acceptodds
Under review as a conference paper at ICLR 2027

OmniSheaf: Unifying Sheaf Hypergraph Networks via Efficient Universal Approximators

Abstract

Higher-order relations among multiple entities frequently appear in fields such as neuroscience, radio allocation and computer networks. These relations can be modeled via hypergraphs, i.e., topological structures that extend traditional graphs via hyperedges connecting multiple nodes at once. Recent work has introduced Sheaf Hypergraph Networks (SHNs), which further augment this topological structure via a cellular sheaf. The latter attaches d-dimensional vector spaces named stalks to each node and hyperedge, as well as learnable linear transformations called restriction maps between them. The key limitation of current sheaf methods for hypergraphs is that they compute the graph clique, which scales quadratically with the number of elements of each hyperedge. Ultimately, this causes loss of valuable topological information and makes computation prohibitive. In this paper, we introduce OmniSheaf (OS), a new Sheaf Neural Network (SNN) that for the first time works directly over hypergraphs via multiset functions over stalk spaces. We mathematically prove that OS generalizes SNNs, SHNs, AllSet, DHNN, and Dir-GNN via its universal approximators and that it can be made directionality-aware over hyperedges via both a complex phase and tail-head indices. Extensive experiments across 25 benchmarks against hypergraph networks, set-based architectures, and sheaf models demonstrate that OS achieves up to +14.37 higher accuracy on undirected node classification benchmarks (on Senate, while reaching 95.58% on Congress) and +4.11 on directed ones, while on Roman-Empire it trains 7.1× faster than DSHNLight, uses 3.6× less peak GPU memory and 14.9× fewer training FLOPs at equal stalk dimension.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Deep Learning, Cellular Sheaves, Hypergraph Neural Networks, Sheaf Neural Networks

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

This paper introduces OmniSheaf, a framework for sheaf learning on hypergraphs through two-stage node-to-hyperedge and hyperedge-to-node aggregation. Learnable restriction maps align stalk features before multiset aggregation, avoiding explicit clique expansion. The authors develop DeepSets and Transformer instantiations, along with directed variants based on complex phases or separate tail/head processing. They present reductions to existing hypergraph and sheaf architectures and universal approximation results for suitable parameterizations. Experiments span 25 benchmarks, with notable gains on committee, directed, and molecular tasks.

Critical strengths

S1. A unified framework with practical computational benefits. OmniSheaf combines sheaf transport with incidence-based multiset aggregation, connecting AllSet, sheaf diffusion, and directed architectures. It preserves hyperedge identities and avoids explicit clique expansion, offering both conceptual value and potential savings on large hypergraphs.

S2. Broad evaluation with substantial gains. Experiments cover node classification, directionality tests, and molecular hyperedge classification, with variability reported over ten runs. Gains of 14.37 points on Senate and 4.11 on Roman-empire, alongside resource savings over DSHN at large scale, support the framework's practical value. The paper also acknowledges cases where simpler baselines perform better or cost less.

Critical weaknesses

W1. Linear complexity is not established for all instantiations. Theorem 4 counts incidence messages without accounting for aggregator costs. Yet Theorem 6 recovers SheafHyperGCN through pairwise distances within hyperedges, potentially requiring quadratic work. Please identify which instantiations remain linear and include stalk dimension, feature width, and aggregator size in the bound to clarify the scope of the efficiency claim.

W2. The headline speedup uses the wrong baseline. The abstract claims a 7.1x speedup over DSHNLight, but Table 3 supports this figure against DSHN. At d = 2, ODS is only about 1.12x faster than DSHNLight and uses more memory (5.20 vs. 2.99 GB). Since the best accuracy comes from costlier directed variants, please correct the attribution and compare accuracy and cost for the same configurations.

W3. The contribution of learned sheaf transport remains unclear. Comparisons with AllSet do not isolate gains from sheaf transport versus added capacity or other architectural changes. Please compare learned and fixed restriction maps under matched parameter and training budgets to establish whether stalk alignment drives the improvements.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer WpaK

W1. Linear Complexity

Comment: Theorem 4 counts incidence messages without accounting for aggregator costs. Yet Theorem 6 recovers SheafHyperGCN through pairwise distances within hyperedges, potentially requiring quadratic work. Please identify which instantiations remain linear and include stalk dimension, feature width, and aggregator size in the bound to clarify the scope of the efficiency claim.

Response: Thank you for pointing this out. The reviewer is right that the proof of Theorem 4 counts incidence messages: it states that each stage of OmniSheaf exchanges exactly K=∑e∈E∣e∣K = \sum_{e \in E} \lvert e \rvert messages, against O(∑e∣e∣2)O(\sum_{e} \lvert e \rvert^2) for clique-based implementations, and does not include the cost of computing each message. We will make this scope explicit and add the following remark after Theorem 4.

Remark (cost per message). With stalk dimension dd, feature width ff and hidden widths O(f)O(f), one layer costs O(K⋅(cmap+cagg))O\big(K \cdot (c_{\mathrm{map}} + c_{\mathrm{agg}})\big) time and O(Kdf)O(K d f) memory. Here cmap=O(f2+df)c_{\mathrm{map}} = O(f^2 + df) for Diag and O(f2+d2f)O(f^2 + d^2 f) for General and Ortho maps is the cost of predicting and applying one restriction map, and caggc_{\mathrm{agg}} is the cost of the multiset functions per element: O(df2)O(d f^2) for row-wise DeepSets and O(d2f2)O(d^2 f^2) for PMA, where the number of seeds and heads adds a constant factor.

After factoring all the above, the cost remains linear in KK whenever the multiset functions process each element at a cost independent of ∣e∣\lvert e \rvert. This holds for all instantiations used in our experiments (ODS, OST, OCS, OCS-T, OPS and OPS-T), which pool or attend once over the members of each hyperedge, and for the reductions to AllSet, SheafHyperGNN, sheaf networks on graphs and DSHN (Theorems 5, 6, 7 and 11). The only exception is the reduction to SheafHyperGCN (Theorem 6). As stated in its proof, fV→Ef_{V\to E} evaluates all pairwise distances within Ve\mathcal{V}_e, which costs O(∣e∣2df)O(\lvert e \rvert^2 d f) per hyperedge. [If the maximal pair is selected after a random scalar projection, as in HyperGCN, this cost drops to O(∣e∣df)O(\lvert e \rvert d f).] We will state this exception by adding this remark and we will clarify Theorem 6 accordingly.

W2. Headline Speedup Claim

Comment: The abstract claims a 7.1x speedup over DSHNLight, but Table 3 supports this figure against DSHN. At d = 2, ODS is only about 1.12x faster than DSHNLight and uses more memory (5.20 vs. 2.99 GB). Since the best accuracy comes from costlier directed variants, please correct the attribution and compare accuracy and cost for the same configurations.

Response: We agree with the reviewer that the speedup claim is imprecise, we will clarify the attribution and compare accuracy and cost for the same configurations throughout the paper in the final version of the manuscript.

W3. Identity & Fixed Random Restriction Maps

Comment: Please compare learned and fixed restriction maps under matched parameter and training budgets to establish whether stalk alignment drives the improvements.

Response: Thank you for the suggestion. We ran this ablation with matched parameter counts and training budgets, comparing the learned maps of our main results with identity and fixed random orthogonal maps.

Table R2. Test accuracy (%). The “Learned” column shows the result reported in our Tables 1, 4 and 10. The identity and fixed random orthogonal variants keep the configuration of the learned model. Δ is relative to learned maps, while † marks learned maps significantly better than that variant (two-sided Welch t-test over 10 runs, Holm-Bonferroni over the 20 comparisons, p < 0.05).

DatasetModelLearnedIdentityIdentity: ΔRandom orth.Random orth.: Δ
Roman-empireODS85.20 ± 0.7880.46 ± 0.71−4.74†72.97 ± 0.68−12.23†
Roman-empireOCS-T93.35 ± 0.2782.66 ± 1.05−10.69†25.03 ± 0.71−68.32†
Molecular-1ODS79.44 ± 0.4675.03 ± 0.78−4.41†66.05 ± 1.22−13.39†
Molecular-1OCS-T90.40 ± 0.1186.90 ± 0.38−3.51†50.19 ± 1.56−40.21†
email-EUODS64.10 ± 2.0244.86 ± 12.18−19.24†58.58 ± 3.89−5.51†
email-EUOCS-T76.65 ± 2.2572.96 ± 4.09−3.69†26.44 ± 2.87−50.21†
SenateODS83.10 ± 1.9972.54 ± 2.97−10.56†77.32 ± 3.80−5.77†
SenateOST81.69 ± 2.3068.73 ± 6.84−12.96†62.82 ± 7.56−18.87†
HouseODS78.33 ± 1.2672.63 ± 2.21−5.70†75.08 ± 2.85−3.25†
HouseOST78.43 ± 1.1772.66 ± 1.74−5.77†73.68 ± 2.32−4.75†

Learned maps outperform both alternatives in all 10 settings, by up to 19.2 points against identity and 68.3 points against fixed random maps, and all 20 differences are significant (two-sided Welch t-test, Holm-Bonferroni, p < 0.05).

We will add this in the final version of the manuscript.

Thank you for the authors' response. I appreciate the clarifications and encourage the authors to incorporate the relevant revisions into the revised manuscript. I will maintain my current score.

Reviewer_vNjnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The paper proposes OmniSheaf, a framework combining cellular sheaves with multiset-based message passing over hypergraphs. It introduces DeepSets and Transformer implementations, alongside directional variants based on complex phases and head-tail partitions.

Critical strengths

The paper addresses the computational efficiency of sheaf-based learning on hypergraphs. Combining restriction maps with incidence-based message passing is interesting, particularly for large hypergraphs.

Critical weaknesses

My main concern is that I could not clearly understand what is fundamentally new about OmniSheaf compared to existing hypergraph and sheaf architectures.

The paper repeatedly motivates its approach through clique expansion, but this is only one way of representing hypergraph interactions through a graph, rather than an intrinsic requirement of hypergraphs or cellular sheaves. From my understanding, the method essentially combines AllSet-style multiset aggregation with learnable sheaf restriction maps. Moreover, the linear sheaf diffusion operator in Theorem 6 can already be expressed through hyperedge aggregation, making the actual novelty unclear.

Another related concern is the choice of hypergraph construction. Several benchmarks are originally graphs converted into hypergraphs using a predefined lifting. However, many other constructions are possible, such as clique-based, cycle-based, k-hop, or even learnable liftings. The paper does not investigate how these choices affect the results, making it difficult to distinguish improvements from the proposed architecture and those arising from the constructed higher-order domain.

Questions for authors
  1. What is the main novel contribution beyond combining existing multiset aggregation with learnable sheaf restriction maps? How does this differ from an efficient implementation of existing sheaf operators?
  2. Why was the particular graph-to-hypergraph construction chosen? How would alternative or learnable liftings affect the results, especially on datasets originally represented as graphs?
9 comments
AuthorsOpenReview ↗

Response to Reviewer vNjn, Question 1, (Part 1/2)

Q1/W1. Clarifications on Novelty (1/2)

Comment: What is the main novel contribution beyond combining existing multiset aggregation with learnable sheaf restriction maps? How does this differ from an efficient implementation of existing sheaf operators?

Response: Thank you for giving us the opportunity to clarify this point. While we agree that the clique expansion is a property of existing implementations, not an intrinsic requirement of hypergraphs or cellular sheaves, we point out that while prior sheaf networks [1,2] learn the sheaf but propagate with a fixed normalized sheaf Laplacian, OmniSheaf makes the operator itself learnable, defining such operator directly on the stalk spaces as a pair of multiset functions (Definition 1), of which the sheaf Laplacians of prior work are particular choices (Theorems 6 and 11).

Specifically, while AllSet [4] learns multiset functions over hypergraphs, it does so on scalar features and with no transport between stalks. Adding restriction maps to AllSet recovers sheaf diffusion only if the backward stage can also normalize each node by its sheaf degree, which depends on the restriction maps of all its incident hyperedges. Definition 1 provides the sheaf degree locally by pairing each transported message with its transport block in the multiset Eu\mathcal{E}_u of Eq. (1), which is what allows OmniSheaf to recover both AllSet (Theorem 5) and the sheaf hypergraph networks (Theorems 6 and 11). As such, while [1,2] learn the restriction maps, they apply a fixed linear diffusion, which is the instance of OmniSheaf with mean pooling and a fixed linear backward function. OmniSheaf learns both the restriction maps and the multiset functions over stalk spaces, and its instances are universal approximators of these functions (Theorem 8 and Corollary 9).

We performed two additional controlled experiments to separate OmniSheaf from each of the two components mentioned by the reviewer:

  • Multiset aggregation without sheaf transport. Replacing the learned restriction maps with the identity (AllSet-style aggregation over stalks), under matched parameter and training budgets, significantly lowers accuracy in all 10 settings, by up to 19.2 points (Table R2 below).

Table R2. Test accuracy (%). The “Learned” column shows the result reported in our Tables 1, 4 and 10. The identity and fixed random orthogonal variants keep the configuration of the learned model. Δ is relative to learned maps, while † marks learned maps significantly better than that variant (two-sided Welch t-test over 10 runs, Holm-Bonferroni over the 20 comparisons, p < 0.05).

DatasetModelLearnedIdentityIdentity: ΔRandom orth.Random orth.: Δ
Roman-empireODS85.20 ± 0.7880.46 ± 0.71−4.74†72.97 ± 0.68−12.23†
Roman-empireOCS-T93.35 ± 0.2782.66 ± 1.05−10.69†25.03 ± 0.71−68.32†
Molecular-1ODS79.44 ± 0.4675.03 ± 0.78−4.41†66.05 ± 1.22−13.39†
Molecular-1OCS-T90.40 ± 0.1186.90 ± 0.38−3.51†50.19 ± 1.56−40.21†
email-EUODS64.10 ± 2.0244.86 ± 12.18−19.24†58.58 ± 3.89−5.51†
email-EUOCS-T76.65 ± 2.2572.96 ± 4.09−3.69†26.44 ± 2.87−50.21†
SenateODS83.10 ± 1.9972.54 ± 2.97−10.56†77.32 ± 3.80−5.77†
SenateOST81.69 ± 2.3068.73 ± 6.84−12.96†62.82 ± 7.56−18.87†
HouseODS78.33 ± 1.2672.63 ± 2.21−5.70†75.08 ± 2.85−3.25†
HouseOST78.43 ± 1.1772.66 ± 1.74−5.77†73.68 ± 2.32−4.75†

continues below*

AuthorsOpenReview ↗

Response to Reviewer vNjn, Question 1, (Part 2/2)

Q1/W1. Clarifications on Novelty (2/2)

  • Efficient implementation of an existing sheaf operator. OS-Linear, the DSHN operator evaluated through incidences with the same restriction maps and tuning budget, is significantly outperformed by learned multiset functions on email-EU, Molecular-1, Roman-empire and Squirrel, by 2.7 to 9.2 points (Table R1 below).

Table R1. Test accuracy (%; F1 for Molecular-1). DSHN and DSHNLight are taken from [1]; ODS and Best OS are the results reported in our Tables 1, 7, 8 and 10, under the same experimental setup as OS-Linear. Best OS is the best OmniSheaf variant in Table 1. All tests are two-sided Welch t-tests over 10 runs with Holm-Bonferroni correction within each family (p < 0.05). † marks results significantly better than the best published baseline (family: the 10 comparisons of OS-Linear and Best OS). Δ is Best OS minus OS-Linear; * marks a significant Δ (family: the 5 datasets).

DatasetDSHNDSHNLightOS-LinearODSBest OSΔ
email-EU78.62 ± 2.5082.67 ± 1.2977.61 ± 3.4164.10 ± 2.0285.56 ± 0.83† (OPS-T)+7.95*
Molecular-1OOM82.32 ± 0.5681.95 ± 0.5579.44 ± 0.4691.11 ± 0.30† (OPS-T)+9.16*
Roman-empireOOM89.24 ± 0.5789.26 ± 0.7285.20 ± 0.7893.35 ± 0.27† (OCS-T)+4.09*
Telegram88.65 ± 5.5481.15 ± 4.1987.88 ± 3.7566.67 ± 4.8089.10 ± 4.80 (OCS-T)+1.22
Squirrel43.55 ± 2.8744.09 ± 2.3642.62 ± 1.8542.68 ± 1.6445.35 ± 0.72 (OCS-T)+2.73*

In the final version, we will rewrite the Introduction and the list of contributions to stress this distinction.

[1] E. Mule et al., "Directional Sheaf Hypergraph Networks", ICLR 2026.

[2] I. Duta et al., "Sheaf Hypergraph Networks", NeurIPS 2023.

[3] S. Fiorini et al., "Sheaves Reloaded: A Direction Awakening", ICLR 2026.

[4] E. Chien et al., "You are AllSet: A Multiset Function Framework for Hypergraph Neural Networks", ICLR 2022.

AuthorsOpenReview ↗

Response to Reviewer vNjn, Question 2

Q2/W2. Graph-to-Hypergraph Construction

Comment: Why was the particular graph-to-hypergraph construction chosen?

Response: Thank you for giving us the opportunity to clarify this point. We chose this construction to compare with prior work [1, 2]. More precisely, we adopted the same splits, datasets and construction as [1] in the directed case (Section 4.1) and [2] in the undirected one (Section 4.3). As such, we reported the same performance results claimed in their papers as baselines. Finally, we used the synthetic directional benchmarks of [3] to validate OmniSheaf's ability to preserve hyperedge orientation under controlled levels of inter-class connectivity (Section 4.2), using the same baselines and results reported in [1].

Comment: How would alternative or learnable liftings affect the results, especially on datasets originally represented as graphs?

Response: Thank you for your comment. To isolate the effect of the lifting, we rebuilt the hypergraphs of three graph datasets from their directed edges with two alternative liftings [5] and kept features, labels and splits unchanged. 2-uniform uses one hyperedge per directed edge. 2-hop uses tail vv with heads all nodes within two out-hops of vv. The default is the construction of [1] (tail vv, heads Nout(v)N_{\mathrm{out}}(v)). Every model follows the same tuning protocol under every lifting. The results are reported in Table R4.

Table R4. Test accuracy (%). The “Avg. size” column reports the mean hyperedge size. We denote in bold the best model per row, while † marks a best model significantly better than the runner-up (i.e., two-sided Welch t-test over 10 runs, Holm-Bonferroni over the 9 rows, p < 0.05).

DatasetLiftingAvg. sizeOCS-TODSDSHNLightGeDi
Roman-empiredefault2.7393.35 ± 0.27†85.20 ± 0.7889.24 ± 0.5783.87 ± 0.63
2-uniform2.0092.72 ± 0.38†78.43 ± 0.7686.73 ± 0.6479.63 ± 0.53
2-hop4.6588.44 ± 0.65†72.73 ± 0.9177.94 ± 0.3876.88 ± 0.50
Chameleondefault12.1145.60 ± 4.8345.02 ± 2.5846.50 ± 4.0939.29 ± 2.04
2-uniform2.0041.57 ± 2.7040.60 ± 3.8542.96 ± 4.5440.33 ± 2.79
2-hop46.7543.42 ± 5.4345.17 ± 4.9640.78 ± 2.5842.01 ± 2.45
Cora (directed)default4.4786.71 ± 0.6386.95 ± 0.5788.02 ± 1.1185.16 ± 0.94
2-uniform2.0085.42 ± 0.6684.58 ± 0.7480.13 ± 16.9176.54 ± 1.49
2-hop8.6484.19 ± 1.0685.88 ± 0.7986.01 ± 0.7181.86 ± 0.65

Table R4 shows that the conclusions of our evaluation do not depend on the choice of lifting. Specifically, on Roman-empire the ranking does not depend on the lifting: OCS-T is the best model under all three, by at least 4.1 points and always significantly. With the 2-uniform lifting it stays within 0.7 points of the default (92.72 vs 93.35, Welch p < 0.001), while its margin over the runner-up grows to 6.0 points, confirming that the advantage comes from the architecture and is not an artifact of higher-order construction. The 2-hop lifting, with larger and noisier hyperedges, degrades every model. On Chameleon and Cora no difference between the best model and the runner-up is significant under any lifting (adjusted p ≥ 0.093), so the choice of lifting does not change the conclusions there.

Learnable liftings [6] are a promising direction that we will discuss as future work. We will stress this aspect and add these results in the final version of the manuscript.

[1] E. Mule et al., "Directional Sheaf Hypergraph Networks", ICLR 2026.

[2] I. Duta et al., "Sheaf Hypergraph Networks", NeurIPS 2023.

[3] S. Fiorini et al., "Sheaves Reloaded: A Direction Awakening", ICLR 2026.

[4] E. Chien et al., "You are AllSet: A Multiset Function Framework for Hypergraph Neural Networks", ICLR 2022.

[5] Y. Gao et al., "HGNN+: General Hypergraph Neural Networks", IEEE TPAMI vol. 45, no. 3, pp. 3181-3199 (2023).

[6] G. Bernárdez et al., "ICML Topological Deep Learning Challenge 2024: Beyond the Graph Domain", PMLR 251, ICML GRaM Workshop (Proceedings Track), 2024.

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

In this paper, the authors introduce OmniSheaf, a unified framework for sheaf-based learning on directed and undirected hypergraphs that avoids explicitly constructing clique expansions. The proposed approach employs a two-stage message-passing mechanism through node–hyperedge incidences, using learnable restriction maps to project node representations into shared hyperedge stalk spaces and permutation-invariant multiset functions to aggregate information. The framework includes two architectural variants based on multilayer perceptrons and attention mechanisms, along with directional extensions incorporating complex phases and explicit tail–head partitions. The authors provide theoretical results on computational complexity, universal approximation, and the generalization of existing sheaf and hypergraph architectures. The proposed framework is evaluated on 25 benchmark graphs.

Critical strengths
  • The authors introduce a unified framework combining sheaf restriction maps with incidence-based multiset aggregation. The resulting architecture accommodates both directed and undirected hypergraphs and allows more general aggregation functions than conventional linear sheaf diffusion. The connection with existing sheaf and hypergraph architectures is particularly interesting, providing a useful perspective on how different message-passing mechanisms can be expressed within a common framework
  • The proposed architectures achieve strong empirical results across a diverse collection of benchmarks, with particularly significant improvements on Senate, Roman-empire, and molecular classification tasks. The evaluation covers both directed and undirected hypergraphs, demonstrating the flexibility of the proposed framework. The computational experiments also show meaningful efficiency improvements over existing implementations of clique-based sheaf networks on larger hypergraphs
Critical weaknesses
  • In the Introduction and Theorem 4, the authors identify the quadratic complexity of clique expansion as a limitation of existing sheaf hypergraph networks and propose incidence-based message passing as a solution. However, the DSHN operator introduced in [1] already admits an exact incidence-based factorization (Equations 2 and 7): QN=DV−1/2B†DE−1BDV−1/2,Q_N = D_V^{-1/2} B^\dagger D_E^{-1} B D_V^{-1/2}, where BB is the complex-valued node–hyperedge incidence matrix incorporating directional restriction maps, and DVD_V and DED_E are the corresponding degree matrices. For an input representation XX, the operator can be evaluated exactly as
H = D_E^{-1}BZ, \qquad Y = D_V^{-1/2}B^\dagger H = Q_NX.$$ This corresponds to two incidence-based aggregation steps, requiring $O(\sum_{e\in E}|e|)$ incidence operations rather than $O(\sum_{e\in E}|e|^2)$ for fixed stalk and feature dimensions, without explicitly constructing the pairwise propagation matrix. Importantly, this factorization preserves the original DSHN operator, including its directional information and learned restriction maps. Thus, the quadratic complexity attributed to DSHN arises from explicitly materializing the node-to-node propagation matrix rather than from an inherent limitation of its mathematical formulation. This raises concerns about the claimed computational novelty of Theorem 4. Moreover, the absence of a comparison with an incidence-factorized implementation of DSHN makes it difficult to assess whether the reported efficiency gains stem from the proposed architecture or from a more efficient implementation of an existing operator - Although Definition 1 introduces general permutation-invariant aggregation functions, the manuscript does not explicitly define their domains and codomains or provide sufficient details on how they are instantiated in OmniDeepSheaf and OmniSheafTransformer. In particular, the processing of transport matrices in the backward aggregation, the distinction between vectorized and row-wise implementations, and the specific attention mechanisms remain unclear. Furthermore, the initialization of hyperedge representations and their relationship with the features used to predict restriction maps in Section 3.3 are not adequately explained. Similar ambiguities arise in OmniPhaseSheaf, where the initialization of complex-valued hyperedge representations and the construction of attention queries are not specified. These omissions make it difficult to reproduce the proposed architectures and assess how their concrete implementations relate to the theoretical framework [1] Mule, Emanuele, et al. "Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs." International Conference on Learning Representations, 2026.
Questions for authors
  • The DSHN propagation operator [1] already admits an exact incidence-based factorization, allowing the same operator to be evaluated through two node–hyperedge aggregation steps without constructing the pairwise propagation matrix. Given this observation, could the authors clarify in what sense Theorem 4 establishes a novel computational advantage over existing sheaf hypergraph operators?
  • Could the authors provide a more explicit mathematical and architectural definition of the aggregation functions introduced in Definition 1? In particular, how are these functions instantiated in OmniDeepSheaf and OmniSheafTransformer, how are the transport matrices processed in the backward aggregation, and how are the initial hyperedge representations and attention queries in OmniPhaseSheaf constructed?
  • To what extent do the reported improvements depend specifically on learning the restriction maps? Could the authors provide an ablation comparing learned, identity, and fixed random restriction maps while keeping the aggregation functions and other architectural components unchanged?
  • Could the authors clarify the practical efficiency advantages of OmniSheaf over DSHNLight, particularly in terms of training time and GPU memory consumption? In Table 3, the improvements over DSHNLight appear relatively modest on Roman-empire, while OmniDeepSheaf requires more GPU memory at equal stalk dimensions. Additionally, could the authors clarify the discrepancy between the 7.1× speedup attributed to DSHNLight in the abstract and the comparison against DSHN reported in Table 3?

[1] Mule, Emanuele, et al. "Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs." International Conference on Learning Representations, 2026.

7 comments
AuthorsOpenReview ↗

Response to Reviewer oiFd, Question 1

Q1/W1. DSHN Propagation Operator

Comment: Could the authors please clarify in what sense Theorem 4 established a novel computational advantage over existing sheaf hypergraph operators?

Response: We thank you for giving us the opportunity to clarify this point. As discussed in Section 2 (paragraph "Hypergraph Sheaves"), the reported quadratic cost comes from the available implementations of DSHN [1] and SHN [2], which build the graph clique before propagation. As pointed out by the reviewer, the DSHN operator itself admits an exact incidence factorization, and our Appendix B.5 uses it to validate Theorem 11. Thus, Theorem 4 describes an advantage of OmniSheaf in the number of exchanged messages, i.e., O(∑e∣e∣)O(\sum_{e} \lvert e \rvert) instead of O(∑e∣e∣2)O(\sum_{e} \lvert e \rvert^2), over these implementations.

The fundamental difference between OmniSheaf and DSHN lies in which sheaf diffusion operator is computed. Specifically, DSHN implements one fixed choice (Theorem 11) of the two multiset functions in Eq. (1) of Definition 1, and we prove in Theorem 11 that it is a special case of OmniSheaf. In contrast, OmniSheaf learns these functions with universal approximation guarantees (Theorem 8 and Corollary 9), while retaining the same linear cost as the factorized DSHN operator.

Comment: [...] the absence of a comparison with an incidence-factorized implementation of DSHN makes it difficult to assess whether the reported efficiency gains stem from the proposed architecture or from a more efficient implementation of an existing operator

Response: Thank you for the suggestion. To further remark the difference between DSHN and OmniSheaf, we trained the special case of Theorem 11 itself as a model (which we call OS-Linear). It evaluates DV−1/2B†DE−1BDV−1/2D_V^{-1/2} B^\dagger D_E^{-1} B D_V^{-1/2} through incidences, with the complex phase on directed datasets. It uses the same restriction-map predictor, classifier, input projection and residuals as OmniSheaf, DSHN's ReLU and bias between layers, and the same splits, seeds, search space and trial budget, extended with DSHN's normalization choices. As such, the two models differ only in the multiset functions.

Table R1. Test accuracy (%; F1 for Molecular-1). DSHN and DSHNLight are taken from [1]; ODS and Best OS are the results reported in our Tables 1, 7, 8 and 10, under the same experimental setup as OS-Linear. Best OS is the best OmniSheaf variant in Table 1. All tests are two-sided Welch t-tests over 10 runs with Holm-Bonferroni correction within each family (p < 0.05). † marks results significantly better than the best published baseline (family: the 10 comparisons of OS-Linear and Best OS). Δ is Best OS minus OS-Linear; * marks a significant Δ (family: the 5 datasets).

DatasetDSHNDSHNLightOS-LinearODSBest OSΔ
email-EU78.62 ± 2.5082.67 ± 1.2977.61 ± 3.4164.10 ± 2.0285.56 ± 0.83† (OPS-T)+7.95*
Molecular-1OOM82.32 ± 0.5681.95 ± 0.5579.44 ± 0.4691.11 ± 0.30† (OPS-T)+9.16*
Roman-empireOOM89.24 ± 0.5789.26 ± 0.7285.20 ± 0.7893.35 ± 0.27† (OCS-T)+4.09*
Telegram88.65 ± 5.5481.15 ± 4.1987.88 ± 3.7566.67 ± 4.8089.10 ± 4.80 (OCS-T)+1.22
Squirrel43.55 ± 2.8744.09 ± 2.3642.62 ± 1.8542.68 ± 1.6445.35 ± 0.72 (OCS-T)+2.73*

The results show that with identical restriction maps, predictor and tuning budget, learning the multiset functions significantly improves over the fixed DSHN aggregation on four of the five datasets (+2.7 to +9.2 points), while on Telegram the two are statistically tied. Notice that OS-Linear is not a re-implementation of the full DSHN pipeline, since it shares our predictor, input projection and classifier. In the final version, we will revise Theorem 4, the Abstract, the Introduction and Sections 3 and 4.4 accordingly, and add these experiments to Section 4.

[1] E. Mule et al., "Directional Sheaf Hypergraph Networks", ICLR 2026.

[2] I. Duta et al., "Sheaf Hypergraph Networks", NeurIPS 2023.

AuthorsOpenReview ↗

Response to Reviewer oiFd, Question 2, (Part 1/2)

Q2/W2. Mathematical Definition (1/2)

Comment: Could the authors provide a more explicit mathematical and architectural definition of the aggregation functions? [...]

Response: We apologize for the confusion. We take the opportunity to clarify the mathematical definitions below.

Every node vv carries a stalk Xv∈Rd×fX_v \in \mathbb{R}^{d \times f} (Cd×f\mathbb{C}^{d \times f} for OmniPhaseSheaf), obtained from the input features by a linear projection to dfd f channels. Let Xˉv∈Rf\bar{X}_v \in \mathbb{R}^{f} be the mean of the dd stalk rows. One layer is defined as follows.

Restriction maps. For each incidence v∈ev \in e, the map Fv⊴eF_{v \trianglelefteq e} is predicted as in Eq. (3), using the same predictor as SheafHyperGNN [2] and DSHN [1], which follow the standard parameterization of Neural Sheaf Diffusion [3]:

Fv⊴e=Ψ(σ(Φ(Xˉv,∥,∑u∈eψ(Xˉu)))),F_{v \trianglelefteq e} = \Psi\Big(\sigma\Big(\Phi\big(\bar{X}_v \\,\Vert\\, \textstyle\sum_{u \in e} \psi(\bar{X}_u)\big)\Big)\Big),

where ∥\Vert denotes concatenation, Φ\Phi and ψ\psi are single affine layers (with an optional input LayerNorm), and σ∈{sigmoid,tanh⁡,id}\sigma \in \lbrace \mathrm{sigmoid}, \tanh, \mathrm{id} \rbrace is tuned. Φ\Phi outputs dd entries for Diag (placed on the diagonal), d2d^2 entries for General (reshaped into a d×dd \times d matrix), or d(d−1)/2d(d-1)/2 entries for Ortho. For Ortho, the entries fill the strictly lower triangle of a unit lower-triangular matrix whose columns are the Householder vectors vkv_k of Eq. (31), so that Fv⊴e=∏k=1d(Id−2vkvk⊤/∥vk∥22)∈O(d)F_{v \trianglelefteq e} = \prod_{k=1}^{d} \big(I_d - 2 v_k v_k^{\top} / \lVert v_k \rVert_2^2\big) \in O(d). A separate predictor is used at every layer.

Two stages. With {!{⋅}!}\lbrace\\!\lbrace \cdot \rbrace\\!\rbrace denoting a multiset and Fv⊴e←F^{\leftarrow}_{v \trianglelefteq e} the backward map defined below:

he=ϕ({!{Fv⊴eXv:v∈e}!}),Xv′=ρ({!{Fv⊴e←,he:e∋v}!});[+,Xv].h_e = \phi\big(\lbrace\\!\lbrace F_{v \trianglelefteq e} X_v : v \in e \rbrace\\!\rbrace\big), \qquad X'_v = \rho\big(\lbrace\\!\lbrace F^{\leftarrow}_{v \trianglelefteq e}\\, h_e : e \ni v \rbrace\\!\rbrace\big) \\; \big[+\\, X_v\big].

The self term (WselfXuW_{\mathrm{self}} X_u in Eqs. (25), (27) and (37)) is implemented as the identity skip connection [+,Xv][+\\,X_v], which is a tuned hyperparameter (Table 6, "Residual skip connections").

Multiset functions. For DeepSets (ODS, OCS, OPS) we use the row-wise form of Eqs. (26)–(27):

ϕ(S)=ψE[∑s∈SφV[s]],ρ(S)=ψV[∑s∈SφE[s]],\phi(S) = \psi_E\Big[\textstyle\sum_{s \in S} \varphi_V[s]\Big], \qquad \rho(S) = \psi_V\Big[\textstyle\sum_{s \in S} \varphi_E[s]\Big],

where g[A]g[A] denotes row-wise application and φV,ψE,φE,ψV\varphi_V, \psi_E, \varphi_E, \psi_V are MLPs with ReLU outputs, shared across the dd stalk rows, since FF already mixes the stalk coordinates. In all experiments the sum is normalized by ∣S∣\lvert S \rvert (mean pooling), which provides the δe−1\delta_e^{-1} normalization used in the reductions of Theorems 6, 7 and 11. Theorem 8 covers the vectorized form of Eqs. (24)–(25).

For PMA (-T variants), we use the PMA block of AllSet [4] with a single seed. With nhn_h heads, each head hh has a learnable seed θh\theta_h and affine key and value maps WKhW_K^{h} and WVhW_V^{h} of width df/nhd f / n_h:

αsh=softmaxs∈S,LeakyReLU(⟨θh,WKh,vec(s)⟩),Z=LN(θ+∑s∈SαsWV,vec(s)),\alpha^{h}_s = \mathrm{softmax}_{s \in S}\\, \mathrm{LeakyReLU}\big(\langle \theta_h, W^{h}_K\\, \mathrm{vec}(s) \rangle\big), \qquad Z = \mathrm{LN}\Big(\theta + \textstyle\sum_{s \in S} \alpha_s W_V\\, \mathrm{vec}(s)\Big),

ϕ(S)=LN(Z+ReLU(rFF(Z))),\phi(S) = \mathrm{LN}\big(Z + \mathrm{ReLU}(\mathrm{rFF}(Z))\big),

where the heads are concatenated in ZZ. We will update Eq. (28) accordingly. ρ\rho has the same form with its own parameters.

Backward transport.

  • ODS and OST. Fv⊴e←=Fv⊴e⊤F^{\leftarrow}_{v \trianglelefteq e} = F_{v \trianglelefteq e}^{\top}.
  • OmniPhaseSheaf. The forward map is sv,eFv⊴es_{v,e} F_{v \trianglelefteq e}, with charge sv,e=1s_{v,e} = 1 for heads and sv,e=e2πiqs_{v,e} = e^{2 \pi i q} for tails (Eq. (32)). The backward map is its conjugate transpose, Fv⊴e←=sˉv,eFv⊴e⊤F^{\leftarrow}_{v \trianglelefteq e} = \bar{s}_{v,e} F_{v \trianglelefteq e}^{\top}, which recovers DSHN (Theorem 11). DSHN writes the tail phase as e−2πiqe^{-2\pi i q}; the two conventions coincide under q↦−qq \mapsto -q.
  • OmniCopreSheaf. The forward maps are untied by role (FtF^{t} for tails, FhF^{h} for heads). The two roles are pooled by separate multiset functions and combined row-wise by an MLP, he=ReLU(γ(ϕt({!{Fu⊴etXu:u∈Te}!}),∥,ϕh({!{Fu⊴ehXu:u∈He}!})))h_e = \mathrm{ReLU}\big(\gamma\big(\phi_t(\lbrace\\!\lbrace F^{t}_{u \trianglelefteq e} X_u : u \in T_e \rbrace\\!\rbrace) \\,\Vert\\, \phi_h(\lbrace\\!\lbrace F^{h}_{u \trianglelefteq e} X_u : u \in H_e \rbrace\\!\rbrace)\big)\big). The backward maps GtG^{t} and GhG^{h} are independent learned maps, not transposes of FtF^{t} and FhF^{h}. They are predicted as in Eq. (3) with their own parameters and selected by the role of vv in ee, and a single ρ\rho pools {!{Gv⊴erole(v,e)he:e∋v}!}\lbrace\\!\lbrace G^{\mathrm{role}(v,e)}_{v \trianglelefteq e} h_e : e \ni v \rbrace\\!\rbrace. We will update Eq. (37) accordingly.

(continues below)

AuthorsOpenReview ↗

Response to Reviewer oiFd, Question 2, (Part 2/2)

Q2/W2. Mathematical Definition (2/2)

Transport matrices in the backward stage. In the learned instantiations, ρ\rho receives only the transported messages Fv⊴e←heF^{\leftarrow}_{v \trianglelefteq e} h_e. The transport blocks F⊤FF^{\top}F of Definition 1 are used by the linear instances (OS-Linear and the reductions of Theorems 6, 7 and 11). There they form the sheaf degree Dv=∑e∋vFv⊴e⊤Fv⊴eD_v = \sum_{e \ni v} F_{v \trianglelefteq e}^{\top} F_{v \trianglelefteq e} that normalizes each node. Since ∣sv,e∣=1\lvert s_{v,e} \rvert = 1, the charge cancels in these blocks (Eq. (33)).

Initial hyperedge representations. Hyperedges carry no state across layers: heh_e is recomputed at every layer by the first stage from its incident nodes, so the argument He(k)H_e^{(k)} in Eq. (1) is not used by any instantiation. The only hyperedge input that precedes the first stage is the predictor's summary ∑u∈eψ(Xˉu)\sum_{u \in e} \psi(\bar{X}_u) above. In OmniPhaseSheaf, as in DSHN [1], the maps FF are real (the predictor reads the real part of the stalks) and the direction is carried only by the charge. We will correct the sentence in Section 3.4 that states Fu⊴e∈Cd×dF_{u \trianglelefteq e} \in \mathbb{C}^{d \times d}.

Attention queries in OmniPhaseSheaf.

  • Linear layers. All linear layers of OmniPhaseSheaf (input projection, keys, values, rFF and DeepSets MLPs) have real weights WW and biases bb, applied identically to the real and imaginary parts: z↦Wz+(1+i),bz \mapsto W z + (1+i)\\, b.
  • Input. The input xx is lifted as x+ixx + i x, so the initial stalks are (1+i)(Wx+b)(1+i)(W x + b), and the phase is introduced by the charges.
  • Query. In the complex PMA, the query is a learnable complex seed θ=θre+i,θim\theta = \theta_{\mathrm{re}} + i\\, \theta_{\mathrm{im}} per head, shared across hyperedges, with θim\theta_{\mathrm{im}} initialized at 0.
  • Score. The keys and values are the linear maps above applied to the transported stalks sv,eFv⊴eXvs_{v,e} F_{v \trianglelefteq e} X_v. The score is the real part of the Hermitian product:

αv,e=softmaxv∈e,LeakyReLU(Re,⟨θ,WK,vec(sv,eFv⊴eXv)⟩),⟨a,b⟩=a†b.\alpha_{v,e} = \mathrm{softmax}_{v \in e}\\, \mathrm{LeakyReLU}\big(\mathrm{Re}\\, \langle \theta, W_K\\, \mathrm{vec}(s_{v,e} F_{v \trianglelefteq e} X_v) \rangle\big), \qquad \langle a, b \rangle = a^{\dagger} b.

The weights αv,e\alpha_{v,e} are therefore real. The charge enters each message exactly once, through the transported value WV,vec(sv,eFv⊴eXv)W_V\\, \mathrm{vec}(s_{v,e} F_{v \trianglelefteq e} X_v), which keeps the phase convention of DSHN. We will update Eqs. (4) and (34) accordingly.

  • Remaining blocks. The values stay complex. LayerNorm is the complex LayerNorm of Trabelsi et al. [5], and the nonlinearity is the complex ReLU, which zeroes an entry when its real part is negative.
  • Output. After the last layer, the stalks are unwound as (Re,∥,Im)(\mathrm{Re} \\,\Vert\\, \mathrm{Im}) and passed to a real classifier.

Effect on the theory. These updates concern only the specific instantiations used in our experiments. They change neither Definition 1 nor any of our theoretical results or proofs. Propositions 2 and 3 and Theorems 4, 5, 6, 7, 10 and 11 are stated for the framework of Definition 1. The reductions are realized by their own linear instances (Appendix B.5), which do not involve Eqs. (4), (28), (34) or (37). Theorem 8 and Corollary 9 concern the general parameterizations, namely vectorized DeepSets, PMA as in [4], and arbitrary real maps on R2df≅Cd×f\mathbb{R}^{2df} \cong \mathbb{C}^{d \times f}. The row-wise and real-weight layers used in our experiments are special cases of these parameterizations.

We will add these definitions to the final version of the manuscript, and the code will be released upon acceptance for reproducibility.

[1] E. Mule et al., "Directional Sheaf Hypergraph Networks", ICLR 2026.

[2] I. Duta et al., "Sheaf Hypergraph Networks", NeurIPS 2023.

[3] C. Bodnar et al., "Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs", NeurIPS 2022.

[4] E. Chien et al., "You are AllSet: A Multiset Function Framework for Hypergraph Neural Networks", ICLR 2022.

[5] C. Trabelsi et al., "Deep Complex Networks", ICLR 2018.

Comments on the paper

1 comment

We sincerely thank the reviewers for their feedback. We are currently updating the manuscript to incorporate the points raised so far in the discussion phase. The updated manuscript will be uploaded before the end of the rebuttal phase.

Please let us know if you have any additional unresolved concerns. We would be happy to do our best to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.