Research map / Architectures and efficiency / Attention and transformers
Nyströmformer: A Nyström-based Algorithm for Approximating Self-Attention
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 standard self-attention with Nyström approximation | BERT-base fine-tuning · SST-2 validation | accuracy: 90.0 -> 91.4 | Helped |
| Replace standard self-attention with Nyström approximation | BERT-base fine-tuning · QNLI validation | accuracy: 90.3 -> 88.7 | Hurt |
| Replace standard self-attention with Nyström approximation | BERT-base fine-tuning · QQP validation | F1: 87.3 -> 86.3 | Hurt |
| Replace standard self-attention with Nyström approximation | BERT-base fine-tuning · IMDB validation | accuracy: 93.3 -> 93.2 | No clear effect |
| Replace standard self-attention with Nyström approximation | Transformer on LRA · LRA average | average classification accuracy: 58.77 -> 58.95 | No clear effect |
| Use Nyström attention rather than other efficient attention methods | Transformer on LRA · LRA average | average classification accuracy: +3.91%, +3.36%, +5.32% | Helped |
| Use Nyström attention rather than standard self-attention | Self-attention module · Sequence length 8192 | memory (MB): 10233 (1×) -> 450 (22.8×) | Helped |
| Use Nyström attention rather than standard self-attention | Self-attention module · Sequence length 8192 | time (ms): 155.4 (1×) -> 12.3 (12.7×) | Helped |
| Use 32 rather than 64 landmarks (ablation) | Nyströmformer · Self-attention module, sequence length 8192 | memory (MB): 450 (22.8×) -> 383 (26.7×) | Helped |
| Use 32 rather than 64 landmarks (ablation) | Nyströmformer · Self-attention module, sequence length 8192 | time (ms): 12.3 (12.7×) -> 11.5 (13.4×) | Helped |
| Fine-tune using sequence length 1024 rather than 512 (ablation) | Nyströmformer · IMDB reviews | accuracy: 93.0 vs. 93.2 | No clear effect |
Closest papers by meaning
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.88)
- Curse of High Dimensionality Issue in Transformer for Long Context Modeling (ICML 2025 · similarity 0.87)
- Proxyformer: Nyström-Based Linear Transformer with Trainable Proxy Tokens (AAAI 2024 · similarity 0.87)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.87)
- Star Attention: Efficient LLM Inference over Long Sequences (ICML 2025 · similarity 0.87)
- Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · similarity 0.87)
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.