acceptodds
Under review as a conference paper at ICLR 2027

Beyond Dense Tensors: Exact Evaluation of Spectral Hypergraph Layers

Abstract

Spectral hypergraph neural networks encode each hyperedge in a dense adjacency tensor, but their storage grows as the number of vertices raised to the size of the largest hyperedge, which makes them prohibitively expensive even on modest hypergraphs. We show that this cost is avoidable for the T-Spectral layer as released, whose readout sums all output slices of the t-product. We eliminate the tensor intermediates algebraically and evaluate the layer directly from node–hyperedge incidence data, preserving its output, normalisation and treatment of repeated hyperedges. After preprocessing, propagation takes time linear in the number of vertices and incidence entries for a fixed feature width, and neither the adjacency tensor nor its pairwise expansion is ever built. The same derivation reveals an expressivity limit already present in the original layer: on a labelled Pasch configuration, two differently labelled hypergraphs induce identical effective operators, which forces chance-level accuracy on a balanced task whose labels are recoverable from the incidence structure. Exact evaluation inherits this limit. Together, these results remove the dense-tensor barrier to evaluating the layer on larger hypergraphs and identify the structural distinctions that require a richer architecture.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
hypergraph neural networks, tensor t-product, exact linear-time evaluation, expressive power, spectral hypergraph learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_6cuyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper rethink the released implementation of T-Spectral HyperGNN and shows that its tensor-based computation can be evaluated using sparse operations on the hypergraph incidence structure. This leads to computational savings. The same derivation also reveals that the released layer reduces higher-order relationships to weighted pairwise interactions, making certain distinct hypergraph structures indistinguishable.

Critical strengths

-interesenting and well delimited research question -Important computational improvement -reproducibility -interesting study on efficiency, but also on the limitation of expressiveness

Critical weaknesses
  1. The main result are on the released code, not the full published method. The reference implementation does not use the symmetrisation described in Wang et al. (2025). However, the symmetrised version does not have the same collapse and reaches 100% on the Pasch task. For me the significance depend a lot on how much this released code is used in the community, and this is not discussed.

  2. The main simplification follows from a known property of circular convolution. the authors explain that the resulting operator is a special case of the TTSV2 contraction from Aksoy et al. (2024). The matrix formulation is also related to classical clique expansion methods (Zhou et al., 2006; Agarwal et al., 2006). the fact that different hypergraphs can have the same weighted clique expansion was in (Aksoy et al., 2024). So while the result is useful the main novelty seems to be the application of these ideas to one specific layer.

  3. once prop. 1 is established, thm 3 follows directly from the fact that the model only sees the effective pairwise operator. The interesting part is the Pasch example. However, the two hypergraphs considered are isomorphic under a permutation of nodes. The limitation is therefore about distinguishing labelled relations, rather than non-isomorphic hypergraph structures. This should be considered when discussing the general significance of the result.

  4. thm 2 is proved only for M = 3. The hypergraphs used in the paper have larger order (for example House has M=81M = 81), and for M≥4M \geq 4 there is only exact enumeration on very small NN and a conjecture.

  5. the released layer is shown to be equivalent to a model with one weighted pairwise propagation step, similar in structure to SGC or APPNP. I think it would be interesting to see a direct comparison with a simple one-hop graph model using the clique expansion and the same tuning budget.

  6. the proposed model uses a larger hyperparameter search than most baselines. the trained model uses an absorbed parameterisation and sometimes an additional normalisation that is not present in the released layer.

  7. I found the introduction difficult to follow. Its purpose should be to help the reader understand the problem, instead it starts directly from more technical aspects without giving the big picture of the work.

-Aksoy, S. G., Amburg, I., and Young, S. J. (2024). Scalable tensor methods for nonuniform hypergraphs. SIAM Journal on Mathematics of Data Science, 6(2):481–503.
-Zhou, D., Huang, J., and Schölkopf, B. (2006). Learning with hypergraphs: Clustering, classification, and embedding. NeurIPS.
-Agarwal, S., Branson, K., and Belongie, S. (2006). Higher order learning with graphs. ICML.

Questions for authors

see weaknesses.

3 comments

Thank you for the detailed review; we agree with most points and changed the paper. We answer your seven Critical Weaknesses in the order you wrote them. Abar is the paper's A-bar.

Weakness 1, released code vs published method. Agreed; this is the main change. Theorem 1 evaluates the published (symmetrised) form exactly from incidence data in O(M2 nnz(H) d)O(M^2\,\mathrm{nnz}(H)\,d) time. Proposition 2 bounds it against the implemented form: published = implemented ×(1−θ)\times (1-\theta), 0≤θ≤1−(1−ρ/2)M−20 \le \theta \le 1-(1-\rho/2)^{M-2} (θ\theta: damping; ρ\rho: largest share of a channel held by one node). Both forms match the authors' code to single precision; for the published form this uses the padding functions the code defines but does not call (Section 5.1). Up to a constant scale, the implemented form, the published form and Wang et al.'s T-Spatial layer are the full sum, the half sum and slice 0 of one symmetrised t-product (Section 2). On use: the T-HyperGNNs line is built on in later work: DHG-Bench ships its message-passing network T-MPHN, which Taimoor et al. (2025) speed up. T-Spectral itself is not in DHG-Bench, which is part of why running it at scale matters (now stated in Section 6).

Weakness 2, novelty. We agree the ingredients are known: the zero-frequency coefficient of a circular convolution, TTSV2 (Aksoy et al.), clique expansions (Zhou et al.; Agarwal et al.), and distinct hypergraphs with one clique expansion (Aksoy et al.). Section 6 ("What is known, and what we add") says so. We add: the implemented form is exactly such an operator, with the builder's normalisation and duplicate conventions; the published form is evaluated exactly with the same generating functions at a feature-dependent vector; the gap between the forms is bounded.

Weakness 3, labelled vs non-isomorphic. Correct: the Pasch pair is one labelled trade, a node-level limit, which is the setting of transductive node classification. We agree the node-level limit follows from the collapse, the submission's Proposition 1 (now eq. (4)); its Theorem 3 is now Remark 1. We added a non-isomorphic pair on six nodes, {13, 15, 135, 146, 236, 245} and {13, 15, 136, 145, 235, 246}, with the same Abar and degrees (Section 4.1; Example 1, Appendix G.3). With the same features, every stack of implemented layers gives them the same output. The published form separates them: in exact arithmetic its outputs differ by 1.6e-3 to 4.7e-2 of the output scale with random positive features (Appendix G.3).

Weakness 4, Theorem 2 only at M = 3. Theorem 2 is now proved for every M≥3M \ge 3 whenever N≥2M−3N \ge 2M-3 (Appendix C), which every dataset in the paper satisfies (House: N = 1,290, M = 81). Section 4.3 gives the M = 3 argument. Exact enumeration at 21 pairs (M, N) with 3≤M≤73 \le M \le 7 and M≤N≤10M \le N \le 10, ten of them with N<2M−3N < 2M-3, outside the theorem's hypotheses, finds that the equivariant weightings are exactly the constants. A second check, exact ranks modulo a prime on a random sample of the equations, reaches N = 1,000 for M = 3, N = 100 for M = 4 and N = 12 for M = 7 (for M≥5M \ge 5 it did not reach N = 100), with no failure at any of its 12 pairs; all have N≥2M−3N \ge 2M-3, so it is a cross-check (Appendix E.2).

Weakness 5, one-hop clique model at the same budget. Added (Table 12, Appendix J): the same model with Abar, HGNN's normalised operator, the unweighted clique and SGC, each tuned with our search. Abar is within 1.12 points of HGNN's operator on all eight sets. Both lead the unweighted clique and SGC by 22.4 to 23.8 points on House-committees, 15.8 to 17.2 on Actor and 4.4 to 26.8 on Walmart-trips. Remark 3 (Appendix D) says why: a hyperedge of size c gives each pair a weight between 1/c and 1/(c-1). So Abar is, up to a diagonal, HGNN's operator before its node-degree normalisation, with hyperedge weights between 1 and c/(c-1), the largest within 4.2% of 1 on every set but Cora.

Weakness 6, search budget, absorbed form, normalisation. We compared the budget in both directions (Table 11, Appendix J). At one default configuration with no search, T-Spectral (exact) loses 0.7 to 11.6 points, so its main rows depend on the search. TF-HNN and UniGCNII retuned with our selection rule are level with it on the homophilic sets and 4.6 to 4.7 points ahead on House-committees and Actor. Table 1 now discusses the row without normalisation; the row with it differs by about one point or less on every set but Yelp. The original factor, trained in double precision with its own search, trails the absorbed form by 6.2 to 52.9 points on DHG-Bench, where it reaches 1050510^{505} at initialisation (a fixed log-domain shift keeps it finite on PubMed). Both forms represent the same functions, so this gap measures training, not the model (Section 5.3).

Weakness 7, introduction. Rewritten to start from the problem: why tensor hypergraph networks, why they have not run at scale, and then the readout as the idea (Section 1).

Thank you for your response and for the changes. I appreciate the effort. I think that some of my concerns have been addressed, like the one on theorem 2 and the additional experiments.

However, I decided to keep my original score for two main reasons:

  1. Novelty (my Weakness 2). The new results on the symmetrised layer improve the paper but many of the main ideas are already in previous work especially Aksoy et al. I still have doubts about whether the novelty is enough for iclr.

  2. Original model (my Weakness 6 and Reviewer 7u7Q's Weakness 1). The new experiments show large differences between the original and absorbed parameterisations with accuracy differences from 6.2 to 52.9 points on DHG-Bench. Even if the two versions can represent the same functions, their training behaviour is very different. For this reason, I think this concern is still important.

  3. Clarity. I tried to read again the paper, with the new parts that you added. I still think that the clarity of the writing is not good enough for this venue, especially abstract and introduction which should help the reader understand the problem. Instead, I had to read them many times to understand their contents. I think the paper would be improved if it were rewritten, making the abstract and introduction simpler and ensuring a better logical flow.

Thank you for reading the revision again.

Clarity. We rewrote the abstract and introduction around one question (can a tensor layer tell apart teams that have the same weighted coauthor pairs?) and one idea (the readout).

Original scale factor. This is a property of the original layer, which exact evaluation at scale exposes. The layer multiplies each channel by S_c^(M-2). On Wang et al.'s hypergraphs (M = 4) this is harmless (-5.65 to +1.67 points); at DHG-Bench sizes it reaches 10^505 at initialisation and sets the logit scale, so the layer as written trains 6.2 to 52.9 points lower. Absorbing the factor keeps the same functions (Appendix E.3) and is what makes the layer trainable (Section 5.3).

Novelty. The ingredients are known and credited where used (Aksoy et al.'s generating functions). New are the exact evaluation of both forms, the bound between them, and what each form can separate.

Reviewer_kv7EOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies spectral hypergraph neural networks, specifically the T-Spectral layer. It introduces a novel and rigorously motivated strategy to evaluate the layer, thereby drastically decreasing storage requirements and achieving propagation time that is linear in the number of vertices and incidence entries. The paper further derives a theoretical expressivity limit for both the original (dense) T-Spectral layer and their novel evaluation route. The paper further provides memory statistics and training time measurements on five standard hypergraph datasets, as well as node classification results on three homophilic and three heterophilic hypergraphs.

Critical strengths
  • Theoretical grounding: Proposition 1 and Theorem 1 provide rigorous motivation for the proposed evaluation algorithm in section 3.3. The presentation of the algorithm itself is clear and relatively good to follow. Additional results in sections 3.4 and 5 provide further theoretical context. The theoretical derivations presented in the main text seem correct.
  • Time and memory savings: the peak memory and training time measurements in Figure 2(a) are very impressive evidence for the success of the proposed evaluation algorithm. Figure 2(b) further nicely visualizes the scalability to larger hypergraphs that becomes possible with the new evaluation approach.
Critical weaknesses
  • My main concern for the paper are the rather limited and somewhat weak node classification results in section 6. While this is clearly not a benchmark paper, some evidence for the empirical performance of the T-Spectral layer would be highly desirable to further motivate why we should even care about that particular layer being scalable to larger hypergraphs. It is possible that I missed some of the reasoning behind the experimental design for section 6, but to me it seems that the paper would clearly benefit from more datasets - especially if the hypergraph structure is indeed useful for the task, so that a simple MLP cannot perform competitively with methods that actively use the hypergraph.
Questions for authors
  • Why this particular choice of datasets in section 6? Why these baseline models? Could the authors present additional results on datasets where hypergraph information is beneficial to model performance?
  • Why only cover the one-layer model in the experiments in section 6?
2 comments

Thank you for the positive assessment and the clear request. We answer your Critical Weakness, then your two Questions. The submission's Section 6 (node classification) is now Sections 5.2-5.3.

Critical Weakness, more datasets where the hypergraph matters. We added three DHG-Bench sets: co-authorship Cora and DBLP, and Yelp (Table 1 now has nine sets). On co-authorship Cora T-Spectral (exact) reaches 84.92, against 84.62 for the best baseline (UniGCNII) and 75.61 for the MLP; on co-authorship DBLP it reaches 91.80, against 91.75 and 85.77. So it matches the best baseline there and is 9.31 and 6.04 points above the MLP (differences of unrounded means). On Yelp five of the eight baselines run out of memory on our 8 GB GPU; with a row-wise normalisation T-Spectral (exact) matches HGNN in accuracy (34.14 vs 33.50) and is above the MLP (31.46). On Walmart-trips, another set where structure carries the signal, the MLP reaches 63.47, HGNN 77.49 and ours 77.25. Over the five homophilic sets T-Spectral (exact) matches the best baseline (-0.03 to +0.63 points); on the three heterophilic sets it trails by 1.49 to 4.85 points (Section 5.2). One caveat: on House-committees and Actor the MLP is 3.4 and 4.3 points ahead of T-Spectral (exact), as it is of most propagation models there.

Critical Weakness, "why we should even care about that particular layer being scalable". Wang et al. (2025) ran the layer on hypergraphs with at most 89 nodes; on House (M = 81) its dense tensor would need about 7×102527 \times 10^{252} bytes. The exact evaluators make both forms of Wang et al.'s layer usable and comparable at benchmark scale (with the scale factor absorbed, the implemented form trains on 88,860 nodes), and show what each form can see (Section 4).

Question 1, choice of datasets and baselines (the request for more datasets is answered above). We use the node-classification sets and baselines of DHG-Bench, a third-party benchmark, at its shipped configurations and protocol, with homophilic and heterophilic sets (Section 5.2, Appendix J). Because the implemented form is one propagation with a weighted clique matrix, we added the same model with other one-hop matrices, tuned with our own search (Table 12): Abar (the paper's A-bar, the pair matrix of the implemented form) is within 1.12 points of HGNN's clique operator on all eight sets, and both lead the unweighted clique and SGC by 22.4 to 23.8 points on House-committees and 15.8 to 17.2 on Actor. TF-HNN and UniGCNII retuned with our selection rule are level with T-Spectral (exact) on the homophilic sets and 4.6 to 4.7 points ahead on House-committees and Actor (Table 11).

Question 2, only one layer. Wang et al. define T-Spectral as one layer, so the main rows use one. We now report stacks of two and three layers with ReLU, searched at the same budget (Section 5.3). Validation picks two layers over three on all six submission sets; in test accuracy the stack gains 1.9 points on Walmart-trips (79.01 vs 77.15) and trails one layer by 0.7 to 11.0 points on the other five, where its validation accuracy is also lower. Stacks of two and three layers also keep the limit of Remark 1: exactly 50% on the Pasch task (Section 4.1). The time and memory figure you mention (your Figure 2) is now Figure 4 (Appendix F.4); Section 5.1 keeps its numbers.

I would like to thank the authors for the detailed response. I remain in favor of accepting this paper and will maintain my original score.

Reviewer_7u7QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes the T-spectral layer from [1]. The authors of this paper mathematically prove that the two versions of the layer proposed in [1] can be implemented much more efficiently than in the original work, both in terms of memory footprint and computational complexity. This work shows a fundamental limitation of the "released" layer of [1], which is not more expressive than using the clique expansion, meaning that it discards higher-order information. The layer is tested on various datasets against well-established HNNs, and in practice achieves results on par with other architectures. The "printed" layer, which was not implemented in the original work, is derived in closed polynomial form, and it is shown that it retains higher-order information. Still, this implementation can only be used for hypergraphs with very few nodes (N=6 in the executed experiments).

References

[1] Fuli Wang, Karelia Peña-Peña, Wei Qian, and Gonzalo R. Arce. T-HyperGNNs: Hypergraph neural networks via tensor representations. IEEE Transactions on Neural Networks and Learning Systems, 36(3):5044–5058, 2025. doi: 10.1109/TNNLS.2024.3371382.

Critical strengths

The paper presents exact, sparse evaluators that eliminate the O(NM)\mathcal{O}(N^M) dense tensor barrier of the models in [1], backed by experimental verification.

The theoretical analysis cleanly exposes the structural collapse of the "released" layer, substantiated by experimental results.

References

[1] Fuli Wang, Karelia Peña-Peña, Wei Qian, and Gonzalo R. Arce. T-HyperGNNs: Hypergraph neural networks via tensor representations. IEEE Transactions on Neural Networks and Learning Systems, 36(3):5044–5058, 2025. doi: 10.1109/TNNLS.2024.3371382.

Critical weaknesses

Setting W1:=diag(S)M−2⋅WW_1 := \text{diag}(S)^{M-2} \cdot W and using W1W_1 as a trainable weight matrix fundamentally changes the training dynamics. I am aware that this is mentioned as a limitation by the authors, but it is a major issue when the goal is to mimic the behavior of the original model.

Adding a normalization layer for PubMed and the heterophilic datasets changes the underlying architecture. It would have been better to keep the results without the normalization layer in the main paper to show the behavior of the original model.

The paper is lacking when it comes to its presentation. The abstract does not mention the closed form of the "printed" layer, which is later listed as a core contribution. The "printed" and "released" layers are only defined in Section 2, after being extensively used in Section 1. Related work, future work, and conclusions are severely lacking.

Questions for authors

Did you try to keep the original diag(S)M−2⋅W\text{diag}(S)^{M-2} \cdot W formulation to observe the training behavior of the model?

You show the collapse of the released layer when using a single layer. Have you considered whether this limitation persists when stacking multiple layers, potentially with non-linearities in between? Even preliminary experimental data on this would be interesting.

2 comments

Thank you for the precise review. We answer your three Critical Weaknesses and two Questions in the order you wrote them.

Weakness 1 and Question 1, the factor W1=diag(S)M−2WW_1 = \mathrm{diag}(S)^{M-2} W. Yes: we trained the original form with S recomputed at every step and gradients through it, in double precision, each form with its own search on the same grid (Section 5.3). On Wang et al.'s five hypergraphs (M = 4, their grid, 10 splits ×\times 3 seeds), original minus absorbed is -5.65, -0.83, +1.67, -3.33 and -1.33 points on co-authorship Cora and DBLP and co-citation Cora, Citeseer and PubMed. On DHG-Bench, with hyperedges of up to 171 nodes, the factor reaches 1050510^{505} at initialisation on PubMed (there a fixed log-domain shift keeps it finite; the other five sets train it as written, Appendix H.2), and the original form trails the absorbed one by 6.2, 52.9, 48.1, 30.1, 7.7 and 17.8 points on Cora, Citeseer, PubMed, House-committees, Walmart-trips and Actor. On a fixed graph the two represent the same functions (Appendix E.3), so the gap reflects optimisation: at these orders the factor sets the logit scale. We therefore train the absorbed form and list as a limitation that the original factor does not train at DHG-Bench orders (Section 7).

Weakness 2, the normalisation. Agreed. Table 1 now shows both selections, and the text discusses the one without normalisation. Re-selected without it, T-Spectral (exact) changes by -0.53, 0.00, +0.15 and -1.12 points on PubMed, House-committees, Walmart-trips and Actor, and by -0.10 on co-authorship DBLP (unrounded differences; Appendix J). On Yelp the gap is larger (29.95 without, 34.14 with), and Table 1 shows both.

Weakness 3, presentation. Agreed on all three points. The abstract now states the closed form of the published (symmetrised) form and its exact evaluator. Both forms are defined in the introduction before they are used, and formally in Section 2. Related work (Section 6, with what is known and what we add), limitations and future work (Section 7: learned or feature-dependent readouts, training the original factor, inductive use, a cheaper published-form evaluator) and the conclusion are rewritten. We renamed "released" to "implemented (unsymmetrised)" and "printed" to "published (symmetrised)"; "T-Spectral" alone now means both forms, and we name the form wherever we compare or measure.

Paper Summary, "this implementation can only be used for hypergraphs with very few nodes (N=6 in the executed experiments)", said of the printed layer. This no longer holds. Theorem 1 evaluates the published form exactly from incidence data in O(M2 nnz(H) d)O(M^2\,\mathrm{nnz}(H)\,d) time, with no tensor, and it matches the authors' code to single precision (relative error 5.0e-7 over 370 cases; the padding functions the code defines but does not call, scaled as in the paper; Appendix H.1). On House (N = 1,290, M = 81) one evaluation takes 2.5 s on one CPU thread at width 1 (Section 5.1). Trained, untuned, at the implemented form's three best configurations on the five of nine DHG-Bench sets where its cost fits our budget, it matches the implemented form within about 0.2 points, at 30 to 1,100 times the cost per epoch (Sections 3.2 and 5.3).

Question 2, stacking layers. The limit persists at any depth, with any row-wise nonlinearity or normalisation in between (Remark 1, Section 4.1): every layer sees the hypergraph only through the same pair matrix, so a node permutation that fixes this matrix and the features is a symmetry of the whole stack. Trained stacks of two and three layers with ReLU score exactly 50% on the Pasch query nodes in all 10 worlds, and 100% on the pairwise control (Table 8, Appendix G.5). On the six DHG-Bench sets of the submission, searched with L in {2, 3}, validation picks two layers over three on every set; in test accuracy the stack trails one layer by 0.7 to 11.0 points on five sets, where its validation accuracy is also lower, and gains 1.9 on Walmart-trips (79.01 vs 77.15) (Section 5.3).

A short follow-up on your Weakness 1. Section 5.3 now states what the experiments show: the factor S^(M-2) is harmless on Wang et al.'s hypergraphs (M = 4), but at DHG-Bench sizes it reaches 10^505 at initialisation and sets the logit scale, so the layer as written trains 6.2 to 52.9 points lower. This is a property of the original layer that appears only at benchmark scale; absorbing the factor keeps the same functions and makes the layer trainable. As you suggested, the abstract and introduction now define both forms before use and state the published form's closed form.

Reviewer_hzKROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to reduce the storage capacity required for the adjacency tensor in presenting scalable spectral hypergraph neural networks. Based on an observation that the sum of a circular convolution is the product of the sums, the layer reduces to a Nx N matrix, applied to the features and scaled by a factor. Then, the authors calculate this matrix from incidence data with Stirling numbers, and show. They also report DHG-Bench node classification with a trained one-hop model.

Critical strengths

In my view, the observation that summing the circular convolution reduces the released layer to a one-hop clique-expansion model with a per-channel scale is valid, but it is not significant, in particular with respect to the ICLR level. This paper could be more suitable for a workshop than ICLR.

Critical weaknesses

As stated in the paper, the authors used LLM for many mathematical materials, such as Proposition 1, Proposition 1, Theorem 1, Theorem 2, and Theorem 3. But some of them seem to be quite easy (and basic). So, the mathematical work is not interesting at all. The evaluation seems to have some issues. The paper states that the printed (symmetrised) layer of Wang et al. (2025) does not collapse, so the proposed method applies only to an unsymmetrised layer, which is not covered in the paper.

Questions for authors

Does the exact evaluator reproduce the accuracies reported in Wang et al. on their datasets?

1 comment

Thank you for the review. We use the terms of the general response: your "printed (symmetrised)" layer is the published form, your "released" one the implemented form. We agree that the submission covered only the implemented form; the revision now covers both forms. Why it matters: both forms of Wang et al.'s layer are now evaluated exactly from incidence data, the implemented form trains on 88,860 nodes (scale factor absorbed) where the layer was run on at most 89, and what each form can see is analysed exactly.

Critical Strengths, "not significant ... more suitable for a workshop". The paper is no longer about one code path. (a) Up to a constant scale, the implemented form (public code), the published form (Wang et al.'s paper) and their T-Spatial layer are three readouts of one symmetrised t-product: full sum, half sum and slice 0 (Section 2). (b) Both forms are evaluated exactly from incidence data, so the layer, which Wang et al. (2025) ran on hypergraphs with at most 89 nodes, now trains (implemented form, scale factor absorbed) on hypergraphs with 88,860 nodes; on House (N = 1,290, M = 81) the dense tensor would need about 7×102527 \times 10^{252} bytes (Sections 1 and 5.2). (c) The published form separates two non-isomorphic six-node hypergraphs that the implemented form, given the same node features, cannot separate at any depth (Section 4.1).

Weakness 1, mathematics "quite easy (and basic)" and LLM use. We agree that the collapse is short: your Proposition 1 is now the identity (4) in Section 3.1 and your Theorem 3 is Remark 1. The results that carry the paper are Theorem 1 (new: the published form from incidence data), Proposition 2 (new) and Theorem 2, now proved for every order M whenever N≥2M−3N \ge 2M-3, which every dataset in the paper satisfies (House: N = 1,290, M = 81). Section 4.3 gives the M = 3 argument and Appendix C the general proof. The statement is also checked by exact enumeration at 21 pairs (M, N) with 3≤M≤73 \le M \le 7 and M≤N≤10M \le N \le 10, ten of them with N<2M−3N < 2M-3, outside the theorem's hypotheses. A second check, exact ranks modulo a prime on a random sample of the equations, reaches N = 1,000 for M = 3, N = 100 for M = 4 and N = 12 for M = 7 (for M≥5M \ge 5 it did not reach N = 100), with no failure at any of its 12 pairs; all have N≥2M−3N \ge 2M-3, so it is a cross-check (Appendix E.2). On LLM use, the LLM section (p. 11) and Appendix P say which results an LLM drafted first and correct the submission's statement; an author proved Theorem 2 at M = 3 by hand, and the authors read and checked every proof.

Weakness 2, "The evaluation seems to have some issues." If you meant the absorbed scale factor diag(S)M−2\mathrm{diag}(S)^{M-2} or the added normalisation, both raised by other reviewers: Table 1 now discusses the model without the normalisation, and we trained the original factor, each form with its own search (Section 5.3; on DHG-Bench it reaches 1050510^{505} and the original form trails by 6.2 to 52.9 points, a training effect, since both forms represent the same functions). The search budget is now compared in both directions (Table 11). We would be glad to address any other issue you have in mind.

Weakness 3, the symmetrised layer "is not covered in the paper". It is now. Theorem 1 evaluates the published form exactly from incidence data in O(M2 nnz(H) d)O(M^2\,\mathrm{nnz}(H)\,d) time, without the tensor. Proposition 2 gives published = implemented ×(1−θ)\times (1-\theta), where θ\theta is the damping, 0≤θ≤1−(1−ρ/2)M−20 \le \theta \le 1-(1-\rho/2)^{M-2} (an upper bound), and ρ\rho is the largest share of a channel held by one node. Both evaluators match the authors' code to single precision (relative error 5.9e-7 and 5.0e-7; for the published form, the padding functions the code defines but does not call, scaled as in the paper; Appendix H.1). Trained, untuned, on the five of nine DHG-Bench sets where its cost fits our budget, the published form matches the implemented one within about 0.2 points (Section 5.3).

Question, "Does the exact evaluator reproduce the accuracies reported in Wang et al. on their datasets?" At the operator level, yes: it returns the output of the authors' code (Section 5.1), so differences in accuracy come from training. In accuracy it depends on the reporting rule. With their grid and 50 balanced splits (Appendix K, Table 14), at the validation-selected epoch accuracy is 6.4 to 13.7 points below the published value on four sets; on PubMed it is 87.68, 1.01 points above the published 86.67. At the best test accuracy over epochs, which is how their released training script reports accuracy, it is within one published standard deviation on three of five sets (e.g. co-citation Cora 72.29 vs 71.59), above it on PubMed (92.96 vs 86.67) and below it on co-authorship DBLP. The test sets hold about a quarter of 59 to 89 nodes, so one node moves accuracy by about 4 to 7 points.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the four reviewers. Blue text in the revised PDF marks new content that answers the reviews; the rest is rewritten for clarity. We now call the two forms of T-Spectral "implemented (unsymmetrised)", the form in the public code (the submission's "released"), and "published (symmetrised)", the form in Wang et al.'s paper (the submission's "printed"). "T-Spectral" alone means both forms; "T-Spectral (exact)" is the implemented form trained with our exact evaluator. Numbering of the submission: Prop. 1 (collapse) is now eq. (4), Thm 1 is Prop. 1, Thm 3 is Remark 1 (now any depth), Fig. 2 is Fig. 4 (App. F.4), and Section 6 (node classification) is now Sections 5.2-5.3. Thm 2 now covers all M; Thm 1 (published form) and Prop. 2 are new. T-Spatial, which the submission called a slice-wise product, is by evenness slice 0 of the padded t-product (Section 2).

What the paper now shows:

(a) The readout of the t-product decides what a t-product layer costs and what it sees. Up to a constant scale, the implemented form, the published form and Wang et al.'s T-Spatial layer are three readouts of one symmetrised t-product: the full sum of the output slices, the half sum, and slice 0.

(b) The layer can now be run on large hypergraphs. In Wang et al. (2025) the layer was run on hypergraphs with at most 89 nodes. Both forms are evaluated exactly from incidence data: on House (N = 1,290, M = 81) the dense tensor would need about 7×102527 \times 10^{252} bytes, while both exact routes read its 11,809 incidences, and the implemented form, with the scale factor absorbed, trains on hypergraphs with up to 88,860 nodes (Sections 1, 3 and 5.2).

(c) The published form sees structure that the implemented form provably cannot: it separates two non-isomorphic six-node hypergraphs with the same pair matrix, which no stack of implemented layers separates given the same features (Section 4.1, Appendix G.3). The difference is small and bounded (Proposition 2), so at benchmark scale the forms behave alike.

(d) Theorem 2 (up to weightings that compute zero, the sum is the only fixed weighting of the output slices that respects node relabelling) is now proved for every order M whenever N≥2M−3N \ge 2M-3 (Appendix C), and cross-checked with exact ranks modulo a prime on a random sample of the equations at 12 pairs (N up to 1,000 for M = 3, 100 for M = 4, 12 for M = 7; for M≥5M \ge 5 not N = 100), with no failure (Appendix E.2).

Points raised by several reviewers.

Shared point 1, the published (symmetrised) layer was not covered (hzKR, 6cuy, 7u7Q). It is now. Theorem 1 evaluates it exactly from incidence data in O(M2 nnz(H) d)O(M^2\,\mathrm{nnz}(H)\,d) time, without the tensor. Proposition 2 relates the two forms: published = implemented ×(1−θ)\times (1-\theta), with 0≤θ≤1−(1−ρ/2)M−20 \le \theta \le 1-(1-\rho/2)^{M-2}, where ρ\rho is the largest share of a channel held by one node. Both evaluators reproduce the authors' code to single precision (for the published form, the padding functions the code defines but does not call): largest relative error 5.9e-7 (implemented, 400 cases) and 5.0e-7 (published, 370 cases) (Section 5.1; Appendix H.1). Trained, untuned, on the five of nine DHG-Bench sets where its cost fits our budget, the published form matches the implemented form within about 0.2 points, at 30 to 1,100 times the cost per epoch (Sections 3.2 and 5.3).

Shared point 2, fidelity of the trained model (7u7Q, 6cuy, hzKR). Table 1 now discusses T-Spectral (exact) without the row-wise normalisation and shows the normalised selection beside it. We also trained the original factor diag(S)M−2W\mathrm{diag}(S)^{M-2} W in double precision, each form with its own search on the same grid (Section 5.3). On Wang et al.'s five hypergraphs, original minus absorbed is -5.65 to +1.67 points. On DHG-Bench, with hyperedges of up to 171 nodes, the factor reaches 1050510^{505} at initialisation (a fixed log-domain shift keeps it finite on PubMed) and the original form trails by 6.2 to 52.9 points. On a fixed graph both represent the same functions (Appendix E.3), so the gap reflects optimisation: at these orders the factor sets the logit scale. We therefore train the absorbed form and list as a limitation that the original factor does not train at DHG-Bench orders (Section 7).

Shared point 3, empirical scope (kv7E, 6cuy, 7u7Q). kv7E asked for more datasets: we added co-authorship Cora and DBLP, and Yelp; Table 1 now has nine sets. On Yelp, where five of the eight baselines do not fit in 8 GB, T-Spectral (exact) matches the best of the rest in accuracy with a row-wise normalisation. 6cuy asked for a one-hop clique model (Table 12) and the search budget (Table 11); kv7E and 7u7Q asked about depth (Section 5.3).

Use of LLMs (hzKR). The LLM section (p. 11) and Appendix P say which results an LLM drafted first, and correct the submission's statement.

Dear reviewers,

A reminder that the discussion period will end on November 18 at 11:59pm AoE. I encourage you to interact with the authors' rebuttal to clarify any remaining issue (and how well the rebuttal addressed your concerns) before the next phase.

Thanks, The AC

We thank the reviewers, and kv7E and 6cuy for their follow-up. The second revision changes presentation only; no result changed.

  1. Abstract and introduction (6cuy, 7u7Q): rewritten around one question (can a tensor layer tell apart teams with the same weighted coauthor pairs?) and one idea (the readout). Both forms are defined before use.

  2. Original scale factor (7u7Q, 6cuy): Section 5.3 now presents it as a finding. S^(M-2) is harmless at M = 4 but reaches 10^505 at DHG-Bench sizes and sets the logit scale. Absorbing it keeps the same functions and makes the layer trainable.

Blue marks content added for the reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.