Can Graph Foundation Models Generalize Over Architecture?

A team from the University of Oxford and Imperial College London has published a paper challenging a fundamental assumption in Graph Foundation Models (GFMs). These models, designed to work across diverse graph-structured data like social networks and molecules, typically use a single, fixed neural architecture. The researchers prove this approach has a critical flaw: it fails when tasks require different computational "ranges" of message-passing between nodes than what the model was trained on. This makes current GFMs non-robust for true zero-shot generalization.

To solve this, the team introduced a novel framework that makes architecture adaptivity a core requirement. Instead of a fixed backbone, their model can discover and dynamically combine task-specific linear graph operators during inference. This allows a single pre-trained model to handle tasks with vastly different architectural needs without any retraining. They validated the approach on both controlled synthetic tasks and real-world benchmarks, showing improved performance and robustness over existing domain-agnostic GFMs.

The work, accepted to the GRaM Workshop at ICLR 2026, argues that for GFMs to live up to their promise of universal applicability on any graph, they must move beyond static architectures. This shift could enable more reliable AI systems for complex, real-world problems in drug discovery, recommendation systems, and network analysis, where data structures are unpredictable and heterogeneous.