A new framework lets one graph AI model handle wildly different types of network data without retraining for each new domain.
Graph foundation models are supposed to work like language models: train once, apply everywhere. But graphs vary wildly in shape, size, and what their features even mean, which has made a truly universal graph model hard to pull off. The new method, called SCGFM-ART, maps any graph onto a shared reference system built from a fixed set of "relational landmarks," then predicts each graph's coordinates directly instead of solving a costly alignment problem at test time. The researchers tested it on 14 cross-domain classification tasks, covering both graph-level and node-level predictions, and it topped the rankings in both.
The bigger story isn't the accuracy - it's the speed. Because the model skips the iterative optimization most alignment methods depend on, inference on new domains runs 44 to 85 times faster. That's the difference between a promising benchmark result and something a real production pipeline could actually use.
Graph AI has lagged behind text and image models partly because "one model, many domains" is a harder promise to keep - this result is a reminder that the gap is narrowing, not that it's closed.