Research map / Architectures and efficiency / Attention and transformers
AdaRoPE: Not All Attention Heads Should Rotate and Scale Equally
Shaowen Wang, Yuke Zheng, Tansheng Zhu, Shuang Chen, Shaofan Liu, Suncong Zheng, Jian Li
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 AdaRoPE instead of RoPE | LLaMA 430M · NLU avg | mean accuracy: 50.65 -> 51.23 | Improved |
| add AdaRoPE instead of RoPE | LLaMA 430M · pretraining | loss: 2.692 -> 2.663 | Improved |
| add AdaRoPE instead of RoPE | Qwen3 Gated Attention 430M · WikiText-103 | perplexity: 26.20 -> 20.08 | Improved |
| add AdaRoPE instead of RoPE | LLaMA 1.3B · NLU avg | mean accuracy: 56.71 -> 57.94 | Improved |
| add head-wise frequencies (vs shared frequency) (ablation) | AdaRoPE LLaMA 430M · NLU avg | mean accuracy: 50.90 -> 51.23 | Improved |
| add AdaScale (ablation) | AdaRoPE LLaMA 430M · NLU avg | mean accuracy: 50.77 -> 51.23 | Improved |
| add learned AdaScale instead of fixed log scaling (ablation) | AdaRoPE LLaMA 430M · NLU avg | mean accuracy: 50.58 -> 51.23 | Improved |
| add head-wise frequency parameterization instead of learned base (ablation) | AdaRoPE LLaMA 430M · NLU avg | mean accuracy: 50.77 -> 51.23 | Improved |
| add AdaRoPE instead of YaRN scaling (extrapolation) | LLaMA 8B · RULER 64k | accuracy: 12.80 -> 51.87 | Improved |
| add AdaFreq (ablation) | AdaRoPE extrapolation · RULER 64k | accuracy: 13.13 -> 15.83 | Improved |
| add AdaScale (ablation) | AdaRoPE extrapolation · RULER 64k | accuracy: 9.97 -> 15.83 | Improved |
| add AdaScale (ablation) | AdaRoPE extrapolation SmolLM 1.7B · NLU | accuracy: 64.01 -> 62.73 | Hurt |
Closest papers by meaning
- Round and Round We Go! What makes Rotary Positional Encodings useful? (ICLR 2025 · similarity 0.90)
- Fourier Position Embedding: Enhancing Attention’s Periodic Extension for Length Generalization (ICML 2025 · similarity 0.89)
- Towards Long-window Anchoring in Vision-Language Model Distillation (AAAI 2026 · similarity 0.89)
- Rope to Nope and Back Again: A New Hybrid Attention Strategy (NeurIPS 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 0.89)
- Wavelet-based Positional Representation for Long Context (ICLR 2025 · similarity 0.89)
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.
