must be resolved; otherwise, data will ultimately become the biggest bottleneck in the implementation of these technologies. This is the conclusion drawn by a former Microsoft researcher from Asia. As for the other , they are not directly related to the main topic but provide additional context.
Author: Deng Zheming,
Editor: Qi Chengyong,
Earlier this autumn, technologies such as Skild S1, Generalist GEN-1.5, and RoboTTT made their debuts, all bringing context learning (ICL) into the field of machine learning. AI Technology Review (Leifeng.com) noted that a domestic team also released a world model capable of context learning around the same time, led by Professor Cao Ting from the Institute of Intelligent Industry at Tsinghua University (AIR). In 2009, Cao Ting went to Australia to pursue her Ph.D., where she studied the trade-off between power consumption and performance between asymmetric multi-core processors and managed languages like Java under the guidance of two leading scholars in the field of programming languages and systems. Her doctoral thesis was selected as a highlight of CACM research and was specifically recommended by David Patterson, the winner of the Turing Award, who also included it in his classic textbook “Computer Architecture: A Quantitative Approach.” It is worth mentioning a paper published in ISCA titled “The Yin and Yang of Power and Performance for Asymmetric Hardware and Managed Software.” The argument in this paper is quite poetic: additional overheads such as garbage collection and just-in-time compilation in virtual machines can be offloaded to the smaller cores, using the “yang” (positive aspects) of the smaller cores to compensate for the “yin” (negative aspects) of the virtual machine’s overhead. Conversely, the complexity inherent in asymmetric hardware is precisely alleviated by the abstraction provided by managed languages. In our view, this experience laid the foundation for Cao Ting’s academic career, particularly her intuition about the concept of “yin and yang”: constraints are not just limitations; they can also be the starting point for fostering innovation and creativity in systems. After completing her Ph.D., Cao Ting joined Huawei’s Compiler and Programming Languages Laboratory, where she was responsible for the memory management of the Ark compiler, working to make programs run faster and more efficiently within limited memory resources.In 2018, she joined the Heterogeneous Computing Group at Microsoft Research Asia (MSRA), where she stayed for seven years. Her work focused on developing AI inference algorithms for mobile devices. She managed to integrate neural networks with billions of parameters into smartphones, PCs, and IoT devices, enabling them to function locally without relying on cloud computing, thereby reducing power consumption and storage requirements. Technologies such as low-bit quantization, LUT (Look-Up Table)-based inference, and integrated computing-acceleration systems were incorporated into Microsoft Office, Windows, and Bing, benefiting millions of users. In the summer of 2025, Cao Ting joined Tsinghua AIR. This time, she faced even more challenging constraints: a severe shortage of training data and a constantly changing physical environment. She quickly realized that robots could perform tasks correctly under fixed conditions, but would fail when the environment changed slightly; the valuable information generated by these failures was often discarded by traditional systems. Although embodied intelligence (intelligence embodied in physical devices) had already been demonstrated in prototypes, it couldn’t be successfully deployed in practical applications. Cao Ting’s approach was to turn these constraints into opportunities for innovation. In October of that year, she and her former colleague from MSRA, Professor Liu Yunxin (the deputy director of Tsinghua AIR), initiated the Z-Trans project, which focused on the self-evolution of embodied systems. They subsequently released several research papers, including Z0, Zeva, Zeva Ego, Zetta, and Z-infra. Over the past decade or so, Cao Ting’s work has consistently focused on enabling intelligent systems to function effectively under limited conditions. Her efforts have ranged from reducing power consumption in software on heterogeneous hardware to improving the computing efficiency of models on edge devices, and ultimately to enabling intelligent systems to learn and adapt in the real world. The following is a conversation between AI Technology Review and Cao Ting, with some editorial adjustments made to maintain the original meaning: **AI Technology Review:** “The overall technological paradigm has not yet stabilized and is changing rapidly—almost every month. However, the concept of offline and online learning paradigms is widely recognized.” **Cao Ting:** “Indeed, the technology is evolving rapidly. In the first half of this year, the industry was still focused on building large-scale data collection facilities, hoping to achieve breakthroughs with massive amounts of data. But in the second half of the year, many studies focused on embodied intelligence and context-based learning. Have the previous research findings provided the necessary insights?” **AI Technology Review:** “Yes, previous research has shown that traditional approaches based solely on offline training are insufficient. Robots need to be able to learn from new demonstrations and feedback in real-time environments.” **Cao Ting:** “That’s exactly right. We realized that offline data alone cannot cover all real-world scenarios, so we need to develop systems that can adapt to changing conditions.”
▎AI Technology Review: Have everyone reached a consensus on the technical aspects?
Cao Ting: Overall, there are two types of learning paradigms: one is offline, and the other is online.
The concept of learning in context first emerged from large language models (LLMs); GPT demonstrated that context examples are sufficient for generalization. In the field of embodied systems, people also hope that by demonstrating scenarios to robots in a contextual setting, the robots can generalize to various situations. Currently, the overall technical paradigm has not yet converged and is changing very rapidly—almost every month there are new developments. However, the offline and online learning paradigms are generally recognized by everyone.
▎AI Technology Review: The ICL (In-Context Learning) of large language models works through emergent behavior; given a few text examples, the model can learn. Could you help clarify the differences between ICL in the embodied systems and outline the current research directions in this area?
Cao Ting: ICL in the embodied systems is much more complex than that in language models. There are currently three different research directions in the industry.
The first approach hopes that abilities will emerge spontaneously, similar to what happens with LLMs. For example, some models claim their abilities arise naturally after being trained with a large amount of human video data and ego-related information; then, given a new video, the model can perform tasks on its own. However, I think this approach is quite challenging and may be subject to many limitations.
The second approach is explicit training, similar to Skild, where the model is taught to predict the next action based on a given video. This is not an emergent process but rather a result of explicit training. It’s similar to the previous VLA approach, where language prompts were used for input and actions were the output.
Our approach is the third one, called ICCL (In-Context Causal Learning). We believe that directly providing a video to the model and asking it to generate an action is very difficult, as the video may contain moving people, irrelevant objects, and reflections, which can make it hard for the model to extract useful action information from the noise.
▎AI Technology Review: Does your approach involve removing all the “noise” from the video data?
Cao Ting: Yes, we have added a module that extracts the action and its effects, reducing the interference from irrelevant visual elements. We only include the action and its resulting effects in the context. From our observations, this method shows better generalization than other models, especially in scenarios with complex backgrounds.
▎AI科技评论：如果ICL这条技术路线真的能够成功实现，那么后续的“后训练”步骤是否就变得没有太大意义了？
曹婷：这需要分别来看。ICL的优势在于其极高的泛化能力——即使面对从未见过的新任务，它也能迅速学会如何处理这些任务，因此特别适合家庭环境等对泛化能力要求很高的场景。
然而，在工厂的生产线上，人们更关心的是模型的精度和响应速度；客户通常要求模型具有99%甚至更高的精度。目前，ICL还远未达到这一水平（例如Skild的报告指出，在面对未知任务时，ICL的步骤成功率仅约为60%）。因此，在变化不那么频繁的柔性工业生产环境中，收集几百条数据并进行后训练仍然是保证模型精度最经济、最可靠的方法。
家庭环境依赖ICL的泛化能力，而工厂则通过后训练来确保系统的稳定性——这两种逻辑是完全不同的。
02
何谓因果关系？
“模型是由数据驱动的，它们从大量数据中学习规律并发展出相应的能力。”
▎AI科技评论：之前提到过，你们的技术路线中包含一种用于提取动作与效果之间的关联的技术，这是否就是你们的核心技术？
曹婷：没错。我们最初开发的是Zeva，它基于Cosmos架构；不过我们的技术并不局限于特定的基础模型（如World Model或VLA），也可以应用于其他模型。后来我们还推出了Zeva-Ego，这是一种基于VLA数据和Ego数据进行的模型训练方法。
这两项工作是相互关联的：Zeva侧重于上下文中的因果关系学习，它能够在线提取动作与效果并将其融入到模型中；而Zeva-Ego则从Ego数据中提取这些信息来进行模型训练，从而显著提升模型的泛化能力。
从根本上说，我们的核心技术就是用于提取动作与效果之间因果关系的工具。
▎AI Technology Review: Did you collect the data used by Zeva-Ego yourself?
Cao Ting: We used open-source data, but we had to clean it first because most of it was not suitable for direct use in action learning. In the end, we only kept about one-tenth of the data.
We will train the causal extractor in two stages. In the first stage, the model learns about environmental changes that are not related to the actions. There is a dedicated network that extracts information that is not related to the actions. In the second stage, we use the environmental representations learned in the first stage to train the causal extractor to extract only information related to the actions. After training, we use the ego data without action annotations to extract causal relationships, which will be used in subsequent training.
▎AI Technology Review: Using unlabeled data directly can save a lot of annotation work and allow for more efficient data usage.
Cao Ting: Yes, we used 10,000 hours of unlabeled ego data to train the model. We used this data to convert it into action-related supervision signals through our encoder.
AI Technology Review: Regarding causality, when the world models first became popular, many people claimed they were causal world models. However, there is still no solid evidence to prove that robots have truly learned physical laws. How do you prove this to others?
Cao Ting: There is a lot of discussion about “causality” in this field. By “causality,” we mean the relationship between an action and its effect: the action is the cause, and the effect is the result. For example, we determine how much a robotic arm needs to push a bottle to move it.
How can we prove that our approach is effective? Firstly, Zeva performs better on public evaluation datasets. In real-world scenarios, our model learns more efficiently compared to other models in the industry, and it requires fewer attempts. For example, while some models may need five to six attempts to pick up a coffee cup, our model can usually succeed on the first or second try.
▎ AI Technology Review: The phrase “understanding physical laws” might suggest a rational understanding based on scientific principles, but in reality, it’s more akin to the chaotic, empirical way humans understand things, right?
Cao Ting: That’s right. Models are data-driven; they learn patterns from large amounts of data and develop their capabilities through that process. You could also directly explain the principles of gravity to them, but they need to experiment on their own. Just like children who, even before learning physics, know how to behave in certain situations—this knowledge comes from a lot of trial and error.
We believe the most reasonable approach is to use extensive pre-training combined with ongoing online learning, so that robots can learn the patterns on their own. I’ve drawn a roadmap showing that robots will eventually become like physicists, capable of formulating new laws on their own.
▎ AI Technology Review: AI for Science?
Cao Ting: Yes, that’s the direction we’re heading towards. But for now, it’s already quite impressive that they can perform these tasks well.
03
The cycle of learning and memory is essential: “Memory and learning are inseparable. When you accumulate enough experience, you can generalize it to handle new, unseen tasks—this is what learning is all about.”
AI Technology Review: When discussing ICL (Intelligent Cognitive Learning), it’s inevitable to mention long-term memory and RSI (Recursive Self-Improvement). What’s your view on the relationship between these three concepts?
Cao Ting: I’ve previously discussed the relationship between memory and learning within AIR. Memory can be divided into several layers. The shortest-term memory, known as working memory, provides context for the robot while it is performing a task. Long-term memory, on the other hand, stores knowledge acquired from past tasks, which can be retrieved and used to assist in decision-making when facing similar tasks in the future.
Memory and learning are indeed inseparable. When you accumulate enough experience, you can apply that knowledge to new, unfamiliar situations—this is learning in its essence. In other words, learning happens when you have enough experience to apply it broadly to new tasks. Embodied RSI, which is the physical world's self-improvement (PSI) that we have defined and integrated, is a more macroscopic closed-loop. Online learning allows robots to accumulate data from successes or failures, and then use this data to retrain the model itself, thereby improving the base model's capabilities. With the base model's capabilities improved, it can complete new tasks more quickly and accumulate more data. This flywheel is the core logic of PSI. ▎AI Tech Review: Currently, we can't directly complete new tasks without trying, but in the long run, people definitely hope robots can achieve the desired outcome in one go. In the future, will work related to the current context give way to work focused on consolidating long-term memory? ▎Cao Ting: You are absolutely right. All our data flows back, and we will soon release a new piece of work where after robots try themselves, we retrain the model with their data, and the model's performance improves significantly. Just like when humans try many times, they convert all that experience into the cerebral cortex as long-term memory. ▎AI Tech Review: After the release of GPT-6, some people worry that embodied models will be "eaten up," how do you view this issue? ▎Cao Ting: Yes. GPT-6 has strong capabilities, but being able to understand tasks, plan steps, and call tools doesn't mean it inherently possesses reliable physical execution capabilities. A real-world issue is latency, especially in operations that require quick responses. By the time the model returns a result, the on-site situation may have already changed. ▎AI Tech Review: Is this latency issue solvable? ▎Cao Ting: It's not realistic to rely solely on optimizing GPT itself to meet the requirements for real-time control of robots. Large models like GPT are definitely aimed at pursuing the upper limit of intelligence, aiming for AGI where a single model can do anything, and the model parameters are getting larger and larger. Moreover, when performing complex tasks, it also needs a thinking process and considers network transmission, and these overheads are hard to reduce. ▎AI Tech Review: GPT is like a comprehensive library, while embodied models are a specialized small library. ▎Cao Ting: Yes. We have long determined that models like World Model and VLA, which are smaller brains, definitely need a brain similar to GPT on top of them, so we have always been moving forward with agent combined with models, where the model is Zeva and Zeva Ego, and the agent (or harness) is Zetta.
Specifically, we extract the embodied capabilities and train them on a small VLM model, which can be considered as an edge-side GPT model. Now we have trained it to tens of B parameters, and the performance is basically the same. This model serves as the brain, responsible for understanding user instructions, decomposing long-term tasks, monitoring task execution, generating recovery actions, and calling lower-level tools and models.
04
Empowering industry is the ultimate goal.
"Although the investment market has been a bit cold recently, the actual industrial demand is particularly large."
▎ AI Tech Review: As a professor at Tsinghua AIR and co-founder and CTO of Domain Transformation, do you have any advantages in terms of industry applications?
Cao Ting: AIR was founded by Professor Zhang Yaqin, who previously served as a global vice president at Microsoft and later as CEO at Baidu. Therefore, most of the people here come from the industry, such as myself, who returned from Microsoft Research last year; Professor Zhou Guyu, who was once one of the early core members of DJI and founded QiuZhi Technology, which recently raised several billion yuan; and Professor Chen Yilun, who previously served as CTO of Huawei's intelligent driving division.
We call it the Intelligent Industry Research Institute, and the word "industry" was suggested by Professor Zhang Yaqin to be added, as he emphasized that we must empower industry. The professors at the institute generally start their own startups, mainly focusing on two areas: biomedicine and embodied intelligence.
▎ AI Tech Review: Does this mean that compared to other institutes like the Fork Institute or the Institute of Automation, AIR is closer to the industry and has more experience in practical implementation?
Cao Ting: The Fork Institute also has many startups. Our strength at AIR is that Professor Zhang Yaqin has a significant influence in the industry, both domestically and internationally, which greatly helps us in acquiring resources and support from the industry sector. AI Tech Review: In addition to the work to be released in November mentioned earlier, what has Domain Transformation been doing recently? Cao Ting: We have placed more emphasis on industrialization. After building the pipeline from a model that couldn't do anything to one that could, we need to validate it in real-world scenarios. In the past few months, we've frequently visited many factories. Although the investment market is a bit cold, the actual industrial demand is very large. Most of the tasks are on assembly lines, such as screwing in screws, and in risky industries, such as chemical plants, robots are very needed to avoid people entering such environments. Additionally, we are developing a home-operating robot. The home environment still has a high ceiling for embodied intelligence, and it's the best way to demonstrate the model's generalization ability. We have set up the entire bedroom, bathroom, and living room of a house in the Xiongan Research Institute, and the robot can perform tasks throughout the entire house. AI Tech Review: Is the team currently mainly composed of doctoral students? Cao Ting: There are also engineers on the industrial line, but mainly doctoral students. Our doctoral students are very strong. They are not only active in generating ideas, but also have very fast hands-on capabilities. Once they have an idea, they can quickly implement it. Moreover, they have a strong grasp of the changes in external technologies, always keeping track of the latest developments. Everyone shares what new work they've done today, and if they find a certain technology particularly promising, they will immediately try it out. One day, a classmate was very excited and told me that he found out that just by tuning GPT, the entire robotic arm could go from knowing nothing to completing a task. I said, that's definitely something we need to do with a self-evolving closed-loop system. Later, we did Zetta, and it took us a month to complete it. Now that it has been released for only two months, it has already been cited extensively. We have seen the influence of this direction, and in the future, this might be the correct path. After that, many companies that saw this work came to us, surprised that they had just realized how important this issue was, while we had already published the paper. I think this is still the credit of the students. AI Tech Review: How many people are there now? Cao Ting: Approximately 20 people. When they find this matter particularly interesting and impactful, they don't need me to urge them; they will move forward on their own with great enthusiasm. In fact, I often ask them if they are tired and whether they need a break.
05
Written at the end
Gao Ting once said with a hint of pride during the exchange: "When the industry suddenly realized the importance of embodied self-evolution, we had already been continuously publishing results." Indeed, Domain Transformation's recent work closely aligns with the industry's hot topics.
In June, Z0 launched the slogan "building continuously self-evolving physical intelligent agents," and before RSI became a mainstream topic, it was the first to consider the overall technical framework, integrating data feedback and model self-iteration into a closed loop. On August 18th, Zetta ζ, which handles the Harness function, was released; on September 1st, Zeva, which achieves embodied contextual learning, and Skild, Generalist, and others doing similar ICL work were released almost simultaneously. Within a month, Zeva-Ego officially made its debut and entered mainstream discussions on the value of ego data.
Tsinghua AIR is not a traditional university laboratory; its positioning has always been integrated research, development, and application. The resource endorsement from the academicians helped the team establish collaborations with institutions like NVIDIA early on, pushing cutting-edge technology to larger scenarios for validation. Gao Ting brought seven years of engineering intuition from the industry research institute, as well as a mindset naturally designed with deployment as the ultimate goal. The doctoral students from top universities are closest to the academic frontier, full of research enthusiasm, quick to absorb and act, and not constrained by path dependencies.
Hot topics are the sunny side, bustling and noisy, visible to all; problems are the shady side, quiet, tough to tackle, and rarely visited.
During the exchange with LeFeng Network(WeChat official account: LeFeng Network), Gao Ting remained smiling and approachable. Outside of work, she climbs mountains and runs, maintaining an excellent overall state. You find it hard to associate this gentle and elegant person with the hardcore researcher who has been working in the shadows for decades. Perhaps the secret to her repeatedly hitting the hot topics lies in her ability to focus deeply in the shadows and step forward confidently in the sunshine.
LeFeng Network's original article; unauthorized转载 is prohibited. Details see转载须知.
