Research map / Architectures and efficiency / Attention and transformers
What Dense Graph Do You Need for Self-Attention?
Yuxin Wang, Chu-Tak Lee, Qipeng Guo, Zhangyue Yin, Yunhua Zhou, Xuanjing Huang, Xipeng Qiu
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 |
|---|---|---|---|
| replace complete attention with hypercube sparse attention | Transformer · LRA average | accuracy: 60.67 -> 62.07 | Improved |
| replace complete attention with hypercube sparse attention | Transformer · LRA at length 4096 | training speedup: 15.8× | Improved |
| add random attention to Longformer-style ring lattice and global attention (ablation) | Longformer-style attention · LRA average | accuracy: 60.00 -> 59.99 | No clear effect |
| increase block size from 16 to 32 (ablation) | Hypercube · LRA Retrieval | accuracy: 81.16 -> 80.74 | Hurt |
| increase block size from 16 to 32 (ablation) | Hypercube · LRA Image | accuracy: 53.79 -> 51.98 | Hurt |
| increase block size from 16 to 64 (ablation) | Hypercube · LRA Retrieval | accuracy: 81.16 -> 80.75 | Hurt |
| increase block size from 16 to 64 (ablation) | Hypercube · LRA Image | accuracy: 53.79 -> 50.75 | Hurt |
| replace BERT128 with CubeBERT128 | BERT · Wikitext103 finetuning | MLM loss: 1.18 -> 1.05 | Improved |
| use hypercube sparse attention instead of standard attention | BERT128 · GLUE average | average score: 82.0 -> 82.3 | Improved |
| use hypercube sparse attention instead of standard attention | BERT · GLUE average | average score: 83.0 -> 82.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- BigBird — Big Bird: Transformers for Longer Sequences (NeurIPS 2020) · better on 0 of 1 shared result
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
- Even Sparser Graph Transformers (NeurIPS 2024 · similarity 0.86)
- Primphormer: Efficient Graph Transformers with Primal Representations (ICML 2025 · similarity 0.85)
- Generalizable Insights for Graph Transformers in Theory and Practice (NeurIPS 2025 · similarity 0.85)
- What Improves the Generalization of Graph Transformers? A Theoretical Dive into the Self-attention and Positional Encoding (ICML 2024 · similarity 0.85)
- Towards Principled Graph Transformers (NeurIPS 2024 · similarity 0.84)
- Graph Convolutions Enrich the Self-Attention in Transformers! (NeurIPS 2024 · similarity 0.84)
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.
