arxivcs.CVcs.LG2026-07-16
AlphaWiSE: Adaptive Weight Interpolation for Continual Multimodal Representation Learning
Sarthak Jain, Qiran Hu, Zhen Zhu, Yaoyao Liu
Multimodal models such as CLIP learn a shared embedding space for cross-modal retrieval, but continual adaptation to sequentially arriving data can disrupt the cross-modal alignment acquired from earlier phases. Conventional continual-learning methods return a single checkpoint,…