CHARM: Multimodal Graph Foundation Model for Zero-Shot Transfer
A research paper introduces CHARM, a multimodal graph foundation model designed for zero-shot transfer across graph domains and tasks. Real-world graphs often associate nodes with text, images, and other modalities, yet existing graph foundation models either require downstream adaptation or address only unimodal graphs. CHARM tackles two key challenges: generalizing knowledge from individual modalities while capturing cross-modal relations, and avoiding domain-specific node representation entanglement without target-domain fine-tuning. The model employs hierarchical context modeling to enable zero-shot transfer on multimodal graphs, a setting that remains underexplored. The paper is published on arXiv with ID 2607.26023.
Key facts
- CHARM is a multimodal graph foundation model.
- It focuses on zero-shot transfer across graph domains and tasks.
- Real-world graphs associate nodes with multiple modalities.
- Existing GNN-based graph foundation models require downstream adaptation.
- LLM-based graph methods mainly address unimodal graphs or single-domain tasks.
- CHARM addresses two key challenges: cross-modal generalization and domain-specific entanglement.
- The model uses hierarchical context modeling.
- The paper is available on arXiv (2607.26023).
Entities
Institutions
- arXiv