Research map / Architectures and efficiency / Attention and transformers
Compass-RoPE: Isotropic Rotary Position Embeddings for Vision Transformers
Chengxi Min, Wei Wang, Yao Zhao
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 DFT mixing of angle vectors (ablation) | Compass-RoPE · ImageNet-1K val, avg over 9 resolutions | Top-1 accuracy: 76.5 -> 77.6 | Improved |
| add DFT mixing of angle vectors (ablation) | Compass-RoPE · ImageNet-1K val, eval at 720 | Top-1 accuracy: 66.4 -> 70.2 | Improved |
| add DFT mixing at near-training resolutions (ablation) | Compass-RoPE · ImageNet-1K val, eval at 224 | Top-1 accuracy: 80.8 -> 81.0 | Mixed |
| replace Mix-RoPE with polar parameterization and uniform angle init (no DFT) (ablation) | Mix-RoPE · ImageNet-1K val, eval at 720 | Top-1 accuracy: 62.2 -> 66.4 | Improved |
| replace Mix-RoPE with Compass-RoPE | Mix-RoPE · ImageNet-1K val, avg over 9 resolutions | Top-1 accuracy: 75.4 -> 77.6 | Improved |
| replace Mix-RoPE with Compass-RoPE | Mix-RoPE · ImageNet-1K val, eval at 224 | Top-1 accuracy: 80.9 -> 81.0 | No clear effect |
| replace APE with Mix-RoPE at the largest resolution | ViT (APE) · ImageNet-1K val, eval at 720 | Top-1 accuracy: 66.8 -> 62.2 | Hurt |
| replace Mix-RoPE with Compass-RoPE | Mix-RoPE · Pascal VOC07 test | mAP: 79.1 -> 79.5 | Improved |
| replace Mix-RoPE with Compass-RoPE | Mix-RoPE · ADE20K val | mIoU: 43.2 -> 43.4 | No clear effect |
| add DFT mixing to improve directional coverage (ablation) | Compass-RoPE · Shift alignment, 90 degrees | shift alignment accuracy: 82.0 -> 90.2 | Improved |
| replace Mix-RoPE with Compass-RoPE for frequency coverage | Mix-RoPE · Effective direction count | exp(H): 28.3 -> 35.7 | Improved |
| combine Compass-RoPE with LookHere | LookHere · ImageNet-100 avg over resolutions | Top-1 accuracy: 73.9 -> 74.3 | Improved |
Closest papers by meaning
- ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices (CVPR 2025 · similarity 0.89)
- A Circular Argument: Does RoPE need to be Equivariant for Vision? (NeurIPS 2025 · similarity 0.87)
- Circle-RoPE: Cone-like Decoupled Rotary Positional Embedding for Vision-Language Models (ICML 2026 · similarity 0.87)
- LieRE: Lie Rotational Positional Encodings (ICML 2025 · similarity 0.85)
- RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers (NeurIPS 2025 · similarity 0.84)
- VideoRoPE: What Makes for Good Video Rotary Position Embedding? (ICML 2025 · similarity 0.84)
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.
