Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
This paper systematically studies how language, visual understanding, and visual generation interact during unified multimodal pretraining. Controlled experiments on synthetic and real-world data examine knowledge transfer, modality synergy, and architectural choices. The authors report that data complexity largely determines whether modalities cooperate or compete; shared attention and normalization with modality-specific feed-forward layers encourage synergy. Early joint unification outperforms late alignment or sequential training and avoids a reported “vision laziness” effect. The paper also proposes recipes reaching strong generative performance with 5% of the compute budget, followed by validation using multiple 13.5B MoE models trained on 2T tokens.
Why it's worth reading
As multimodal systems move from late-stage alignment toward native joint pretraining, this work offers testable claims about modality transfer, architectural sharing, visual underuse, and compute-efficient training recipes.