including Physics-IQ, RoboTwin-c2r, MolmoSpaces, and RoboDojo Sim. Taking this opportunity, we also compiled the leaderboard situation across various world model tracks, and we should be the only team able to achieve such results on so many different evaluation benchmarks.
So I also believe that in the end everyone must let their results speak. Relying on hype alone can only last for a while, and soon you won't be able to keep it up.
▎AI Tech Review: Compared with other world models on the same track, what is the biggest feature of the model released this time?
于开丞: Our company has not been established for long—only since the beginning of this year. Competitors include both established teams and new teams. So we take the "Jinggangshan" approach, using a highly focused strategy to amplify our model advantages.
Title: In Conversation with Westlake University's 于开丞: However Strong General Language Models Get, the Physical World "Can't Wait"
Body: Over the past four months, what we have done is: define the model architecture and validate it clearly, build out the model infrastructure, set up the data pipeline, build the evaluation and even the entire data closed-loop pipeline, then get it running, optimize every part of the production line to the extreme, and then push training forward.
Following this approach and this rapid-response mechanism, our model architecture iterates very quickly, and model performance keeps climbing.
In this process, the Westlake World Model Research Center has actually been deeply involved. Before the research center was officially announced, we had already started collaborating with some faculty at Westlake University, and we already had output, all of which provided the foundation for improving model capabilities. And academics, through our model, have also validated the value of their academic work.

But It Tests Whether a Model Can Sit at the "Card Table"
▎AI Tech Review: World models are still at a fairly early stage, and what gets submitted is mostly scorecards on leaderboards. But leaderboards aren't that mature either; even some leaderboard founders have said that models that climb the boards still have a considerable gap from "actually being usable." How do you view the existing problems with leaderboards?
于开丞: Indeed, the current problem with leaderboards is mainly the gap between the virtual and the real: the basic capabilities a model is evaluated on are not consistent between the virtual world and real space. So leaderboards are not yet 100% trustworthy. And precisely because of this, we have done a lot of testing and evaluation on real robots.
But on the other hand, it is also a very fair validation tool. At this stage models need these leaderboards; at the next stage they will need the next batch of leaderboards. Everyone is still climbing, and there is still a lot worth exploring. Moreover, as models keep scaling and iterating, climbing leaderboards will only get easier.
But one thing model teams must be alert to: if you make climbing leaderboards a goal and fixate on one leaderboard, no matter how high a score you get, it actually has little practical value. Especially some teams that throw their entire workforce into all kinds of dirty, exhausting work to climb one or two leaderboards—that is a very easy way to "cheat," but it consumes a lot of time and energy.
So we chose another approach that is hard to cheat on: climbing multiple leaderboards at once—single-arm, dual-arm, operation-only without video, pure video-side without operation—we test on leaderboards of different types.
And our average manpower investment in leaderboards is only half a person to one person. So I can say with confidence that the results we get on leaderboards are a test of the true capabilities of our base model.
But it cannot be denied that leaderboards test basic capabilities, not all capabilities.
▎AI Tech Review: So you're saying leaderboards don't represent everything, but they can test whether a model has "the basic capability to sit at the card table"?
于开丞: Yes. Although leaderboards are always criticized, to some extent they can still reflect this capability. More critically, they are relatively fair: leaderboards are public, which avoids the self-serving boasting of many companies—"my model does extremely well in such-and-such scenario."
Including many demos that get released—you don't even know whether they're staged, there's no way to verify, and sometimes you can only rely on an investor's personal guarantee.
Leaderboards are indeed not the only thing. Does a high score mean it's necessarily good? Not necessarily. But conversely, if you can't even climb the leaderboard, then what the model's capabilities really are becomes even more of a question mark. Because a truly general model that solves the real world well should solve the virtual space, as a simplified version, even better.

▎AI Tech Review: In September, GPT-6 Astra topped the RoboDojo leaderboard, which dealt a big shock to the world model track, and even some investors began to question "the necessity of world models." What's your view?
于开丞: I understand their logic very well, which is why I emphasized one point from the start: for world model technology, you have to look at what direction it fits. Language models are good at chain of thought; world models are good at real-time interaction with the physical world. The two are not a substitute relationship.
At a superficial level, language models were designed from the start to serve chain of thought: they first walk through the reasoning process step by step, then give a result. Even with image chain of thought added, so they can look at images, the underlying process is still a complex thinking process.
In pure text tasks, a model can think for two or three hours and it doesn't matter; but robot control is real-time, and a language model takes two or three seconds just to "think." A robot simply cannot afford to wait those two or three seconds—the task has already failed.
At a deeper level, when language models achieve embodied operation, the underlying logic is code as policy, that is, writing code. Writing code is essentially advanced industrial automation: a human breaks the task into steps, writes a program, and lets the machine execute it.
The strength of industrial automation is that it can handle some unconventional scenarios, but its weakness is also obvious: the pattern is very fixed. Every time the environment changes, you have to write a new set of code, then re-debug and re-control. In pure text tasks, a model can slowly write code and slowly debug; but the physical world changes in real time, and a robot simply cannot wait for the model to finish writing code before changing its actions.
Therefore, language models do help the evolution of embodied intelligence—that cannot be denied; but wanting them to directly replace embodied models is impossible.

With a "millet plus rifles" approach,
▎AI Tech Review: Your industry company Shuzhi was founded in January this year, but only made its public appearance in August. The track is so crowded and you started so early—why choose a relatively late time to rush into the market?
于开丞: We are rather academically self-disciplined and wanted to make the official announcement after some things had landed. Tracks have hot and slow periods; taking some impatient money is not necessarily a good thing, and a company's valuation suddenly shooting up very high may also be a problem—every choice has two sides.
So we did an internal assessment and decided to advance at our own pace: when we have solid results, we take a step forward; when there are new results, we take another step. Capital money is not free; once you take it, you bear a lot of pressure.
▎AI Tech Review: What is the relationship between Shuzhi and AutoLab, which you run at the university?
于开丞: Shuzhi was incubated from AutoLab. Right now most students are interning at the company. This is also what I wanted to do as soon as I joined Westlake University: joint industry-academia-research training. I think relying on just one lab is a bit thin; you need industry forces to participate together.
In world model research, AutoLab and Shuzhi are in a cooperative relationship: the university provides talent, the company provides compute, as well as connecting data and resources outside the university and overseas.
▎AI Tech Review: Is the company team now fully built?
于开丞: It was only around June or July that the team's structure became relatively complete. Currently 40 people, including 20 students and 20 full-time employees. Apart from a few functional staff, the full-time employees are all technical staff, basically from industry veterans and young talents. Overall it is still a research-and-development-intensive team.
▎AI Tech Review: Does Shuzhi have a benchmark company?
于开丞: A company that is a complete benchmark—there really isn't one right now. This is because of Westlake University's uniqueness: we have a great many experts and top labs in life sciences and chemistry. After the base model is done, the next thing we will do right away is empower them to do research well. Later we will announce more technical and product details, and by then the industry will definitely directly feel our uniqueness.
▎AI Tech Review: You have both deep academic experience and rich industry experience. How do you think these experiences will help you break out on the world model track?
于开丞: It is precisely because of these experiences that I have this essential understanding of world models, and can produce verifiable technical results.
By the end of October, we had completed closed-loop validation, proving that this is feasible and effective, and proving it with such a small resource footprint. We are much poorer than many companies. Frankly speaking, to build a world model and validate it as effective with so few resources—we can do it, others cannot.
This is like the Long March back then: with a "millet plus rifles" approach, you can still carve out a path. We often use this history to encourage ourselves, recalling hardships and savoring sweetness.
Moreover, we are not fighting domestic companies; we are fighting Europe and America. Looking at the domestic track, we would rather find "one round table" of partners to go together and take on European and American AI in the ring.
And precisely because of this, over the past period we have been gathering momentum: the logic with the university, the relationship with the government, the capability building of the technical model, the relationships with partners—all were prepared during this period.
Next, we are going to scale up infrastructure. We have moved relatively slowly because we want to move more steadily, to truly lay a solid foundation, and to truly scale the entire technical progress of world models.

Why Set Up a Dedicated Research Center?
▎AI Tech Review: You have AutoLab at Westlake University, you've founded the industry company Shuzhi, and now you're also leading the establishment of the Westlake World Model Research Center. What is it that the current lab and company cannot carry and that needs a separate dedicated research center to take on?
于开丞: Setting up a dedicated research center is to make industry-academia-research more solid, deeper, and more influential.
Previously it was AutoLab combined with Shuzhi, and the cooperation was good, but the force was very small. If we can use this opportunity to unite all of Westlake University's strength to do this, it will be different.
So we deliberately named the research center "Westlake World Model Research Center," hoping that world model research will take Westlake University as its starting point and subsequently attract more people to join. The research center's first batch of partners includes German institutions, as well as domestic companies and startup teams, and it has already begun to "step out" of Westlake.

▎AI Tech Review: Roughly when did you have the idea of setting up a research center?
于开丞: I had this idea at the end of March. At that time President Shi Yigong had actually already agreed, but the specific establishment required a series of approval processes: from our Department of Artificial Intelligence and the School of Engineering to the university council, all needed level-by-level review and advancement. The entire university leadership knew we were going to set up a world model research center. We also widely solicited opinions, and this research center can be said to have the unanimous support of the whole university.
▎AI Tech Review: Beijing has the General World Model Lab, and there is also the Joint World Model Lab jointly established by Peking University and Inverse Matrix. Westlake University is essentially the first world model center set up by a purely academic institution. How do you view its role and positioning?
Title: In Conversation with Westlake University's 于开丞: However Strong General Language Models Get, the Physical World "Can't Wait"
Body: 于开丞: We hope the research center will first take Westlake University as its starting point and build solid basic research; then, in the next stage, apply to provincial and even national-level platforms to take on vertical research missions.
Looking at the industry, world models are currently mainly used for embodied intelligence, but in terms of application scenarios, we believe there are still many areas that urgently need exploration—for example, can world models be used to do scientific research and drive progress in disciplines such as biology and chemistry?
Our greatest uniqueness is that what we want to build is a general world model, so that this general-purpose technology can be applied to all kinds of basic disciplines and different industry sectors.
▎AI Tech Review: You now hold three roles at once—you run a lab at the university and supervise students, you are a company founder, and you head the World Model Research Center. How do you coordinate these different roles well?
于开丞: This benefits from Westlake University's talent mechanism, which ensures that my roles do not conflict with one another. In a sense, one major characteristic of Westlake University is that it can shorten the conversion cycle from research to industry as much as possible. This is also the foundation that allows me to make research and collaboration between academia and industry work.
These three roles are essentially all about doing one thing: building a general world model, truly making the technology work, empowering interdisciplinary academic questions and scenario applications across various fields, and truly bringing value to society.

In research, "not pretending to work hard" matters
▎AI Tech Review: Your two PhD advisors, Pascal Fua and Mathieu Salzmann, are both leading figures in geometric vision. What influence have they had on you?
于开丞: Pascal is a very romantic Frenchman, very smart, and also very hard-driving—even now, in his sixties, he is still very hard-driving. He has excellent judgment in academia and insists on starting from intuition to grasp the essence; I quite admire that about him.
Mathieu is now deputy director of the Swiss Data Science Center. My deepest impression of him is that he can compress a day's work efficiently and precisely into 8 hours. He arrives punctually at 8 a.m., coming right after dropping off his kids; he leaves punctually at 6 p.m. He can maintain a full mental state the whole time—his time-control ability is very strong.
This state is very reasonable. Working hours and actual work efficiency are often inversely related: the longer the working hours, the lower the work efficiency. If you work 16 hours a day, your brain stops turning. For writing code and truly thinking through and solving problems, 8 hours a day is already the upper limit.
The ones who work the most overtime are PhD students in the United States. They wear themselves out all day, but their contributions are not high. Chinese students who go to the United States for a 5-year PhD—don't they also publish just three to five papers? There is no essential difference from the output of European PhD students. So whether to engage in rat-racing ultimately comes down to whether you face the real you.
Overall, the European PhD experience and my two advisors gave me a lot of inspiration, gave me more judgment about the diversity of life, gave me another layer of thinking about the completeness of life, and subtly influenced my entire research rhythm and persistence.

A photo of 于开丞 with the two advisors (photo provided by the interviewee)
▎AI Tech Review: One important piece of work during your PhD was discovering that "a network designed by AI itself stops working when you change the dataset," and this problem was listed by the NAS workshop as one of the three core problems of the entire field. How did this problem-oriented way of thinking take shape?
于开丞: You've done very thorough background research on me (laughs). Problem orientation is indeed my basic approach to academia. Looking back now, why my academic research went so smoothly—the key lies in this problem-oriented academic approach.
In fact, my first paper in my first year of the PhD was rather trivial; it was a topic given by my advisor, not so problem-oriented, so working on that topic was quite painful. Although I could produce decent results, it didn't feel very good.
The second piece of work was the NAS problem you just mentioned. At the time, we first followed a piece of basic research from a certain big company and wanted to continue optimizing on that basis. But in the process, we quickly found that no matter what ideas we tried, it was very difficult to improve the long-term and stable performance of the searched model. At that time it was quite despairing. Because a large number of our experiments showed that many baseline works performed even worse than random search.
But then I thought, could it be that the research foundation itself has an underlying flaw? At first, because this came from a big company's research, we didn't think in the direction of "the basic assumption may be problematic." But once this thought emerged, I worked with an intern in the lab to do a large number of experiments, and finally discovered and pointed out the underlying problem of this work—that is, a basic assumption of the previous work may fundamentally not hold. This discovery triggered a lot of discussion in the field at the time and led to many subsequent works.
Therefore, through this work, I understood more deeply: doing research should not presuppose that some premise definitely holds; you must go back to the most fundamental logic and dig deeper: does this premise actually hold?
It was also from that time that I developed this fundamental-level way of thinking. Before starting a study, I first think: is this thing actually worth doing? Now that I supervise PhD students myself, I also advise them to use this fundamental-level thinking to judge the value of a topic. But not everyone can accept this view.
(Author's note: 于开丞's work in the NAS direction during his PhD was a chain of "first find the problem, then give the solution." The ICLR 2020 paper "Evaluating the Search Phase of Neural Architecture Search" pointed out that mainstream NAS algorithms perform about the same as random sampling in the search phase, and that weight sharing disrupts the ranking of candidate architectures, causing search to fail; this finding was listed by the NAS workshop as one of the three core problems of the field. The CVPR 2021 paper "Landmark Regularization: Ranking Guided Super-Net Training in Neural Architecture Search" proposed a ranking-guided supernet training method to improve the robustness of various NAS algorithms, providing a new line of thinking for the above problem. The Qualcomm Innovation Fund corresponds precisely to this NAS research line.)
▎AI Tech Review: The research community is quite competitive now, and problem-oriented methods often produce results less quickly than tuning methods.
Answer: I both agree and disagree; I think this is more of an excuse for "not thinking."
Many people often make themselves very busy and very tired, saying every day, "Teacher, I worked hard," but I think they have not faced the real themselves: they keep filling up their time, as if they are busy, but when you ask about results, there are simply no results.

In the world-model startup track, which is full of anxiety everywhere, 于开丞 has a rare sense of ease.
As a "number one" who holds three roles across industry, academia, and research, and an entrepreneur who has just led a forty-person team through six leaderboards, when asked whether he gets very tired, 于开丞's answer is: "No, my life is very regular, and I also really enjoy this process."
In his view, when a person is under excessive pressure, it absolutely means the underlying operating mechanism is not running smoothly, and the core energy is not being placed on the most valuable things. For him, "tiredness" is not proof of hard work.

A photo of 于开丞 in daily life (photo provided by the interviewee)
Where does this state come from? The answer lies in his five years in Europe during his PhD. From his two advisors, Pascal Fua and Mathieu Salzmann, he learned a lesson he still benefits from today: grasp the essential contradiction.
This lesson was first engraved into his research path. In his second year of the PhD, he identified an industry-level core problem in NAS, proposed a new solution, and won the Qualcomm Innovation Fund; after returning to China and joining DAMO Academy, he became famous in the autonomous driving circle with BEVFusion; and then, because of a real-world dilemma in autonomous driving, he relied on research intuition to plunge headfirst into the uncharted territory of world models.
From NAS to autonomous driving, and then to world models, every time he changed direction it was not about chasing hot spots, but about being pushed forward by a concrete problem.
What supports this research path is another kind of cultivation: not engaging in involution, not "pretending to work hard." He dares to face himself, repeatedly asking himself "does what I am doing actually have value," and does not do things without value. In his view, many students now "look busy and tired, yet cannot produce results," which is precisely an excuse for "not thinking," and even more the most wasteful state in research.
Putting these two things together, his later seemingly "unreasonable" choices all become logical.
Why did he walk in the uncharted territory of world models for three years and wait until August this year to let the industry company appear publicly? Because before technological results are verified, appearing publicly is meaningless.
Why, amid doubts everywhere asking "is there still any need for world models," did he still firmly push forward the establishment of the Westlake World Model Research Center? Because he had asked himself this question thousands of times, and the answer was "100% valuable."
This certainty may also be related to his study experience in Europe. There, he saw the diversity of life and gained another layer of thinking about the completeness of life. Academia is not the only path, and success does not have only one shape. People who have seen a broader world are, instead, less likely to panic.
Thus, amid the clamor in which "fast research" has become the default rhythm, he waited until everything was ready before releasing the product: the general world model Awomo-0.5, with 4 billion parameters, simultaneously topping six mainstream international evaluation leaderboards, including embodied intelligence and physical world understanding.
And the battle he wants to fight has never been only domestic. The "people around one round table" in the domestic world-model track need to be brought together, to fight on the arena controlled by AI forces in Europe and America.
"The right thing is what is worth doing." This sentence, which he has said many times, is now not only a belief he tells himself, but also a public answer given by "number one on six leaderboards." Leiphone Leiphone (official account: Leiphone) Leiphone
Leiphone original article, unauthorized reproduction prohibited. For details, see the reprint notice.
