Research map / Architectures and efficiency / Attention and transformers
Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention
Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, François Fleuret
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 softmax attention with positive feature-map linear attention | Autoregressive image generation · MNIST | images/sec: 0.45 -> 142.8 | Improved |
| Replace softmax attention with positive feature-map linear attention | Autoregressive image generation · MNIST | bits/dim: 0.621 -> 0.644 | No clear effect |
| Replace softmax attention with positive feature-map linear attention | Autoregressive image generation · CIFAR-10 | images/sec: 0.004 -> 17.85 | Improved |
| Replace softmax attention with positive feature-map linear attention | Autoregressive image generation · CIFAR-10 | bits/dim: 3.47 -> 3.40 | Improved |
| Replace softmax attention with positive feature-map linear attention | Transformer · WSJ | Validation PER: 5.12 -> 8.08 | Hurt |
| Replace softmax attention with positive feature-map linear attention | Transformer · WSJ | Time/epoch (s): 2711 -> 824 | Improved |
| Use linear attention instead of LSH attention | Autoregressive image generation · MNIST | images/sec: 0.68 -> 142.8 | Improved |
| Use linear attention instead of LSH attention | Transformer · WSJ | Validation PER: 9.33 -> 8.08 | Improved |
| Cache and reuse softmax keys and values across autoregressive steps | Softmax transformer · MNIST image generation | images/sec: 0.45 -> 7.56 | Improved |
| Use linear attention instead of cached softmax attention | Autoregressive image generation · CIFAR-10 | images/sec: 0.32 -> 17.85 | Improved |
Closest papers by meaning
- When Do Transformers Outperform Feedforward and Recurrent Networks? A Statistical Perspective (NeurIPS 2025 · similarity 0.88)
- How do Transformers Perform In-Context Autoregressive Learning ? (ICML 2024 · similarity 0.88)
- Fixed-Point RNNs: Interpolating from Diagonal to Dense (NeurIPS 2025 · similarity 0.87)
- On Mesa-Optimization in Autoregressively Trained Transformers: Emergence and Capability (NeurIPS 2024 · similarity 0.87)
- Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators? (ICLR 2024 · similarity 0.87)
- kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers (ICLR 2025 · 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.
