acceptodds
Under review as a conference paper at ICLR 2027

The Role of Edge Geometry in Graph Neural Networks

Abstract

While much work on graph neural networks has focused on designing node features, some structural information, such as curvature, is naturally defined on edges. Such information can help in two ways: by supplying structural information that standard message passing does not recover, or by improving message propagation. We separate the two and ask when each helps, and for which architectures and tasks. We introduce curvature profiles that combine degree-based curvature and counts of cycles of different lengths. These profiles can be used as edge features in standard graph networks or to guide message propagation in our curvature-aware RicciMP layer. Exact constructions show how profiles can improve filtering and how neighboring context reveals curvature information absent from one edge's profile. At the same time, longer cycles yield a strict expressivity hierarchy: for each fixed curvature-profile refinement is incomparable with the coordinate-wise -Weisfeiler-Leman test once the cycle order reaches an explicit -dependent threshold, each distinguishing graph pairs the other misses. Because profiles are computed before any message passing, this reach is obtained by changing the labels refinement reads, not the refinement procedure. Empirically, we show that whether curvature is most useful as a feature or for propagation depends on the task. In particular, curvature profiles substantially improve molecular prediction with GINE, achieving scores close to matched subgraph-counting models, and benefit baseline classification benchmarks and community-recovery tasks. RicciMP also improves image-region labeling over GINE with the same structural features and unweighted multi-hop propagation. On long-range tasks, augmentation effects are mixed, and RicciMP trails the strongest tested baselines without curvature profiles. Together, these results distinguish when structural augmentation helps prediction and what the tested propagation choices add.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, message passing, expressivity, Weisfeiler-Leman, discrete curvature, Forman curvature, Ollivier-Ricci curvature, edge encodings

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8CsSOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors analyze the role and the benefits of adding curvature to Graph Neural Networks (GNNs). After reviewing the related work and previous results in the literature, the manuscript proceeds to illustrate the methodology. In particular, the authors present a two-sided method. After defining an edge curvature, made of the unweighted Forman curvature and the number of cycles of each length through the edge, they investigate whether adding this information to the edge features of edge-aware GNNs (like GINE or GPS) helps learning. Moreover, they use these features to implement a new MPNN layer, RICCIMP, whose edge weights are initialized with these values. The authors test both methods in several scenarios, and their experimental section is broad, covering ZINC, LRGB, TU, node classification, planted-community recovery, and a synthetic denoising task.

Critical strengths
  1. The research question is relevant and the analysis is interesting. I also think that this problem is relevant to the GNN literature.
  2. Propositions and theorems are, as far as I checked, well posed. The proofs also seem correct.
  3. The experimental setting is broad, and the experiments have been conducted rigorously.
Critical weaknesses
  1. The contribution is only incremental, and the relevance of the evaluated architecture is insufficiently clear. Under the manuscript's framing, it appears that the introduced Forman curvature ultimately reduces to including node degrees in the edge features (Lemma 12 acknowledges that it adds no stable separation power). Moreover, Theorem 3, although correct and carefully derived, extends known results on the expressiveness of message-passing GNNs [3, Lemma 2, Theorem 3]. The authors should clearly distinguish the specific contributions of their method from existing expressiveness results for edge-labeled MPNNs. My main concern is the methodological novelty and the practical relevance of the expressiveness results.

  2. Experimental weaknesses beyond the strengths mentioned above. I acknowledge that the experiments are thoughtful and well executed. However, the reported results do not make clear whether the RICCIMP layer provides an actual benefit over the tested baselines. For example, curvature-profile GINE outperforms RICCIMP on ZINC, while the long-range ablations show no clear additional benefit from adaptive conductances. These findings make me question the actual need for the whole RICCIMP layer (which has a notably higher computational cost) and for the conductance machinery the authors introduce. In addition to the weak experimental evidence, I think the authors should also compare their methodology with well-established and widely used positional encodings. In particular, they should show how their Forman curvature labeling and RICCIMP layer compare against adding LapPE or RWPE to the MPNN baselines (see [1,2,8]), which are used to inject global and topological information into the graph; in the manuscript's current state, RWPE is only adopted for the LRGB tasks.

  3. Presentation. The manuscript is really hard to follow: sentences are dense and heavily hedged, notation is overloaded (ρ\rho, σ\sigma, CrC_r), and much of the evidence supporting the main claims is relegated to the appendix, whose writing suffers from the same issues. The writing would need substantial revision before the manuscript can be considered for publication.

  • [1] Dwivedi et al., Benchmarking Graph Neural Networks, JMLR 2023
  • [2] Kreuzer et al., Rethinking graph transformers with spectral attention. NeurIPS 2021
  • [3] Xu et al. How Powerful are Graph Neural Networks? ICLR 2019
  • [4] Tönshoff, J., et al. Where Did the Gap Go? Reassessing the Long-Range Graph Benchmark. arXiv preprint arXiv:2309.00367, 2023.
  • [5] Bamberger, J. et al. On Measuring Long-Range Interactions in Graph Neural Networks. ICML, 2025.
  • [6] Mathys, J. et al. LRIM: a Physics-Based Benchmark for Provably Evaluating Long-Range Capabilities in Graph Learning. ICLR, 2026.
  • [7] Miglior, L. et al. Can You Hear Me Now? A Benchmark for Long-Range Graph Propagation, ICLR 2026
  • [8] Dwivedi et al., Graph Neural Networks with Learnable Structural and Positional Representations, ICLR 2022
  • [9] Yan et al. Cycle Invariant Positional Encoding for Graph Representation Learning. arXiv preprint arXiv:2311.14333, 2023.
Questions for authors
  1. Can the authors clarify which expressiveness results are specific to RICCIMP? Theorem 3 shows that its propagation components (conductances, gates, walk filter) add no separation power w.r.t. 1-WL, and Appendix C notes that Theorem 5 does not separate RICCIMP from GSN-e. Which results hold only for RICCIMP, and which hold for any edge-aware GNN (e.g., GINE or GSN-e)?
  2. Could the authors compare their methods (both RICCIMP and 1–2 baselines such as GINE) with Laplacian PE, RWSE, or other cycle-based PEs (e.g., [9]), including the profile combined with RWSE? This would show whether the profile adds information beyond degree and existing encodings, helping to better position the paper's empirical novelty. I think that an extensive ablation with controlled feature selection would improve the narrative of the results and help sharpen the empirical benefits of using RICCIMP.
  3. Could the authors provide results on additional state-of-the-art, controllable long-range benchmarks such as LRIM [6] or ECHO [7]? Recent work [4,5] questions whether performance on LRGB alone establishes long-range modeling capability. Such an experiment would clarify whether the observed gains arise from propagating distant information and whether adaptive conductances contribute to improved propagation.
Reviewer_XNGwOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this work the authors present a study on exploring curvature profiles that use degree-based curvature and counts of cycles of different lengths as inputs to a layer towards controling edge weights and propagation range. The paper also demonstrates that curvature profiles can outperform baselines in some settings but trail stronger models in long-range tasks without curvature. The work bridges geometry with GNN design, offering a framework to separate structural augmentation from propagation mechanisms. In general, the paper is well-written and presented clearly.

Critical strengths
  • a simple way to create a new layer for exploting graph structural info
  • the study shows which specific structures are to be used ideally for models.
  • the paper explicitly ties edge curvature to geometric properties (e.g., Forman curvature), providing a link between graph structure and neural net design. This bridges abstract geometric concepts with practical GNN applications.
Critical weaknesses
  • the paper shows what the architecture can distinguish but does not state what it cannot. The theoretical analysis focuses on finite graphs with fixed cycle lengths, but the paper does not address how curvature profiles perform on real-world graphs with highly variable or infinite cycle structures. This could limit the applicability of the framework to dynamic or large-scale graphs.
  • the paper does not thoroughly explore the trade-offs between propagation range (K) and layer depth. For example, while longer walks (higher K) reduce the need for additional layers, the paper does not clarify how this affects computational efficiency or generalization in practice.
Questions for authors
  • How do curvature profiles handle graphs with extremely long cycles that exceed the theoretical thresholds shown in the paper? What modifications would be required to extend the framework to analogous cases?
  • Did the authors observe anything that may resemble to over-smoothing or over-squashing?
Reviewer_r1SKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper asks whether edge geometry helps GNNs by providing structural information or by changing message propagation. It uses edge profiles containing Forman curvature and simple-cycle counts, both as inputs to existing models and within a new RicciMP layer. RicciMP combines positive edge conductances with a learned propagation polynomial to seperate edge weighting from propagation range. The paper analyzes expressivity through edge-labeled 1-WL and gives examples where profile information improves filtering. Experiments show a substantial benefit from adding profiles to GINE on ZINC, but less consistent evidence for the proposed propagation mechanism.

Critical strengths
  • The theory takes the details of the architecture seriously. In particular, the expressivity argument accounts for LayerNorm, and the filtering example is connected to a Jacobian that the actual layer can realize. These are useful details that are often left implicit. The assumptions and limitations are also stated clearly.

  • I apreciate the range of controls and the willingness to report negative results. The comparisons with cycle-only GINE, matched GSN-e, and polynomial GINE help explain what the different components contribute. The ZINC improvement from cycle-only to full-profile GINE holds across all ten seeds. Equally useful is the finding that static conductances capture much of the benefit of adaptive ones, and that the denoising advantage can disappear after small topology changes.

Critical weaknesses
  • I am not yet convinced that the contribution beyond existing work is substantial enough. Edge-rooted substructure counts are already central to GSN-e, and the incomparability argument builds on existing cycle-counting results used in Loopy WL. The adaptation to static edge labels is useful, but the forward separation already follows from global cycle totals. Its first threshold, R=8R=8, also lies beyond the implemented R≤6R\le6. The practical results leave a similar concern: on ZINC, cycle-only GSN-e achieves 0.08899 MAE, compared with 0.09035 for full-profile GINE and 0.09490 for RicciMP. The polynomial GINE comparision gives RicciMP a clear win on Pascal, but losses on both peptide tasks and higher runtime. Together, these results make the additional value of the proposed layer and theoretical characterization appear fairly limited.

  • The explanation for the positive conductance results remains incomplete. In Eq. (12), changing conductance changes both relative edge weights and the scale of the aggregate. The current controls do not separate these effects, so the large denoising improvement cannot yet be attributed specifically to geometry-dependent reweighting. The authors acknowledge this limitation, but it matters to the paper’s main question. The strongest result also uses the same topology for training and unperturbed testing, up to relabeling. Static weights nearly match adaptive weights, and four degree-preserving switches make both worse than unit conductance. On LRGB, none of the 24 conductance contrasts survives the reported Holm correction. I therefore find the evidence for when learned edge geometry improves propagation less convincing than the evidence for structural features.

Questions for authors
  • Since conductance changes both relative edge weights and aggregate scale, what evidence supports attributing the denoising gain to edge-specific reweighting? A small scalar-only control in the existing denoising setup would help, if feasible. Alternatively, please clarify how this limitation affects the interpretation of the result.

  • Could you report an additonal cycle-only versus full-profile GINE comparison on ZINC using the same learning rate and training recipe? Table 11 uses 5×10−45\times10^{-4} for the former and 2×10−32\times10^{-3} for the latter. The independently tuned comparison is reasonable, but a fixed-recipe comparison would clarify how much of the reported 11.3% improvement comes from the Forman coordinate itself.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.