Ying-Cong Chen

Frontier VLMs excel at precise scene estimation but falter in executing coordinated, goal-directed actions, highlighting a critical bottleneck for robot generalism.

The real driver of on-policy distillation gains in multi-turn agents is counterfactual rollback at the first fatal mistake—an insight leveraged here to achieve double-digit RLVR gains by internalizing rollback dynamics directly into gradient updates without resetting the environment.

ReWorld achieves unprecedented control fidelity and video generation quality by elegantly separating short-term control from long-term memory, outperforming existing models in interactive environments.