Graph Foundation Models

Graph Foundation Models study how to build reusable models that learn from graphs and transfer to new datasets and tasks, such as node classification, link prediction, clustering, and graph generation. Common approaches pretrain GNNs or graph transformers on real or synthetic graphs, encode structure with tokens or geometric representations, and adapt with prompts, in-context learning, or fine-tuning; some also align graphs with language or tabular models. Work increasingly targets transfer across different feature spaces, graph types, and domains, including knowledge graphs, molecules, and dynamic networks. Results are promising, but fair benchmarks show that carefully tuned task-specific GNNs often remain competitive or stronger, while transfer reliability, inference cost, and robustness are unresolved concerns.