acceptodds
Withdrawn from ICLR 2027

GD-Mamba: Adaptive Cross-Domain Alternating Graph Diffusion Model with Selective State Space Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph generation is a critical task in deep learning, aimed at learning and modelling the structural distribution of graphs, with broad applications in areas such as social network analysis and molecular design. Current mainstream graph generation methods adopt diffusion models as the foundational architecture. Graph data typically consists of two modalities: the discrete graph modality based on topological structures, and the continuous latent space modality based on vector representations. Diffusion models for the graph modality focus on denoising discrete nodes and edges and lack unified control over the global structure, resulting in jump-like changes in the generation process. In contrast, diffusion models for the latent space modality perform denoising in a continuous space, resulting in gradual changes throughout the generation process. To incorporate global semantic guidance from the latent space modality into graph modality denoising, while injecting discrete priors from the graph modality into latent space denoising, we propose GD-Mamba, an Adaptive Cross-Domain Alternating Graph Diffusion Model with Selective State Space Model. GD-Mamba alternately performs denoising between the continuous and discrete modalities, namely the graph and latent space modalities. To stabilize the dual-modality generation process, a two-level Mamba structure (G-Mamba and S-Mamba) is designed to align features across both modalities. Additionally, a multiscale wavelet positional encoding attention (MSW-PE) mechanism is introduced in the cross-domain mapping between both modalities. This mechanism effectively extracts the core structure of graphs by locating neighbourhoods of nodes and edges at different frequency scales, improving the accuracy and robustness in the alternating denoising process. The experimental results demonstrate that the proposed GD-Mamba method has superior performance on both generic graph datasets and molecular graph datasets.

Primary area:
generative models
Keywords:
Graph Generation, Diffusion Models, Deep Generative Models

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

This paper presents GD-Mamba, a graph generation framework that alternates denoising between the graph domain and a continuous latent space. A graph-latent mapping module (GLMM) encodes node and edge features into latent representations and decodes them back into graphs. The framework uses a graph-domain Mamba denoiser (G-Mamba) and a latent-space diffusion model with selective state-space blocks (S-Mamba/SSDM); intermediate latent features are used to condition graph denoising. The authors also introduce multiscale wavelet positional encoding (MSW-PE) in the GLMM attention layers.

Critical strengths

1.The paper addresses a relevant challenge in graph generation: combining discrete structure topology and attributes with continuous representations 2. The overall structure of the paper is quite clear.

Critical weaknesses

1.This paper is extremely poorly written and disorganized. For instance, "where changes occur as discrete jumps between entities" The internal "entity" is usually used for kg nodes rather than graph nodes. Moreover, the term "discret jump" is very confusing. I understand it to refer to a kind of edge that is more like a discrete transition probability. However, they often represent regular changes rather than sudden mutations. Additionally, DDPM_reverse does not mention citing the DDPM paper, but the author only states that it is the inverse process of the diffusion model. However, there are many different versions of diffusion model. Therefore, this is not rigorous. Also, S-Mamba/SSDM appear multiple times, making it difficult to determine exactly what they should be called. For models, a mathematical symbol rather than a name is usually used to represent them, and the author's use of symbols is also very poor. 2.The figures in the paper are difficult to read. For instance, in Fig. 1, the font size of the formula on the left side is too small, and the color combination on the right side is very unclear. 3.The baselines compared in the paper are all extremely outdated. They are mainly published before 2024. All related papers for graph generation published at 2025 and 2026 did not make any comparisons. Overall, I think there are very serious problems in all aspects of this paper and it does not meet the standards of ICLR.

Questions for authors

See Weakness.

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

This paper proposes GD-Mamba, a graph generation framework that alternates denoising between discrete graph and continuous latent spaces. It introduces a two-level Mamba architecture to coordinate the denoising processes and a multiscale wavelet positional encoding mechanism (MSW-PE) to preserve structural information during cross-domain mapping. Experiments on general and molecular graph datasets demonstrate competitive generation quality, particularly in molecular validity.

Critical strengths
  1. Interesting motivation for dual-space graph generation. Combining discrete graph-space denoising with continuous latent-space denoising is a promising direction, as the two representations offer complementary advantages in preserving structural details and capturing global information. The alternating design provides an intuitive way to exploit this complementarity.
  2. Promising empirical results. GD-Mamba achieves strong performance across general and molecular graph generation benchmarks. In particular, its high molecular validity and competitive distribution-matching metrics suggest that the proposed framework can generate high-quality graphs.
Critical weaknesses
  1. The novelty of the alternating denoising mechanism is insufficiently established. Related works, such as TreeDiff and CGDM, have already explored combining graph-space structural refinement with latent-space diffusion. Although their specific mechanisms differ, the paper does not adequately distinguish its core idea from these existing approaches. A clearer explanation of what is fundamentally new in GD-Mamba, beyond the particular architectural combination, would be helpful to establish its contribution.
  2. The central design lacks sufficient justification beyond an empirical heuristic. Even if the proposed alternating procedure is considered a distinct contribution, the paper does not yet provide a convincing explanation of why this specific design is necessary or generally effective. In particular, the claimed role of the two-level Mamba architecture in aligning denoising processes remains largely intuitive, while the existing ablations mainly demonstrate performance sensitivity to step allocation and improved mapping reconstruction. Stronger evidence that isolates the benefits of alternating denoising and Mamba-based coordination would help elevate the proposed technique from an interesting heuristic to a more substantial methodological contribution.
Questions for authors
  1. How does the proposed alternating denoising mechanism fundamentally differ from existing dual-space approaches such as TreeDiff and CGDM? What is the key methodological contribution beyond combining existing components?
  2. What specifically enables the two-level Mamba architecture to align discrete and continuous denoising processes? Could the authors clarify the underlying mechanism and provide evidence that this coordination is beneficial compared with simpler alternatives?
  3. Could the authors provide a clearer specification of the alternating sampling procedure, including the graph-space noise process, synchronization of diffusion steps, and conditioning information exchanged between the two denoisers?
Reviewer_iJMSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GD-Mamba alternates denoising between graph representations and a continuous latent space. Two Mamba modules coordinate through latent features, with multiscale wavelet positional encoding in the mapping module. Experiments cover four generic graph datasets and the QM9 and ZINC250k molecular datasets.

Critical strengths

Section 3.1, Eqs. (1)–(3), explicitly couples the two representations through intermediate features. Table 1 achieves the lowest average MMD on three generic graph datasets. Table 2 reports high validity without valence correction and improves several distribution metrics over the latent-only variant.

Critical weaknesses
  1. Section 3 insufficiently defines the graph forward process, discretization rules, and noise-time correspondence across spaces.
  2. These omissions prevent checking training–sampling consistency and require complete algorithms and noise schedules.
  3. Section 4.4 ablates positional encoding and step allocation, but does not isolate Mamba coordination through architecture controls with matched frameworks and budgets.
Questions for authors
  1. How are graph states represented, and what forward noise process and discretization rules are used?
  2. How are encoded graph states matched to latent diffusion times, and how are conditioning features obtained during the initial graph stage?
  3. Do existing experiments replace Mamba or remove conditioning while retaining the same alternating framework?

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

Related papers

Discussion (0)

Sign in to comment.