Research map / Architectures and efficiency / Attention and transformers
SAC: Accelerating and Structuring Self-Attention via Sparse Adaptive Connection
Xiaoya Li, Yuxian Meng, Mingxin Zhou, Qinghong Han, Fei Wu, Jiwei 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 |
|---|---|---|---|
| increase the number of learned attention edges from 2N to 10N (ablation) | SAC base · WMT14 English-German | BLEU: 18.3 -> 27.7 | Improved |
| increase the number of learned attention edges from 10N to 15N (ablation) | SAC base · WMT14 English-German | BLEU: 27.7 -> 27.4 | Hurt |
| increase attention heads from 8 to 16 (ablation) | SAC base with 10N edges · WMT14 English-German | BLEU: 27.7 -> 28.1 | Improved |
| increase depth from 6 to 12 blocks (ablation) | SAC base with 16 heads and 10N edges · WMT14 English-German | BLEU: 28.1 -> 28.4 | Improved |
| incorporate dependency-tree distance information into edge prediction (ablation) | SAC Large · WMT14 English-German | BLEU: +0.14 BLEU | Improved |
| use head-specific edge predictors (ablation) | SAC · WMT14 English-German | BLEU | No clear effect |
| use head-specific edge predictors (ablation) | SAC · Text8 | BPC: 1.07 -> 1.06 | Improved |
| use head-specific edge predictors (ablation) | SAC · Enwiki8 | BPC: 1.02 -> 1.00 | Improved |
| use head-specific edge predictors (ablation) | SAC · Cora | accuracy: 84.8 -> 84.7 | Hurt |
| use head-specific edge predictors (ablation) | SAC · Citeseer | accuracy: 73.8 -> 74.0 | Improved |
| use head-specific edge predictors (ablation) | SAC · Pubmed | accuracy: 79.7 -> 80.1 | Improved |
| use head-specific edge predictors (ablation) | SAC · PPI | micro-F1: 98.4 -> 98.4 | No clear effect |
Closest papers by meaning
- Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning (NeurIPS 2025 · similarity 0.88)
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.87)
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.87)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.87)
- Low-Rank Approximation for Sparse Attention in Multi-Modal LLMs (CVPR 2024 · similarity 0.87)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 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.
