S$^3$GNN: Efficient Global Mixing and Local Message Passing for Long-Range Graph Learning

Researchers introduced S$^3$GNN, a graph neural network designed to improve long-range dependency learning without restrictive assumptions.

The model reduces oversquashing by efficiently combining global mixing with local message passing, reintroducing omitted components at low cost.

It delivers strong results on standard benchmarks, knowledge graph question answering, and mesh-based fluid dynamics tasks.

S$^3$GNN also uses fewer parameters and less compute, making it a more efficient option for graph learning.