Research & Papers

Fair Evaluation Finds Only Prior-data Fitted Network GFMs Beat GNNs

New rigorous study of 9 Graph Foundation Models reveals most fail to outperform tuned GNNs

Deep Dive

A new paper titled 'A Fair Evaluation of Graph Foundation Models for Node Property Prediction' (arXiv:2606.24509), accepted at the ICML 2026 Workshop on Graph Foundation Models, systematically re-evaluates nine recent GFMs for node property prediction—a key task in graph machine learning with real-world uses in fraud detection, social network analysis, and e-commerce recommendations. The authors, Oleg Platonov, Gleb Bazhenov, Dmitry Eremeev, and Liudmila Prokhorenkova, note that the field lacks a unified evaluation setting, making it difficult to compare GFMs against each other or against strong baselines.

Their rigorous experiments show that among the nine GFMs tested, only those built on the Prior-data Fitted Networks (PFNs) approach outperform well-tuned GNNs in predictive accuracy. However, this performance gain comes with higher inference costs. The study underscores the importance of standardized benchmarks and reveals that many GFMs currently fail to justify their complexity over simpler GNN architectures. This finding has direct implications for practitioners selecting models for production systems where both accuracy and computational efficiency matter.

Key Points
  • Only GFMs using the Prior-data Fitted Networks (PFNs) paradigm beat well-tuned GNNs in predictive performance
  • PFN-based GFMs incur higher inference costs than GNN baselines
  • The field currently lacks a unified evaluation protocol, preventing reliable cross-model comparisons

Why It Matters

Guides graph ML practitioners in choosing between GFMs and GNNs, highlighting accuracy-cost trade-offs for real-world deployments.

📬 Get the top 10 AI stories daily