single mention · not yet filed
Causal video-representation distillation for temporal supervision
Causal video-representation distillation (DINO-Video) supervises a VLA's query states to predict motion-aware video features extracted from a causal video representation model.1 Unlike per-frame image features, DINO-Video applies causal temporal attention, so each time step's feature depends only on current and past observations, letting the future query learn to anticipate future representations.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
the current query is trained to predict the DINO-Video feature of the current frame, while the future query is trained to predict the feature of the future frame
— From Foundation to Application: Improving VLA Models in PracticeUnlike image-level features extracted independently for each frame, DINO-Video produces motion-aware visual representations with causal temporal attention, such that the feature at each time step depends only on the current and past observations
— From Foundation to Application: Improving VLA Models in Practice
Also called DINO-Video.