acceptodds
Under review as a conference paper at ICLR 2027

Alignment Matters Inside and Out in Equivariant Graph Flow Matching

14 views

Abstract

Graphs are invariant under node permutations, motivating permutation-equivariant architectures in generative models. In flow matching, however, symmetry also affects the source–target coupling: one must choose both which graphs to pair and which node representatives to align. We study these two forms of alignment, termed outer and inner alignment, and their effect on equivariant graph flow matching. We connect inner alignment to transport on the graph quotient space, whose Euclidean quotient metric coincides with Gromov–Monge distance, and show that quotient couplings admit aligned representative lifts while symmetrization yields equivariant flow-matching minimizers, including for categorical endpoints. In practice, we compare random augmentation, permutation-blind minibatch optimal transport, approximate Gromov–Wasserstein alignment, and combinations of inner and outer alignment. Across continuous graph and molecular generation, alignment can substantially simplify trajectories and improve few-step generation, while its benefits depend on the alignment and computational budget. Our results highlight that effective graph flow matching benefits from alignment both inside and across graph pairs.

Primary area:
generative models
Keywords:
Flow Matching, Gromov--Wasserstein, Optimal Transport, Graph Generative Models, Equivariance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7zYsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies graph flow matching by considering graph symmetry. Since any graph can be mapped to multiple equally valid edge feature matrices, the authors introduce a distance metric over the graphs in the quotient space of feature matrices. Showing that the quotient space metric can be defined in terms of representative feature matrix, authors propose two inner alignment methods that selects representative matrix—GW and BoA. Outer alignment further matches mini-batch samples using the metric computed by the inner alignment methods, which helps straightening flow paths. Experiments were conducted on cycle-graphs for continuous targets and on two molecular graphs for categorical targets.

Critical strengths

The explicit decomposition into node- and graph-level pairings provides a conceptual framework for graph flow matching.

Critical weaknesses

Presentation

  • The paper is overly complicated and notation-heavy compared to the presented ideas.
  • Some notations are used without proper definitions, so I can only infer the meaning of the symbols from the context. E.g. O(.)O(.) for orthogonal group, CvC_v and CeC_e for the channel dimensions, σ^GM−1⋅F\widehat\sigma_{\mathrm{GM}}^{-1}\cdot F or ρNgraph(σ)E\rho_N^{\mathrm{graph}}(\sigma) E for permuted feature matrices, etc.
  • The paper does not follow formatting instructions, such as "Paragraphs are separated by 1/2 line space." The fontsize in the table is too small.
  • Limitation of the paper is not discussed.
  • Related work is deferred to the Appendix, making difficult to position the paper within broader context. The related work section is also omits several graph generative paradigms.

Quality of experiments

  • Experimental evaluation is relatively weak. Cycle-graphs, while illustrative, is too simple, and Table 1 includes no baselines. Table 2 also compares within variants of the proposed method.
  • Table 3 only compares flow matching-based graph generative methods, omitting other generative paradigms. For example, GEEL [1] achieves better validity and FCD on both QM9 and Zinc250k datasets. While pushing the boundaries of flow matching may be meaningful, the authors do not justify this choice.

Significance

  • I think the originality of the paper is somewhat limited, as the outer alignment corresponds to minibatch OT, so the paper's novelty appears to lie primarily in the inner alignment formulation and its evaluation.
  • While feasible for small graphs, the compute cost could become a bottleneck for large graphs, as inner alignment step is required for each pair. As Zinc250k contains graphs with up to N=40N = 40 nodes, the authors could evaluate their model on larger graphs.

[1] Jang et al., A Simple and Scalable Representation for Graph Generation, ICLR, 2024.

Questions for authors

Can you provide accessible summary of your work, the significance of your paper, and why it matters for the broader graph generative models?

4 comments

We thank the reviewer for the detailed feedback and for appreciating the decomposition into node- and graph-level pairings. We respectfully disagree with some of the assessments and clarify them below. Please also see our general comment. Sections, figures and tables refer to our updated manuscript.

Q: Accessible summary and novelty

Graphs have no node order, so flow matching must decide which source node is interpolated to which target node. When graphs are interpolated in their tensor representation space (e.g. one-hot node and edge features), doing this correctly requires a quadratic assignment over edges and features. To our knowledge, no existing graph flow model accounts for this geometry, and we are the first to connect it to standard optimal transport on the quotient space. This results in equivariant flow matching with straighter paths, cheaper and better transport estimators (Prop. 4.1), and improved few-step generation.

Presentation

  • Notation. Thank you, we streamlined the notation. In the original version, the permutation action ρNgraph(σ)\rho^{\text{graph}}_N(\sigma) was introduced in the first displayed equation, and CvC_v, CeC_e as the dimensions of node and edge features. The orthogonal group O(⋅)O(\cdot) is standard notation, which we now also state explicitly. Moreover, we replaced ρNgraph(σ)E\rho^{\text{graph}}_N(\sigma)E by the simpler σ⋅E\sigma\cdot E and state all definitions where they are first used.
  • Formatting. We fixed the paragraph spacing and the table font sizes.
  • Limitations. We added a limitations paragraph to the conclusion (Sec. 6).
  • Related work. We clarified the positioning in the introduction and added Fig. 1, which illustrates the difference to random node orderings, canonicalization and outer OT. The extended related work remains in App. D.

Quality of experiments

  • Focus on the coupling. Our contribution is the coupling, not a new architecture or a molecule-specific model. Inner and outer alignment are drop-in replacements for established methods, namely continuous flow matching for continuous data and CatFlow-style endpoint prediction for categorical data. The natural baselines are thus the standard couplings under an identical architecture. In Tab. 1 and Tab. 2, "Random" is standard flow matching and "MBOT" is standard OT flow matching, so both tables do contain baselines.
  • Beyond cycle graphs. Besides QM9 and ZINC250k, we already had two continuous-SBM experiments (unconditional and conditional, three training seeds each, App. E), whose continuous targets cannot be modeled directly by discrete diffusion. We now also added GuacaMol (N≤88N \le 88, Tab. 2), where GW improves FCD over Random by ≈25%\approx 25\% at 5 steps. Moreover, we extended our method to 3D QM9 by adding rotations (Sec. 5.3, Tab. 4), where our quadratic GW+rotation alignment improves 5-step molecule stability over linear alignment from 0.62 to 0.75.
  • Other paradigms and GEEL. Tab. 3 already covers autoregressive models (GraphAF), normalizing flows (MoFlow), continuous diffusion (GDSS), discrete diffusion (DiGress), and discrete and categorical flow matching and flow maps (DeFoG, GGFlow, CatFlow, now also CFM-CSD). Thank you for pointing us to GEEL [1], an autoregressive model we were not aware of. We added it to Tab. 3, and it indeed reaches better absolute scores. However, we never claimed state-of-the-art results, and Tab. 3 provides context only, since network sizes and evaluation protocols differ (App. G). Under an identical setup, GW improves our CatFlow reimplementation on ZINC250k from 1.240 to 0.966 FCD.

Significance

  • Originality. We agree that outer alignment corresponds to minibatch OT and do not claim it as our contribution (see general comment). Our contributions are the quadratic assignment geometry of graphs for equivariant flow matching (Thm. 3.1, 3.2) and cheaper estimators that exploit it. Notably, this give better transport cost estimates than standard MinibatchOT at less compute (Prop. 4.1, Fig. 3).
  • Compute. Inner alignment runs on CPU, decoupled from gradient steps, since data-loader workers align batch k+1k+1 while the GPU trains on batch kk (Sec. 4.1). With spare CPU cores, this hides the alignment cost, also on GuacaMol with up to 88 nodes. In addition, BoA trades alignment quality for compute (Fig. 3), and on ZINC250k we observe a better trade-off between transport cost and compute than on QM9.

If our answers address your concerns, we would be grateful if you considered raising your score.

[1] Jang et al. A Simple and Scalable Representation for Graph Generation.

Dear Reviewer,

As the discussion period is coming to an end, we would greatly appreciate your feedback on our response and updated manuscript.

Thank you for your time!

I thank the authors for their rebuttal. I appreciate the extended experiments (3D molecules) and the new discussion on computational cost. However, I am not convinced that the method is 'competitive with recent graph generators across paradigms.' If the authors believe that evaluation protocols differ across methods, they should use more standardized benchmarks, such as GuacaMol and MOSES; while I found results on GuacaMol in Table 2, they are missing from Table 3. While the paper improves upon CatFlow (which offers transferable insights), the proposed method remains limited in scope to graph generation. The authors would benefit from more thorough experiments and a clearer presentation of the method's advantages compared to more general generative models.

We thank the reviewer for the follow-up and for acknowledging the 3D experiments and compute discussion. In response, we have just uploaded a further updated version of the manuscript and kindly refer the reviewer to this latest version.

GuacaMol in Tab. 3. We had already started the GuacaMol experiments at the beginning of the rebuttal period and have now added full-budget results to Tab. 3 of the updated manuscript, evaluated with the official GuacaMol distribution-learning suite (normalized FCD score):

MethodParadigmValidity ↑FCD score ↑
DiGressDiscrete diffusion85.268.0
DeFoGDiscrete flow matching99.073.8
GEELAutoregressive88.271.5
AutoGraphAutoregressive91.679.2
CatFlow (our impl.)Categorical flow matching98.971.8
GW-CatFlow (ours)Categorical flow matching98.678.3

Since all methods use the same official evaluation suite, these results provide a standardized comparison. GW-CatFlow outperforms all methods in FCD score except AutoGraph, which has substantially lower validity (91.6 vs. 98.6). Due to discussion-period time constraints, both our models were trained for only 100 epochs, leaving additional training-budget scaling unexplored.

Competitiveness. We believe our claim is supported across benchmarks. On ZINC250k, GW-CatFlow reaches the second-best FCD in Tab. 3 (0.966), behind only GEEL and ahead of all diffusion and flow models (VBFN 1.307, DeFoG 1.425, GGFlow 1.455, DiGress 3.482). Moreover, App. G (Tab. 8) verifies that these gains are not an artifact of the evaluation protocol: GW remains ahead of Random under all tested ZINC250k conventions (charge handling, reference-set size), and the dominant reporting difference, our smaller PyG test set, biases our absolute FCD upward (by ≈0.22), i.e., against us. On GuacaMol, it achieves the best FCD score among diffusion and flow models in our comparison and is within 0.9 points of autoregressive AutoGraph, with clearly higher validity. Moreover, across datasets and tasks, we consistently demonstrate the benefits of quadratic alignment for both standard and categorical graph flow matching in Euclidean space.

Scope. We are confused about the statement that we are "limited in scope to graph generation". Graph generation is the explicit focus of our paper, and it is a large and important field in its own right (molecular design, materials, networks). Its central difficulty is node-permutation symmetry, which is exactly what our coupling addresses. Our method also extends beyond pure permutations to graphs in 3D space, as shown by the joint permutation-rotation alignment on 3D QM9 (Sec. 5.3).

Experimental scope. Our experiments are deliberately controlled: within each coupling comparison, architecture, training, and seeds are kept identical, and results are averaged over three evaluations. We cover continuous targets (cycle graphs, conditional and unconditional SBM), categorical targets (QM9, ZINC250k, GuacaMol), and 3D molecules. For context, 7 of the 11 published baselines in Tab. 3 (GraphAF, MoFlow, GDSS, CatFlow, CFM-CSD, GGFlow, VBFN) do not seem to report GuacaMol results in their original papers.

We hope the additional standardized benchmark addresses the remaining concerns. Given the breadth of experiments and competitive results, we find it difficult to understand which specific shortcomings still justify the current score. If specific concerns remain, we would appreciate clarification. Otherwise, we would be grateful if the reviewer reconsidered their assessment.

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

The paper studies how source–target coupling should account for node-permutation symmetry in graph flow matching. It separates two decisions: inner alignment, which chooses node correspondences within a graph pair, and outer alignment, which chooses which source and target graphs to pair within a minibatch. The authors connect these choices to transport on a quotient space, establish results on lifting quotient couplings and symmetrizing the flow-matching objective, and extend the symmetry argument to categorical endpoint prediction. Experiments compare random coupling, minibatch OT, best-of-augmentation, and approximate Gromov–Wasserstein alignment on cycle graphs, stochastic block models, QM9, and ZINC250k. The clearest empirical result is that inner alignment improves generation with few Euler steps. Additional outer alignment gives more dataset-dependent benefits, and the differences after flow-map distillation are smaller and less consistent.

Critical strengths

Strengths

  1. The paper gives a useful mathematical explanation of alignment under equivariance. Choosing a target permutation based on the source can change the labelled target distribution, so its effect on training needs care. The quotient-space formulation and Theorem 3.2 explain how the quotient target distribution is preserved and how the equivariant training objective relates to the symmetrized coupling. This provides a nice foundation for the procedure, including its use with categorical endpoint prediction.
  2. The controlled experiments show meaningful improvements at low integration budgets. Keeping the architecture and training schedule fixed makes the effect of changing the coupling reasonably clear. At five Euler steps on ZINC250k, inner GW improves validity from 56.41% to 64.31% and FCD from 18.448 to 15.024. Adding outer GW increases validity to 72.80%, although FCD becomes slightly worse at 15.377 (Table 2a). The continuous-SBM experiments, which include three independent training seeds, strengthen the evidence beyond molecular validity. These results support the usefulness of alignment while showing that additional outer matching does not improve every metric.
Critical weaknesses

Weaknesses

  1. I had limited time to work through the mathematics in detail, so apologies if I am missing a distinction here. When reading Theorem 3.1, much of it seems to relate to the quotient-transport results of Galaz-García et al., On quotients of spaces with Ricci curvature bounded below. Their results seem to cover some of the same ground, particularly the relationship between quotient Wasserstein distances and invariant lifts, and the lifting of optimal couplings. I did not see this connection discussed in the paper. Could the authors explain which parts of Theorem 3.1 build on these results and what their formulation adds, especially when both the source representative distribution and the quotient coupling are prescribed? This would help me better understand the theoretical contribution.
  2. In the appendix I noticed the authors evaluate ZINC250k FCD against a 5,000-molecule PyG test split and specify largest-fragment extraction and deduplication. These details are clear for the proposed model, but if I understand the setup their compatibility with the quoted Table 3 baselines is not established. Moreover, the full-budget model adds capacity, longer training, structural features, self-conditioning, and other changes. Every row of the ablation already uses GW, and that ablation uses the smaller architecture and training budget. I appreciate that the paper explicitly distinguishes the controlled coupling study from the competitiveness benchmark. Still, an otherwise identical Random or MBOT control would establish whether alignment remains useful once the other additions are included.
Questions for authors
  • Which parts of Theorem 3.1 extend the quotient-transport results discussed above? In particular, what additional content follows from prescribing both the quotient coupling and the source representative distribution?
  • Were the quoted Table 3 FCD scores obtained with compatible reference splits, sample counts, fragment handling, and deduplication? If the protocols differ, can the main comparison be reconciled under a common protocol? Do you have an otherwise identical Random or MBOT control for the enhanced model? This would clarify alignment’s contribution under that setup, beyond what is established by the smaller controlled sweep.
  • What are the end-to-end training times for the actual Table 2 configurations? Figure 2 and Table 10 provide useful alignment-cost measurements, but a matched training-time comparison would help show when the additional overhead pays off. Do the smaller outer-alignment and distillation gains persist across independent training runs? Table 2 clearly labels its uncertainty as evaluation variability; training variability would help determine how much weight to put on the smaller differences.
2 comments

We thank the reviewer for the careful and constructive review, and for appreciating both the mathematical foundation and the controlled experiments. Sections, figures and tables refer to our updated manuscript.

W1/Q1: Relation to Galaz-García et al. (2018)

Thank you very much for pointing us to this important reference, which we had indeed missed in the first version. We now cite it in the introduction, the related work, after Thm. 3.1 and in App. A, and added the following remark:

Our construction builds on their quotient-Wasserstein geometry and lifting results. We explicitly formulate the cost-preserving lift for arbitrary quotient couplings, not only optimal ones, while keeping any prescribed source representative fixed and selecting a suitable target representative instead of symmetrizing both marginals. For optimal couplings, we additionally connect these lifts to Euclidean flow-matching trajectories and their kinetic energy.

Prescribing the source representative is exactly what flow matching needs. The source (e.g. a Gaussian) is fixed, and only the target representative can be chosen. Thm. 3.2 then shows that this non-invariant lift still yields the correct equivariant minimizer.

W2/Q2: Table 3 protocols and an identical Random control

  • Identical control. Table 3 now contains our CatFlow reimplementation (Random) under the same full-budget setup as GW-CatFlow, where only inner alignment differs (App. G). GW improves ZINC250k validity from 98.63% to 99.01% and FCD from 1.240 to 0.966, and QM9 FCD from 0.277 to 0.115.

  • Protocols. We do not aim for state-of-the-art claims, and Table 3 only provides context. ZINC250k papers differ in splits, charge decoding, fragment handling and deduplication, so a fully harmonized comparison is not possible. We follow CatFlow, the closest method to ours, which uses a PyG split in its repo [1] and its OpenReview supplement. We differ by predicting formal charges in {−1,0,+1}\{-1,0,+1\} so that charged molecules can pass valence checks. App. G (Tab. 8) quantifies all differences for ZINC250k:

    • Reference-set size has the largest effect and works against us. FCD is biased upward by roughly 1/nref1/n_{\text{ref}}, and our PyG test set has around 5 000 unique molecules vs. around 25 000 for GDSS. Fitting a+b/nrefa + b/n_{\text{ref}} on random PyG reference subsets and extrapolating to the GDSS size gives estimated FCDs of 1.036 (Random) and 0.752 (GW).
    • GDSS test split. Scoring directly on it gives FCD 1.009 (Random) and 0.724 (GW). This holds under data leakage since 88% of these molecules are in the PyG training split, but is consistent with our statistical correction above.
    • Charge conventions ("+1" or charge-free) shift FCD by only 0.06–0.11.
    • Deduplication (2–6 duplicates per 10,000 samples) has negligible effect.

    GW remains ahead of Random in all variants.

Q3: Training times and seed variability

  • Training times. Since our runs were performed on different machines and under varying load, we cannot provide reliable end-to-end runtime comparisons. Alignment runs on CPU and is decoupled from gradient steps, since data-loader workers align batch k+1k+1 while the GPU trains on batch kk (Sec. 4.1, "Parallelization"). On our primary workstation (AMD Ryzen 9 9900X, 12 cores, single RTX 5090), this hides the cost of inner GW alignment, and we observed no noticeable difference in wall-clock time to Random in our full-budget setting (Tab. 3). Outer GW alignment does add a noticeable overhead, which we discuss as a limitation. Per-pair CPU costs are in Fig. 3 and App. K.

  • Seed variability. We added training-seed repeats on QM9 for both the direct flow and the distilled flow maps (App. L, Tab. 14–15), and the gains persist. The two continuous-SBM experiments (unconditional and conditional) already use three training seeds each (App. E).

Thank you again for the helpful questions. If our answers address your concerns, we would be grateful if you considered raising your score.

[1] PyG split as in the CatFlow code: https://github.com/vincenttsai2015/VFM_CatFlow

Dear Reviewer,

As the discussion period is coming to an end, we would greatly appreciate your feedback on our response and updated manuscript.

Thank you for your time!

Reviewer_SRb7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses approximate Gromov-Wasserstein (GW) matching for inner (node) alignment on graph quotient space, and outer alignment (graph pairing) in equivariant flow matching. Results on 2D small molecular graph generation benchmarks show improement in few-step generation.

Critical strengths
  1. Sound formulation. The quotient construction separates equivariance from alignment; categorical marginal posteriors correctly recover the required conditional mean.

  2. Controlled gains. At fixed architecture/training settings, joint GW improves five-step ZINC validity from 56.41% to 72.80% and FCD from 18.448 to 15.377.

Critical weaknesses
  1. Lack of novelty and discussion on related work: the outer alignment through minibatch OT has been widely adopted (e.g., [1, 2]); the prior work on inner alignment of quotient-space perspective [3] and the connection to canonicalization [2] is also insufficiently discussed. Please explain your novelty.

  2. Theory: Appendix E's input noise changes the endpoint posterior, beyond Eq. 6's clean-path justification.

  3. Missing baselines and evaluations on large/3D benchmarks: there are substantially missing baselines, including [4] studying the same discrete distillation over the same task. There is no experiments on 3D molecules, whose SO(3) equivariance is not considered in this paper -- please show experiments on 3D benchmarks (e.g., QM9 and Geom-DRUGs) and compare with other distillation methods for 3D molecules, including but not limited to [5-7]. Even for 2D graphs, results on larger datasets including GuacaMol and MOSES are missing.

[1] SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching

[2] Rethinking diffusion models with symmetries through canonicalization with applications to molecular graph generation

[3] Quotient-space diffusion models

[4] Categorical flow maps

[5] EC-Conf: A ultra-fast diffusion model for molecular conformation generation with equivariant consistency

[6] Accelerating 3D molecule generative models with trajectory diagnosis

[7] FlashMol: High-quality molecule generation in as few as four steps

Questions for authors

See weaknesses

2 comments

We thank the reviewer for the careful review and for acknowledging both the sound formulation and the controlled gains of our approach. To avoid misunderstandings, we would first like to clarify that

  1. while we are aware of related work, to our knowledge nobody has focused on the quadratic assignment geometry that is relevant for continuously interpolating graphs with standard equivariant graph networks.
  2. we focus on general 2D graphs, not 3D molecules.
  3. our contribution is not distillation. We merely show that our flows are compatible with and potentially beneficial for distillation methods such as [4].

We keep the reviewer's numbering [1]–[7]. Sections, figures and tables refer to our updated manuscript.

W1: Novelty and related work

  • Outer alignment via minibatch OT [1, 2]. We agree that this is widely adopted and do not claim it as our contribution. For graphs, however, standard outer alignment alone is blind to the quadratic permutation geometry of graphs. Accounting for geometry allows for estimators that give better transport cost estimates than standard MinibatchOT at less compute (Fig. 3, newly added Prop. 4.1).

  • Inner alignment via permutation [1]. [1] permutes node orderings via linear assignment, whereas the quotient metric requires a quadratic assignment. With node-wise costs alone, two non-isomorphic graphs with same node features (e.g. same atoms, different bonds) have distance zero.

  • Canonicalization [2]. [2] couples either on the canonicalized space M:=Ψ(RD)M:=\Psi(\mathbb R^D), with Ψ\Psi constant on GG-orbits, or via a symmetrized independent coupling. Neither yields a process that descends to a Wasserstein geodesic on the quotient. Fig. 1 illustrates the difference, and App. D.1 gives an example where canonicalization is incompatible with a standard Gaussian source and an equivariant graph network.

  • Quotient-space diffusion [3]. [3] projects a GG-equivariant SDE to the quotient and constructs a canonical lift, which has shorter trajectories than the original one in the deterministic setting. However, no Wasserstein geodesic property on the quotient is established.

  • Our Thm. 3.1: Building on the quotient-Wasserstein geometry of [11] (App. A), we show that any coupling on the quotient, not only an optimal one, lifts with equal cost to a coupling of representatives starting from a prescribed, not necessarily invariant source. For optimal couplings, the lifted flow matching energy equals the quotient Wasserstein distance, giving a natural minimal-cost probability path on the quotient.

  • Our Thm. 3.2: Unlike [2], which requires training pairs in M×MM\times M or from a jointly GG-invariant coupling, Thm. 3.2 holds for any coupling in the lifted space: the optimal GG-equivariant vector field induces a curve on the quotient ending at the projected data distribution.

W2: Input noise in App. E (now App. G)

This is correct. However, most experiments do not use input noise, and its effect is ablated in App. G (Tab. 9). The noise only acts as a regularizer: for discrete data, interpolants are very sparse, which makes it hard for the flow matching network to learn to correct ODE errors. This is common for continuous-space methods that sample discrete data [9, 10]. We added this as a limitation to our conclusion.

W3: Benchmarking

  • Larger 2D datasets. We added GuacaMol (N≤88N \le 88) to Tab. 2: GW still improves FCD over standard flow matching ("Random") at all step budgets, e.g. by ≈25%\approx 25\% at 5 steps.

  • 3D benchmarks and [5–7]. Our focus is general 2D graph generation, not 3D molecules, e.g., [5–7] are not directly applicable to our continuous SBM experiments (App. E). Unlike our setting, SemlaFlow [1] uses discrete flows for atoms and bonds, so our quotient geometry does not apply directly. Still, the extension is possible by adding rotations (Sec. 4.2). Adapting SemlaFlow to continuous atom and bond paths on 3D QM9 (Sec. 5.3, Tab. 4), our quadratic GW+rotation alignment improves 5-step molecule stability over its linear, coordinate-only alignment from 0.62 to 0.75.

  • Distillation. Regarding [4], our contribution is not distillation. Better couplings are complementary to few-step distillation: [8] shows that semidiscrete OT couplings improve flow matching when using mean-flow models, i.e. few-step flow maps. With more steps (App. F), the benefit becomes clearer: on QM9, inner-outer GW improves FCD by ≈10%\approx 10\% over the default flow at 20 steps (Tab. 7). We added the results of [4] to Tab. 3.

We hope this clarifies our contribution and would appreciate a reconsideration of the score.

References

[8] A. Mousavi-Hosseini et al. Flow Matching with Semidiscrete Couplings.

[9] M. Ning et al. Input Perturbation Reduces Exposure Bias in Diffusion Models.

[10] CatFlow code: https://github.com/vincenttsai2015/VFM_CatFlow

[11] F. Galaz-García et al. On quotients of spaces with Ricci curvature bounded below.

Dear Reviewer,

As the discussion period is coming to an end, we would greatly appreciate your feedback on our response and updated manuscript.

Thank you for your time!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarification

We thank all reviewers for their instructive and helpful feedback. Since we have the impression that several reviewers are confused about the novelty of our contribution and its relation to the current literature, we feel the need to clarify a few things in addition to our more detailed answers below:

  1. Why is our approach novel? Because we clearly distinguish graph representations from the underlying graph. In particular, an optimal interpolation between two graphs in Euclidean space requires a quadratic node alignment jointly based on edge and features. Existing work focus either

    1. on matching representations instead of actual graphs [1, 2],
    2. node alignment via linear problem instead of a quadratic one, missing the edge alignment [3], or
    3. they just consider another group, e.g. rigid motions instead of node permutations [4].

    All of these approaches are potentially helpful, but the underlying geometry is not necessarily aligned with the geometry that an equivariant neural ODE sees -- unlike our approach. Note also that we do flow matching in continuous space, not discrete/categorical flow matching as e.g. in [5]. This is exactly what makes the Euclidean (GW) geometry of the interpolation the relevant one. We illustrated all this by updating* the manuscript with a new explanatory Figure 1.

  2. Is all this not a special case of equivariant flow matching? Sure, but

    1. we identify the relevant quadratic group actions for graphs, unlike related work,
    2. we show that equivariant flow matching is compatible with KL loss here, and
    3. graph alignment is an NP-hard quadratic alignment problem.

    Because NP-hard problems tend to be quite hard, we offer an important observation that is summarized in our Scaling-Figure**: We can get better optimal transport estimators for less compute than classical MinibatchOT due to equivariance. This is hard to quantify mathematically for GW-solvers since these have no convergence guarantee. However, we get a solid answer by generalizing the MinibatchOT estimator to our BoA+BoAoutBoA+BoA_{out}-estimator depending on variables AA and BB: linear scaling in AA vs. cubic scaling in BB.

    Long story short: Since increasing AA monotonically reduces transport cost, there is a point where our BoA-estimator dominates a standard MinibatchOT estimator. This is already the case in our experiments: e.g. (A,B)=(64,512)(A,B)=(64,512) has less compute time and lower transport cost than (A,B)=(1,8192)(A,B)=(1,8192), see Fig. 3 of the manuscript. Note that better minibatch OT estimators have been shown to improve flow matching [6].

    We made this more explicit by replacing the old remark with the new Proposition 4.1.

Thank you for your attention to this matter,

Authors

* FYI: Maybe you have to explicitly download the new version via right click, since some browsers cache the old version.

** old version: Fig. 2, updated version: Fig. 3

References

[1] X. Hou et al. GGFlow: A Graph Flow Matching Method with Efficient Optimal Transport. TMLR, 2026.

[2] K. Jiang, J. Cui, X. Dong, L. Toni. Bures-Wasserstein Flow Matching for Graph Generation. ICLR, 2026.

[3] SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching

[4] Quotient-space diffusion models

[5] Categorical flow maps

[6] S. Zhang, A. Mousavi-Hosseini, M. Klein, M. Cuturi. On Fitting Flow Models with Large Sinkhorn Couplings. TMLR, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.