Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes

By Junlin Han · Paper · cs.CV

Vision offers a critical axis for advancing foundation models, driving a shift towards natively unified multimodal pretraining. Despite this momentum, the design space and the fundamental mechanisms of how modalities interact during unified training remain underexplored. We provi

Cs.cv

View original

HomeResourceLoading…