modality-centric

Proposes CoMAG, a unified backbone for multimodal attributed graphs that learns task-adaptive reliable contexts and performs modality-preserving alignment, achieving state-of-the-art results on graph-level prediction, modality matching, and graph-conditioned generation.