- 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
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 and sampling budgets (see Fig. 6). Strong low-step performance persists for sufficiently large fixed parallel budgets. With , H-Glow achieves high V.U.N. for , while and achieve high V.U.N. with smaller parallel budgets such as . For small , 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.
We thank the reviewer for their feedback. In the following, we address their comments.
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 ± 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 T, 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
MOSES
We report our results as the mean ± 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
MOSES
GuacaMol
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.
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.
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.