Research map / Architectures and efficiency / Attention and transformers
Spiral RoPE: Rotate Your Rotary Positional Embeddings in the 2D Plane
Haoyu Liu, Sucheng Ren, Tingyu Zhu, Peng Wang, Cihang Xie, Alan Yuille, Zeyu Zheng, Feng Wang
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 Spiral RoPE instead of Axial 2D RoPE | DeiT-L · ImageNet-1k val (448) | top-1 accuracy (%): 85.42 -> 85.67 | Improved |
| use Spiral RoPE instead of Axial 2D RoPE | DeiT-L · ADE20k val | mIoU (%): 47.90 -> 49.12 | Improved |
| use Spiral RoPE instead of Axial 2D RoPE | DiT-B/2 · ImageNet 256 generation | FID: 39.88 -> 37.74 | Improved |
| use Spiral RoPE instead of RoPE-Mixed (learnable frequencies) on DeiT-S 224 | DeiT-S · ImageNet-1k val (224) | top-1 accuracy (%): 80.48 -> 80.39 | Hurt |
| use Spiral RoPE instead of Axial RoPE on diagonal OOD placements | DeiT-Tiny from scratch · synthetic distance task OOD | error rate (%): 70.38 -> 41.72 | Improved |
| add Spiral RoPE direction groups to ViT for rotated images | DeiT-B · ImageNet-1k val rotated 45 deg | top-1 accuracy gap vs Axial: +2.65 at 45 deg; +0.18 at 0 deg | Improved |
| increase number of directions K from 16 to 32 (ablation) | Spiral RoPE · ImageNet-1k DeiT-B 224 | top-1 accuracy (%): 83.39 -> 83.32 | Hurt |
| use K=4 directions vs axial K=2 (ablation) | Spiral RoPE · ImageNet-1k DeiT-B 224 | top-1 accuracy (%): 83.21 -> 83.31 | Improved |
| scale frequencies by 1.5 vs 1.0 (ablation) | Spiral RoPE · ImageNet-1k DeiT-B 224 | top-1 accuracy (%): 83.33 -> 83.39 | No clear effect |
| increase frequency scale to 2.0 (ablation) | Spiral RoPE · ImageNet-1k DeiT-B 224 | top-1 accuracy (%): 83.39 -> 83.23 | Hurt |
| train Spiral RoPE DiT-XL/2 for 7M steps vs SiT | Spiral RoPE · ImageNet 256 generation | FID: 2.06 -> 1.74 | Improved |
Closest papers by meaning
- Circle-RoPE: Cone-like Decoupled Rotary Positional Embedding for Vision-Language Models (ICML 2026 · similarity 0.91)
- A Circular Argument: Does RoPE need to be Equivariant for Vision? (NeurIPS 2025 · similarity 0.91)
- ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices (CVPR 2025 · similarity 0.91)
- LieRE: Lie Rotational Positional Encodings (ICML 2025 · similarity 0.89)
- What Rotary Position Embedding Can Tell Us: Identifying Query and Key Weights Corresponding to Basic Syntactic or High-level Semantic Information (NeurIPS 2024 · similarity 0.88)
- VideoRoPE: What Makes for Good Video Rotary Position Embedding? (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.
