Research map / Architectures and efficiency / Attention and transformers
Vision-MoR: Scaling Vision Transformer via Patch-Level Mixture-of-Recursions
Yunhong He, Zhengqing Yuan, Weixiang Sun, Yiyang Li, Yixin Liu, Yanfang Ye, Lichao Sun
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 |
|---|---|---|---|
| use Vision-MoR Small instead of SHViT-S1 | Vision-MoR · ImageNet-1K val (224px, 300 epochs, from scratch) | Top-1 accuracy (%): 72.8 -> 74.6 | Improved |
| use Vision-MoR X-Large instead of SHViT-S4 | Vision-MoR · ImageNet-1K val (224px, 300 epochs, from scratch) | Top-1 accuracy (%): 79.4 -> 80.4 | Improved |
| use Vision-MoR X-Large backbone instead of SHViT-S4 | Vision-MoR · COCO 2017 detection (RetinaNet) | AP: 38.8 -> 39.1 | Improved |
| use Vision-MoR instead of SHViT at smallest scale | Vision-MoR · GH200 GPU, smallest scale (batch 1) | latency (ms): 7.03 -> 3.63 | Improved |
| replace Linear router with spatial router (shifted-window attention) (ablation) | Vision-MoR · ImageNet-1K ablation, expert-choice router | Top-1 accuracy (%): 70.1 -> 77.1 | Improved |
| use auxiliary loss instead of auxiliary router (ablation) | Vision-MoR · ImageNet-1K ablation, expert-choice router | Top-1 accuracy (%): 55.4 -> 77.1 | Improved |
| add Z-loss (ablation) | Vision-MoR (expert-choice) · ImageNet-1K ablation, expert-choice router | Top-1 accuracy (%): 76.5 -> 77.1 | Improved |
| add Z-loss (ablation) | Vision-MoR (token-choice) · ImageNet-1K ablation, token-choice router | Top-1 accuracy (%): 74.2 -> 73.5 | Hurt |
| replace sigmoid with tanh (ablation) | Vision-MoR (expert-choice) · ImageNet-1K ablation, expert-choice router | Top-1 accuracy (%): 77.1 -> 76.4 | No clear effect |
| increase input resolution from 224 to 1120 (ablation) | Vision-MoR Small · ImageNet-1K val | Top-1 accuracy (%): 74.6 -> 79.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MobileFormer-508M — Mobile-Former: Bridging MobileNet and Transformer (CVPR 2022) · better on 1 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
- Edge-RecViT: Efficient Vision Transformer via Semantic-Refined Dynamic Recursion (CVPR 2026 · similarity 0.91)
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.90)
- AdaViT: Adaptive Vision Transformers for Efficient Image Recognition (CVPR 2022 · similarity 0.90)
- A Survey on Efficient Vision Transformers: Algorithms, Techniques, and Performance Benchmarking (TPAMI 2024 · similarity 0.89)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.88)
- Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More (ICML 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.
