Research map / Probabilistic and scientific modeling / Molecules and proteins
nD-RoPE: A Generalized RoPE for n-Dimensional Position Embedding
Boyang Li, Yulin Wu, Sizhe Xu, Nuoxian Huang, Zhonghang Yuan, Shangyi Guo, Shu Yang, Takahiro Yabe
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 nD-RoPE instead of RoPE-Mixed + APE | ViT-S · ImageNet-1K 224 | top-1 accuracy (%): 80.92 -> 81.07 | Improved |
| use nD-RoPE instead of RoPE-Mixed + APE under 30 degree rotation | ViT-S · ImageNet-1K rotated 30 degrees | top-1 accuracy (%): 71.12 -> 78.51 | Improved |
| use nD-RoPE instead of RoPE-Mixed + APE at 0 degree rotation | ViT-S · ImageNet-1K unrotated | top-1 accuracy (%): 81.13 -> 80.81 | No clear effect |
| add YaRN to nD-RoPE at 1024 resolution | nD-RoPE ViT-S · ImageNet-1K eval 1024 | top-1 accuracy (%): 35.51 -> 68.46 | Improved |
| add YaRN to RoPE-Mixed at 1024 resolution | RoPE-Mixed ViT-S · ImageNet-1K eval 1024 | top-1 accuracy (%): 16.63 -> 43.48 | Improved |
| use nD-RoPE with vector attention instead of dot-product attention | Point Transformer · ModelNet40 2048 points | instance mIoU (%): 85.07 -> 85.97 | Improved |
| use nD-RoPE with vector attention instead of dot-product attention | Point Transformer v2 · SemanticKITTI grid 0.05 | instance mIoU (%): 67.12 -> 71.91 | Improved |
| use nD-RoPE instead of learnable relative PE | Point Transformer · ModelNet40 2048 points | instance mIoU (%): 82.58 -> 85.97 | Improved |
| use balanced scales per head (6 scales x 10 heads) instead of 64 scales x 1 head (ablation) | nD-RoPE · ImageNet-1K 224 | top-1 accuracy (%): 70.99 -> 81.07 | Improved |
| use balanced scales per head (6 scales x 10 heads) instead of 1 scale x 64 heads (ablation) | nD-RoPE · ImageNet-1K 224 | top-1 accuracy (%): 65.93 -> 81.07 | Improved |
| use frequency base 100 instead of 10^4 (ablation) | nD-RoPE vector attention · ModelNet40 256 points | instance mIoU (%): 45.94 -> 55.04 | Improved |
| use frequency base 100 instead of 10^6 (ablation) | nD-RoPE vector attention · ModelNet40 1536 points | instance mIoU (%): 82.67 -> 85.61 | 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 0.88)
- RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers (NeurIPS 2025 · similarity 0.85)
- LieRE: Lie Rotational Positional Encodings (ICML 2025 · similarity 0.85)
- VideoRoPE: What Makes for Good Video Rotary Position Embedding? (ICML 2025 · similarity 0.85)
- Circle-RoPE: Cone-like Decoupled Rotary Positional Embedding for Vision-Language Models (ICML 2026 · 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.
