How Far Are We from Graph Foundation Models? A Controlled Cross-Paradigm Benchmark

Author Name : Aditya Agarwal

DOI: https://doi.org/10.56025/IJARESM.141026211

 

ABSTRACT

Graph foundation models (GFMs) aspire to match the pre-train-then-adapt success of language and vision foundation models, yet competing paradigm families have been evaluated on disjoint benchmarks with incomparable protocols. We present the first controlled crossparadigm comparison, systematically evaluating 11 methods from 4 paradigm families— GNN message passing, heterophily-specialized GNNs, self-supervised pre-trained GNNs, and Graph Transformers—on 11 datasets spanning homophilic citation networks, heterophilic web graphs, large-scale OGB graphs, molecular graphs, biological graphs, and social networks. Across 1,400+ runs with 10 seeds each, we find that no single paradigm dominates: five different methods achieve rank 1 on at least one dataset. Heterophily-specialized methods such as H2GCN achieve 80.2% on Wisconsin, vastly outperforming standard GNNs, while LINKX emerges as the most homophily-robust method with only a 1.0 pp drop between regimes and dominates the large-scale ogbn-arxiv benchmark (67.3%). Method rankings completely invert between homophilic and heterophilic graphs, with GAT dropping from rank 1–2 to rank 8–11 on heterophilic datasets. Self-supervised pre-training via GRAPHMAE never achieves best-in-class on any of the 11 benchmarks despite full fine-tuning. Deeper models significantly help on structuresensitive tasks, yet oversmoothing remains critical, with GCN accuracy dropping from 79.6% (2 layers) to 50.7% (8 layers) on Cora. These findings challenge the assumption that a single pre-training recipe will yield a universal GFM and demonstrate that graph structure properties, particularly homophily, are the primary determinant of optimal architecture choice. Code and all experimental artifacts are publicly available.