Research map / Architectures and efficiency / Attention and transformers
SeqPE: Transformer With Sequential Position Encoding
Huayang Li, Yahui Liu, Hongyu Sun, Deng Cai, Leyang Cui, Wei Bi, Peilin Zhao, Taro Watanabe
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 distance regularization L_delta (ablation) | SeqPE (language modeling) · Wikitext-103 (avg 512-4K) | perplexity: 57.24 -> 18.89 | Improved |
| add OOD distillation L_OOD on top of L_delta (ablation) | SeqPE (language modeling) · Wikitext-103 (avg 512-4K) | perplexity: 18.89 -> 18.86 | Improved |
| add L_OOD alone without L_delta (ablation) | SeqPE (language modeling) · Wikitext-103 (avg 512-4K) | perplexity: 57.24 -> 90.69 | Hurt |
| add L_delta (ablation) | SeqPE (image classification) · ImageNet-1k (avg 224-448) | accuracy: 79.7 -> 80.7 | Improved |
| add L_OOD on top of L_delta (ablation) | SeqPE (image classification) · ImageNet-1k (avg 224-448) | accuracy: 80.7 -> 81.3 | Improved |
| add L_OOD alone without L_delta (ablation) | SeqPE (image classification) · ImageNet-1k (avg 224-448) | accuracy: 79.7 -> 79.5 | Hurt |
| use AttnMul instead of AttnSum integration (ablation) | SeqPE · ImageNet-1k (avg 224-448) | accuracy: 80.0 -> 81.4 | Improved |
| use AttnMul instead of AttnBias integration (ablation) | SeqPE · ImageNet-1k (avg 224-448) | accuracy: 80.2 -> 81.4 | Improved |
| use SeqPE instead of ALiBi | GPT-2 · Wikitext-103 (16K context) | perplexity: 21.39 -> 19.37 | Improved |
| use SeqPE instead of ALiBi after QA fine-tuning | GPT-2 · RULER-SQuAD (avg 1K-8K) | perplexity: 36.06 -> 12.34 | Improved |
| use SeqPE instead of RoPE2D-Mix | ViT-S · ImageNet-1k (672 resolution) | accuracy: 68.3 -> 76.7 | Improved |
| use SeqPE instead of RoPE2D-Mix | ViT-S · ImageNet-1k (320 resolution) | accuracy: 82.2 -> 81.8 | Hurt |
Closest papers by meaning
- DAPE: Data-Adaptive Positional Encoding for Length Extrapolation (NeurIPS 2024 · similarity 0.90)
- Functional Interpolation for Relative Positions improves Long Context Transformers (ICLR 2024 · similarity 0.88)
- PaTH Attention: Position Encoding via Accumulating Householder Transformations (NeurIPS 2025 · similarity 0.88)
- Wavelet-based Positional Representation for Long Context (ICLR 2025 · similarity 0.88)
- Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability (NeurIPS 2025 · similarity 0.88)
- What Rotary Position Embedding Can Tell Us: Identifying Query and Key Weights Corresponding to Basic Syntactic or High-level Semantic Information (NeurIPS 2024 · similarity 0.87)
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.
