Research map / Architectures and efficiency / Attention and transformers
On Scalar Embedding of Relative Positions in Attention Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 learnable bucketing and an MLP to model dependencies between scalar bucket embeddings instead of fixed T5 bucketing and independent embeddings | Transformer relative positional encoding · Process-50 | accuracy: 0.808 -> 0.843 | Helped |
| Use learnable AT5 positional encoding instead of T5 positional encoding | Transformer · SST | accuracy: 0.849 -> 0.863 | Helped |
| Use learnable AT5 positional encoding instead of T5 positional encoding | Transformer · MR | accuracy: 0.806 -> 0.812 | Helped |
| Use learnable AT5 positional encoding instead of T5 positional encoding | Transformer · TREC | accuracy: 0.925 -> 0.94 | Helped |
| Use learnable AT5 positional encoding instead of T5 positional encoding | QANet · SQuAD 1.0 | F1: 80.36(0.12) -> 80.93(0.21) | Helped |
| Use learnable AT5 positional encoding instead of T5 positional encoding | Transformer · newstest2016 | BLEU: 33.64(0.23) -> 34.26(0.25) | Helped |
| Remove bucketing from T5 positional encoding (ablation) | Transformer · Process-50 | accuracy: 0.808 -> 0.772 | Hurt |
| Add bucketing to AT5 positional encoding (ablation) | Transformer · Adding-100 | accuracy: 0.961 -> 1.0 | Helped |
| Add bucketing to AT5 positional encoding (ablation) | QANet · SQuAD 1.0 | F1: 79.88(0.52) -> 80.93(0.21) | Helped |
| Add bucketing to AT5 positional encoding (ablation) | Transformer · newstest2016 | BLEU: 33.58(0.27) -> 34.26(0.25) | Helped |
| Vary T5 bucket-number and ramping-parameter configurations (ablation) | QANet with T5 positional encoding · SQuAD 1.0 | F1: 80.36(0.12); 80.36(0.21); 80.14(0.22) | No clear effect |
| Use AT5 instead of T5 on text classification | Transformer · SUBJ, CR, and MPQA | accuracy: SUBJ: 0.939 -> 0.943; CR: 0.845 -> 0.851; MPQA: 0.903 -> 0.901 | Mixed |
Closest papers by meaning
- Stable, Fast and Accurate: Kernelized Attention with Relative Positional Encoding (NeurIPS 2021 · similarity 0.88)
- Positional LSH: Binary Block Matrix Approximation for Attention with Linear Biases (NeurIPS 2026 · similarity 0.87)
- Positional Attention: Expressivity and Learnability of Algorithmic Computation (ICML 2025 · similarity 0.86)
- Parabolic Position Encoding: Vision-Centric, Principled, Extrapolatable, General (NeurIPS 2026 · similarity 0.85)
- Toward Relative Positional Encoding in Spiking Transformers (NeurIPS 2025 · similarity 0.85)
- RelFlexformer: Efficient Attention Transformers for Integrable Relative Positional Encodings (NeurIPS 2026 · 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.