Research map / Architectures and efficiency / Attention and transformers
Your Transformer May Not be as Powerful as You Expect
Shengjie Luo, Shanda Li, Shuxin Zheng, Tie-Yan Liu, Liwei Wang, Di He
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 URPE-based Attention to Transformer-XL | Transformer-XL Base · WikiText-103 validation and test | perplexity: 23.1 -> 22.4; 24.0 -> 23.2 | Improved |
| Add URPE-based Attention to Graphormer | Graphormer · ZINC-Subset test | MAE: 0.122$\pm$0.006 -> 0.086$\pm$0.007 | Improved |
| Add URPE-based Attention to Graphormer | Graphormer · ZINC-Full test | MAE: 0.052$\pm$0.005 -> 0.028$\pm$0.002 | Improved |
| Add URPE-based Attention to Graphormer | Graphormer · PCQM4M validation | MAE: 0.1264 -> 0.1238 | Improved |
| Add URPE-based Attention to Graphormer | Graphormer · MAG240M validation | accuracy: 0.7013 -> 0.7074 | Improved |
| Add URPE-based Attention (ablation) | Transformer-XL · WikiText-103 validation, L=4 | perplexity: 29.6 -> 28.7 | Improved |
| Add URPE-based Attention (ablation) | Transformer-XL · WikiText-103 validation, L=8 | perplexity: 26.0 -> 25.2 | Improved |
| Add URPE-based Attention (ablation) | Transformer-XL · WikiText-103 validation, L=16 | perplexity: 23.1 -> 22.4 | Improved |
| Use URPE-based Attention instead of RPE (ablation) | Transformer · Inference, N=128 | runtime (ms in log base 2): 4.55 -> 4.59 | Hurt |
| Use URPE-based Attention instead of RPE (ablation) | Transformer · Inference, N=256 | peak memory (GB): 1.12 -> 1.17 | Hurt |
| Use URPE-based Attention instead of RPE (ablation) | Transformer · Inference, N=512 | peak memory (GB): 1.86 -> 2.04 | Hurt |
| Use URPE-based Attention instead of T5-style RPE | Vanilla Transformer · Synthetic Position Identification and Even Token Prediction | token-level accuracy: less than 60\% -> 100\% | Improved |
Closest papers by meaning
- DAPE: Data-Adaptive Positional Encoding for Length Extrapolation (NeurIPS 2024 · similarity 0.87)
- Toward Relative Positional Encoding in Spiking Transformers (NeurIPS 2025 · similarity 0.86)
- Prompting a Pretrained Transformer Can Be a Universal Approximator (ICML 2024 · similarity 0.86)
- Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators? (ICLR 2024 · similarity 0.85)
- A unified framework for establishing the universal approximation of transformer-type architectures (NeurIPS 2025 · similarity 0.85)
- Approximation Rate of the Transformer Architecture for Sequence Modeling (NeurIPS 2024 · similarity 0.85)
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.
