Research map / Architectures and efficiency / Attention and transformers
RNN as Linear Transformer: A Closer Investigation into Representational Potentials of Visual Mamba Models
Timing Yang, Feng Wang, Guoyizhe Wei
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 |
|---|---|---|---|
| replace linear attention with Mamba-2 mixer | DINO ViT-B · ImageNet-1k linear probing | top-1 accuracy: 74.7 -> 78.1 | Improved |
| replace softmax self-attention with Mamba-2 mixer | DINO ViT-B · ImageNet-1k linear probing | top-1 accuracy: 78.2 -> 78.1 | No clear effect |
| replace linear attention with Mamba-2 mixer | DINO ViT-B · ADE20K | mIoU: 29.2 -> 38.0 | Improved |
| replace softmax self-attention with Mamba-2 mixer | DINO ViT-B · ADE20K | mIoU: 43.2 -> 38.0 | Hurt |
| replace linear attention with Mamba-2 mixer | DINO ViT-B · COCO 2017 | box AP: 37.1 -> 42.8 | Improved |
| replace softmax self-attention with Mamba-2 mixer | DINO ViT-B · COCO 2017 | box AP: 44.8 -> 42.8 | Hurt |
| replace linear attention with Mamba-2 mixer | DINO ViT-B · ImageNet-A | top-1 accuracy: 9.6 -> 14.2 | Improved |
| replace softmax self-attention with Mamba-2 mixer | DINO ViT-B · ImageNet-Sketch | top-1 accuracy: 25.5 -> 27.6 | Improved |
| add register tokens (Mamba-Reg) and heavier head | DinoVim-B · ImageNet-1k linear probing | top-1 accuracy: 78.3 -> 78.5 | Improved |
| use self-supervised DINO instead of supervised training | Vision backbones · feature-map Binary-AUC | AUC | Improved |
| increase model size | DinoVim · ImageNet-1k linear probing | top-1 accuracy: 73.7 -> 77.4 | Improved |
Closest papers by meaning
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.92)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.90)
- Mamba-Reg: Vision Mamba Also Needs Registers (CVPR 2025 · similarity 0.89)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.89)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 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.
