acceptodds
Under review as a conference paper at ICLR 2027

Coefficient-Depth Blindspots in Topological Message Passing

Abstract

Learning tasks can involve interactions among groups of entities. Graphs represent pairwise relations; simplicial and cellular complexes also represent larger groups. Topological message passing learns through local exchanges between their cells. Homology supplies global summaries that local computations can miss. Yet separate limits of these two channels do not characterize their combined expressivity. For every prime , integer and prescribed order , we construct finite pure two-dimensional full-incidence pairs with identical order- observations at every finite depth. Their homology agrees over every field in every degree. For , one chain isomorphism matches all permitted finite diagrams inside -Mod, up to abstract isomorphism. Integral first homology and depth- modules distinguish different cyclic lengths under the common torsion exponent . Full cubical realization from a height-four lattice preserves the separation, adding a common free summand. Boundary computations and shared-weight sum/attention forwards check three source pairs and one dependent cube realization. An additional source pair extends the neural checks to matching mod- and distinct mod- information. An order-four simulation covers the standard scalable multicellular network interface. A depth- homology readout separates each pair; fixed structural order, depth or width cannot uniformly replace this coefficient precision.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
topological message passing, expressivity, integer homology, coefficient precision, counting logic, cubical complexes

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_q9zdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As I understand it, the paper tries to define a rather convoluted version of some kind of Weisfeiler-Leman-style ``message-passing’’ isomorphism test on 2-complexes and shows that it is not complete.

Critical strengths

Understanding the expressive power of isomorphism tests has been an important direction in trying to understand the computational complexity of the isomorphism problem. As an example, Cai-Fuhrer-Immerman graphs and the technique used in them have been useful in studying the complexity of several graph logics, and understanding polynomial-time properties of graphs in general.

Critical weaknesses

The message-passing algorithm proposed in the paper is convoluted, involving higher-order interactions in 2-complexes, homology over Q and finite fields, etc. First of all, it is really, really hard to follow, on which I elaborate more below. But besides that, it is not at all clear why this specific version is so important to understand. It is not so hard to come up with some isomorphism test, involving several different techniques, putting layers over layers, and it would probably be rather difficult to find a counter-example of two non-isomorphic structures that it does not distinguish. However, how such a counter-example (in the case of this paper, an LLM-generated one) would advance our understanding in some important problem? The current paper never stops to elaborate on that, it just goes off, defining the message-passing algorithm, and then proceeding to the construction, without reflecting what is the value in all these, probably highly non-trivial, constructions.

Maybe even a more important weakness is the writing style, which makes the technical content of the paper impossible to follow. Honestly, there are two options – either it makes no mathematical sense, or requires some super-human level of technical understanding. On the surface, the style has an LLM-smooth look, short sentences without complicated structures, like when you ask an LLM to generate an essay. But if you look closely, the sentences are filled with pseudo-scientific, intelligence-sounding phrases that involve terminology that was never defined before, and thus just unnecessarily complicates the text, making it barely readable.

For example, let’s take a phrase The structure records equality and strict geometric containment Sub between faces, including vertex-to-trianglec containment. Why not simply say something like Let Sub denote the inclusion relation on faces. There is a specific minimalistic style that written mathematics has to follow in order to avoid thinking what phrases like ‘’the structure records equality’’ formally mean.

More examples of the same phenomena that made it impossible for me to understand actual technical contribution of the paper (if there is one, actually):

For simplicial and cubical inputs, Face denotes the same containment – simplicial and cubical inputs are some other kind of inputs? Or is it a subset of the set of inputs? Why in any case we introduce a new name for the same thing?

Line 90, ‘’unmarked inputs’’ - what are you trying to say in this sentence? Inputs can be marked somehow? In which way? If it is important, define it properly, if not, delete this phrase.

Definition 2 – so what object is a paired profile?

Paragraph at line 114 – what is the purpose of this paragraph? Does it define something?

Line 122 – what is a mark predicate?

Proposition 3 – for example, what is the displayed neighborhood? Are there non-displayed ones? What is a ‘’mathematical output’’ – there can be a non-mathematical output? Parameters of what are we talking about in this proposition? Is this some kind of GNN we are talking about – you have never defined. Overall, it is impossible to understand this statement.

Line 141 – when encoder is covered? What do you mean by express as neural maps? What are you trying to do in this paragraph?

Lines 145-150 none of this notation has not been previously defined.

Line 160 – what are boundary matrices, cellular bases? No idea what are you trying to say here.

Questions for authors

Overall, I would like to understand the motivation of this specific construction and why one should care in trying to construct counter-examples for it.

Reviewer_KhKmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper is about the expressivity of topological message passing architectures (in simplicial and cellular domains) when equipped with global homology summaries. Specifically, authors consider whether adding Betti numbers (persistent homology) and torsion orders can help address the expressivity "blindness" of message passing alone. The main result is that it does not, uniformly--authors show this by constructing a pair of rank 2 complexes that cannot be distinguished. Specifically, they have identical order-k colored histograms at every depth/width, isomorphic homology over every field and over Z/p^j for all j < s, and one fixed chain isomorphism over A = Z/p^{s−1} matching every finite diagram in A-Mod including connecting maps. Computational evidence shows that untrained higher-order message-passing and attention encoders support the predicted gaps.

Critical strengths

The main theorem here appears to provide a clean and definitive answer to the question: "Can structural order k substitute expressivity for some coefficient precision"? I thin this is a relevant question for Topological Deep Learning and agree this is a novel contribution compared to previously proposed limits.

Prop 8 is compelling as it goes beyond matching Betti numbers and instead considers finite diagram of A-modules.

The combinatorial topology result of Prop 7 is interesting outside of usefulness for machine learning.

Critical weaknesses

I really struggled to read the paper as it is extremely dense and heavily relies on a 40 page appendix that I am not able to properly dedicate time to read. Just the abstract and introduction are basically impossible to read without having the full context of the rest of the paper. This style of writing 9which I suspect is heavily AI generated) makes it really hard to learn something new from the paper, which is why I rated both presentation and Informativeness poorly

The other most critical weakness here is the authors' failure to contextualize the usefulness of the work from a practical ML perspective. Characterizing how often this gap in expressivity in higher order message passing models arises and, when it does, how much it affects performance would have been a straightforward way to do that. I understand that the nature of the contribution is not to propose a new SOTA approach, but it would nevertheless be helpful to characterize some practical application of this.

Questions for authors

Clarifications

Can you explain whether this expressivity problem actually occurs in standard TDL benchmarks? As in, for standard TDL datasets ( MANTRA, TUD, and so on), does H1 ever carry p^s-torsion with s ≥ 2, and if so, do pairs of instances in the same class differ only at depth s?

What is the intended takeaway for a practitioner? Is there a specific feature (homology summaries over Z/p^s for some fixed s) they should always consider adding?

The witness pairs are very large (40k cells at a minimum, from my understanding). Why do they have to be so big? If the smallest pair is necessarily this large, does that suggest the gap is unlikely to matter in practice?

Suggestions for future work

Could you reframe the introduction (and abstract) to broadly motivate the work without heavily relying on the reader already understanding the contribution at hand?

Reviewer_e2rfOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies the joint expressivity of topological message passing on cell complexes and homology-based global readouts. For every prime p, every s ≥ 2 and every fixed structural order k, the authors construct pairs of two-dimensional complexes that are indistinguishable by order-k message passing at any depth and have identical homology over every field, but differ in the torsion of their integral first homology. They show that a single homology readout over Z/p^s separates these pairs, extend the construction to cubical complexes, and complement the theory with computational checks on sum and attention architectures. The authors conclude that increasing structural order, depth or width cannot replace this coefficient precision.

Reviewer Note: I would like to note that the theoretical foundations of this paper (algebraic topology, integer homology and counting-logic expressivity) lie outside my core area of expertise. My review therefore focuses on the standard criteria for ML papers at venues such as ICLR, namely clarity, presentation, positioning with respect to prior work, experimental support and relevance to the ICLR audience. I was not able to verify the proofs in detail or to fully assess the novelty of the theoretical contributions, and my confidence score reflects this.

Critical strengths

Relevant and well-motivated question. The paper addresses the expressivity of topological message passing combined with homology-based readouts, a combination that prior work has studied mainly in isolation. It positions itself clearly against existing results: prior blindspots of higher-order message passing (Eitan et al., 2025) are resolved by ordinary homology, whereas the pairs constructed here also agree in homology over every field and differ only in torsion.

Technically solid in appearance. Although I was not able to verify all proofs in detail, the theoretical development appears to be rigorous and carefully worked out. Definitions, propositions and proofs are developed systematically and in considerable depth in the appendix, and the constructions are additionally checked computationally.

Critical weaknesses
  1. Presentation, readability and suitability of the format. The paper addresses an interesting question and appears to be technically solid. However, it is very hard to follow because it is written in an extremely compressed style. Several paragraphs are so dense that they are difficult to parse, e.g. l. 050–053. This is aggravated by frequent forward and backward references that interrupt the reading flow, e.g. l. 161 refers ahead to Section 4.4 and l. 173 to Proposition 8. In addition, central notation is introduced without definition: in l. 146–150, the functions E_{r(x)}, AGG_{l,j}, U_{l,r(x)} and D_θ are used but not defined. More fundamentally, the amount of material exceeds what a conference paper can accommodate. The appendix contains an extensive sequence of definitions and results (up to Proposition 127), and the main text relies on it so heavily, as the numerous references to the appendix show, that the main paper does not stand on its own. Within a conference reviewing period, a contribution of this scope cannot be reviewed with adequate care, and it cannot be presented adequately within the page limit of the main paper. Taken together, this makes the paper largely inaccessible to readers outside algebraic topology, which includes a substantial part of the ICLR audience, and suggests that a journal format would be more appropriate for this work.

  2. Related work is deferred to the appendix. The main paper does not contain a related work section; the comparison with related torsion and persistence settings is only provided in Appendix I. Since the contribution is defined largely by its relation to existing expressivity results for topological message passing and homology-based readouts, this positioning belongs in the main text. Without it, the reader cannot assess the novelty and significance of the results from the main paper alone.

  3. Missing discussion of the results. The paper does not contain a discussion section that relates the results to the state of the art, explains their implications, and states their limitations.

  4. Unclear purpose and baselines of the computational checks. It is not clear what the computational checks are compared against. According to the figure caption, only four configurations are evaluated, and no comparison with other methods or baselines is reported. The authors should clarify whether the checks are intended only as a verification of the theoretical constructions or as an empirical evaluation, and in the latter case, against which baselines.

  5. (very minor) Figures 1 and 2. The fonts used within the figures differ from each other and from the main text, which is distracting and inconsistent with the overall presentation of the paper.

Questions for authors
  1. Practical relevance. Which applications or learning tasks would actually require the separation shown in the paper? Concretely, are there datasets or domains in which complexes that agree in homology over every field, but differ in torsion, occur in practice, so that a depth-s homology readout over Z/p^s would make a difference to model performance?
  2. Computational complexity. What is the computational cost of the proposed depth-s homology readout over Z/p^s, and how does it compare to existing approaches, such as higher-order message passing and homology readouts over a field (e.g. Betti numbers)? How does this cost scale with the coefficient depth s and with the size of the complex?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.