Tsinghua AIR and Domain Transformation Release Embodied Model Zeva, Achieving In-Context Causal Learning without Weight Updates

Tsinghua University Institute for AI Industry Research and Yubianhuan have released Zeva, an embodied model that uses In-Context Causal Learning (ICCL) to learn from interaction experience without updating its weights. ICCL combines causal interaction extraction, dual-timescale causal memory, and context policy injection.

The team reports that on RoboCasa365, Zeva's cumulative success rate rises from 26% to 73%, and in a real chemistry lab it shows monotonic improvement. The model can also learn a new operation from a single human demonstration.

Original sources (Chinese)