Paper recorded by Signals 4 on 2026-10-21 in cs.CV. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-21 on arXiv · recorded by Signals 4 on 2026-10-22
Category: cs.CV · 计算机视觉 · first seen 2026-10-22
World models turn visual representations into predictions and actions. We argue that their visual foundation should combine physical-state information, native language alignment, and pixel reconstruction, so that a single representation can support prediction, instruction, and inspection. However, existing encoders typically specialize in only one or two of these capabilities; we therefore introdu