Tsinghua AIR's Cao Ting on Embodied ICL: ICCL Causal Extraction, On-Device Brains and the Limits of Post-Training
Leiphone published on November 12, 2026 an interview with Cao Ting, a professor at Tsinghua University's Institute for AI Industry Research (AIR) and co-founder and CTO of the startup Yubianhuan, whose name is given here in romanized form. She says embodied in-context learning has three routes, and her team is pursuing ICCL, in-context causal learning, which uses a two-stage action-effect extractor to turn unlabeled first-person data into action supervision signals; she also says home scenarios rely on ICL generalization while factory lines still rely on small-data post-training to hold precision. The report does not mention open weights, licensing or an API.

What changed
The report opens by saying that in early autumn this year, work such as Skild S1, Generalist GEN-1.5 and RoboTTT appeared in quick succession, bringing in-context learning into the embodied field. According to the interview, embodied ICL currently has three routes: first, expecting capabilities to emerge online the way they do in language models, which the report illustrates with Generalist; second, explicit training that has the model predict the next action directly from a video prompt, illustrated with Skild; third, ICCL (in-context causal learning), proposed by Cao Ting's team. ICCL starts from the observation that videos contain moving people, irrelevant objects and reflections, making it hard for a model to distill action information from them, so the team adds an action-effect extraction module that puts only actions and their resulting effects into the context. According to the interview, the team's earliest model, Zeva, started from Cosmos, but the method is not tied to a base model and both world models and VLAs can use it; the later Zeva-Ego is built on a VLA and uses first-person (ego) data for mid-training. The causal extractor is trained in two stages: the first stage has a dedicated network learn environment changes unrelated to actions, and the second stage, drawing on the environment representation from the first, extracts only action-related information and then converts ego data without action labels into supervision signals. The scale given in the report is 10,000 hours of unlabeled ego data; open-source data needs cleaning, and roughly only one tenth survives cleaning. On the timeline, the report says Z0 was released in June, Zetta, which plays the harness role, was released on August 18, Zeva was released on September 1 alongside Skild, Generalist and others, and Zeva-Ego appeared within a month; a new piece of work that feeds robot self-collected data back to retrain the base model will also be released soon. She also describes memory as two layers: working memory for the current task, and a retrievable, reusable cross-task long-term memory; and says her definition of PSI (physical-world self-evolution) is a closed loop in which online learning accumulates success and failure data, the data flows back to retrain the model, and a stronger base model in turn accumulates data faster. On architecture, the team works on two layers: distilling part of the embodied capability into a small VLM trained to a few tens of billions of parameters, which Cao Ting says performs basically the same as a larger model, and this small model serves as an on-device brain responsible for understanding user instructions, decomposing long-horizon tasks, monitoring task execution, generating recovery actions and calling lower-layer tools and models; the lower layer consists of world models and VLAs such as Zeva and Zeva-Ego, while Zetta belongs to the agent or harness side. Her judgment on having large models perform control directly comes down to latency: by the time the model returns a result, the state on site may already have changed, and large models pursue the upper bound of capability, so the cost of their reasoning process and of network transmission is hard to reduce. The interview also relays Skild's reported figures, saying the ICL step success rate on unseen tasks is only slightly above 60 percent.
What it means for developers
For developers working on embodied models and data pipelines, the most direct effect of these claims lands on the data side. The two-stage causal extraction described in the report turns unlabeled first-person video into action supervision signals, shifting the center of gravity of the budget from labeling toward cleaning and extraction; but "only one tenth survives" is the interviewee's rule of thumb, not a universal constant. If your ego data comes from more varied sources and has lower action density, you need to measure the retention rate and its usefulness yourself first. A practical way to check this step is to fix a small unseen-task set with cluttered backgrounds and compare step success rates with and without the extractor, along with the number of attempts from first try to success; the reference given in the report is the team's own account of succeeding in one or two attempts against five or six for comparison models, which is a self-reported order of magnitude and can only serve as a reference. According to Skild's report as relayed in the interview, the ICL step success rate on unseen tasks is only slightly above 60 percent; that number is a relay of a third-party report, and if you want to use it as a threshold you should go back to the original report to check it. Evaluation criteria need to be separated by scenario. The report places ICL on the side of frequently changing tasks that demand generalization, and post-training on the side of fixed form factors that demand above 99 percent precision and low latency. That amounts to saying the same set of models faces different acceptance standards on the two sides: the former looks at unseen-task success rate and number of attempts, the latter at precision and per-step latency, and the two cannot be used to refute each other from a single scorecard. Latency decides the layered structure. The report states plainly that a large model being able to understand a task, plan steps and call tools does not mean it has reliable physical execution capability, and that on-device latency and network overhead are both hard to compress; the team therefore distills capability into an on-device small VLM of a few tens of billions of parameters, making it the brain, while real-time operation is left to the lower-layer world models and VLAs. For developers, this turns "planning in the cloud or on the premises" from a matter of preference into a matter of latency budget: if the control loop requires fast responses, it is worth first falsifying, with end-to-end latency, whether a cloud-side large model should take part in every step of the loop. The last point is behavioral drift from data flowing back. The report says all data will flow back, and that self-collected data will soon be trained back into the base model. Once such a loop is live, model behavior keeps changing with deployment. This is an inference from the report rather than a reported fact: without version pinning and regression testing, online performance can change on days when you have not touched the code, and setting observation criteria and rollback options in advance is less trouble than attributing a cause afterward.
Integration and evaluation
The following three items are recommendations and conditional analysis based on the report, not facts from the report, and they do not mean the team's approach will necessarily hold in your scenario. Decision one: whether to make online in-context demonstration or post-training the main path. The change the report describes is the industry moving from offline data-collection factories toward online in-context learning after deployment; Cao Ting says that no matter how much offline data you collect, it cannot cover all the situations in the real world. The choice is whether to wire the online demonstration and context pipeline into the loop first, or to run a round of small-data post-training first. The triggers are how often tasks change, whether the site can reliably provide demonstrations, and the precision requirement: if tasks change every shift and trial and error is acceptable, starting with the context pipeline pays off more; if the task form is fixed and requires above 99 percent precision and low latency, the report's position is that collecting a few hundred samples for post-training is still the most economical and reliable approach. The quantifiable criterion is the unseen-task step success rate and the number of attempts needed for first success on a self-built cluttered-background test set, treating the relayed third-party figure of slightly above 60 percent as a lower-bound reference rather than a target. Decision two: whether to build the action-effect extraction and data-cleaning stage first. The change the report describes is using a two-stage causal extractor to turn unlabeled ego data into action supervision signals; the team used 10,000 hours of unlabeled data. The choice is whether to invest in an extractor and cleaning pipeline, or to invest directly in labeling. The precondition is that you already have a large amount of unlabeled first-person video and can accept a retention rate on the order of one tenth after cleaning; if neither holds, the premise of the path fails, and manual labeling or switching to data that carries action labels is more realistic. The quantifiable criteria are the number of usable samples after cleaning and extraction, the increment in downstream unseen-task success rate, and whether success rate on cluttered backgrounds is clearly higher than a baseline that does no extraction. Decision three: following the division of labor in the report, treat the on-device small model as the upper-layer brain rather than putting long-horizon understanding in a cloud large model. The change the report describes is distilling part of the embodied capability into an on-device brain, a small VLM of a few tens of billions of parameters, responsible for understanding instructions, decomposing long-horizon tasks, monitoring execution and generating recovery actions, with lower-layer world models and VLAs handling actual operations, on the grounds that a large model's reasoning process and network transmission overhead struggle to meet real-time control. The choice is whether to design around this "on-device brain plus lower-layer execution models" layering, or to keep using a cloud large model for planning inside the loop. The triggers are how sensitive the control loop is to latency and the network conditions on site: the heavier the fast-response operational tasks, the more you should first verify whether the on-device brain can reach acceptable capability; if the task pace is slow and replanning can tolerate second-level round trips, the cloud approach may not need replacing. The quantifiable criteria are the end-to-end latency of the same control loop, especially high-percentile latency, and the recovery success rate when the brain generates recovery actions after an execution failure and keeps the task moving. One addition: "a small model of a few tens of billions of parameters performs basically the same as a larger model" is the interviewee's own account and needs comparison on your own task set before you adopt it.
Limits and open questions
This report is an edited interview, and almost all the results cited are the team's own account or second-hand relay: "better results" on public benchmarks, "it can pick it up in one or two attempts", and training a small VLM to a few tens of billions of parameters with "basically no difference in results" are all given without benchmark names, baseline lists or reproduction scripts; Skild's slightly-above-60-percent figure comes through the interviewee and is a relayed account. The three ICL routes are the interviewee's division, not a unified industry evaluation framework. The "only one tenth survives cleaning" figure is a self-reported rule of thumb; the specific rules and the composition of the data are not public. The interview also contains a great deal of personal and industry material, such as the researcher's background, her department's resources, other companies' funding, and the team being about 20 people, mostly doctoral students. None of that speaks to model capability and none of it should be used as a basis for choosing a model. On availability, the report discusses only papers and methods; it does not say whether Z0, Zeva, Zeva-Ego, Zetta or Z-infra are released with open weights or under what license, nor does it give callable interfaces, input and output formats, or minimum deployment hardware. Open weights and an open-source license are two different things and need to be verified separately. To be clear, none of these versions are currently listed in KimiSeek's model catalog, and this report mentions no API launch; if interfaces appear later, their availability and terms are up to the vendor, and vendor announcements govern. The new work Cao Ting mentions as being released "soon" had not been announced when this article was written, so neither its capabilities nor its degree of openness can be verified.
Illustrations from the original report











