Graph Hierarchical Recurrence for Long-Range Generalization

In three linesGraph Hierarchical Recurrence (GHR) is a novel framework for GNNs and Graph Transformers that captures long-range dependencies through hierarchical abstraction via pooling. GHR outperforms existing models on long-range benchmarks using only 1% of SOTA parameters, and improves out-of-distribution generalization.

Summary generated by Claude — human-verified