Rethinking the Gold Standard: Why Discrete Curvature Fails to Fully Capture Over-squashing in GNNs?

The paper challenges the prevailing assumption that negative discrete curvature is both necessary and sufficient for detecting over-squashing in graph neural networks. It sits within the Curvature-Based Over-Squashing Characterization leaf, which contains four papers providing formal analysis of how discrete curvature measures characterize bottlenecks. This is a moderately populated research direction within the broader Theoretical Foundations and Analysis branch, suggesting active but not overcrowded investigation into the fundamental relationship between geometric properties and information flow in GNNs.

The taxonomy reveals closely related work in sibling leaves: Over-Squashing and Over-Smoothing Trade-offs examines interactions between competing failure modes, while Geometric and Topological Perspectives explores alternative embedding spaces beyond standard graphs. Neighboring branches address practical interventions—Curvature-Guided Rewiring Methods develops topology modification algorithms, and Curvature-Constrained Message Passing integrates curvature into neural architectures. The paper's theoretical focus on characterization boundaries distinguishes it from these application-oriented directions, though its proposed WAF3 measure bridges toward the rewiring and architecture branches.

Among twelve candidates examined, none clearly refute the three main contributions. The theoretical claim that negative curvature is insufficient examined two candidates without finding overlapping prior work. The MOSR metric examined ten candidates, again with no refutations detected. The WAF3 curvature measure examined zero candidates, suggesting either limited semantic overlap in the search or a genuinely novel formulation. Given the small search scope—twelve papers from a field with thirty surveyed works—these statistics indicate potential novelty but cannot rule out relevant prior work outside the top-ranked semantic matches.

The analysis covers a focused slice of the literature: top-ranked semantic neighbors plus citation expansion within a thirty-paper taxonomy. The absence of refutations across contributions suggests the work introduces fresh perspectives on curvature-based diagnostics, particularly the counterexamples demonstrating insufficiency and the WAF3 approximation algorithm. However, the limited search scope means closely related theoretical results or alternative curvature formulations may exist in venues or subfields not captured by this retrieval strategy.