Zeva: Frozen Embodied Model Improves 26% to 73% Success Without Weight Updates

量子位 · wechat · 2026-09-01

Tsinghua AIR and its spinoff Yubianhuan released Zeva, an embodied manipulation model that first implements In-Context Causal Learning (ICCL): the model's weights stay completely frozen while it continuously improves by consuming its own interaction experience as context. On RoboCasa365, cumulative success rises from 26% to 73% across attempts; in a real chemistry lab, three atomic tasks (pick up test tube, place beaker, pour water) improve monotonically, reaching up to 100%. A single human demonstration can initialize its persistent causal memory for one-shot skill acquisition without fine-tuning, adding up to 20 percentage points.

Technically, Zeva uses a Causal Transition Encoder to capture action-to-state-change effects, a Dual-timescale Causal Memory to accumulate reusable experience across attempts, and In-Context Policy Injection to feed this causal context into a frozen Cosmos action generation model with zero gradient updates. The team frames this as a new scaling axis: instead of scaling parameters, scale the causal context consumed at deployment time.

Original post →

More from Embodied

Embodied channel →