Get a GRIP, this will be a long TRIP: A Quantifiable Long-Range Framework for Verifying Over-squashing
The paper proposes axioms for proving whether a graph benchmark really tests long-range GNN behavior.
The authors argue that existing long-range benchmarks can be solved by tuned short-range models or depend too much on specific graph topologies. They introduce TRIP and GRIP, procedures meant to turn arbitrary graphs into provably long-ranged tasks using stable feature distributions. GRIP also gives a closed-form per-range maximum-likelihood oracle, yielding an a priori lower bound on test error. They use the framework to audit four common benchmarks and report several failure modes. ArXiv · AI/CL/LG's note
The authors argue that existing long-range benchmarks can be solved by tuned short-range models or depend too much on specific graph topologies. They introduce TRIP and GRIP, procedures meant to turn arbitrary graphs into provably long-ranged tasks using stable feature distributions. GRIP also gives a closed-form per-range maximum-likelihood oracle, yielding an a priori lower bound on test error. They use the framework to audit four common benchmarks and report several failure modes. ArXiv · AI/CL/LG's note
score 5