Research map / Graph learning / GNN architectures
Structural GNN variants
Develop new graph architectures using geometry, trees, templates, and commute-time structure.
65 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in GNN architectures, keyword cluster “message passing · equivariant · mpnns”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 4 of 6 tracked benchmarks.
29 papers in 2023–24 and 21 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.1% to 4.1%.
Open questions the data points to
- Citeseer (accuracy): best went from 77.33 (up to 2024) to 81.88 in 2025 — Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations)
- Cora (accuracy): best went from 87.93 (up to 2024) to 90.82 in 2025 — Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations)
- PubMed (accuracy): best went from 89.96 (up to 2024) to 91.36 in 2025 — Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations)
- CS (accuracy): best went from 95.51 (up to 2024) to 96.58 in 2025 — Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations)
No longer reported by papers here: ZINC (MAE, 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
- 2020 · Random Walk Graph Neural Networks (NeurIPS 2020 · 50 citations)
- 2021 · Rethinking Graph Neural Architecture Search From Message-Passing (CVPR 2021 · 49 citations)
- 2022 · Equivariant Graph Mechanics Networks with Constraints (ICLR 2022 · 18 citations)
- 2023 · Graph Ordering Attention Networks (AAAI 2023 · 36 citations)
- 2024 · Physics-Informed Graph Neural Networks for Water Distribution Systems (AAAI 2024 · 28 citations)
- 2025 · Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · 25 citations)
- 2026 · Permutation Equivariant Framelet-based Hypergraph Neural Networks (AAAI 2026 · 2 citations)
Earliest papers here
- Random Walk Graph Neural Networks (NeurIPS 2020 · 50 citations)
- Pointer Graph Networks (NeurIPS 2020 · 18 citations)
- Towards Scale-Invariant Graph-related Problem Solving by Iterative Homogeneous GNNs (NeurIPS 2020 · 12 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.
Citeseer · accuracy (higher is better)
8 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 73.7 | Equivariant Hypergraph Diffusion Neural Operators (ICLR 2023 · 6 citations) |
| 2024 | 77.33 | Mitigating Oversmoothing Through Reverse Process of GNNs for Heterophilic Graphs (ICML 2024 · 0 citations) |
| 2025 | 81.88 | Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations) |
Cora · accuracy (higher is better)
7 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 84.9 | Graph Ordering Attention Networks (AAAI 2023 · 36 citations) |
| 2024 | 87.93 | Mitigating Oversmoothing Through Reverse Process of GNNs for Heterophilic Graphs (ICML 2024 · 0 citations) |
| 2025 | 90.82 | Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations) |
PubMed · accuracy (higher is better)
6 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 89.03 | Equivariant Hypergraph Diffusion Neural Operators (ICLR 2023 · 6 citations) |
| 2024 | 89.96 | Mitigating Oversmoothing Through Reverse Process of GNNs for Heterophilic Graphs (ICML 2024 · 0 citations) |
| 2025 | 91.36 | Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations) |
CS · accuracy (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 94.21 | Graph Ordering Attention Networks (AAAI 2023 · 36 citations) |
| 2024 | 95.51 | Polynormer: Polynomial-Expressive Graph Transformer in Linear Time (ICLR 2024 · 5 citations) |
| 2025 | 96.58 | Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · 0 citations) |
Photo · accuracy (higher is better)
4 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 94.38 | Graph Ordering Attention Networks (AAAI 2023 · 36 citations) |
| 2024 | 96.11 | Polynormer: Polynomial-Expressive Graph Transformer in Linear Time (ICLR 2024 · 5 citations) |
ZINC · MAE (lower is better)
4 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 0.242 | Rethinking Graph Neural Architecture Search From Message-Passing (CVPR 2021 · 49 citations) |
| 2022 | 0.136 | Automatic Relation-Aware Graph Network Proliferation (CVPR 2022 · 9 citations) |
| 2024 | 0.063 | Graph As Point Set (ICML 2024 · 0 citations) |
What has been tried
56 papers here report 606 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 7%, hurt 24%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Include demands reconstructed by the physics-informed algorithm as input node features | WDS experiments · performance | No clear effect | Physics-Informed Graph Neural Networks for Water Distribution Systems (AAAI 2024 · 28 citations) | |
| Add principal graph learning in place of complete-graph learning | Amazon recommendation datasets · Recall@20 | Helped | Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · 25 citations) | |
| Add neighbor node features | ogbl-ppa · Hits@100 (%) | 63.22±1.74 -> 61.88±4.33 | Hurt | Structural Information Enhanced Graph Representation for Link Prediction (AAAI 2024 · 20 citations) |
| Increase maximum degree from 1 to 6 | 5-body · MSE (× 10⁻²) | 0.52 -> 0.69 | Hurt | Are High-Degree Representations Really Unnecessary in Equivariant Graph Neural Networks? (NeurIPS 2024 · 7 citations) |
| Use neighborhood-based masked attention instead of adjacency-based masking or unmasked full attention | Cora, Citeseer, Pubmed, Arxiv · Accuracy | Cora: 83.12 ± 0.24, 83.41 ± 0.38 -> 84.03 ± 0.15; Citeseer: 71.13 ± 0.35, 72.45 ± 0.33 -> 73.15 ± 0.18; Pubmed: 83.41 ± | Helped | NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (AAAI 2025 · 6 citations) |
| remove PCBA_1328 from pretraining | downstream tasks · performance | Hurt | On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · 6 citations) | |
| add ReLU activation | ogbn-arxiv · accuracy | 71.82 -> 73.46 | Helped | Polynormer: Polynomial-Expressive Graph Transformer in Linear Time (ICLR 2024 · 5 citations) |
| Stack two message-passing layers rather than use the one-step NRI baseline | Spring, ER-50, δt=20 · AUC | 99.84±0.47 -> 51.01±0.64 | Hurt | A Graph Dynamics Prior for Relational Inference (AAAI 2024 · 3 citations) |
| Add equivariant matrix features (L=1) instead of invariant features (L=0) | Cumulene dihedral-angle scan · energy trend capture | Helped | Equivariant Matrix Function Neural Networks (ICLR 2024 · 3 citations) | |
| Replace standard heterogeneous GNN message passing with atomic routes, using GraphSAGE-style aggregation | RelBench classification, regression, and recommendation tasks · task performance | Helped | RelGNN: Composite Message Passing for Relational Deep Learning (ICML 2025 · 1 citation) | |
| reduce diffusion length from the full schedule to 10 or 5 diffusion steps | BendingBeam, ImpactPlate, and DeformingPlate · displacement RMSE | Hurt | Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics (NeurIPS 2025 · 1 citation) | |
| pre-train with G²PM rather than use G²PM without pre-training | ogbn-arxiv · accuracy | 69.64 ± 0.08 -> 72.31 ± 0.07 | Helped | Generative Graph Pattern Machine (NeurIPS 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 16 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Citeseer · 6 of 16 papers
- Cora · 5 of 16 papers
- Pubmed · 5 of 16 papers
- Roman-Empire · 4 of 16 papers
- Ogbn-Arxiv · 2 of 16 papers
- Squirrel · 2 of 16 papers
- Chameleon · 2 of 16 papers
- AM-Photo · 2 of 16 papers
Baselines
- GCN · 8 of 16 papers
- GAT · 7 of 16 papers
- GIN · 5 of 16 papers
- GraphSAGE · 4 of 16 papers
- DirGNN · 3 of 16 papers
- MLP · 3 of 16 papers
- GraphGPS · 2 of 16 papers
- Exphormer · 2 of 16 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 50 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- University of California, Los Angeles · 3 papers
- Purdue University · 3 papers
- Stanford University · 3 papers
- Nanyang Technological University · 3 papers
- University of Oxford · 2 papers
- Shanghai Jiao Tong University · 2 papers
- University of Cambridge · 2 papers
- University of British Columbia · 2 papers
- Amazon · 2 papers
Use it
Download BibTeX of 37 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
- Expressive equivariant message passing (109 papers)
- Graph rewiring and scalable message passing (72 papers)
- Equivariant networks for molecules (43 papers)
Elsewhere in the topic
- Graph structure learning (201 papers)
- GNN generalization theory (187 papers)
- Simple and scalable GNNs (131 papers)
- Robustness, privacy, and fairness of GNNs (94 papers)
- Theory of graph convolutional networks (83 papers)
- Deep graph convolution and simplification (82 papers)
- Physics-inspired GNNs (81 papers)
- Weisfeiler-Leman expressivity hierarchies (66 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).