The Zeva team contributed
The model is not being updated, yet its capabilities continue to grow.
This is the most counterintuitive aspect of Zeva, and also the new answer brought by the collaboration between Tsinghua AIR and Domain Adaptation.
On multiple typical embodied benchmark tasks, the cumulative success rate of the frozen model was increased from 26% to 73%; in a real chemistry lab, the robotic arm became increasingly stable over multiple attempts.
More critically, a single human demonstration can enable the model to "learn" a new operation.
Zeva is the first embodied WAM to achieve In-Context Causal Learning (ICCL).
This enables the model to continuously learn from its interactive experiences without updating its weights. As this capability matures, embodied intelligence will usher in its own "GPT moment".
Project Information
Project Name: Zeva Issuing Institution: Tsinghua AIR, Domain Adaptation Paper Title: Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation Project Homepage: https://air-embodied-brain.github.io/Zeva/
The self-evolution of embodied models, why was it difficult in the past?
The real bottleneck for embodied intelligence is whether models can keep improving once deployed in the field.
In traditional thinking, to enable robots to adapt to new environments, it is usually necessary to re-collect data, fine-tune models, or perform test-time training.
However, these methods rely on gradient updates, are slow and costly, and may also destroy existing capabilities when adapting to new tasks.
Another type of in-context learning method can enable models to adjust their behavior based on demonstrations or task descriptions, but primarily addresses "task understanding" and does not allow models to learn physical causality from the consequences of their own actions.
From "Task Generalization" to "Causal Transfer"
In the past, when discussing cross-domain applications, we often referred to sim-to-real, cross-ontology migration, and cross-task generalization. Most of these methods rely on data alignment or representation alignment, essentially transferring "task experience" between different domains.
Zeva takes it a step further: it migrates the "causal relationship between actions and state changes".
The same model can bring physically causal knowledge learned in simulations to real robotic arms; the experience of failure accumulated in one attempt can guide another similar physical process attempt; a human demonstration can also directly become the causal context for a machine.
This cross-domain migration is achieved through context. The model infers environmental causal structures in the context and applies causal knowledge to generate the next action, which is what Zeva defines as In-Context Causal Learning.
Example 1: Extract causal relationships from misoperations such as incorrectly placing weighing paper, complete context-based learning, and improve task accuracy.
Example 2: Extract causal relationships from misoperations such as taking reagents to weighing paper, complete context-based learning, and improve task accuracy.
Example 3: In-Context Causal Learning from a Single Human Demonstration.
Zeva: Integrating "Causal Learning" into Model Context
The overall framework of △Zeva. The model achieves self-evolution during deployment without updating weights through Causal Interaction Extraction, Dual-timescale Causal Memory, and In-Context Policy Injection. Image source: paper
Zeva is a model. Its Causal Transition Encoder, Dual-timescale Causal Memory, and In-Context Policy Injection are all part of the model's inference chain.
At each environment step, Zeva extracts three types of signals: visual state, action encoding, and observed state change.
They are recursively encoded as Causal Interaction State and further projected as Phase Token and Causal Interaction Signal.
These causal signals are organized into dual-timestamp contexts: Brief Interaction Trace is responsible for coherent execution within the current attempt, while Persistent Interaction Memory accumulates reusable experiences across attempts.
During decision-making, Zeva retrieves interactive evidence matching the current task stage, constructs a Causal Prompt, and injects it into the frozen Cosmos3 action generation model, with the entire process involving zero gradient updates.
Technical Dissection: How Zeva Achieves "Great Things with Small Means"
Highlight 1: Causal Transition Encoder, allowing the model to see "action→result".
The model explicitly learns the state changes caused by actions. The training objectives include causal effect prediction, task clustering, and phase progression, allowing the Causal Interaction Signal to truly carry physical causal information.
Highlight 2: Dual-timestamp causal memory, allowing the model to self-evolve across attempts.
BIT ensures the coherence of the current execution, while PIM distills one failure, one correction, and one success into a retrievable context for subsequent use, making the same model stronger with repeated attempts.
Highlight three: In-Context Policy Injection, allowing frozen models to directly consume causal context.
By injecting task, phase, and interaction evidence into the generation model through Causal Prompt, without adding extra noise or participating in flow loss, but only as a conditional influence on subsequent actions, this is the key design for "evolving without weight adjustment".
Compared to traditional Test-Time Training, Zeva's difference is straightforward:
Data Speaks: Self-Evolution + One-Shot Human Demonstration Enhancement
On RoboCasa365-Atomic5, Zeva had an average success rate of 76.8%, achieving the optimal effect.
More importantly, the self-evolution curve after deployment: the cumulative success rate increased from 26% for Evolve 1 to 73% for Evolve 4.
On the real-world chemistry laboratory simulation platform ChemLab-Evo, three atomic tasks also showed steady growth: Pick Up Test Tube from 65% to 100%, Place Beaker from 25% to 70%, and Pour Water from 30% to 80%.
Zeva also supports one-shot human demonstration enhancement: after a single human demonstration initializes the PIM, the model can replicate key operations without fine-tuning.
Quantitative results show that human warm-up brings up to a 20 percentage point improvement on Place Beaker and a 15 percentage point improvement on Pour Water.
On RoboCasa365, Zeva's cumulative success rate increased from 26% in Evolve 1 to 73% in Evolve 4, with its model weights completely frozen. Image source: paper
In a real-world chemistry lab, the cumulative success rate of three atomic tasks increases with trial milestones. Model parameters were not updated. Image source: paper
△One-shot human demonstration warm-up impact on three atomic tasks. Dashed lines represent human demonstration initialization, solid lines represent pure self-evolution. Image source: paper
Why Zeva is a new paradigm?
Zeva's profound implication is to extend the scaling of embodied models from "parameter space" to "context space".
Traditional scaling boosts model capabilities by increasing data, parameters, and computing power.
Another type of scaling demonstrated by Zeva is: continuously increasing the causal context available for the model to consume during deployment.
Each attempt generates new interaction evidence, which is compressed into a Causal Interaction Signal and enters dual-timestamp memory.
The model performs an additional "implicit system identification" of the current physical environment with each interaction.
The theoretical basis here is that the model can take the environmental attributes of each new scenario as latent variables, and infer them from past action-effect evidence during reasoning.
Object mass, joint constraints, contact properties, and friction conditions can all be implicitly estimated through interactive signals. The freezing strategy can thus continuously approach the true dynamics of the current physical system without changing parameters.
Furthermore, Zeva's Causal Interaction Signal is a compact representation in the effect space, directly encoding physical effect information.
Cross-task effect retrieval experiments show that equivalent physical effects—observed across different objects, viewpoints, and task instructions—cluster together in that space.
This means the model learns "how this type of physical effect is produced", while also covering the methods for specific tasks. This representation naturally supports cross-domain migration, making "domain transformation" a computable, retrievable, and reusable technical concept.
If VLA is viewed as the infrastructure for "task understanding and action generation", and the world model is seen as the infrastructure for "experience imagination", then ICCL, as defined by Zeva, is the infrastructure for "on-site causal learning".
It transforms deployment from "capability consumption" to "capability accumulation", and turns each failure from noise into retrievable causal evidence.
If ICCL is established, the scaling of embodied intelligence will have an additional curve independent of parameter size: the more interactions, the richer the causal context, the more accurate the model's recognition of the current physical domain, and the higher the task success rate. This is the true meaning of "domain transformation" as a new paradigm.
In conclusion
Zeva, launched jointly by Tsinghua AIR and Domain Transformation, redefines self-evolution for embodied models:
Without adjusting weights, it can become stronger with more use; without retraining, it can learn new operations from a single human demonstration.
The activation of domain transformation may open up a new scaling path, enabling robots to learn causality and complete evolution on their own in the real physical world.
Reference links: [1] Paper: Zeva: In-Context Causal Interaction Memory for Embodied Action Generalization [2] Project homepage: https://air-embodied-brain.github.io/Zeva/
Regarding Tsinghua AIR and Domain Transformation
Tsinghua AIR (Institute for AI Industry Research, Tsinghua University), the university's international, intelligent, and industrialization-focused research arm built for the Fourth Industrial Revolution, is regarded as a first-tier force in China's AI industry-academia-research integration. Yubianhuan, a company incubated by Tsinghua AIR for the era of self-evolving physical intelligence, was founded by Professors Liu Yunxin and Cao Ting. The launch of Zeva marks a key milestone in Tsinghua AIR's transition of embodied intelligence from frontier research to industrial deployment.
Liu Yunxin
Vice President of the Research Institute of Intelligent Industry at Tsinghua University, doctoral supervisor, Wanxiang Data Distinguished Professor, and Chief Researcher, a national high-level talent and IEEE Fellow.
Former Chief Research Supervisor at Microsoft Research Asia, with long-term research focus on AIoT and on-device intelligence.
Received awards from multiple top international academic conferences including MobiCom, MobiSys, and SenSys, as well as the CCF Science and Technology Progress First Prize.
Cao Ting
Tsinghua University Professor, Beijing's High-Level Introduced Talent
Former Chief Research Officer of Microsoft Research Asia, with research focuses including edge intelligence and embodied intelligence.
Won the best paper award at top conferences four times, with work selected as ACM/Microsoft Research Highlights, including an ACM Research Highlight personally written and recommended by Turing Award winner David Patterson.
This article is published with authorization from Quantum Bit, and the views expressed are those of the original author only.