Research map / Graph learning / Graph self-supervised learning
NAR-Former V2: Rethinking Transformer for Universal Neural Network Representation Learning
Yun Yi, Haokui Zhang, Rong Xiao, Nannan Wang, Xiaoyu Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP same distribution (All) | MAPE: 3.47% -> 3.07% | Improved |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP same distribution (All) | Acc(10%): 95.25% -> 96.41% | Improved |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP unseen network type, average | Acc(10%): 54.62% -> 62.70% | Improved |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP unseen network type, average | MAPE: 11.55% -> 10.55% | Improved |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP same distribution, NasBench201 | MAPE: 1.46% -> 1.82% | Hurt |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP unseen VGG | MAPE: 7.60% -> 15.40% | Hurt |
| use graph-aided Transformer instead of GNN (NNLP) | NNLP · NNLQP unseen AlexNet | MAPE: 10.64% -> 24.28% | Hurt |
| add GNN-inspired adjacency-masked attention and TA-Enhance to NAR-Former | NAR-Former · NAS-Bench-101 0.1% train, all test | Kendall's Tau: 0.765 -> 0.773 | Improved |
| NAR-Former V2 instead of NAR-Former | NAR-Former · NAS-Bench-101 1% train | Kendall's Tau: 0.871 -> 0.861 | Hurt |
| NAR-Former V2 instead of NAR-Former | NAR-Former · NAS-Bench-201 5% train | Kendall's Tau: 0.849 -> 0.874 | Improved |
| NAR-Former V2 instead of NAR-Former | NAR-Former · NAS-Bench-201 10% train | Kendall's Tau: 0.901 -> 0.888 | Hurt |
| encode op type with one-hot instead of PE (GNN) (ablation) | GNN baseline · NNLQP ablation | Acc(10%): 95.11 -> 95.57 | Improved |
Later papers that compared with this method
1 later accepted paper reports NAR-Former V2 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- NN-Former: Rethinking Graph Structure in Neural Architecture Representation (CVPR 2025 · better on 5 of 5 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- NN-Former: Rethinking Graph Structure in Neural Architecture Representation (CVPR 2025 · similarity 0.86)
- Encodings for Prediction-based Neural Architecture Search (ICML 2024 · similarity 0.86)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.85)
- Equivariant Neural Functional Networks for Transformers (ICLR 2025 · similarity 0.85)
- Transitivity-Preserving Graph Representation Learning for Bridging Local Connectivity and Role-Based Similarity (AAAI 2024 · similarity 0.85)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
