arxivcs.CVcs.RO2026-07-09
Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio
Utkarsh A. Mishra, Yongxin Chen, Danfei Xu, Yang Liu, Xi Chen, Jiayuan Mao
Generative video foundation models exhibit strong compositional priors, yet world-action models (WAMs) and video-action models (VAMs) often lose these priors after finetuning on robotic action data. We refer to this discrepancy as the video-action generalization gap. In this pape…