Research map / Graph learning / Graph self-supervised learning
Diffusion-based graph generation
Generate graphs with discrete and score-based diffusion and flow matching.
138 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Graph self-supervised learning, keyword cluster “molecular · generation · molecules”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
35 papers in 2023–24 and 77 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 10.2% to 13.2%.
Open questions the data points to
- Planar (V.U.N.): best went from 90 (up to 2024) to 100 in 2025 — Graph Generative Pre-trained Transformer (ICML 2025 · 1 citation)
- SBM (V.U.N.): best went from 85 (up to 2024) to 100 in 2025 — Graph Generative Pre-trained Transformer (ICML 2025 · 1 citation)
- Papers report DiGress on QM9 (FCD) anywhere from 0.095 to 0.36: comparisons against it depend on whose number you take.
- Papers report DiGress on Planar (orbit MMD) anywhere from 0.0009 to 0.008: comparisons against it depend on whose number you take.
What looks settled
- ZINC250k (FCD): best up to 2024 1.93, best in 2025–26 2.116 from 3 papers, at most 0.2% better, if at all.
- Planar (Orbit MMD): best up to 2024 0, best in 2025–26 0 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: Grid (Orbit MMD, last in 2024).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Regularized Diffusion Process for Visual Retrieval (AAAI 2017 · 66 citations)
- 2018 · GraphGAN: Graph Representation Learning With Generative Adversarial Nets (AAAI 2018 · 604 citations)
- 2019 · Efficient Graph Generation with Graph Recurrent Attention Networks (NeurIPS 2019 · 168 citations)
- 2020 · Reinforcement Learning Based Graph-to-Sequence Model for Natural Question Generation (ICLR 2020 · 102 citations)
- 2021 · HDMapGen: A Hierarchical Graph Generative Model of High Definition Maps (CVPR 2021 · 45 citations)
- 2022 · Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations (ICML 2022 · 25 citations)
- 2023 · DiGress: Discrete Denoising diffusion for graph generation (ICLR 2023 · 73 citations)
- 2024 · Hyperbolic Graph Diffusion Model (AAAI 2024 · 9 citations)
- 2025 · GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph Generation (AAAI 2025 · 9 citations)
- 2026 · FLAG: Foundation model representation with Latent diffusion Alignment via Graph for spatial gene expression prediction (ICML 2026 · 1 citation)
Earliest papers here
- Regularized Diffusion Process for Visual Retrieval (AAAI 2017 · 66 citations)
- GraphGAN: Graph Representation Learning With Generative Adversarial Nets (AAAI 2018 · 604 citations)
- GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models (ICML 2018 · 478 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ZINC250k · FCD (lower is better)
9 papers, 2022–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 14.66 | Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations (ICML 2022 · 25 citations) |
| 2024 | 1.93 | Graph Generation with $K^2$-trees (ICLR 2024 · 0 citations) |
QM9 · Validity (%) (higher is better)
8 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 96.95 | Multi-resolution Spectral Coherence for Graph Generation with Score-based Diffusion (NeurIPS 2023 · 0 citations) |
| 2024 | 100 | A Simple and Scalable Representation for Graph Generation (ICLR 2024 · 0 citations) |
Planar · V.U.N. (higher is better)
6 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 90 | Graph Generation with Diffusion Mixture (ICML 2024 · 0 citations) |
| 2025 | 100 | Graph Generative Pre-trained Transformer (ICML 2025 · 1 citation) |
SBM · V.U.N. (higher is better)
6 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 74 | DiGress: Discrete Denoising diffusion for graph generation (ICLR 2023 · 73 citations) |
| 2024 | 85 | Graph Generation with Diffusion Mixture (ICML 2024 · 0 citations) |
| 2025 | 100 | Graph Generative Pre-trained Transformer (ICML 2025 · 1 citation) |
Grid · Orbit MMD (lower is better)
5 papers, 2018–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 0.0001 | GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models (ICML 2018 · 478 citations) |
| 2024 | 0 | Graph Generation with $K^2$-trees (ICLR 2024 · 0 citations) |
Planar · Orbit MMD (lower is better)
5 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0 | Graph Generation with $K^2$-trees (ICLR 2024 · 0 citations) |
What has been tried
81 papers here report 810 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 7%, hurt 19%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove edge perception enhancement | GSDiff ablation · FID | 4.83 -> 4.96 | Hurt | GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph Generation (AAAI 2025 · 9 citations) |
| Add a Euclidean autoencoder embedding before GDSS diffusion | QM9 · FCD | 2.813 -> 2.615 | Helped | Hyperbolic Graph Diffusion Model (AAAI 2024 · 9 citations) |
| use exponential rather than normal distribution for graph-level latent space | Twitter16 · MMD² | 0.207 -> 0.134 | Helped | DAG-Aware Variational Autoencoder for Social Propagation Graph Generation (AAAI 2024 · 8 citations) |
| Remove 1D CNN edge feature extraction | B-rep models up to 120 faces (50K models) · Valid (%) | 58.8 -> 50.3 | Hurt | BrepGiff: Lightweight Generation of Complex B-rep with 3D GAT Diffusion (CVPR 2025 · 6 citations) |
| Vary FairWire λ during link prediction training | Cora link prediction · ΔSP | 27.01 ± 1.38; 16.17± 4.50; 4.76± 0.24; 4.54± 0.24 | Helped | FairWire: Fair Graph Generation (NeurIPS 2024 · 4 citations) |
| use quantized rather than unquantized latent nodes | QM9 · FCD | 1.207 -> 0.201 | Helped | GLAD: Improving Latent Graph Generative Modeling with Simple Quantization (AAAI 2025 · 3 citations) |
| add autoregressive block generation (K_h=1) instead of full diffusion without AR | QM9 · Validity | 93.8 -> 97.1 | Helped | Pard: Permutation-Invariant Autoregressive Diffusion for Graph Generation (NeurIPS 2024 · 3 citations) |
| Add radial popularity diffusion constraint | Three real-world datasets · node classification performance | Helped | Hyperbolic Geometric Latent Diffusion Model for Graph Generation (ICML 2024 · 2 citations) | |
| reduce sampling steps from 500 to 30 | QM9 · Valid (%) | 99.3 ± 0.6 -> 97.9 ± 1.2 | Hurt | Discrete-state Continuous-time Diffusion for Graph Generation (NeurIPS 2024 · 2 citations) |
| Add the residual hybrid layer | Beijing · Accessibility | up to a 105.1% improvement | Helped | FAP-CD: Fairness-Driven Age-Friendly Community Planning via Conditional Diffusion Generation (AAAI 2025 · 1 citation) |
| Add SMTP pre-training instead of no pre-training with decoder | PCQM4Mv2 valid · MAE | 0.0978 -> 0.0807 | Helped | GraphGPT: Generative Pre-trained Graph Eulerian Transformer (ICML 2025 · 1 citation) |
| pre-train G2PT on GuacaMol before MoleculeNet fine-tuning | MoleculeNet average · ROC-AUC | 64.9 -> 73.3 | Helped | Graph Generative Pre-trained Transformer (ICML 2025 · 1 citation) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| DiGress | QM9 | FCD | 0.095 – 0.36 | 5 papers, e.g. ICML 2024 and AAAI 2025 |
| DiGress | Planar | orbit MMD | 0.0009 – 0.008 | 2 papers, e.g. ICML 2024 and ICLR 2024 |
| DiGress | Planar | Clus. MMD | 0.0178 – 0.0372 | 2 papers, e.g. ICLR 2025 and ICML 2024 |
| GDSS* | ZINC250k | FCD | 14.66 – 16.62 | 2 papers, e.g. ICML 2023 and AAAI 2024 |
| DiGress | Planar graphs | V.U.N. | 70.02 – 77.5 | 2 papers, e.g. NeurIPS 2024 and ICLR 2024 |
| DiGress | QM9 | Novelty | 33.4 – 35.31 | 2 papers, e.g. ICML 2023 and NeurIPS 2023 |
What reviewers will expect
Datasets and baselines named in the 51 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- QM9 · 13 of 51 papers
- ZINC250k · 7 of 51 papers
- Planar · 7 of 51 papers
- SBM · 7 of 51 papers
- MOSES · 6 of 51 papers
- Community-small · 4 of 51 papers
- Ego-small · 4 of 51 papers
- Proteins · 4 of 51 papers
Baselines
- DiGress · 16 of 51 papers
- GraphRNN · 11 of 51 papers
- GDSS · 11 of 51 papers
- GruM · 7 of 51 papers
- GRAN · 6 of 51 papers
- GraphDF · 5 of 51 papers
- MoFlow · 5 of 51 papers
- SPECTRE · 5 of 51 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 112 papers; from OpenAlex and the papers' affiliations).
- Pohang University of Science and Technology · 5 papers
- Renmin University of China · 4 papers
- Amazon · 4 papers
- Swiss Federal Institute of Technology Lausanne · 3 papers
- Technische Universität München · 3 papers
- Northeastern University · 3 papers
- EPFL · 3 papers
- Microsoft · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Northwestern Polytechnical University · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Molecular graph pretraining (101 papers)
Elsewhere in the topic
- Semi-supervised node representation learning (214 papers)
- Heterophily and imbalance on graphs (178 papers)
- LLMs and graph foundation models (117 papers)
- Graph anomaly detection (101 papers)
- Graph domain adaptation (82 papers)
- Brain and biological graph learning (78 papers)
- Multi-view graph clustering (70 papers)
- Graph contrastive learning analysis (61 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).