The AI Industry Research Institute (AIR) at Tsinghua University, in collaboration with Domain Transform, has introduced an embodied manipulation method called Zeva. This method keeps the weights of the base model frozen and improves during deployment solely through in-context causal memory.
Across several tested embodied manipulation benchmarks, the aggregate success rate increased from approximately 26% to 73%, representing a gain of 47 points without needing to go through the standard cycle of data collection, labeling, and fine-tuning. The paper, titled 'Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation,' utilizes the Cosmos3 model as its foundation, with authors Liu Yunxin and Zhao Ting from Tsinghua AIR being mentioned.
The core idea of the approach is not to perform more extensive fine-tuning of vision, language, and action models, but rather to add a memory layer that teaches a fixed model how its own actions have changed the scene. Zeva consists of three main components.
A causal transfer encoder links each action to the subsequent state change, forming the raw material for subsequent queries. Then, causal memory operates on two time scales: a short interaction trajectory for the current attempt and a longer persistent interaction memory that accumulates filtered experience between attempts. The deliberate separation of these stores prevents immediate contamination of long-term memory by noise from failures while still feeding the next command within the same episode.
Finally, in-context policy injection inserts causal prompts into the input of the frozen model, with no gradient updates occurring during deployment. This means that adaptation to the environment transforms from a training operational tuning problem into a context engineering problem.
On the RoboCasa365-Atomic5 platform, the average success rate reached 76.8%. More pronounced results were achieved on ChemLab-Evo, a chemical laboratory task set: picking up a test tube increased from 65% to 100%, placing a glass from 25% to 70%, and pouring water from 30% to 80%. The increase was more significant where the baseline scores were weaker, aligning with the idea that frequent failures generate more useful causal pairs.
A single human demonstration action added about 20 points to glass placement and approximately 15 points to pouring—a gain the authors attribute to the introduction of memory rather than parameter changes. This is an attractive point when labeling budgets are limited.
The chemical domain was chosen because state changes, such as liquid level and vessel posture, are relatively clean, allowing the encoder pairs to remain low-noise; however, the paper notes that tasks involving deformable objects, like laundry, might prove more challenging. Practically, Zeva trades weeks of site-specific fine-tuning for longer contexts and higher per-step inference cost, which improves with use. Sites that rarely change may still prefer classical fine-tuning for economic reasons. Since independent robotics platforms outside the configured Cosmos3 system were not included in the released package, claims of transfer remain theoretical until other labs reproduce the memory stack on different hardware and task taxonomies.
