Research map / Architectures and efficiency / Attention and transformers
Softmax-Free Linear Transformers
Jiachen Lu, Junge Zhang, Xiatian Zhu, Jianfeng Feng, Tao Xiang, Li Zhang
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 |
|---|---|---|---|
| add symmetric attention normalization (SOFT -> SOFT++) and test at larger input | SOFT-Small · ImageNet-1K val, 512x512 test input | top-1 accuracy (%): 0.14 -> 80.2 | Improved |
| add attention normalization (SOFT -> SOFT++) at 384x384 test | SOFT-Small · ImageNet-1K val, 384x384 test input | top-1 accuracy (%): 33.7 -> 82.6 | Improved |
| add attention normalization to the Large model | SOFT-Large · ImageNet-1K val | top-1 accuracy (%): 83.6 -> 84.1 | Improved |
| add attention normalization to the Medium model | SOFT-Medium · ImageNet-1K val | top-1 accuracy (%): 83.2 -> 83.7 | Improved |
| add attention normalization to the Small model | SOFT-Small · ImageNet-1K val | top-1 accuracy (%): 82.5 -> 82.6 | No clear effect |
| use convolution sampling instead of biased sampling of first m tokens (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 79.9 -> 80.9 | Improved |
| use convolution sampling instead of random sampling (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 80.8 -> 80.9 | No clear effect |
| use convolution sampling instead of average pooling (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 80.8 -> 80.9 | No clear effect |
| use pyramidal multi-stage architecture (ablation) | SOFT++-Small · ImageNet-1K val (Tiny) | top-1 accuracy (%): 80.1 -> 82.4 | Improved |
| use overlapped convolution in patch embedding (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 78.4 -> 80.9 | Improved |
| increase bottleneck tokens from 36 to 49 (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 80.5 -> 80.9 | Improved |
| increase bottleneck tokens from 49 to 81 (ablation) | SOFT++-Tiny · ImageNet-1K val (Tiny) | top-1 accuracy (%): 80.9 -> 80.4 | Hurt |
Closest papers by meaning
- Breaking the Low-Rank Dilemma of Linear Attention (CVPR 2025 · similarity 0.89)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention (NeurIPS 2025 · similarity 0.89)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.88)
- MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity (AAAI 2025 · similarity 0.88)
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.
