What's this about?
A study of the "physics" of multimodal pretraining: how knowledge flows between modalities (text, image, etc.), when modalities help each other vs. interfere, how "unification" emerges early in training, and practical recipes drawn from the findings (11↑).
What exactly is "modality synergy"?
Whether training on multiple data types together helps or hurts. Sometimes learning images improves text understanding (synergy); sometimes they compete for capacity (interference). Understanding when each happens tells you how to mix modalities during pretraining — currently more art than science.
Why does it matter?
Unified multimodal models are everywhere this week (Hunyuan3D, ToolArtist, AURORA-LM). A paper explaining the underlying dynamics — rather than just shipping another model — is the kind of foundational work that makes the next generation of unified models less trial-and-error.