We present a principled recipe for building graph foundation models that generalize across arbitrary graphs, features, and label spaces.

We create a new MPNN with Boundary Conditions of Riemannian Dynamics to combat oversquashing, while allowing for MPNN going deeper.