Graph neural networks that generalize to longer distances

Graph Neural Networks and Transformers struggle to capture correlations between distant nodes and fail when test graphs contain longer-range interactions than training data. Graph Hierarchical Recurrence (GHR) addresses this by operating jointly on the original graph and hierarchical pooled abstractions. Across long-range benchmarks, GHR matches or exceeds state-of-the-art performance while using as little as 1% of baseline parameters. The work demonstrates that parameter efficiency—not scaling capacity—may be key to out-of-distribution generalization on graph tasks.