- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This work proposes an alternative to the simplicial WL test where nodes of the simplicial complex are initialized with different colors (using their coordinates) and the higher simplices (edges, triangles, etc.) are initialized with a permutation-invariant function of their vertex coordinates (this can even be empty). The result is that after dim(K) rounds of color message-passing, each simplex’s color fully describes the coordinates of the nodes bounding it, because info flows upward through the boundary messages one dimension at a time. So what does this mean? It means that GSWL is complete for coordinate-preserving isomorphism of embedded complexes (in a fixed frame). So message-passing is at most as expressive as GSWL. The authors also use this to derive a sufficient MLP width, H ≥ 3m, for the architecture to match GSWL on a fixed finite family (just like GIN), and from that show it can output exact sampled ECT values on that family. The coboundary channel isn’t needed for this recovery and is only justified by an ablation. Authors use datasets that share the same abstract complex (same connectivity, different embeddings) to show that the GSWL model can successfully discriminate pairs of data and basic SWL cannot, since it literally outputs the same thing for every input.
Critical strengths
Strength 1: a straightforward extension to an existing expressivity test
GSWL provides a clearcut, straightforward extension of the simplicial WL test by using node coordinates at initialization. The theoretical framework is laid out clearly and simply enough to understand. The depth requirement (L ≥ dim K) is a meaningful prediction, and the theorems are well written (and reflect limitations).
Strength 2: demonstrates the contribution with well controlled experiments.
The experimental design reflects the contribution and provides the evidence one would expect The ladder of DeepSets to GCN to GIN to simplicial exposes the empirical gains at each structural level. The authors also include some interesting sanity checks with Mantra and with the flattened coordinate MLP. Experiments seem reproducible.
Critical weaknesses
The main weakness of the paper is the practical value of the contribution, which is mostly a (well verified, see Strength 2) theoretical framework. To my understanding, most simplicial and mesh networks already use coordinates at initialization. E(n) equivariant simplicial networks (Eijkelboom et al. 2023, Liu et al. 2024) and mesh architectures like MeshCNN (Hanocka et al. 2019) and MeshGraphNets (Pfaff et al. 2021) are examples of this (some of these are referenced in the paper but not compared against).
The paper's "formalization" of this practice is a fine idea, but it gives a trivially complete test which leaves practitioners with little new guidance. Also, why would a practitioner be interested in fixed-frame equivalence for meshes, compared to, for instance, rotation augmentation (Eijkelboom et al.)? Also, I think the baselines are misisng a model that does appear in topological practitioner pipelines, E(n)-equivariant TNNs (Battiloro et al.)-- could authors include a simplicial E(n)-TNN in the experiments?
Questions for authors
Clarifications
-
Why does the paper focus on fixed-frame equivalence? Mesh tasks typically require E(3) invariance, and it seems like the FAUST setup uses rotation augmentation. So I am struggling to understand what the fixed-frame target gets us that an E(n)-invariant formulation (Eijkelboom et al.) doesn’t.
-
I am confused about some of the synthetic MSE numbers: Table 1b reports 0.157 for the geometry-aware model at L=4 (Appendix B), but the depth ablation reports 0.121 at L=1 and 0.099 at L=6. Could you clarify? is it a different split?
-
You report combinatorial SMP on FAUST at 0.062 with 0 variance. I find that really surprising with 10 balanced classes and stratified folds (seems like there should be some variance). Can you help me understand why its 0?
Suggestions for future work
Per the weakness I raised above, could you include an E(n)-equivariant simplicial baseline on FAUST and the synthetic tasks?
3 comments
Thank you for providing these new results and clarifications. While I am eager to see the theoretical revisions described implemented in the paper, I trust the authors will do this of their own accord and raise my score.
Thank you very much for your support! We really appreciate it and will update the paper accordingly.
We thank the reviewer for recognizing the depth characterization and for raising the practical and reporting questions.
Fixed-frame equivalence characterizes the geometric information message passing retains before imposing task-specific invariances. This matters for orientation-dependent targets and the direction-indexed ECT: an invariant predictor cannot distinguish rotation-equivalent inputs with different orientation-dependent labels.
For invariant tasks, we support both augmentation and canonicalization. Centering and PCA frame averaging preserve every paired prediction under rotation on three unseen planar mesh templates, achieving 74.4 ± 3.1% accuracy before and after transformation. Thus, coordinate-preserving expressivity supports invariant prediction pipelines while retaining the flexibility to address orientation-dependent tasks.
The text retained an earlier single-seed depth sweep using nonlinear messages before aggregation, whereas Table 1b reports a five-seed average using summed-message layers. We recovered the scripts and dated figures: reruns closely reproduce the earlier sweep and reproduce Table 1b’s 0.157 ± 0.062. Seed 0 uses exactly the same dataset and split across the scripts; the table additionally averages four independently generated datasets and splits.
A matched five-seed sweep with training-only normalization gives 0.165 ± 0.081, 0.145 ± 0.068, and 0.146 ± 0.069 at L=1,4,6. The improvement from L=1 to L=4 occurs in all five seeds, with no consistent further improvement at L=6. We have corrected the monotonic-improvement claim. This global regression uses richer initialization; the separate local experiment directly tests the minimal-initialization depth bound.
The folds are contiguous, not stratified: five folds of 16 test meshes from the 80-mesh pool. Each training-majority class occurs seven times in training but only once in the corresponding test fold.
Combinatorial SMP receives identical connectivity and constant attributes, forcing a constant prediction within each fitted fold. Any training-majority prediction therefore achieves 1/16 = 0.0625 in every fold, yielding zero fold variance even if different folds select different classes. This accounts for how the reported result can arise; uniform-random chance remains 10%. We will clarify the fold construction and variance convention.
We added an EMPSN-based simplicial baseline with an invariant classification readout on both synthetic data and FAUST. On synthetic deformation classification, it achieves 87.2 ± 6.3% after 400 epochs across five seeds, compared with 97.3 ± 1.2% for EGNN and 98.4 ± 1.3% for rotation-augmented SMP.
We will prioritize finalizing the focused theoretical revisions described above and, time permitting, further investigate the FAUST EMPSN fitting issue. We will update the reviewers if validated results become available before the deadline.