Microsoft’s Graphormer: The Transformer That Finally Beats GNNs

Researchers from Microsoft Research Asia, Princeton University, Peking University, and Dalian University of Technology introduce Graphormer, a Transformer redesigned for graph data. By adding centrality encoding, shortest-path spatial bias, and edge-aware attention, it outperforms classic GNNs and wins the Open Graph Benchmark Large-Scale Challenge. Graphormer matches or exceeds GIN and GCN across PCQM4M-LSC, MolHIV, MolPCBA, and ZINC—proving Transformers can dominate graph learning.