acceptodds
Under review as a conference paper at ICLR 2027

CCMamba: Rank-Aware State-Space Modeling from Causal Scans to Incidence-Structured Mixing on Combinatorial Complexes

Abstract

Selective state space models (SSMs) rely on ordered scans, whereas combinatorial complexes, a general framework encompassing graphs, hypergraphs, simplicial complexes, and cellular complexes, have no canonical ordering of their cells. Imposing a linearization, as in existing Graph Mamba approaches, can introduce order dependence because causal selective scans induce position-dependent, block lower-triangular mixing. We study this mismatch by separating the support of a mixing operator from its factorized content-dependent interactions, and introduce Combinatorial Complex Mamba (CCMamba), a rank-aware framework that combines incidence-based cross-rank lifting with two complementary structured mixers. CCMamba-S preserves bidirectional selective SSM scans and efficient fused sequence processing, while remaining sensitive to the chosen within-rank ordering. CCMamba-K instead replaces causal sequence support with incidence-supported, index-free structural gating and factorized kernel mixing, providing permutation-equivariant mixing without requiring a cell linearization. We characterize the order sensitivity of causal scans, establish the permutation equivariance and structural factorization of CCMamba-K, and give a conditional characterization of its expressive power under explicit assumptions. Experiments across graph, hypergraph, simplicial-complex, and cellular-complex benchmarks show that CCMamba achieves competitive predictive performance, with memory and runtime advantages over the evaluated attention-based baselines in several higher-order settings. Code: https://anonymous.4open.science/r/CCMamba .

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Deep Learning, Combinatorial Complexes, Higher-Order Graph Learning, State Space Models

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_NyYiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to adapt selective state-space models to combinatorial complexes, where cells do not have a natural ordering. It proposes CCMamba with two variants: CCMamba-S uses bidirectional Mamba scans over an imposed rank-wise ordering and therefore remains order-sensitive, while CCMamba-K replaces the scan with incidence-based structural kernel mixing, making it permutation equivariant. The paper provides theoretical results on order sensitivity, equivariance, and a conditional connection to 1-CCWL, and evaluates the models across graph and higher-order benchmarks with several ablation and efficiency studies.

Critical strengths
  1. The permutation experiments are a clear strength. They directly test the main distinction between the two variants: CCMamba-S remains sensitive to the imposed ordering, while CCMamba-K is permutation equivariant up to numerical precision.
  2. The experimental evaluation is broad and includes useful ablations. The paper covers graphs, hypergraphs, simplicial and cellular complexes, as well as native higher-order datasets. The matched-mixer ablation is particularly useful because it compares CCMamba-K with message-passing and attention-based alternatives under a controlled setup.
Critical weaknesses
  1. My main concern is the novelty relative to closely related work. The ordering problem in Graph Mamba is already well known. For example, Behrouz and Hashemi (2024) already use graph-dependent ordering together with bidirectional state-space processing. TopoMamba (Montagna et al., 2025) also applies Mamba-style state-space modeling to simplicial complexes. More importantly, Huang, Miao, and Li (2024) already study the lack of a canonical ordering for graphs and propose permutation-equivariant aggregation based on factorized graph kernels. In this context, CCMamba-S seems mainly to extend existing Graph Mamba ideas to a rank-aware combinatorial-complex setting. CCMamba-K is more distinctive because of the combination of multi-rank incidence support, structural gating, and factorized kernel interactions, but I think the paper should explain much more clearly how this differs from the approach of Huang et al. and from generalized combinatorial-complex architectures such as TopoTune (Papillon et al., 2025).

The most relevant references are:

  • Ali Behrouz and Farnoosh Hashemi. “Graph Mamba: Towards Learning on Graphs with State Space Models.” Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 119–130, 2024.
  • Yinan Huang, Siqi Miao, and Pan Li. “What Can We Learn from State Space Models for Machine Learning on Graphs?” arXiv preprint arXiv:2406.05815, 2024.
  • Marco Montagna, Simone Scardapane, and Lev Telyatnikov. “Topological Deep Learning with State-Space Models: A Mamba Approach for Simplicial Complexes.” 2025 International Joint Conference on Neural Networks (IJCNN), pp. 1–8, IEEE, 2025.
  • Mathilde Papillon, Guillermo Bernardez, Claudio Battiloro, and Nina Miolane. “TopoTune: A Framework for Generalized Combinatorial Complex Neural Networks.” International Conference on Machine Learning, 2025.
  1. I am also not fully convinced by the long-range/state-space framing of CCMamba-K. The K variant is not a selective SSM block, and its neighborhood is restricted to immediate incidence relations, so longer-range information still requires stacking multiple layers. I therefore think the paper should be more precise about which advantages associated with state-space models are actually retained by CCMamba-K. The 1-CCWL result is also proved for an idealized version of the model under several assumptions, including injective mappings and a simplified aggregation setting, so I think its implications for the practical architecture should be stated more carefully.
  2. The battery of experiments is vast, but the gains are overall incremental, on mostly saturated benchmarks, excluding Mantra.
Questions for authors
  1. The main point I would like clarified is the relation between CCMamba-K and GSSC (Huang, Miao, and Li, 2024). Both approaches address the lack of a canonical ordering by moving away from purely sequential processing and using factorized, permutation-equivariant interactions. What is the main methodological difference between CCMamba-K and a GSSC-style kernel mixer applied over combinatorial-complex incidence relations? In particular, what does CCMamba-K enable that could not be obtained by combining the GSSC idea with an existing combinatorial-complex architecture?
  2. What exactly is meant by long-range modeling in CCMamba-K? As I understand it, the K variant only mixes over immediate incidence neighborhoods, so information from more distant cells is obtained by stacking layers. If this is correct, I think the distinction between the long-range behavior of CCMamba-S and the local mixing of CCMamba-K should be made clearer. If CCMamba-K has a different mechanism for capturing non-local information, could the authors clarify it?
  3. Could the authors clarify the scope of the 1-CCWL result for the practical CCMamba-K model? The theoretical result is proved under an idealized setting with assumptions such as simplified structural coefficients and injective mappings. Moreover, the correlation between WL expressivity and downstream performance seems to be fundamentally refuted by recent literature. How directly should this result be interpreted as a statement about the learned model used in the experiments?
  4. Could the authors clarify the experimental protocol? In the appendix, different seed sets seem to be reported in different places, and some of the default lifting choices do not appear to match exactly the corresponding results in the main tables and lifting ablations. Which seeds and lifting configurations were used for the final reported numbers? C.1 reports seeds {0,3,5,7,9}, whereas C.4 reports {1,3,5,7,9}; moreover, C.4 specifies KHop and Clique as the default hypergraph/simplex lifts, but the corresponding values in Table D.2 do not match the main Table 1 values.
  5. Tables 2 and 3 are clearly not adhering to the correct paper format.

Overall, given the idea, this paper is well written and tested. However, the idea per se is simple, partly incremental and the battery of experiments is exhaustive but not particularly insightful or modern in terms of employed benchmarks.

5 comments

W1/Q1:....,the novelty relative to closely related work,..., should explain much more clearly how this differs from the approach of Huang et al. and from generalized combinatorial-complex architectures such as TopoTune, ... methodological difference between CCMamba-K and a GSSC-style kernel mixer applied over combinatorial-complex incidence relations?

We thank the reviewer for highlighting these connections. We address this concern in four respects.

(1)Methodological novelty and relation to prior work. Graph Mamba (Behrouz and Hashemi, 2024) combines graph-dependent tokenization and ordering with bidirectional selective SSMs. TopoMamba (Montagna et al., 2025) constructs node-specific sequences from rank-wise summaries of incident simplices. GSSC (Huang et al., 2024) achieves global, permutation-equivariant graph aggregation through factorized spectral kernels. TopoTune (Papillon et al., 2025) provides a general framework for combining configurable CC neighborhoods with neural network backbones.

MethodDomainSSM RecurrenceMixing SupportStructural MechanismPermutation Equivariance
Graph MambaGraph√Ordered neighborhood tokensGraph-dependent tokenization and orderingNot guaranteed in general
TopoMambaSimplicial Complex√Rank-wise neighborhood summariesSimplicial incidence aggregation and bidirectional SSMConditional on sequence construction
GSSCGraph×Global graph kernelFactorized spectral structural kernel√, under stated conditions
TopoTuneCombinatorial Complex×Relational neighborhoodsHasse-GraphBackbone-dependent
CCMamba-SCombinatorial Complex√Rank-specific global scansMulti-rank incidence lifting + bidirectional selective SSMNot guaranteed in general
CCMamba-KCombinatorial Complex×Adjacent-rank incidence neighborhoodsStructural gating and factorized content kernel√, under index-free conditions

CCMamba-S extends selective recurrence to multi-rank cell representations through incidence lifting and rank-specific scans. CCMamba-K uses incidence relations directly in the mixing weights, as detailed below. Appendix B (Table B.1) summarizes these distinctions.

(2) What distinguishes CCMamba-K from GSSC and GSSC-TopoTune?

In GSSC, Laplacian-derived representations zuz_u and zvz_v determine the structural kernel for global graph aggregation. Its update can be written schematically as:

hu=∑v∈V⟨zuWQ,zvWK⟩d⏟spectral structural kernel⊙WOxv+Self⁡(u).h_u = \sum_{v\in\mathcal V} \underbrace{ \langle z_uW_Q,z_vW_K\rangle_d }_{\text{spectral structural kernel}} \odot W_Ox_v +\operatorname{Self}(u).

GSSC-TopoTune incorporates the GSSC mixer into TopoTune's relational framework.

In contrast, CCMamba-K defines interactions directly over typed adjacent-rank incidence neighborhoods. Its value-weighted mixing matrix is

MCCMamba−K=LCC(X)⊙(ΨQΨK⊤)M_{\mathrm{CCMamba-K}}=L_{\mathrm{CC}}(X)\odot(\Psi_Q\Psi_K^\top),

and its cell-wise update is:

h_\sigma = \left[ \sum_{\tau\in\mathcal N(\sigma)} a_{\sigma\tau} \psi(q_\sigma)^\top\psi(k_\tau)v_\tau \\;\middle\\|\\; z_\sigma \right].

Here, LCC(X)L_{\mathrm{CC}}(X) encodes typed adjacent-rank incidence support and learned structural coefficients, while ΨQΨK⊤\Psi_Q\Psi_K^\top represents factorized content-dependent kernel interactions. The second output block zσz_\sigma retains the uncontracted neighborhood key aggregate.

(3) Direct empirical comparison with GSSC-TopoTune. We implemented GSSC-TopoTune by incorporating the GSSC mixer into TopoTune and evaluated it alongside the methods below. We report accuracy(%) for classification and PM25CC (MAE↓). Here, “—” indicates that the model cannot be directly applied to this structure.

MethodGraph(MUTAG)Graph(PROTEINS)Hypergraph(House)Hypergraph(Cora-Citation)Hypergraph(Citeseer-Citation)Simplex(MANTRA-genus-400)Simplex(Manifold40)Cellular(PM25CC)
GMN76.34±3.4975.77±0.65——————
GSSC77.81±4.8575.70±1.68——————
TopoTune78.72±6.4977.06±2.8448.85±6.8468.33±1.3067.32±0.7831.30±2.49 54.70±2.14 1.368±0.058
GSSC-TopoTune77.45±3.2375.70±1.2844.70±2.1969.39±0.8567.87±1.5666.40±6.2352.50±3.751.402±0.0107
CCMamba-S79.75±2.4077.42±4.1254.49±4.6683.65±1.5873.79±3.1879.40±6.3962.00±2.321.327±0.054
CCMamba-K80.85±3.3674.62±1.8054.80±4.7779.04±1.0073.91±2.1987.19±4.5864.21±2.071.362±0.067

CCMamba-K achieves higher mean accuracy than GSSC-TopoTune with absolute improvements of 10.10, 9.65, and 20.79 percentage points on House, Citation-Cora, and MANTRA Genus, respectively. However, GSSC-TopoTune performs better on PROTEINS.

(4)Matched Mixer Ablation. To control for differences in lifting and relational inputs, we also compared the mixers under shared incidence support and experimental settings. In this comparison, CCMamba-K achieved higher mean accuracy than Incidence-GSSC on all four datasets, with gains of 3.05–4.25 percentage points. We moved this ablation to the main text (Table 4).

W2/Q2:What exactly is meant by long-range modeling in CCMamba-K?...,the distinction between the long-range behavior of CCMamba-S and the local mixing of CCMamba-K should be made clearer.

We thank the reviewer for highlighting the need to clarify how information propagates in the two variants.

(1)Distinction between CCMamba-S and CCMamba-K. CCMamba-S retains genuine bidirectional selective SSM recurrence and efficient fused scanning, enabling information propagation across rank-specific cell sequences within a single scan block, although its representations remain order-sensitive. In contrast, CCMamba-K does not retain selective SSM recurrence, fused scanning, or single-layer global propagation. Instead, it employs permutation-equivariant, incidence-supported factorized kernel mixing. Its computational efficiency arises from sparse structural aggregation rather than state-space recurrence.

(2)What does long-range modeling mean in CCMamba-K? We agree that CCMamba-K directly mixes information only within immediate adjacent-rank incidence neighborhoods: $ \mathcal N(\sigma)= \mathcal N_{k-1\rightarrow k}(\sigma) \cup \mathcal N_{k+1\rightarrow k}(\sigma)$. Stacking layers allows information from more distant cells to reach σ\sigma. Higher-order lifting can also shorten paths by introducing additional incidence relations. A single kernel mixer remains local to its incidence neighborhood. Our depth-sensitivity study evaluates 2–64 layers across graph, hypergraph, simplicial, and cellular representations (Appendix E.6, Fig. E.1). The table shows CCMamba-K test accuracy (%) on CiteSeer.

Topology248163264
Graph72.8973.1372.5671.5771.0471.40
Hypergraph70.5469.8470.1870.2070.9570.83
Simplex74.2973.3372.8273.8373.8373.83
Cellular74.8675.3774.4374.1774.0773.21

We have explicitly clarified this distinction in the revised Related Work (Appendix B, Table B.1 and the accompanying discussion) and reiterated it in Section 6. CCMamba-S provides rank-wise sequence-global selective scanning, whereas CCMamba-K performs incidence-local structural mixing with depth-dependent multi-hop propagation.

W2/Q3:Could the authors clarify the scope of the 1-CCWL result for the practical CCMamba-K model?..., the correlation between WL expressivity and downstream performance seems to be fundamentally refuted by recent literature. How directly should this result be interpreted as a statement about the learned model used in the experiments?

We thank the reviewer for highlighting the distinction between idealized theoretical expressivity and the realized expressive power of the practical CCMamba-K model.

(1)Scope of the 1-CCWL result. Conventional Mamba relies on order-dependent selective scans, which do not guarantee permutation equivariance and do not directly admit a standard WL-style multiset analysis. CCMamba-K aggregates over unordered incidence neighborhoods. Its index-free, permutation-equivariant updates permit a multiset-based analysis. Theorem 1 establishes that an idealized CCMamba-K specialization matches the distinguishing power of 1-CCWL under the stated assumptions, including the following.

PropertyIdealized CCMamba-KPractical CCMamba-K
Structural coefficients(Theorem 1)aστ=1a_{\sigma\tau}=1Learned structural gating
Kernel representation(Lemma 5, Fig. E.2)Injective multiset embeddingFinite-dimensional feature mapping
Aggregation (Table E.9)Unnormalized zσz_\sigma retainedStructural aggregation with learned coefficients
Update, Readout(Theorem 1)Assumed injectiveLearned MLP, injectivity not guaranteed
Expressivity(Appendix C.4)Conditional 1-CCWL matchingNo unconditional matching guarantee

Importantly, permutation equivariance enables the structural multiset analysis but does not guarantee 1-CCWL matching, which additionally requires injectivity and specific aggregation assumptions. We have clarified these limitations in Section 4 and Appendix C.4.

(2) Relationship Between WL Expressivity and Downstream Performance. Appendix E.4 evaluates downstream accuracy and structural discrimination separately. Table E.7 compares CCMamba with 12 expressive GNN baselines across ten graph benchmarks. In Table E.8, CCMamba-K separates 6.25% of BREC pairs and 9.20% of SR25 pairs, compared with 0% on both for the evaluated GIN baseline. On CSL, it achieves 11.60% classification accuracy, close to the 10% chance level. These results show a modest discrimination advantage over GIN on the two pair-separation benchmarks.

We agree that stronger WL distinguishability does not necessarily imply better downstream predictive performance or generalization (Franks et al., 2024; Maskey et al., 2026). We distinguish theoretical structural discrimination from empirical predictive utility. Line2039 reports these experiment is the evidence for the benefit of incidence-aware higher-order representation rather than stronger expressive power.

W3: The battery of experiments is vast, but the gains are overall incremental, on mostly saturated benchmarks, excluding Mantra.

We thank the reviewer for this fair assessment, and we agree with part of it. On mature, near-saturated benchmarks (e.g., PubMed, Minesweeper, Roman-Empire), differences between CCMamba and the strongest baselines are small relative to seed variance, and we do not claim otherwise. We address this concern in two respects.

(1)Statistical Significance. We have strengthened the evaluation with paired statistical tests and Holm correction for multiple comparisons(Appendix E.8, Tables E.10–E.11) Significant differences are marked with ∗,∗∗*, **, and ∗∗∗*** for adjusted (p<0.05), (p<0.01), and (p<0.001), respectively, while non-significant comparisons are explicitly identified.

For example, the supplementary comparisons provide statistical support for CCMamba-S on 20NewsW100 and IMDB-B.

Domain/DatasetBaselineCCMambaGain (pp)Adjusted pSig.
Hypergraph/20NewsW100HJRL: 80.88±0.22CCMamba-S: 81.47±0.31+0.590.0022** ↑
Hypergraph/20NewsW100HJRL: 80.88±0.22CCMamba-K: 82.09±0.63+1.210.0198* ↑
Graph/IMDB-BGraphGPS: 72.35±2.60CCMamba-S: 73.20±3.56+0.850.0340* ↑

Interpretation of Incremental Gains. These results confirm that performance differences are dataset-dependent: some modest improvements receive statistical support, while others are not significant, and certain comparisons favor the baselines. We therefore do not claim consistent or substantial improvements across saturated benchmarks.

(2)Scope of claims. We do not claim consistent accuracy gains on saturated benchmarks. Our contributions are the order-sensitivity analysis of causal scans, a permutation-equivariant incidence-supported mixer with conditional expressivity analysis, and favorable memory/runtime scaling. We have revised Section 5 to state that gains on saturated benchmarks are within variance and to foreground the MANTRA, hypergraph, and efficiency results. Larger native combinatorial-complex benchmarks remain needed (Section 6, Appendix F).

Q4: Could the authors clarify the experimental protocol? In the appendix, different seed sets appear to be reported in different places, and some of the default lifting choices do not appear to match exactly the corresponding results in the main tables and lifting ablations...

We thank the reviewer for identifying these inconsistencies. We have standardized the training seeds to {1, 3, 5, 7, 9} throughout the manuscript and clarified the distinction between training seeds and data-split seeds. For graph-origin benchmarks, all methods use matched data partitions, with training randomness controlled by the five reported seeds. Native hypergraph benchmarks follow their dataset-specific splitting protocols, with the same partitions shared across the compared methods.

We also clarify that the default graph-to-complex lifting configurations are KHop for hypergraphs, Clique for simplicial complexes, and Cycle for cellular complexes. We have corrected the default-lifting entries in Appendix E.2 (Table E.5) to match the corresponding CCMamba-S results in Table 1 across all ten graph-origin benchmarks.

Q5: Tables 2 and 3 are clearly not adhering to the correct paper format.

We thank the reviewer for pointing out the formatting issues. We have reformatted Tables 2 and 3 to comply with the ICLR guidelines and improve readability. We reorganized Table 2 by moving selected benchmarks to the appendix and adjusted the layout, spacing, and alignment of both tables while preserving the original results.

We hope this response addresses your concerns. If our clarifications address your question, we would sincerely appreciate your reconsideration of the score. Otherwise, we remain happy to discuss further.

Reviewer_RxdiOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CCMamba, a framework for higher-order structural modeling on combinatorial complexes. The method integrates information across different ranks through incidence lifting and introduces two mixing mechanisms, CCMamba-S and CCMamba-K, focusing on efficient sequence processing and permutation equivariance, respectively. Experiments on graphs, hypergraphs, and other higher-order structure datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Unified Modeling Framework. The paper proposes a unified framework for higher-order structural modeling and introduces two complementary mixing mechanisms, CCMamba-S and CCMamba-K, focusing on computational efficiency and structural symmetry, respectively. This design accommodates different types of higher-order structures, offering methodological and practical value.

  2. Comprehensive Experimental Evaluation. The paper conducts extensive experiments on various datasets, including graphs, hypergraphs, simplicial complexes, and cell complexes, along with comprehensive ablation studies and efficiency evaluations. The results demonstrate that the proposed methods achieve competitive performance across multiple tasks and offer computational efficiency advantages in certain settings, providing solid support for the paper's main conclusions.

Critical weaknesses
  1. Limited Methodological Novelty. The proposed rank-aware aggregation, bidirectional Mamba, and kernel mixing mechanisms share similarities with existing approaches. Although integrating these techniques into a unified framework for higher-order structural modeling offers some value, the novelty of several key design choices remains insufficiently demonstrated. In particular, the distinction between CCMamba-K and existing structured kernel attention methods requires further clarification.

  2. Presentation Issues. The paper contains several formatting and presentation issues. Parts of Tables 2 and 3 extend beyond the column boundaries, while Tables 1–3 contain densely packed text and numerical values. Some labels in Figures 3 and 6 are too small, affecting readability. Additionally, the SSM introduction in Section 2 contains an incomplete sentence. The authors are encouraged to carefully review and improve the formatting and presentation to enhance the overall readability of the paper.

Questions for authors
  1. Figure 5(c) shows that rank-aware ordering performs well on graph, hypergraph, and cellular structures, whereas canonical or degree-based ordering achieves better results on simplicial structures. What do the authors consider to be the main reason for this difference?

  2. In Remark 3 of Appendix B.3, the content-dependent gate gστΔg^\Delta_{\sigma\tau} only uses the target cell features hσh_\sigma and the relation weight wστw_{\sigma\tau}, while the source cell features hτh_\tau are incorporated only through kernel interactions. What motivates this design choice, rather than incorporating the source cell features directly into the gating computation?

5 comments

W1: Limited Methodological Novelty. ...., the distinction between CCMamba-K and existing structured kernel attention methods requires further clarification.

We thank the reviewer for this important observation. We agree that rank-aware aggregation, bidirectional selective scans, and kernel attention are individually established techniques. Our contribution lies in their structured adaptation to multi-rank combinatorial complexes, particularly the design of incidence-conditioned mixing operators.

(1)Methodological Distinction. We clarify the differences between CCMamba and related structured mixing architectures in the following comparison.

MethodDomainMixingStructural SupportSelective SSM RecurrencePermutation EquivarianceMathematical Formulation
Softmax Attention(Vaswani et al., 2017)GeneralSoftmax attentionGlobal×√H′=softmax(QK⊤/d)VH'=\mathrm{softmax}(QK^\top/\sqrt{d})V
Linear Kernel Attention(Katharopoulos et al., 2020)GeneralFactorized kernel attentionMasked×√H′=ΨQ(ΨK⊤V)H'=\Psi_Q(\Psi_K^\top V)
Graph Mamba(Behrouz and Hashemi, 2024)GraphSelective SSMOrdered graph neighborhoods√×H′=SSM(πG(X))H'=\mathrm{SSM}(\pi_G(X))
TopoMamba(Montagna et al., 2025)Simplicial ComplexRank-pooled selective SSMIncident simplices grouped by rank√  Node-levelhu′=SSM([Poolr(Nr(u))]r)h_u'=\mathrm{SSM}([\mathrm{Pool}_r(\mathcal{N}_r(u))]_r)
GSSC(Huang et al., 2024)GraphSpectral kernel aggregationGlobal graph spectral kernel×√H′=ZQZK⊤VH'=Z_QZ_K^\top V
TopoTune(Papillon et al., 2025)Combinatorial ComplexBackbone-dependentRelational neighborhoods-Backbone-dependentH′=AGGR∈RfR(H,AR)H'=\mathrm{AGG}_{R\in\mathcal{R}}f_R(H,A_R)
CCMamba-S (Ours)Combinatorial ComplexBidirectional selective SSMRank-specific cell sequences√×Hr′=BiSSM(πr(Xr))H_r'=\mathrm{BiSSM}(\pi_r(X_r))
CCMamba-K (Ours)Combinatorial ComplexIncidence-conditioned kernel mixingTyped adjacent-rank incidence×√MCC=LCC(X)⊙(ΨQΨK⊤)M_{\mathrm{CC}}=L_{\mathrm{CC}}(X)\odot(\Psi_Q\Psi_K^\top)

CCMamba-S retains genuine selective state-space recurrence over rank-specific cell sequences, whereas CCMamba-K replaces order-dependent scanning with permutation-equivariant, incidence-conditioned kernel aggregation. Unlike GSSC, which primarily employs globally factorized graph spectral kernels, CCMamba-K explicitly combines typed multi-rank incidence support with content-dependent kernel interactions. We emphasize that this formulation is not categorically distinct from general edge-aware kernel attention; rather, its contribution lies in the specific coupling of higher-order structural relations and content-dependent mixing.

(2)Matched Mixer Ablation. To isolate the contribution of the mixing mechanism, we compare CCMamba-K against Gated-Incidence-MP, Incidence-Softmax, Incidence-Kernel-Attention, and Incidence-GSSC under matched structural inputs and interaction support. We have moved the existing matched-mixer ablation from the Supplementary Information to the main text (Table 4), comparing CCMamba-K against four alternatives.

ModelMUTAGPROTEINSCoraCiteSeer
Gated-Incidence-MP84.30±3.1677.70±3.0986.53±1.1275.50±2.13
Incidence-Softmax84.15±3.8173.05±2.5887.74±1.0375.41±2.33
Incidence-Kernel-Attn84.85±3.4275.99±2.1886.77±1.4474.31±2.43
Incidence-GSSC82.98±4.0274.84±3.1086.91±1.1574.65±2.13
CCMamba-K87.23±5.7578.49±2.7689.96±1.8677.79±2.12

These results support the empirical value of combining incidence-conditioned structural gating with factorized content interactions, beyond simply replacing the backbone with an existing kernel-attention mechanism.

The theoretical and empirical analyses clarify that CCMamba's contribution lies in complementary rank-aware structural mixing mechanisms, with CCMamba-K providing a concrete permutation-equivariant operator for multi-rank relational learning.

W2: Presentation Issues,...,Tables 2 and 3 extend beyond the column boundaries,...Some labels in Figures 3 and 6 are too small,....Section 2 contains an incomplete sentence

We thank the reviewer for pointing out these presentation issues. We have carefully revised the manuscript to improve its readability and formatting.

(1)Tables. We have reformatted Tables 1–3 to ensure compliance with the ICLR formatting guidelines. Specifically, we moved selected benchmark results from Table 2 to the Appendix, repositioned Table 3, and adjusted column widths, spacing, and alignment to improve readability and prevent content overflow.

(2)Figures. We have revised Figures 3 and 6 by enlarging labels, legends, and axis annotations to improve visual clarity.

(3)Text. We have corrected the incomplete sentence in Section 2 and refined the SSM introduction to clarify continuous-to-discrete state-space modeling.

We have also reviewed the manuscript for consistency in notation, typography, and overall presentation.

Q1:Figure 5(c)...canonical or degree-based ordering achieves better results on simplicial structures,...What do the authors consider to be the main reason for this difference?

We thank the reviewer for this insightful question. This difference arises from the interaction between simplicial structure and sequence ordering.

The rank-aware ordering uses the incidence-based signature sk(σ)=(dk−(σ),dk+(σ))s_k(\sigma)=(d_k^-(\sigma),d_k^+(\sigma)). However, due to the downward-closure property of simplicial complexes, every kk-simplex has exactly k+1k+1 immediate faces. For example, in a simplicial complex, every kk-simplex has exactly k+1k+1 immediate lower-dimensional faces, yielding dk−=k+1d_k^-=k+1 for k≥1k\geq1. In a 2-dimensional simplicial complex, all top-rank triangles satisfy d2−=3d_2^-=3 and d2+=0d_2^+=0, resulting in the identical rank-aware ordering signature s2=(3,0)s_2=(3,0). Consequently, this ordering strategy cannot distinguish these triangles based on their structural signatures and must rely on tie-breaking.

Canonical or degree-based ordering may preserve alternative indexing or connectivity cues that produce more favorable sequences for selective SSM processing on certain datasets. This provides a plausible explanation for their better performance in Fig. 5(c), although the ordering advantages remain dataset-dependent.

The rank-aware information propagation and rank-aware sequence ordering are distinct. The former explicitly models incidence-defined cross-rank interactions, whereas the latter determines the sequence arrangement for order-sensitive selective scans. Thus, the observed differences do not undermine rank-aware structural modeling but highlight the limitations of a particular ordering heuristic(see Appendix E.3 for a detailed discussion).

Q2:In Remark 3 of Appendix B.3, .... while the source cell features (h_\tau) are incorporated only through kernel interactions. What motivates this design choice, rather than incorporating the source cell features directly into the gating computation?

We thank the reviewer for this insightful question. Our design separates structural gating from source-target content interaction.

(1)Motivation. The structural gate gστstrg_{\sigma\tau}^{\mathrm{str}} captures relation importance through incidence weights and degree information, while gστΔg_{\sigma\tau}^{\Delta} adaptively regulates incoming information according to the target representation hσh_\sigma.

Source features are not omitted from the mixing process. They are incorporated through both the kernel key kτ=WKXτk_\tau=W_KX_\tau and value vτ=WVXτv_\tau=W_VX_\tau. The query-key interaction ψ(qσ)⊤ψ(kτ)\psi(q_\sigma)^\top\psi(k_\tau) captures source-target content compatibility, while vτv_\tau carries source information into the aggregated representation.

(2)Empirical Results. To examine this design choice, we compare the proposed target-conditioned gate with a source-target-conditioned alternative, while keeping the incidence support and kernel mixer unchanged.

DatasetOriginal Gating:Target Featurew/ Source FeatureGains
MUTAG87.23±5.7586.15±5.51+1.08
PROTEINS78.49±2.7673.12±2.44+5.37
IMDB-B72.40±2.2268.64±2.71+3.76
Cora89.96±1.8686.65±1.77+3.31
CiteSeer77.79±2.1275.71±2.46+2.08

The results show that target-conditioned gating achieves higher mean accuracy across all five datasets. These findings support our design choice of separating target-conditioned information reception from source-target kernel matching, rather than introducing additional source-dependent gating. We have clarified the motivation for target-conditioned gating in Appendix C.3 (Remark 3) and included a controlled comparison with source-target-conditioned gating in Appendix E.10 (Table E.13).

Thank you again for your constructive feedback. We hope these revisions clarify the methodological contribution and address your concerns. We welcome any further questions.

Dear Reviewer,

As the rebuttal period is coming to an end, I would like to kindly ask you to read the authors' response and reply as soon as possible.

Thank you for your prompt attention.

Best regards,

Area Chair

Reviewer_DG9sOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to develop a framework for rank-aware state-space modeling, moving from causal scans to incidence-structure mixing using combinatorial complexes. Combinatorial complexes are mathematical objects that capture relationships among neighboring elements, where nodes, edges, and faces correspond to ranks 0, 1, and 2, respectively. This representation helps capture structural proximity. The paper proposes CCMamba, building on the existing Mamba approach. The central idea is to model structured mixing by separating interaction support from content-dependent mixing, enabling application to graphs, hypergraphs, simplicial complexes, and cellular complexes while aiming to achieve order insensitivity. The paper introduces two variants: CCMamba-S, which uses selective state-space scans for sequence processing, and CCMamba-K, which employs a structure-kernel mixer with incidence-defined support to reduce dependencies inherited from linearization in state-space.

Critical strengths
  1. Modeling combinatorial complexes provides a broader representation for a network and can generalize beyond conventional graphs to more generic structured domains.

  2. In contrast to conventional sequence-based approaches which impose an ordering requirement on structured data, CCMamba-K reduces dependence with incidence-defined structural support.

Critical weaknesses
  • Overall, the paper is difficult to follow for many reasons. First, several symbols are introduced without sufficient explanation leaving reader for guess work. For example, in Lines 103–104, the meanings of (\sigma) and (\tau) in this context are unclear. The meaning of “rank” and its relationship to the broader graph structure is unclear. Citations appear to be missing parentheses, and the definition of “large-scale” is insufficiently precise ; is it few tens of nodes or a few hundreds or in millions. It would be beneficial to define what “large-scale” mean here. There are also missing or incorrect symbols in Equation 1, a typo in Line 241 (“M(x)x”), and a vague statement in Lines 257–258 that “(\tau) is an immediate face of (\sigma).” Also, the notation (\operatorname{co}\delta_\sigma) is unclear and requires explanation.

  • The motivation for this work remains vague and would benefit from a clearer explanation. In particular, why is rank awareness necessary in the first place? As a reader it is not clear as to why ranking is the novelty of the approach as combinatorial complexes already encode rank information; what is the additional novelty the proposed method introduces beyond the existing representation. It is not clear how the proposed approach leverages rank information in a different way. If rank awareness is not a central methodological contribution the title should be revised.

  • The order-insensitive claim would benefit with further validation by defining this property and testing the model with different orderings of the same structure to demonstrate that its independent on the choice of ordering.

  • The evaluation is limited to classification tasks and does not consider link prediction as graph has variety of tasks (Section 5.1) and having link prediction task would highlight the capability and limitation of the proposed approach. Moreover, Cellular complexes are missing baselines from table 1. Having those baselines will help make the manuscript stronger.

  • Results are close to the baselines especially when considering variance. This questions the empirical contribution of new method. Given the small performance differences along with reported variance, it would be useful to conduct statistical significance tests.

  • The paper contains numerous formatting issues. Table 1 extends beyond the page margins, and its font sizes are inconsistent. Tables 2 and 3 also appear to exceed the margins. As a result the paper does not follow ICLR submission format.

  • The paper also lacks a conclusion section.

Questions for authors

see above

7 comments

W1:...First, several symbols are introduced without sufficient explanation, ... Lines 103–104, the meanings of σ\sigma and τ\tau in this context are unclear, ..."rank" and its relationship to the broader graph structure is unclear,...,a typo in Line 241 ("M(x)x"), and a vague statement in Lines 257–258 ...

We thank the reviewer for highlighting the issues. We have made the following revisions:

Notation and Rank Definition. Section 2 now states that σ,τ∈C\sigma,\tau\in\mathcal{C} denote arbitrary cells (target/source depending on context), and Table A.1 (Appendix A) summarizes all notation. In Definition 1, nodes, edges, and faces have ranks 0, 1, and 2, respectively.

Definitions and Examples. We replaced the vague statement in Lines 257–258 with formal definitions: τ≺σ\tau \prec \sigma iff τ⊊σ\tau\subsetneq\sigma and rk(σ)=rk(τ)+1\mathrm{rk}(\sigma)=\mathrm{rk}(\tau)+1 (Definition 6), with the boundary ∂σ\partial\sigma (Definition 7) and coboundary co ∂σ\mathrm{co}\,\partial\sigma (Definition 8) collecting immediate lower- and higher-rank neighbors. Figure A.1 gives a rank-2 example (e.g., co ∂v1={e12,e13}\mathrm{co}\,\partial v_1=\{e_{12},e_{13}\}).

Mathematical Clarifications. Section 2 now defines NC\mathcal{N}_{\mathcal{C}} and N(σ)\mathcal{N}(\sigma) before Eq. (1) and ψN,k\psi_{\mathcal{N},k} immediately after it. The expression M(x)xM(x)x in line 241 denotes the product of the input-dependent mixing matrix M(x)M(x) and the input sequence xx. Proposition 1 now makes this multiplication explicit.

Large-scale. Appendix D.2 now specifies that we use large-scale for networks with at least (10$^5$) nodes. We have also corrected citation formatting and typographical errors throughout the manuscript.

W2:..., why is rank awareness necessary in the first place? ..., what is the additional novelty the proposed method introduces beyond the existing representation. It is not clear how the proposed approach leverages rank information in a different way...

(1)Why is rank awareness necessary? A combinatorial complex (CC) encodes cell ranks and incidence relations, but a conventional SSM operating on a linearized sequence has no built-in mechanism for distinguishing messages from different ranks. CCMamba addresses this limitation by constructing rank-specific representations that distinguish a cell’s own features from information received from lower- and higher-rank cells (Eq. (4) and Fig. 2). Information propagates directly between adjacent ranks through incidence relations. CCMamba-S applies rank-specific bidirectional selective SSMs to the resulting representations, while CCMamba-K uses incidence-supported kernel interactions.

(2) Empirical Evidence. Figure 5 (b) evaluates different input-rank configurations, while Appendix E.9 (Table E.12) further isolates rank-specific encoding (it shares the input encoder across ranks while retaining relation-specific mixing), inter-rank propagation, incidence connectivity, and rank-conditioned mixing.

ModelVariantMUTAGPROTEINSPubMed
CCMamba-SFull (matched)85.11±2.4078.14±1.6889.29 ±1.47
Rank-Agnostic (shared encoder)82.34±1.1775.91±1.6785.81±2.85
w/o Inter-Rank76.17±1.7876.49±0.9085.81±2.85
w/o Cross-Rank (rewired)79.15±2.7776.56±0.8687.44±2.99
Rank-feature84.79±6.2876.27±0.9388.07±1.16
Incidence-only84.34±1.1777.20±1.2088.44±3.91
ModelVariantMUTAGPROTEINSPubMed
CCMamba-KFull (matched)87.23±5.7578.49±2.7688.66 ±0.63
Rank-Agnostic (shared encoder)85.06±1.7875.20±1.1285.05 ±1.03
w/o Inter-Rank78.72±4.7675.70±1.4285.52 ±1.10
w/o Cross-Rank (rewired)77.87±7.6174.62±2.5486.07±1.05
Rank-feature86.92±0.9576.85±1.2887.99±1.79
Incidence-only85.66±8.5675.12±2.0886.06±2.14

On MUTAG, compared to ablation experiments involving other rank-based features, the full results for CCMamba-K validate the benefits of the rank-aware computational approach.

(3)Methodological Novelty and Title. Our contribution lies in how rank information enters the mixing operator. CCMamba-S applies rank-specific selective SSM recurrence to incidence-lifted representations, while CCMamba-K incorporates typed incidence support directly into an index-free, structurally gated kernel operator. TopoMamba (Montagna et al., 2025) pools incident simplices by rank to construct a sequence for each node. TopoTune (Papillon et al., 2025) provides a general framework for relational aggregation with configurable backbones. CCMamba specifies how cells at each rank receive and mix information from incident cells at adjacent ranks. The rank and incidence ablations (Table E.12) and the mixer comparisons in Table 4 provide empirical evidence for this design. We have clarified these distinctions in Appendix B (Table B.1). We retain “Rank-Aware” in the title because rank directly shapes the mixing operation in both variants.

W3: The order-insensitive claim would benefit from further validation by defining this property and testing the model with different orderings of the same structure to demonstrate that it is independent on the choice of ordering.

We thank the reviewer for this valuable suggestion.

(1)Formal Definition. In CCMamba-K, order insensitivity refers to permutation equivariance of cell-level representations under rank- and incidence-preserving cell relabelings. We have added this definition to Appendix C.

Definition 10 (Permutation Equivariance): Let Ik∈Rnk−1×nkI_k \in \mathbb{R}^{n_{k-1}\times n_k} be the incidence matrix and Hk∈Rnk×dH_k \in \mathbb{R}^{n_k\times d} be the rank-kk feature matrix. Under arbitrary cell permutations Pk−1P_{k-1} and PkP_k, let Ik′=Pk−1IkPk⊤I_k'=P_{k-1}I_kP_k^\top and Hk′=PkHkH_k'=P_kH_k. A mapping Φ\Phi is permutation equivariant if Φ(Ik′,Hk′)=Pk−1Φ(Ik,Hk)\Phi(I_k',H_k')=P_{k-1}\Phi(I_k,H_k).

(2)Sensitivity to cell ordering. Our original manuscript included an ordering sensitivity study of CCMamba-S in Appendix E.3 (Table E.6), considering Canonical, BFS, Degree, Random, and Feature-norm orderings, which revealed its sensitivity to input ordering. We further evaluate CCMamba-K(cellular) under the same five strategies on Cora and CiteSeer.

OrderingCoraCiteSeer
Canonical88.65±0.7276.40±1.45
BFS88.68±0.7676.30±1.43
Degree88.68±0.7676.30±1.35
Random88.60±0.8376.35±1.47
Feature norm88.65±0.7176.35±1.40

Across the five orderings, mean accuracy varied by only 0.08 percentage points on Cora and 0.10 percentage points on CiteSeer.

(3)Validation of Permutation Equivariance. We further evaluate rank-wise representation equivariance. Section 5.2 (Table 5) reports rank-wise equivariance errors and graph-level prediction gaps on MUTAG under 100 rank-preserving permutations. Smaller equivariance errors indicate better consistency with the expected permutation transformation (Eq. 11).

ModelRank-0$\epsilon_{\mathrm{eq}}$Rank-1\epsilon_{\mathrm{eq}}\Prediction gap
GAT(Veličković et al., 2018)(1.92±0.66)e−7(1.92±0.66)e{-7}—(3.71±2.51)e−7(3.71±2.51)e{-7}
GSSC(Huangetal.,2024)(4.07±0.24)e−6(4.07±0.24)e{-6}(1.47±0.34)e−6(1.47±0.34)e{-6}(5.11±0.42)e−6(5.11±0.42)e{-6}
TopoTune(Papillonetal.,2025)(3.46±1.80)e−6(3.46±1.80)e{-6}0.00±0.000.00±0.00(4.61±3.27)e−6(4.61±3.27)e{-6}
Graph-Mamba(Wangetal.,2024)1.07±0.181.07±0.180.58±0.160.58±0.161.10±0.901.10±0.90
TopoMamba(2.61±1.16)e−8(2.61±1.16)e{-8}(2.69±0.94)e−8(2.69±0.94)e{-8}(6.00±5.73)e−8(6.00±5.73)e{-8}
CCMamba-S1.22±0.161.22±0.160.80±0.130.80±0.133.59±0.943.59±0.94
CCMamba-K(1.24±0.14)e−6(1.24±0.14)e{-6}(8.70±0.11)e−6(8.70±0.11)e{-6}(2.01±0.18)e−6(2.01±0.18)e{-6}

CCMamba-K exhibits small numerical errors at both ranks, consistent with the equivariance analysis in Appendix C.3. CCMamba-S and Graph-Mamba show substantially greater sensitivity to the same permutations.

W4: The evaluation ..., does not consider link prediction as graph has variety of tasks (Section 5.1) and having link prediction task would highlight the capability and limitation of the proposed approach. ..., Cellular complexes are missing baselines from Table 1. Having those baselines will help make the manuscript stronger.

We thank the reviewer for these suggestions. In addition to node-level regression on PM25CC (Table 3), we have added link prediction and hyperedge prediction experiments (Table E.1-E.2). Appendix D.4 details the data splits, negative sampling, decoders, and training protocols for both prediction tasks.

(1)Graph Link Prediction:

MethodCiteSeer(AUC)CiteSeer(AP)Minesweeper(AUC)Minesweeper(AP)
GCN75.50±0.7476.81±0.9179.93±0.1077.35±0.16
GAT75.81±1.5476.39±0.9676.78±0.6675.02±0.61
GIN69.60±2.3773.28±1.7967.58±8.5165.95±7.41
GraphSAGE73.96±2.0074.70±1.4279.78±0.1378.14±0.20
SGC70.94±7.1474.44±4.5775.26±0.3972.78±0.42
Graph-Mamba77.12±2.0176.99±1.5676.62±0.7174.99±0.64
GraphGPS81.57±0.6876.98±0.7369.47±0.5264.71±0.66
MaskGAE73.59±1.1975.21±1.3878.75±0.3276.09±0.35
CCMamba-S83.55±0.6984.41±0.9485.06±6.9984.16±6.04
CCMamba-K75.71±1.4977.04±1.6086.29±0.7285.85±0.81

(2)Hyperedges prediction:

MethodGerman (AUC)German (AP)Citation-Cora (AUC)Citation-Cora (AP)
HGNN71.03±1.7069.54±1.5163.15±1.5462.27±0.84
HyperGCN71.93±1.7068.66±1.5764.17±1.6763.13±2.73
HNHN70.42±2.4669.00±2.5759.32±1.3059.46±1.80
HCHA71.34±3.8569.38±3.3664.19±1.6264.15±1.14
UniGCNII50.95±4.5550.80±2.9762.21±2.5762.65±3.44
AllSetTransformer67.31±3.0264.24±2.8160.14±1.8359.37±2.83
EDGNN68.25±1.4765.87±0.7661.32±1.3060.51±1.06
HyperGT63.00±5.7061.38±4.9553.74±0.8253.50±1.07
CCMamba-S (hypergraph)73.05±2.0677.01±2.6067.43±1.8166.70±1.43
CCMamba-K (hypergraph)71.96±2.2774.21±2.9367.07±2.7666.13±1.62

(3)Cellular-Complex Baselines: We added CCIN to Table 1 alongside CCNN, CCXN, CWN, and TopoTune. Table 3 provides additional comparisons on native higher-order benchmarks.

W5: Results are close to the baselines especially when considering variance. This questions the empirical contribution of new method. Given the small performance differences along with reported variance, it would be useful to conduct statistical significance tests.

We thank the reviewer for this important observation. We agree that several performance differences are modest relative to the reported variance. To assess their statistical reliability, we have supplemented Tables 1 and 2 with statistical significance tests (Appendix E.8,Tables E.10–E.11).

(1)Statistical Significance. Appendix E.8 (Tables E.10–E.11) reports raw and Holm-adjusted p-values for comparisons between the two CCMamba variants and selected baselines across four structural domains. Representative results from the supplementary tests are shown below.

Domain/DatasetBaselineCCMambaGain (pp)Adjusted pSig.
Graph/IMDB-BGraphGPS: 72.35±2.60CCMamba-S: 73.20±3.56+0.850.0340* ↑
Graph/MUTAGGraphGPS: 78.85±3.31CCMamba-S: 79.75±2.40+0.901.0000ns
Hypergraph/20NewsW100HJRL: 80.88±0.22CCMamba-S: 81.47±0.31+0.590.0022** ↑
Hypergraph/20NewsW100HJRL: 80.88±0.22CCMamba-K: 82.09±0.63+1.210.0198* ↑

These comparisons indicate that some modest improvements receive statistical support, whereas others, such as MUTAG, do not. We therefore acknowledge that CCMamba does not consistently achieve statistically significant improvements across the evaluated benchmarks. Non-significant results are not interpreted as evidence of equivalence.

(2)Beyond Predictive Accuracy. We also evaluate the contribution of rank-aware mixing through controlled ablations (Appendix E.9), measure permutation equivariance under cell relabeling (Table 5), and compare training time and peak GPU memory (Section 5.4). These analyses assess the structural and computational properties of the two variants alongside predictive performance.

We have revised the experimental discussion to distinguish statistically supported improvements from non-significant numerical differences and to avoid overstating performance gains on established benchmarks.

W6:formatting issues.

We thank the reviewer for identifying these formatting issues. For Table 1, we adjusted column widths, font sizes, and spacing to address the margin and readability concerns. For Table 2, we moved the Mushroom, Bail, and Actor results to Appendix E.1 (Table E.3) to reduce its width. We also adjusted the placement and spacing of Table 3 to reduce crowding.

W7:The paper also lacks a conclusion section.

We thank the reviewer for this suggestion. We have retitled Section 6 “Conclusion” and revised it to summarize the main findings, the trade-off between selective scanning in CCMamba-S and permutation-equivariant mixing in CCMamba-K, and directions for future work.

We hope this response alleviates your concerns. If our clarifications address your question, we would sincerely appreciate your reconsideration of the score. We welcome any further questions and discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

We sincerely thank all three reviewers (DG9s, Rxdi, and NyYi) for their careful evaluations and constructive feedback. To facilitate the ongoing discussion, we summarize the main revisions and supporting evidence.

Methodological Novelty and Distinction from Existing Methods

  • Methodological comparison (DG9s, Rxdi, NyYi). Appendix B (Table B.1) provides a systematic design-space comparison with Graph Mamba, TopoMamba, GSSC, and TopoTune. CCMamba-S retains rank-specific selective SSM recurrence, while CCMamba-K combines typed incidence support, learned structural gating, and factorized content interactions.
  • Direct comparison (NyYi). We have additionally implemented GSSC-TopoTune to evaluate the integration of existing graph kernel mixing with generalized CC architectures. CCMamba-K achieves higher mean performance, as reported in our response.
  • Matched mixer ablation (Rxdi, NyYi). The existing mixer comparison has been moved from the appendix to the main text (Table 4), comparing CCMamba-K against four alternative mixers under shared structural support.

Permutation Equivariance, Long-Range Modeling, and Theoretical Expressivity

  • Permutation equivariance (DG9s). Appendix C establishes the permutation equivariance and exact support-weighted kernel factorization of CCMamba-K under the stated conditions. Table 5, previously in the appendix, evaluates 100 cell permutations on MUTAG. CCMamba-K exhibits equivariance errors on the order of 10−610^{-6}, whereas CCMamba-S remains order-sensitive.
  • Long-range modeling (NyYi). CCMamba-S supports rank-wise, sequence-global propagation through selective recurrence, whereas CCMamba-K performs incidence-local mixing, with broader propagation obtained through stacking. Appendix E.6 examines depth-dependent performance across 2–64 layers.
  • Expressivity (NyYi). Theorem 1 establishes that idealized CCMamba-K matches 1-CCWL distinguishing power under explicit assumptions. Table C.1 contrasts these assumptions with the practical model, which does not guarantee them. Thus, the theoretical result does not imply unconditional 1-CCWL matching for practical CCMamba-K.

Rank-Aware Modeling and Architectural Design Choices

  • Rank and incidence ablations (DG9s). We have added controlled ablations (Appendix E.9, Table E.12) examining rank-specific encoding, cross-rank propagation, and incidence connectivity. Removing inter-rank interactions or perturbing incidence relations generally reduces predictive accuracy.
  • Ordering analysis (Rxdi). Appendix E.3 examines different ordering strategies and explains why identical structural signatures limit rank-aware ordering on simplicial complexes.
  • Structural gating (Rxdi). The additional controlled comparison (Appendix E.10, Table E.13) shows that target-conditioned gating achieves higher mean accuracy than joint source-target gating across five datasets.

Broader Empirical Evaluation and Experimental Protocols

  • Additional tasks (DG9s). We have extended the evaluation to graph link prediction and hyperedge prediction, reporting AUC and AP across multiple datasets (Appendix E.1, Tables E.1–E.2).
  • Cellular-complex baselines (DG9s). Table 1 now includes CCIN alongside the existing CCNN, CCXN, CWN, and TopoTune baselines.
  • Statistical analysis (DG9s, NyYi). Appendix E.8 (Tables E.10–E.11) reports statistical comparisons with Holm correction, distinguishing significant improvements and degradations from nonsignificant differences.
  • Reproducibility (NyYi). We have standardized the training seeds to {1,3,5,7,9}\{1,3,5,7,9\}, clarified dataset-split protocols, and aligned the default lifting configurations: KHop for hypergraphs, Clique for simplicial complexes, and Cycle for cellular complexes.

Mathematical Clarity and Manuscript Presentation

  • Notation and definitions (DG9s). Section 2 and Appendix A provide formal definitions and examples of cell ranks, face and coface relations, incidence operators, and input-dependent SSM mixing, supported by the notation table (Table A.1).
  • Presentation (DG9s, Rxdi, NyYi). Tables 1–3 feature revised layouts and formatting, with selected results relocated to the appendix; figure labels have also been adjusted.
  • Conclusion (DG9s). Section 6 now summarizes the main findings, methodological trade-offs, limitations, and future directions.

We hope this summary provides a useful overview of the revisions and supporting evidence. We appreciate the reviewers' time and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.