acceptodds
Under review as a conference paper at ICLR 2027

H-Glow: Accelerating Graph Generation via Hazard-Guided Discrete Flow Matching

Abstract

Discrete flow matching (DFM) offers a promising approach to graph generation; however, generating high-quality graphs with few steps remains challenging due to the strong dependencies between nodes and edges. Although DFM decouples training from sampling, enabling flexible sampling strategies without retraining, existing graph DFM samplers often rely on task-specific heuristics, substantially expanding the hyperparameter space. Moreover, graph DFM models still require sampling budgets comparable to those of diffusion-based approaches, leaving their potential for accelerating graph generation only partially realized. In this work, we introduce H-Glow, a hazard-guided framework for accelerating graph generation built on the decomposition of Continuous-Time Markov chain (CTMC) generators into hazards and conditional jump distributions. H-Glow leverages these quantities to plan graph edits and dynamically adjust the sampling time to graph activity. H-Glow can be adapted to pretrained graph DFM models, substantially improving the generation quality with fewer steps. We further introduce a novel graph DFM model based on the conditional generator matching framework, which directly predicts hazards and conditional jumps, and show that its training objective decomposes exactly into hazard matching and hazard-weighted jump prediction. Experiments on molecular and synthetic graph benchmarks demonstrate improved generation quality with substantially fewer sampling steps.

Primary area:
generative models
Keywords:
discrete flow matching, discrete diffusion, graph generation, molecule generation, CTMC

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_6ks8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes H-Glow, a novel graph generation framework. The overall idea is to decompose component-wise (nodes and edges respectively) CTMC generators into exit hazards and conditional jump distributions. From my understanding, the main advantage of the decomposition is that, it can be used to improve the efficiency. Specifically, it can first select the component for update in each diffusion steps and then prioritize graph edits in specific areas, adapt model-query times, and stop sampling early. This design can sufficiently mitigate redundant computation budget. An important feature is that the sampler can be applied to pre-trained graph flow models. The paper also introduces a hazard–jump parameterization trained through a decomposed generator-matching objective. Finally, the experimental results validate the effectiveness of the design.

Critical strengths

Overall, I believe this paper proposes an interesting idea of first predict hazard, then jump, which can significantly boost the model efficiency. In addition to the novelty, the strengths include:

  • The sampler addresses a concrete practical problem: maintaining graph-generation quality under a small computation budget.
  • The idea can be seemly integrated into the framework of Defog, where the empirical evidences show that the improvement does not only come from retraining but also in extending from Defog.
  • The same-model sampler comparisons and component ablations support the empirical value of the sampling strategy.
  • The evaluation is rather comprehensive and convincing.
Critical weaknesses
  • Actual acceleration needs stronger measurement. Table 10 reports realized NFEs, appropriately accounting for early stopping. However, a convincing acceleration claim also requires matched-hardware wall-clock measurements, including graph-feature computation and sampling overhead. Quality–NFE and quality–runtime curves would be more informative than nominal step counts alone.

  • Molecular results generally lack uncertainty estimates. This matters when comparing small FCD differences and assessing the independent benefit of the new parameterization.

Questions for authors
  • The first question is for clarification. I wish to understand more clearly if the design can be utilized for improving training efficiency, or just inference efficiency?
  • While this paper is mainly focus on graph generation, the idea of hazard seems to be generalizable to other domains such as discrete diffusion (e.g. diffusion language models. Not sure if the authors consider extending such an idea to other fields, and what can be the challenges?
  • The end-to-end runtime seems missing and I am curious to see how much could be the improvement. What is the end-to-end runtime improvement under matched hardware and batch sizes?
  • Does the low-step advantage persist when the parallel edit budget is held fixed?
3 comments

We thank the reviewer for their feedback. In the following, we address their comments.

W1. Actual acceleration needs stronger measurement.

Q3. The end-to-end runtime seems missing

We report the sampling runtime (hours) for generating 10,000 molecules on GuacaMol and 25,000 on MOSES, using a single NVIDIA H100 80 GB GPU, with a batch size of 128. Parentheses report per-molecule mean NFE ±\pm standard deviation. We also provide quality-NFE and quality-runtime curves in our updated manuscript; see Appendix D.6. H-Glow terminates the generation before exhausting the entire sampling budget TT, while maintaining strong performance, unlocking accelerated sampling compared to prior discrete diffusion and flow approaches. Importantly, H-Glow is adaptive to the graph activity and adjusts the sampling trajectory to the graph hazard.

GuacaMol

MethodT=50T=100T=200
Cometh0.1492 (50$\pm$0)0.2960 (100$\pm$0)0.5893 (200$\pm$0)
DeFoG0.1485 (50$\pm$0)0.2963 (100$\pm$0)0.5932 (200$\pm$0)
H-Glow0.1236 (40.69$\pm$8.01)0.1622 (49.72$\pm$13.76)0.2856 (78.8$\pm$35.6)

MOSES

MethodT=50T=100T=200
Cometh0.1119 (50$\pm$0)0.2203 (100$\pm$0)0.4374 (200$\pm$0)
DeFoG0.1173 (50$\pm$0)0.2339 (100$\pm$0)0.4683 (200$\pm$0)
H-Glow0.0977 (36.03$\pm$6.29)0.1672 (62.24$\pm$20.08)0.2838 (85.97$\pm$33.35)

W2. Molecular results generally lack uncertainty estimates.

We report our results as the mean ±\pm standard deviation across five sampling runs for each dataset. Each run generated 10,000 molecules on QM9, 25,000 on MOSES, and 10,000 on GuacaMol. We updated Table 1 and Table 2 in the manuscript.

QM9

TValidityRelaxed validityUniquenessFCD
2099.1$\pm$0.199.5$\pm$0.194.8$\pm$0.30.37$\pm$0.01
5099.1$\pm$0.199.5$\pm$0.196.5$\pm$0.10.10$\pm$0.00

MOSES

TValidityUniquenessNoveltyFiltersFCDSNNScaf.
5089.7$\pm$0.1100.0$\pm$0.094.3$\pm$0.199.3$\pm$0.00.98$\pm$0.010.55$\pm$0.017.8$\pm$1.0
10089.8$\pm$0.2100.0$\pm$0.093.3$\pm$0.199.3$\pm$0.00.97$\pm$0.010.56$\pm$0.015.3$\pm$1.4
20091.8$\pm$0.299.9$\pm$0.093.1$\pm$0.299.5$\pm$0.11.07$\pm$0.020.58$\pm$0.0015.0$\pm$0.4

GuacaMol

TValidityV.U.V.U.N.KLDFCD Score
5094.8$\pm$0.294.3$\pm$0.294.3$\pm$0.293.8$\pm$0.275.5$\pm$0.5
10099.2$\pm$0.199.2$\pm$0.197.9$\pm$0.296.9$\pm$0.377.6$\pm$0.2
20099.4$\pm$0.099.3$\pm$0.098.0$\pm$0.197.1$\pm$0.177.3$\pm$0.3

Q1. I wish to understand more clearly if the design can be utilized for improving training efficiency, or just inference efficiency?

  • Our main goal is to achieve accelerated sampling for graph discrete flow matching approaches, which still require many sampling steps TT. We achieve this by proposing a novel sampler that efficiently leverages the model’s predicted hazards and jump distributions, achieving strong performance within a few NFEs.
  • To assess the training efficiency, we compare two settings: a denoiser network trained with cross-entropy and a hazard–jump decomposed model trained with the decomposed objective (Eq. 22). Both models use the same data splits, graph Transformer backbone, optimizer, and training hyperparameters. We initialize their shared parameters identically. For each model, we report the number of epochs and wall-clock training time required to first reach a validation validity of at least 99.5% on QM9 and 85.0% on MOSES using T=50 steps.

QM9: We train on a single NVIDIA A100 GPU (40 GB) and validate at every epoch using 10,000 molecules. Results are averaged over four independent training runs.

ModelEpochsTraining time (min)Validity
Hazard/jump; Decomposed KL18.00$\pm$1.414.71$\pm$0.4199.15$\pm$0.11
Denoiser; CE52.00$\pm$9.7213.37$\pm$2.3599.05$\pm$0.04

MOSES: We train on a single NVIDIA H200 GPU and validate at every epoch using 25,000 generated samples. Results are averaged over four independent training runs.

ModelEpochsTraining time (min)Validity
Hazard/jump; Decomposed KL13.50$\pm$1.50232.27$\pm$30.2385.13$\pm$0.15
Denoiser; CE15.25$\pm$1.30257.84$\pm$25.8285.10$\pm$0.15

Overall, we find that our model trained with the decomposed objective learns to generate valid molecules faster. We believe that investigating whether this extends to other benchmarks and domains is an interesting direction for future work.

Q2. While this paper is mainly focus on graph generation, the idea of hazard seems to be generalizable to other domains

Our framework is applicable beyond graphs: the hazard–jump decomposition is a general property of CTMC generators, including those describing the reverse process of continuous-time discrete diffusion models. One challenge is adapting the sampling-time allocation to the domain and generator. For example, Lee et al. [1] show that, in continuous flow language models over one-hot representations, larger vocabularies concentrate token resolution within a narrow time interval, motivating a time reparameterization that accounts for vocabulary size. This motivates investigating whether similar effects arise under discrete generators and how effectively hazard-guided clocks capture these critical regions. We believe this is a promising direction for future work.

[1] Lee, Chanhyuk, et al. "Flow map language models: One-step language modeling via continuous denoising." (2026).

Q4. Does the low-step advantage persist when the parallel edit budget is held fixed?

We evaluate H-Glow on Planar across several parallel budgets BB and sampling budgets TT (see Fig. 6). Strong low-step performance persists for sufficiently large fixed parallel budgets. With T=20T=20, H-Glow achieves high V.U.N. for B≥512B\geq512, while T=50T=50 and T=100T=100 achieve high V.U.N. with smaller parallel budgets such as B=128B=128. For small BB, low-step performance deteriorates, demonstrating a trade-off between the parallel budget and the number of sampling steps for large graphs.

We again thank the reviewer for their feedback and are happy to clarify any remaining questions or suggestions.

I thank the authors for their rebuttal. My concerns and questions are properly addressed, and I believe extending the algorithms to general discrete diffusion can be a promising direction. Overall, I stand by the acceptance of the manuscript and would like to maintain my score.

Reviewer_MVP7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Existing samplers rely on parallel sampling, which can lead to the violation of structural constraints, as graph components are highly dependent. The authors propose a hazard-guided sampling strategy for discrete flow matching, where transition rates (hazards) are used to select editing sites and the sampling time is dynamically adjusted. The authors derive two decompositions: one separating the graph generator into node and edge components, and another decomposing the generator into hazards and conditional jump distributions. Based on these decompositions, H-Glow proposes two selection rules and hazard-based time allocation. While the method is applicable to the pretrained models, the authors further propose training with an objective that directly parameterizes the generator. Experiments are conducted on three widely used molecular datasets (QM9, MOSES, GuacaMol), as well as on two synthetic graphs (Planar, Tree).

Critical strengths

(Clarity) The motivation of the paper is well-presented in the introduction. The technical details are solid, and the proofs can be found in the appendix. Figure 1 makes the method concrete, helping the reader understand the overall paper.

(Originality) To the authors' credit, the adaptive time selection, hazard–jump decomposition have been previously explored, as stated in the related work, while their graph specific adaptation and evaluation are new to this work. The proposed method appears to be a reasonable modification, although certain aspects rely on heuristics.

(Significance) Experimental results are extensive and encouraging. Across three molecular benchmarks, H-Glow obtains competitive molecular quality at a low sampling budget. Ablation also provides useful comparisons, including comparisons on parallel generation and selection rules. Experimental details are also provided in the appendix, which is relatively well documented. In addition, discrete flow matching and its application to graphs are important research directions.

Critical weaknesses
  • I think the paper will benefit more from explanatory and suggestive of the underlying intuition. Each mechanism is described, but the relationship between components could be explained more explicitly. Propositions are operationally important, but has limited novelty or almost immediate from the CTMC formulation, which could be stated more briefly.
  • The proposed training loss is only evaluated on QM9 and presented in the appendix. It feels too brief, given the amount of space allocated in the main text.
  • While the experimental setup is described in the appendix, the authors do not provide code.

Minor comments: Section 4 paragraph headings "Sampling complexity" and "Sampling efficiency" are almost the same. Maybe the authors intended sampling accuracy or similar for the second paragraph.

Questions for authors

Could you release the code/config?

4 comments

We thank the reviewer for their feedback.

W1. I think the paper will benefit more from explanatory and suggestive of the underlying intuition. Each mechanism is described, but the relationship between components could be explained more explicitly. Propositions are operationally important, but has limited novelty or almost immediate from the CTMC formulation, which could be stated more briefly.

We thank the reviewer for this suggestion. We agree that the relationships among the components should be explained more explicitly, and we have revised Section 4 to emphasize this more clearly.

We agree that the factorization in Proposition 1 follows directly from standard CTMC formulation. We instantiate it for graphs because it establishes the componentwise structure underlying our framework: it yields the total graph hazard in Proposition 2 and provides the basis for decomposing the graph generator-matching objective in Theorem 1. These propositions therefore serve as critical results connecting the CTMC formulation to our sampling and learning framework.

We also update Section 4.2 to clarify the different components of H-Glow. Building on the hazard-jump decomposition, we design our sampler to fully leverage the model’s own predictions for graph-edit planning. The predicted hazards and conditional jump distributions jointly guide the following decisions: (i) which nodes or edges to update, (ii) which states the selected components should transition to, (iii) at which time step to evaluate the model next, and (iv) when to terminate the generation trajectory.

We thank the reviewer again and are happy to take any further suggestions.

W2. The proposed training loss is only evaluated on QM9 and presented in the appendix. It feels too brief, given the amount of space allocated in the main text.

We would like to clarify that our main results use the H-Glow DFM model trained with the decomposed objective (Section 5). We also sample using the H-Glow sampler (Section 4). The corresponding results are presented on QM9 (Table 1), MOSES and GuacaMol (Table 2). To accelerate pretrained DFM models, we use the pretrained DeFoG model, which trains a denoiser network with the cross-entropy objective, and we sample from it with our H-Glow sampler. For the ablations section, we use the H-Glow model and ablate over different sampling settings. This includes the time distortions, design choices, and sampler (H-Glow or parallel Euler sampling). We apologize if this has caused any confusion.

We updated the manuscript to clearly distinguish between the "H-Glow model" and the "H-Glow sampler". "H-Glow" refers to our general framework that builds on the CTMC generator decomposition.

W3 & Q1. While the experimental setup is described in the appendix, the authors do not provide code.

We have made our code available and added a link to the revised manuscript.

Minor comments: Section 4 paragraph headings "Sampling complexity" and "Sampling efficiency" are almost the same. Maybe the authors intended sampling accuracy or similar for the second paragraph.

We thank the reviewer for their suggestion. We updated the heading of the second paragraph to "Sampling accuracy at low NFE".

Again, we thank the reviewer for their feedback and are happy to address any remaining questions.

Thank you to the authors for their thorough response. All of my concerns have been addressed.

Regarding Reviewer 38ej's point, I would also like to see a broader comparison of H-Glow, and flow matching in general, against other graph generative methods in the Related Work or Discussion section. Looking at the MOSES results in Table 12 of the SNACK paper [1], autoregressive models appear to outperform H-Glow by a large margin. I understand there are many differences that make a direct comparison difficult and do not expect flow matching to surpass those numbers. I would like to hear the authors' thoughts on the advantages of applying flow matching to graph generation, and related perspectives such as whether they view it as a promising direction for future research.

Also, extending the Related Work section to include other graph generative models will greatly help readers better understand the work in a broader context.

[1] SNACK: A Sequential Notation Framework for Probabilistic Graph Generation

.. I would also like to see a broader comparison of H-Glow, and flow matching in general, against other graph generative methods in the Related Work or Discussion section.

We thank the reviewer for this suggestion. We extended the Related Work section to provide a broader overview of graph generative approaches. We categorize these into two classes: autoregressive approaches, which generate nodes and edges sequentially, and one-shot methods that jointly generate or iteratively refine graph components. We also discuss complementary approaches that use string-based graph representations for autoregressive modeling.

Please refer to the updated manuscript.

Looking at the MOSES results in Table 12 of the SNACK paper [1], autoregressive models appear to outperform H-Glow by a large margin. I understand there are many differences that make a direct comparison difficult and do not expect flow matching to surpass those numbers. I would like to hear the authors' thoughts on the advantages of applying flow matching to graph generation, and related perspectives such as whether they view it as a promising direction for future research.

We agree with the reviewer that a direct comparison is difficult. There are two orthogonal design dimensions: Graph representation (nodes/edges vs string-based representations), and the generative process itself (autoregressive vs one-shot). SNACK combines graph serialization with autoregressive modeling, whereas H-Glow applies discrete flow matching directly to node and edge attributes. We also would like to highlight that SNACK incorporates explicit decoding constraints that mask tokens violating syntactic rules, chemical valency, or BFS ordering, injecting structural and chemical prior knowledge directly into generation and contributing to improved generation quality. In contrast, H-Glow relies on learned hazard and conditional jump distributions to guide iterative refinement of node and edge attributes, without explicitly enforcing chemical or structural constraints during sampling.

In general, we believe flow matching is a promising direction for graph generation. First, it supports iterative refinement of the graph: node and edge predictions can be revisited using the evolving graph context, providing opportunities to correct intermediate structures. Second, operating directly on graphs naturally accommodates permutation-equivariant architectures and avoids the need for a predefined graph serialization order. These properties are also shared with graph diffusion models and provide useful inductive biases. An important motivation for discrete flow matching is its continuous-time formulation, which allows sampling strategies and computational effort along the generation trajectory to be adapted using a single trained model. H-Glow exploits this flexibility through predicted hazards that guide component selection and sampling-time allocation, improving the quality–computation trade-off of pretrained graph DFM models without retraining or auxiliary networks. Promising future directions include incorporating structural constraints into iterative graph refinement, better coordinating parallel updates, and developing sparse representations that avoid explicitly modeling all node pairs.

We again thank the reviewer for their thoughtful feedback and are happy to take any further suggestions.

Thank you for your response. The authors have properly addressed my concerns. The paper appears to be novel and can be applied more generally. I stand by my recommendation for acceptance.

Reviewer_38ejOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes H-Glow, a hazard-guided discrete flow matching framework for efficient graph generation. The method decomposes CTMC generators into hazards and conditional jump distributions, which are then used to guide graph edits and adapt the sampling schedule. The authors also propose a corresponding generator-learning objective. Experiments on molecular and synthetic graph datasets show that H-Glow achieves competitive generation quality with fewer sampling steps.

Critical strengths
  1. Graph generation is an important problem, particularly for scientific applications such as molecular design and drug discovery, where both generation quality and computational efficiency matter. The paper addresses a relevant problem in graph-based discrete flow matching, aiming to improve generation quality under limited sampling budgets.

  2. The proposed method is technically solid, with well-developed mathematical formulations and theoretical derivations. In particular, the use of hazard-jump decomposition to guide graph edits and adapt sampling time provides a coherent foundation for the overall framework.

  3. The experimental evaluation is comprehensive, covering both molecular and synthetic graph benchmarks, with comparisons against several representative baselines. The results are encouraging, especially in low-step generation settings. The ablation studies also provide useful evidence for the effectiveness of the main design choices.

Critical weaknesses
  1. I find the motivation somewhat unclear. Several recent works, such as SimGFM, have already explored efficient discrete generation and hazard-based modeling. While the paper discusses some of these studies, I would like to see a clearer explanation of what is still missing in existing approaches and how H-Glow specifically addresses that gap.

  2. There has also been growing interest in autoregressive and LLM-based graph generation, including G2PT and Llamole. Although these methods follow different paradigms, it would be useful to discuss how H-Glow compares with them and in which scenarios the proposed approach might be preferable.

  3. The experimental evaluation is generally solid, but a few additional analyses would make it more convincing. In particular, wall-clock runtime comparisons would provide a clearer picture of the actual efficiency gains. It would also be helpful to examine whether the hazard-guided sampling strategy reduces invalid or conflicting graph updates, as suggested by the motivation.

Questions for authors
  1. What is the main technical distinction between H-Glow and existing hazard-based or adaptive sampling approaches?

  2. How does H-Glow compare with recent autoregressive or LLM-based graph generators in terms of generation quality and efficiency, where the task settings are comparable?

3 comments

We thank the reviewer for their feedback. In the following, we address their comments.

W1. I find the motivation somewhat unclear.

Q1. What is the main technical distinction between H-Glow and existing hazard-based or adaptive sampling approaches?

The main motivation of our framework is to fully leverage the model's predictions to guide graph generation. Building on the hazard-jump decomposition, we design the H-Glow sampler to leverage the predicted hazards and conditional jump distributions for graph-edit planning. These quantities guide the following decisions: which nodes or edges to update, which states the selected components should transition to, when to evaluate the model next, and when to terminate the sampling trajectory, thereby allocating a limited sampling budget more effectively. Our framework substantially reduces the methodological complexity of existing DFM approaches and enables accelerated graph generation.

Existing graph discrete diffusion and flow models rely on parallel samplers. During every generation step, multiple components are updated in parallel. This requires graph DFM models to perform multiple sampling steps to capture correlations between nodes and edges, which limits the possibility of performing low-step generation while maintaining good sample quality. Existing DFM approaches also exploit training-sampling decoupling to improve generation. DeFoG explores stochasticity adjustments, target guidance, and sampling-time distortions, while SimGFM improves efficiency through its rate formulation and scheduler design. These approaches demonstrate the importance of sampling choices. However, they increase methodological complexity by introducing multiple hyperparameters that are tuned for each task and sampling budget TT.

We updated our manuscript to emphasize the motivation of our framework. Please refer to Section 3, where we discuss limitations of existing works, and Section 4, where we highlight the relevance of H-Glow for guiding graph generation.

W2. ..it would be useful to discuss how H-Glow compares with them and in which scenarios the proposed approach might be preferable.

We compare H-Glow to autoregressive (AR) approaches for graph generation in the following:

  • Refinement: AR methods generate tokens sequentially and commit to earlier predictions [1,2]. H-Glow starts from a prior graph and iteratively refines node and edge attributes, allowing earlier predictions to be revised.

  • Generation order: G2PT generates nodes before edges using a prescribed ordering strategy, while AutoGraph uses traversal-based sequences. H-Glow does not impose a predefined order for node or edge generation. H-Glow selects components according to predicted hazards, adapting the refinement order to the evolving graph.

  • H-Glow employs a permutation-equivariant graph Transformer with shared node and edge prediction heads. Its training objective is invariant to node relabeling. During inference, hazard-guided component selection and conditional jump sampling preserve permutation equivariance, while the hazard-guided time allocation and early stopping depend only on permutation-invariant graph-level quantities. Together with a permutation-invariant prior, these properties yield a permutation-invariant distribution over generated graphs. In contrast, AR Transformer decoders do not explicitly enforce node-permutation equivariance.

  • Discrete diffusion and flow-based graph models typically represent graphs using dense adjacency matrices and iteratively denoise node and edge attributes. This requires quadratic memory complexity in the number of nodes, limiting their scalability to large, sparse graphs. AR methods [2] serialize attributed graphs into sequences and employ decoder-only Transformers, offering a scalable alternative for generating sparse graphs.

We note that Llamole uses an AR LLM to generate text and decide when to generate a molecule; the molecular graphs themselves are generated by a graph diffusion model.

W3. .. wall-clock runtime comparisons would provide a clearer picture of the actual efficiency gains.

We report sampling runtime (hours) for 25,000 MOSES and 10,000 GuacaMol molecules on a single NVIDIA H100 80 GB (batch size 128). Parentheses show per-molecule NFE (mean ±\pm std). H-Glow adapts each trajectory to the total graph hazard, terminating the generation trajectory before exhausting the entire sampling budget TT, while maintaining a strong performance.

GuacaMol

MethodT=50T=100T=200
DeFoG0.1485(50$\pm$0)0.2963(100$\pm$0)0.5932(200$\pm$0)
H-Glow(Ours)0.1236(40.69$\pm$8.01)0.1622(49.72$\pm$13.76)0.2856(78.8$\pm$35.6)

MOSES

MethodT=50T=100T=200
DeFoG0.1173(50$\pm$0)0.2339(100$\pm$0)0.4683(200$\pm$0)
H-Glow(Ours)0.0977(36.03$\pm$6.29)0.1672(62.24$\pm$20.08)0.2838(85.97$\pm$33.35)

Q2. How does H-Glow compare with recent autoregressive or LLM-based graph generators in terms of generation quality and efficiency..?

We compare H-Glow with recent AR approaches [1,2], and present results in the table below. We evaluate H-Glow and AutoGraph on the Planar dataset. We use a H100 GPU 80GB and a batch size of 40. Results are reported as mean ±\pm standard deviation over five seeds, each generating 40 graphs. With only 20 sampling steps, H-Glow outperforms AutoGraph on both V.U.N. and Ratio metrics. H-Glow (T=20) also has a lower sampling time. We are unable to re-evaluate G2PT as they do not provide their checkpoints on synthetic datasets.

We update our manuscript to include AR baselines (see Table 14).

SettingPlanar V.U.N. ↑\uparrowPlanar Ratio ↓\downarrowPlanar sampling time (s) ↓\downarrowTree V.U.N. ↑\uparrowTree Ratio ↓\downarrow
G2PT-small*95.0––99.0–
G2PT-base*100.0––99.0–
AutoGraph85.0$\pm$2.72.2$\pm$0.61.451$\pm$0.016––
H-Glow T=2095.5$\pm$2.91.8$\pm$0.21.154$\pm$0.00193.5$\pm$4.62.2$\pm$0.8
H-Glow T=5099.0$\pm$1.21.8$\pm$0.52.874$\pm$0.00198.0$\pm$2.92.7$\pm$0.9

(*) Results are taken from [1].

We thank the reviewer for their feedback and are happy to address any remaining questions.

[1] Chen, Xiaohui, et al. "Graph generative pre-trained transformer." ICML 2025.

[2] Chen, Dexiong, et al. "Flatten graphs as sequences: Transformers are scalable graph generators." NeurIPS 2025.

Reviewer_38ejOpenReview ↗

Official Comment

The rebuttal is meaningful; therefore, I maintain my original positive opinion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.