Tsinghua University's Institute for AI Industry Research (AIR), together with Domain Transform, has released a new embodied manipulation study using a method called Zeva. Its approach stands out in a field where the usual playbook is different: the model's weights stay frozen throughout, with no gradient updates at all, and the robot climbs its own performance curve during deployment purely through context. Across several embodied benchmarks, the cumulative success rate rose from 26% to 73%.
The dominant approach is to collect data, label it, fine-tune a VLA model, and repeat the cycle whenever a new scenario shows up. Zeva moves that whole process to inference time instead — it records the causal relationships between actions and the resulting state changes, then feeds that relationship back to the frozen base model as a prompt. The paper uses Cosmos3 as its backbone.
Three components
The causal transfer encoder extracts state changes from every action, pairing "what I did" with "what changed in the environment." This is the data-gathering step for the whole method, and how accurately it extracts these pairs directly determines whether anything downstream is usable.
Causal memory operates on two timescales. The short one, called the brief interaction trajectory, only tracks the current attempt. The long one, the persistent interaction memory, accumulates across attempts. The former answers "what did I just try this round," and the latter answers "where did I trip up last time." Keeping the two separate is deliberate: a failure signal from the current attempt needs to feed back into the next action immediately, while cross-attempt experience needs to be filtered before it earns a place in long-term memory — mixing them risks letting a single noisy attempt skew the whole thing.
The third piece is in-context policy injection, which slots the causal prompts into the frozen model's input. Zero gradient updates happen anywhere in the deployment process.
Where the numbers land
On the RoboCasa365-Atomic5 benchmark, the average success rate was 76.8%. The more telling numbers come from moving into a real chemistry lab setup, which the team calls ChemLab-Evo: picking up a test tube went from 65% to 100%, placing a beaker went from 25% to 70%, and pouring water went from 30% to 80%.
The three tasks started from very different baselines. Picking up a test tube already had a 65% floor, so pushing it to a perfect score mostly meant smoothing out the edge cases. Placing a beaker started at 25% and nearly tripled — suggesting the memory mechanism pays off more in the low-success-rate range, since a robot that fails more often also generates more causal pairs to learn from.
A single human demonstration pushes things up further still: placing a beaker gains roughly 20 percentage points, and pouring water gains roughly 15. A single demonstration carries limited information on its own, so getting this much lift from it suggests it's going straight into memory rather than touching any parameters — plug-and-play, in other words.
Why a chemistry lab
The choice of setting wasn't arbitrary. Chemistry operations have clearly bounded state changes — liquid level, test tube position, whether something spilled — so the causal signal is clean and the pairs the encoder extracts carry little noise. The paper doesn't shy away from noting that the same mechanism may not hold up as well for tasks like folding laundry, where state is harder to quantify.
For deployment, the practical implication is a different shape of engineering work. Under the fine-tuning approach, every new site means pulling data, retraining, and running regression checks — a pipeline that takes weeks. Zeva starts working the moment it's deployed and gets more accurate with use, at the cost of a longer context and higher per-step overhead at inference time. Which approach pays off more depends on how often the scene changes — if a robot spends years doing the same job, fine-tuning still comes out cheaper.
The paper is titled "Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation," with contributors including Liu Yunxin and Cao Ting from Tsinghua AIR.
Sources: Zeva paper, QbitAI, CocoLoop; success rates and benchmark names verified against the paper's disclosed figures, with gains calculated as direct differences from the reported numbers.