acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Simplicial Message Passing

Abstract

The Weisfeiler–Lehman (WL) test and its simplicial extension (SWL) characterize the combinatorial expressivity of message passing networks, but they are blind to geometry, i.e., meshes with identical connectivity but different embeddings are indistinguishable. We introduce the Geometric Simplicial Weisfeiler–Lehman (GSWL) test, which incorporates vertex coordinates into color refinement for geometric simplicial complexes. In addition, we show that (i) the expressivity of geometry-aware simplicial message passing schemes is bounded above by GSWL, and (ii) that there exist parameters such that the discriminating power of GSWL is matched by these schemes on any fixed finite family of geometric simplicial complexes. Combined with the Euler Characteristic Transform (ECT), a complete invariant for geometric simplicial complexes, this yields a geometric expressivity characterization together with an approximation framework. Experiments on synthetic and mesh datasets illustrate our theory, showing a clear hierarchy from combinatorial to geometry-aware models.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, topological deep learning, higher-order message passing, simplicial message passing, Weisfeiler--Lehman, expressivity, simplicial complex

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vcdiOpenReview ↗
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

  1. 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.

  2. 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?

  3. 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

We thank the reviewer for recognizing the depth characterization and for raising the practical and reporting questions.

Q1: Fixed-frame equivalence characterizes

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.

Q2: Why do the synthetic MSE values differ across depths?

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.

Q3:

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.

Proposed future work:

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.

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.

Reviewer_67BFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper studies the expressive power of message-passing neural networks on geometric simplicial complexes, such as meshes consisting of vertices, edges, and triangles.
  • Its central argument is that existing expressivity frameworks, particularly the Weisfeiler–Lehman (WL) test and its simplicial extension (SWL), characterize structural information but do not account for the geometric embedding of simplicial complexes.
  • For example, two meshes may have identical vertex-edge-triangle connectivity but different vertex coordinates. Conventional SWL considers them indistinguishable despite their geometric differences.
  • To address this limitation, the paper proposes the Geometric Simplicial Weisfeiler–Lehman (GSWL) test and develops a corresponding geometry-aware simplicial message-passing architecture.
  • The paper's primary contributions are theoretical: it establishes the relationship between GSWL and the expressive power of geometry-aware simplicial networks and connects this characterization to the Euler Characteristic Transform (ECT), a complete invariant of geometric support.
  • The authors also conduct experiments on synthetic triangulations, manifold triangulations, and human body meshes to illustrate the benefits of incorporating geometric information and higher-order simplicial interactions.
Critical strengths
  • The paper is easy to follow, which is meritable, especially as a theory paper.
Critical weaknesses

Regarding the concept of GSWL itself

W1. Limited methodological novelty of GSWL. The proposed GSWL test appears to be a relatively straightforward extension of the existing SWL framework. Its primary modification is to initialize simplices with coordinate-derived features, while retaining essentially the same boundary/coboundary refinement mechanism. Moreover, incorporating geometric information into simplicial message passing has already been explored, notably by Eijkelboom et al. (ICML 2023) [already cited by the authors], although their work considers Euclidean equivariance rather than coordinate-preserving isomorphism. The authors should more clearly articulate what fundamentally new insight GSWL provides beyond applying existing simplicial refinement mechanisms to geometry-enriched initial features.

  • Eijkelboom et al. (ICML 2023): E(n) Equivariant Message Passing Simplicial Networks

W2. Limited significance of the theoretical completeness guarantees under injective coordinate initialization. The main completeness result appears to follow rather directly from the assumption that vertex coordinates are injective and are used as initial colors. Since each vertex is uniquely identified by its coordinates, injective boundary aggregation naturally allows each higher-dimensional simplex to recover its constituent vertices after sufficiently many refinement steps. More importantly, Definition 1 explicitly permits initializing higher-dimensional simplices with sorted tuples of their constituent vertex coordinates, in which case the embedded complex can already be identified without any refinement. Furthermore, the finite-family realizability result relies on sufficiently expressive interpolation functions and a hidden dimension that depends on the number of observed colors. These observations raise questions about whether the theoretical guarantees demonstrate a substantive advance in geometric expressivity or largely formalize consequences of uniquely identifying initial features.

Regarding the theoretical results

W3. Limited scope of the ECT realizability and approximation guarantees. The connection to the Euler Characteristic Transform (ECT) is presented as a central theoretical contribution, but the established guarantees appear substantially narrower than the broader motivation. Theorem 4 proves exact computation of sampled ECT values only for a fixed finite family of embedded simplicial complexes, relying on the finite interpolation property of sufficiently expressive neural functions. This does not establish uniform approximation or generalization to unseen embeddings. Moreover, Corollary 2 provides an ECT approximation error bound only under the assumption that the representation already recovers simplex entry times uniformly accurately. The authors explicitly acknowledge that their architecture is not proven to satisfy this assumption on bounded classes. Thus, the result primarily translates an assumed entry-time recovery accuracy into an ECT error bound, rather than establishing that the proposed architecture can uniformly approximate the complete geometric invariant. The authors should more clearly distinguish these conditional and finite-family guarantees from broader claims about geometric expressivity and approximation.

W4. Potential inconsistency between the geometric embedding assumptions and the MANTRA experiments. Section 2.2 defines a geometric simplicial complex through a coordinate map that extends to an injective geometric realization of the entire complex. However, Appendix B states that the MANTRA triangulations are embedded into two-dimensional Euclidean space using the second and third eigenvectors of the graph Laplacian. The evaluated complexes include triangulations of closed surfaces such as spheres, tori, Klein bottles, and real projective planes. Since a compact two-dimensional manifold without boundary cannot be embedded into the plane, these spectral layouts cannot satisfy the injective geometric-realization assumption in Section 2.2. Although the resulting coordinates can still be used as vertex features, this creates a discrepancy between the theoretical setting and the experimental construction. In particular, it is unclear whether the interpretation of the ECT as a complete invariant of geometric support remains applicable to these representations. The authors should clarify whether their framework also covers arbitrary coordinate-annotated abstract complexes with potentially intersecting simplices, and which theoretical results remain valid in that setting.

Regarding the experimental evaluation

W5. Insufficiently controlled comparisons to establish the benefit of higher-order geometric message passing. The experimental comparisons do not adequately isolate the contribution of the proposed higher-order architecture from that of the additional geometric information provided to it. In particular, the geometry-aware SMP receives vertex coordinates along with coordinate-derived edge and triangle features, including lengths, areas, and centroids, whereas GCN and GIN use only vertex coordinates on the 1-skeleton, and DeepSets ignores connectivity. Moreover, combinatorial SMP is inherently unable to distinguish samples sharing identical abstract connectivity, making its poor performance on the deformation tasks largely expected. Even against coordinate-aware baselines, the empirical advantage is relatively modest: on FAUST, the proposed model achieves 0.85 ± 0.03 accuracy compared with 0.80 ± 0.04 for GIN, with overlapping error bars, while the flattened-coordinate MLP matches the proposed method's accuracy on synthetic deformation classification. The evaluation would be more convincing with controlled comparisons that equalize geometric input features, ablations that isolate the contribution of higher-dimensional simplices, and comparisons against existing geometry-aware simplicial architectures, such as Eijkelboom et al. (ICML 2023) and Liu et al. (ICLR 2024), while accounting for their different symmetry assumptions.

  • Eijkelboom et al. (ICML 2023): E(n) Equivariant Message Passing Simplicial Networks
  • Liu et al. (ICLR 2024): Clifford Group Equivariant Simplicial Message Passing Networks

W6. The coboundary ablation does not convincingly establish the necessity of higher-order message passing. Section 6.3 argues that coboundary messages are necessary for geometric expressivity, primarily based on per-vertex curvature prediction. However, under the proposed architecture, a vertex has no boundary faces. Consequently, when coboundary messages are removed, vertex representations cannot receive information from neighboring vertices, edges, or triangles at any depth. Since the curvature target explicitly depends on the geometry of incident triangles, the poor performance of this boundary-only variant is largely expected by construction. The experiment therefore demonstrates the importance of enabling neighborhood information flow to vertices, rather than isolating the specific benefit of coboundary aggregation or higher-order simplicial processing. Furthermore, the coordinate-recovery argument underlying the main theoretical results relies on upward boundary propagation and does not inherently require coboundary messages. A more informative evaluation would compare against vertex-level geometric GNNs that retain neighborhood communication, such as GINE with geometric edge features, EGNN (Satorras et al., ICML 2021), and DiffusionNet (Sharp et al., SIGGRAPH 2022). Furthermore, the authors should conduct targeted architectural ablations that remove triangle-level processing while preserving edge-to-vertex message passing. These comparisons would help determine whether the observed improvements arise specifically from higher-order simplicial interactions or simply from enabling geometric neighborhood aggregation.

  • EGNN (Satorras et al., ICML 2021): E(n) Equivariant Graph Neural Networks
  • DiffusionNet (Sharp et al., SIGGRAPH 2022): DiffusionNet: Discretization Agnostic Learning on Surfaces
Questions for authors

See weaknesses above.

5 comments

We thank the reviewer for the detailed assessment. We address the concerns through a targeted constructive refinement of the existing approximation argument, explicit distinctions between the theoretical settings, and additional controls for geometric inputs and higher-order processing.

W1: Novelty relative to SWL and EMPSNs

Our contribution is the expressivity characterization connecting geometric simplicial message passing to the ECT, including the initialization, depth, and readout conditions needed to compute geometric information. EMPSN addresses Euclidean equivariance; GSWL characterizes coordinate preservation, which also supports invariant prediction through canonicalization. The revision strengthens this existing connection through a targeted constructive refinement of the approximation argument.

W2: Completeness under injective initialization

Completeness provides the foundation; the substantive characterization concerns how geometry becomes available through propagation. Sorted coordinate tuples allow depth-zero identification. For fixed connectivity, with coordinates only at vertices and dimension-only initialization elsewhere, a kk-simplex’s state cannot depend on coordinates before L=kL=k, and GSWL recovers its constituent coordinates at that depth. This is a tight local recovery bound, not a depth requirement for every global prediction.

We distinguish the standard finite-family realization argument, including H≥3mH\geq3m (Theorem 3; proof in the appendix), from uniform approximation on bounded classes (Theorem 5).

W3: Scope of the ECT approximation guarantee

Section 5.2 makes entry-time recovery constructive within the existing message-passing framework. Continuous coordinate moments propagate through boundary messages and enable uniform approximation of simplex entry times.

Fix ambient dimension d, maximal simplex dimension D, a bound N on the number of simplices, and coordinate bound R. For any >0, sufficiently expressive neural maps at depth LD admit one shared parameter choice with uniform entry-time error =/N. Integrating the existing simplexwise ECT decomposition gives F^(ν,t)=∑σ(−1)dim⁡σ[t−a^σ(ν)]+.\widehat F(\nu,t)=\sum_\sigma(-1)^{\dim\sigma}[t-\widehat a_\sigma(\nu)]_+.

The hinge is 1-Lipschitz, so the error is at most N=, uniformly over the bounded class, directions, and thresholds. Raw-ECT approximation in L1 and generalized-support approximation follow as corollaries. This establishes approximation capacity; completeness applies to the exact full cumulative transform.

W4: MANTRA embedding assumptions

MANTRA uses spectral coordinates as attributes of abstract complexes. The refinement/message-passing results and signed-sum approximation construction remain valid in this setting. With distinct vertex coordinates, coordinate-preserving abstract-isomorphism completeness also remains valid.

Here, the signed sum describes the abstract lower-star filtration, which need not equal the ECT of the geometric image when simplices intersect. Genuine geometric embeddings are required for that identification and the resulting support-completeness interpretation. We will make this domain distinction explicit.

W5: Controlled comparisons

We added vertex-only initialization, triangle-removal controls, and comparisons with a common readout and approximately matched parameter counts. On FAUST, accuracies are 85.8 ± 1.2% for SMP, 87.9 ± 2.1% without triangles, 89.6 ± 1.2% for GIN and 88.3 ± 0.6% for GINE. These are mean ± standard deviation across three seed-wise five-fold averages on 80 meshes. They establish strong graph baselines and clarify that FAUST does not demonstrate a triangle advantage. We also added EGNN and EMPSN on synthetic tasks; EMPSN reaches 87.2 ± 6.3% on deformation classification after 400 epochs across five seeds. Our FAUST EMPSN adaptation failed a clean training-subset sanity check, so we are still working on that comparison.

W6: Interpretation of the coboundary ablation

Within this architecture, the boundary-only ablation establishes the information route to vertex outputs: removing coboundary messages eliminates their incoming neighborhood information. Upward coordinate recovery does not require this channel.

We added a triangle-free control retaining vertex–edge communication and replace the planar angle-defect evidence, affected by foldovers and numerical noise, with a 3D surface task using training-only normalization. SMP achieves MSE 0.461 ± 0.071, versus 0.490 ± 0.039 for the parameter-matched triangle-free model, improving in four of five seeds. The larger GINE baseline achieves 0.541 ± 0.037, and EGNN 0.427 ± 0.122. These results show a small average improvement from retaining triangles on this local task.

We are prioritizing final verification of these theoretical revisions and will report further validated baseline results if available before the deadline.

Thank you for the detailed and constructive rebuttal. I appreciate the authors' efforts to address my concerns. The responses have clarified several points:

  • W3: The proposed constructive uniform approximation theorem appears to be a substantial improvement over the original conditional guarantee. If established rigorously, it would address much of my concern regarding the scope of the ECT approximation results.
  • W4: The distinction between abstract complexes with coordinate attributes and genuine geometric embeddings satisfactorily resolves my concern about the MANTRA experiments, provided it is clearly reflected in the revision.
  • W5–W6: The additional graph baselines and triangle-removal controls significantly improve the experimental evaluation. I also appreciate the authors' transparent reporting of results that do not consistently favor the proposed architecture.

However, two main concerns remain.

AQ1. The substantive theoretical contribution of the new uniform approximation result.

The new constructive moment-based argument appears to be a substantial theoretical improvement over the original conditional guarantee. However, since the rebuttal provides only a high-level description, I cannot yet fully assess its validity and implications.

Could the authors provide a more detailed proof sketch explaining how continuous coordinate moments are propagated through the proposed message-passing architecture and how this yields uniform entry-time approximation over the entire bounded class using a single shared parameter choice?

Additionally, I would appreciate clarification of which guarantees apply to the integrated ECT, the raw ECT, and geometric-support recovery, and how the proposed construction differs substantively from existing simplicial message-passing frameworks.

AQ2. The empirical significance of higher-order simplicial processing.

The new controlled experiments are informative, but their results also suggest that the benefit of triangle-level processing is limited.

On FAUST, the proposed SMP achieves 85.8% accuracy, compared with 87.9% without triangles, 89.6% for GIN, and 88.3% for GINE. Thus, the controlled comparison does not demonstrate an advantage from higher-order processing.

On the new 3D curvature task, retaining triangles yields a relatively small improvement (MSE 0.461 versus 0.490), while EGNN achieves a lower mean MSE of 0.427. Although these results are encouraging in part, they do not yet establish a consistent empirical advantage from triangle-level processing.

Could the authors clarify the strength of the conclusions supported by these results, and ideally provide paired seed-wise comparisons for the triangle-removal ablation? I would also encourage revising the experimental discussion to reflect these mixed outcomes rather than presenting a consistent hierarchy favoring simplicial message passing.

Thank you for the constructive follow-up. We have now uploaded the revised manuscript; we apologize that it was not available with our earlier response.

AQ1. Constructive uniform approximation

Section 5.2 states the guarantees, and full proofs are in Appendix A.1. This refinement uses the existing boundary-propagation argument of Lemma 1 and signed simplexwise readout to establish uniform neural approximation, which injective initialization alone does not provide.

Fix ambient dimension dd, maximal simplex dimension DD, simplex-count bound NN, and coordinate-radius bound RR. Use a vertex encoder approximating q(x)=(xα)1≤∣α∣≤D+1q(x)=(x^{\alpha})_{1\leq|\alpha|\leq D+1}, initialize nonvertex feature blocks to zero, and retain dimension tags. Preserving vertex states, the boundary update (Eq. 12)

=\frac{1}{k}\sum_{\tau\in\partial\sigma}p_{\tau}^{(\ell)}, \qquad k=\dim\sigma\geq1,$$ propagates the encoded vertex sums to each simplex after $L\geq D$. The moments determine the projected power sums and, by Newton’s identities, the entry time. Since the maximum is continuous and equal moments and directions give equal maxima, compactness gives a continuous decoder, including at degeneracies. Dimension tags distinguish simplex dimensions; extension and universal approximation give one shared decoder MLP. As specified in Eq. (11), first choose this decoder with error $\delta/2$ and Lipschitz constant $C_A$, then choose uniform encoder error $$\eta\leq\frac{\delta}{2(D+1)\max\\{1,C_A\\}}.$$ The entry-time error is at most $\delta/2+C_A(D+1)\eta\leq\delta$. Thus **one parameter choice works across the entire bounded class and all query directions**. The guarantees are: - **Cumulative ECT (Theorem 5):** the signed hinge readout has error at most $N\delta$, uniformly over complexes, directions, and thresholds. - **Raw ECT (Corollary 2):** a separate ramp readout of width $\tau$ gives threshold-$L^1$ error at most $N(\delta+\tau/2)$, uniformly over complexes and directions (Eq. 13), rather than uniform approximation across threshold crossings. - **Generalized support function (Corollary 3):** uniform approximation follows from $h_K(\nu)=F_K(-\nu,R)-R\chi(K)$ (Eq. 10); this summary is generally incomplete. Completeness applies to the exact full cumulative transform. **AQ2. Empirical significance of triangles** The results support a **modest, task-dependent benefit**. The paired curvature results are: | Seed | Full SMP | Matched triangle-free SMP | MSE difference | |---|---:|---:|---:| | 0 | 0.3495 | 0.4326 | −0.0830 | | 1 | 0.5393 | 0.5287 | +0.0105 | | 2 | 0.4251 | 0.4797 | −0.0546 | | 3 | 0.4583 | 0.4732 | −0.0149 | | 4 | 0.5331 | 0.5380 | −0.0049 | Differences are full minus triangle-free, computed before rounding. Pairs share data, splits, and normalization; the control retains vertex–edge communication with approximately matched active parameters. The paired difference is **−0.0294 ± 0.0344** (mean ± population SD), favoring full SMP in **4/5 seeds**. EGNN retains the lowest mean MSE. On FAUST, full-minus-triangle-free accuracy differences are **−5.00, +2.50, and −3.75 percentage points** across three seed-wise five-fold averages: **−2.08 ± 3.28 points** overall. This does not demonstrate a triangle advantage; folds are not treated as independent replicates. On MANTRA, coordinate SMP achieves 91.2–99.2% accuracy across four fixed templates, compared with 20.4% for combinatorial SMP (Appendix B.7, Table 13). This shows that coordinates help on these tasks; the triangle-removal experiments above test the contribution of triangles. Sections 6.4–6.5 (**Tables 4–5**) and the **Scope** paragraph discuss these outcomes. Our theory establishes geometric approximation capacity, without implying a consistent empirical advantage from triangle-level processing.
Reviewer_o3sEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

the paper introduced an extension of simplicial WL, and focuses on the geometry aspect through lower/upper expressivity results and coordinate recovery linked to the ect. Experiments on synthetic, manifold, and body meshes support the claims using geometry aware models.

Critical strengths
  • theoretical results linking simplicial WL expressivity and embedded geometry.
  • coordinate recovery and the ect connection provide insights in what geometric information MP networks can recover
Critical weaknesses
  • the claims are not surprising since it's obviously to use geometric information when it's available and necessary
  • the experiments are small-scale and not practical
Questions for authors
  • any results regarding the geometric invariants, like wrt rotations and translations
  • possible to apply this to a larger benchmark with stronger point cloud or equivariant geometric NN baselines?
2 comments

We thank the reviewer for recognizing the theoretical connection between geometric message passing and the ECT. We have clarified what this characterization adds and expanded the evaluation with stronger geometric baselines, controlled comparisons, and explicit rotation handling.

Claims are not surprising

We thank the reviewer for their comment. We believe that the fact that geometry helps when available is expected. However, it is not clear a priori what geometry a message-passing algorithm can provably compute and how. Our paper answers this by connecting the ECT to a GSWL test. In our revision, we make this connection now more explicit (see updated Section 5.1 in the paper). Specifically, we show that integrating the ECT with respect to tt yields the support function of the simplicial complex and, for 2D data, its perimeter. These quantities both decompose as signed sum overs implices, with each summand being a Lipschitz continuous function of the vertex coordinates of a single simple. Hence, we can show that an appropriate readout function, like the one introduced by us, approximates these quantities uniformly over all complexes of bounded. The revised experiments directly illustrate this characterization. Triangle-perimeter MSE drops from 1.105 to 0.096 at the predicted coordinate-access depth (Table 1). On nonconvex embedded meshes, SMP’s cumulative-ECT error is 18% of the corresponding constant-baseline error, versus 70% for the raw ECT (Table 2).

Q1: results on geometric invariants (rotations / translations)

We handle translations by centering each mesh. For planar rotations, we use PCA frame averaging: align the principal axes and average logits over the two sign-consistent frames. This gives rotation invariance when the covariance eigenvalues are distinct.

On three unseen mesh templates, PCA-SMP preserves every paired prediction before and after rotation, achieving 74.4 ± 3.1% accuracy in both settings. Rotation-augmented SMP reaches 81.1 ± 2.4% on rotated meshes, compared with 77.8 ± 1.4% for augmented GIN. Results use a common readout, approximately matched parameter counts, and three training seeds (Table 3).

Q2: larger benchmarks

We expanded the FAUST comparison on meshes with 6,890 vertices and 13,776 triangles, using a common readout and approximately matched parameter counts. SMP achieves 85.8 ± 1.2%, compared with 89.6 ± 1.2% for GIN, 88.3 ± 0.6% for GINE, and 87.9 ± 2.1% for SMP without triangles. Results are mean ± standard deviation across three seed-wise five-fold averages on 80 meshes. These controls establish strong geometric graph baselines and clarify where higher-order processing provides an empirical benefit. FAUST establishes strong graph baselines without demonstrating a triangle advantage.

We also added EGNN and EMPSN on synthetic deformation classification. After 400 epochs, they achieve 97.3 ± 1.2% and 87.2 ± 6.3%, respectively, versus 98.4 ± 1.3% for rotation-augmented SMP across five seeds. Our FAUST EMPSN adaptation failed a clean training-subset sanity check, so we are still working on that comparison.

These revisions connect the theoretical characterization to concrete choices of initialization, depth, readout, and symmetry handling. We will prioritize final checks of the focused theoretical revisions and update this response if validated FAUST EMPSN results become available before the deadline.

We believe that your questions helped us clarify and improve the framing of our paper and we moreover believe that we have addressed your concerns. As the deadline for the discussion period is approaching, please let us know whether you have any other questions or comments.

Comments on the paper

1 comment

We thank all reviewers for the constructive feedback and discussion. As the discussion deadline approaches, please let us know if any concerns remain so that we have an opportunity to address them.

We would like to emphasize the novelty of the work: to our knowledge, this is the first constructive connection between the ECT and the neural realizability and approximation capabilities of simplicial message passing.

We hope our revisions and responses have addressed the concerns raised. If they have strengthened your assessment of the work, we would be grateful if you would consider raising your scores accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.