arXiv:2611.05406v1 Announce Type: cross Abstract: Contrastive pretraining has advanced image-text understanding, yet image, video, and audio encoders are still trained independently, with cross-modal alignment realized mainly through post-hoc binding to a frozen anchor space.
