ts-net: Graph foundation model zero-shots to real-world spreading dynamics

Network dynamics tasks—such as spreading, influence maximization, and epidemic modeling—have long been trapped in the transductive paradigm: models must be retrained for each new graph. In a new paper, Czuba et al. argue that inductive cross-network generalization is essential for true Graph Foundation Models (GFMs). They propose four design properties to achieve this goal and present ts-net (TopSpreadersNetwork) as a proof of concept. ts-net is trained exclusively on synthetic multilayer networks (MLNs) yet demonstrates zero-shot generalization to real-world MLNs of varying size and layer count. It beats classical heuristics and transductive baselines on three out of four metrics, marking a significant step toward deployable GFMs for network dynamics.

The paper does not stop at ts-net's performance. Based on its results, the authors outline five open challenges: scaling to larger networks, generalization to many-layer graphs, self-supervised pretraining, cross-task transfer, and integration of node attributes. These challenges define a roadmap for the community. ts-net itself focuses on super-spreader identification—a critical task in epidemiology and social network analysis—and proves that synthetic data can effectively pretrain models for real-world graph tasks. The work, available on arXiv (2606.08306), offers both a practical baseline and a strategic direction for building foundation models that understand complex networked systems.