Research map / Graph learning / GNN architectures
GNN generalization theory
Analyze size generalization and statistical properties of GNNs, with benchmarks.
187 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in GNN architectures, keyword cluster “gnn · explanations · trained”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
53 papers in 2023–24 and 75 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 11.2% to 14.8%.
Open questions the data points to
- Citeseer (accuracy): best went from 80.25 (up to 2024) to 94.9 in 2025 — Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs (ICLR 2025 · 1 citation)
- Papers report GCN on PubMed (accuracy (%)) anywhere from 76.3 to 86.48: comparisons against it depend on whose number you take.
- Papers report GAT on Citeseer (accuracy) anywhere from 68.4 to 80.13: comparisons against it depend on whose number you take.
What looks settled
- Cora (accuracy): best up to 2024 93.68, best in 2025–26 91.14 from 9 papers, at most 0.2% better, if at all.
- Pubmed (accuracy): best up to 2024 90.55, best in 2025–26 90.41 from 9 papers, at most 0.2% better, if at all.
- ogbn-arxiv (accuracy): best up to 2024 74.3, best in 2025–26 73.26 from 5 papers, at most 0.2% better, if at all.
- Cornell (accuracy): best up to 2024 92.1, best in 2025–26 91.4 from 4 papers, at most 0.2% better, if at all.
- Chameleon (accuracy): best up to 2024 79.91, best in 2025–26 43.9 from 2 papers, at most 0.2% better, if at all.
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
- 2019 · Position-aware Graph Neural Networks (ICML 2019 · 78 citations)
- 2020 · EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs (AAAI 2020 · 1,234 citations)
- 2021 · Do Transformers Really Perform Badly for Graph Representation? (NeurIPS 2021 · 378 citations)
- 2022 · Robust Optimization As Data Augmentation for Large-Scale Graphs (CVPR 2022 · 94 citations)
- 2023 · Boosting Graph Neural Networks via Adaptive Knowledge Distillation (AAAI 2023 · 31 citations)
- 2024 · GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs (CVPR 2024 · 39 citations)
- 2025 · Graph Coarsening via Supervised Granular-Ball for Scalable Graph Neural Network Training (AAAI 2025 · 10 citations)
- 2026 · Backward Oversmoothing: why is it hard to train deep Graph Neural Networks? (ICML 2026 · 1 citation)
Earliest papers here
- Position-aware Graph Neural Networks (ICML 2019 · 78 citations)
- Understanding Attention and Generalization in Graph Neural Networks (NeurIPS 2019 · 55 citations)
- Guided Dropout (AAAI 2019 · 24 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.
Cora · accuracy (higher is better)
29 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 87.81 | Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · 171 citations) |
| 2022 | 88.6 | Hierarchical Graph Transformer with Adaptive Node Sampling (NeurIPS 2022 · 30 citations) |
| 2024 | 93.68 | Graph Neural Networks Need Cluster-Normalize-Activate Modules (NeurIPS 2024 · 2 citations) |
Pubmed · accuracy (higher is better)
26 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 89.59 | Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · 171 citations) |
| 2022 | 89.62 | Finding Global Homophily in Graph Neural Networks When Meeting Heterophily (ICML 2022 · 37 citations) |
| 2023 | 89.7 | NAGphormer: A Tokenized Graph Transformer for Node Classification in Large Graphs (ICLR 2023 · 21 citations) |
| 2024 | 90.55 | Feature Transportation Improves Graph Neural Networks (AAAI 2024 · 12 citations) |
Citeseer · accuracy (higher is better)
25 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 77.07 | Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · 171 citations) |
| 2022 | 80.25 | Hierarchical Graph Transformer with Adaptive Node Sampling (NeurIPS 2022 · 30 citations) |
| 2025 | 94.9 | Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs (ICLR 2025 · 1 citation) |
ogbn-arxiv · accuracy (higher is better)
15 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 74.3 | Topological Relational Learning on Graphs (NeurIPS 2021 · 1 citation) |
Cornell · accuracy (higher is better)
11 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 82.16 | Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · 171 citations) |
| 2022 | 92.1 | Hierarchical Graph Transformer with Adaptive Node Sampling (NeurIPS 2022 · 30 citations) |
Chameleon · accuracy (higher is better)
10 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 59.39 | Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · 171 citations) |
| 2022 | 69.78 | Finding Global Homophily in Graph Neural Networks When Meeting Heterophily (ICML 2022 · 37 citations) |
| 2023 | 71.58 | Towards Deep Attention in Graph Neural Networks: Problems and Remedies (ICML 2023 · 2 citations) |
| 2024 | 79.91 | Feature Transportation Improves Graph Neural Networks (AAAI 2024 · 12 citations) |
What has been tried
140 papers here report 1,497 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 10%, hurt 19%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use K = 8, 4, 2, 1 instead of K = 16, 8, 4, 2 | ImageNet-1K · top-1 accuracy | 83.5% ... 83.9% | Helped | GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs (CVPR 2024 · 39 citations) |
| Add normalization | large-scale graphs · ROC-AUC | 82.21 -> 77.42; 85.01 -> 80.32 | Helped | Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · 33 citations) |
| Add advection to the full ADR-GNN_S | Chameleon · accuracy (%) | 73.46 -> 79.91 | Helped | Feature Transportation Improves Graph Neural Networks (AAAI 2024 · 12 citations) |
| add the masking matrix | Cora · accuracy | 82.1±1.5% -> 83.1±1.0% | Helped | HONGAT: Graph Attention Networks in the Presence of High-Order Neighbors (AAAI 2024 · 12 citations) |
| Increase initial centers from √N to 2√N or 4√N | Pubmed · coarsening time | 29.7 seconds -> 35.0 seconds; 29.7 seconds -> 78.4 seconds | Hurt | Graph Coarsening via Supervised Granular-Ball for Scalable Graph Neural Network Training (AAAI 2025 · 10 citations) |
| Use SLG rather than LLG for domain-independent heuristic learning | Domain-independent planning · coverage | Helped | Learning Domain-Independent Heuristics for Grounded and Lifted Planning (AAAI 2024 · 10 citations) | |
| Co-train e5-335M with GraphSAGE rather than training GraphSAGE alone | HuggingFace · n-F1 | 77.90 -> 77.87 | No clear effect | Can Graph Learning Improve Planning in LLM-based Agents? (NeurIPS 2024 · 8 citations) |
| Add SIGN feature aggregation instead of GCN aggregation | ogbl-citation2 · MRR | 87.05 -> 87.53 | Helped | On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · 5 citations) |
| add R-SoftGraphAIN | Cora, 32 layers · accuracy | 40.22±5.71 -> 85.12±0.15 | Helped | Curriculum-Enhanced Residual Soft An-Isotropic Normalization for Over-Smoothness in Deep GNNs (AAAI 2024 · 3 citations) |
| use estimated-attention sampling instead of uniform sampling | Computer, Photo, Minesweeper, Tolokers · accuracy / ROC-AUC (×100) | 86.65 ± 0.46 vs 91.09 ± 0.08; 94.21 ± 0.22 vs 95.33 ± 0.49; 84.15 ± 0.22 vs 90.71 ± 0.17; 82.56 ± 0.17 vs 83.34 ± 0.31 | Helped | Even Sparser Graph Transformers (NeurIPS 2024 · 3 citations) |
| Use fewer layers | OGBN-Arxiv · generalization gap | Helped | Generalization of Graph Neural Networks Is Robust to Model Mismatch (AAAI 2025 · 2 citations) | |
| increase embedding size from 64 to 1,024 | MP-intractable two-instance dataset · epochs to reach 10^{-6} error | 16,570 -> 980 | Helped | Rethinking the Capacity of Graph Neural Networks for Branching Strategy (NeurIPS 2024 · 2 citations) |
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 |
|---|---|---|---|---|
| GCN | PubMed | accuracy (%) | 76.3 – 86.48 | 4 papers, e.g. AAAI 2020 and ICML 2025 |
| GAT | Citeseer | accuracy | 68.4 – 80.13 | 3 papers, e.g. AAAI 2024 and NeurIPS 2022 |
| APPNP | Citeseer | accuracy | 69.2 – 75.9 | 3 papers, e.g. AAAI 2024 and ICLR 2025 |
| PNA | ZINC | MAE | 0.188 – 0.32 | 2 papers, e.g. ICML 2021 and ICLR 2022 |
| H2GCN | Texas | accuracy | 72.3 – 84.76 | 2 papers, e.g. ICML 2025 and NeurIPS 2022 |
| GCNII | Pubmed | accuracy (%) | 78.06 – 90.3 | 2 papers, e.g. ICML 2023 and AAAI 2024 |
What reviewers will expect
Datasets and baselines named in the 55 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Cora · 21 of 55 papers
- Pubmed · 21 of 55 papers
- Citeseer · 19 of 55 papers
- ogbn-arxiv · 12 of 55 papers
- Roman-Empire · 10 of 55 papers
- Amazon-Ratings · 10 of 55 papers
- Squirrel · 7 of 55 papers
- Cornell · 6 of 55 papers
Baselines
- GCN · 15 of 55 papers
- GAT · 11 of 55 papers
- GraphSAGE · 8 of 55 papers
- MLP · 8 of 55 papers
- GPS · 5 of 55 papers
- SGC · 4 of 55 papers
- GatedGCN · 4 of 55 papers
- GIN · 4 of 55 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 128 papers; from OpenAlex and the papers' affiliations).
- Massachusetts Institute of Technology · 6 papers
- Peking University · 5 papers
- Korea Advanced Institute of Science and Technology · 4 papers
- Nanyang Technological University · 4 papers
- Chinese Academy of Sciences · 4 papers
- Amazon · 4 papers
- Stanford University · 4 papers
- University of Oxford · 4 papers
- University of Notre Dame · 3 papers
- Ant Group · 3 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
- Simple and scalable GNNs (131 papers)
- Robustness, privacy, and fairness of GNNs (94 papers)
- Physics-inspired GNNs (81 papers)
- GNN explainability (64 papers)
- Graph structure learning (201 papers)
Elsewhere in the topic
- Expressive equivariant message passing (109 papers)
- Theory of graph convolutional networks (83 papers)
- Deep graph convolution and simplification (82 papers)
- Graph rewiring and scalable message passing (72 papers)
- Weisfeiler-Leman expressivity hierarchies (66 papers)
- Structural GNN variants (65 papers)
- Oversmoothing and heterophily (60 papers)
- Substructure-aware GNN expressivity (50 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).