Research map / Architectures and efficiency / Attention and transformers
Making Every Head Count: Sparse Attention Without the Speed-Performance Trade-off
Mingkuan Zhao, Wentao Hu, Jiayin Wang, Xin Lai, Tianchen Huang, Yuheng Min, Rui Yan, Xiaoyan Zhu
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 |
|---|---|---|---|
| replace dense attention with SPAttention | OLMoE 0.25B-1.75B · Average of six metrics | accuracy: 0.3610 -> 0.3684 | Improved |
| replace dense attention with SPAttention | OLMoE 1B-7B · Average of six metrics | accuracy: 0.3402 -> 0.3532 | Improved |
| replace dense attention with SPAttention | OLMoE 1B-7B · Winogrande | accuracy: 0.5040 -> 0.4956 | Hurt |
| replace dense attention with SPAttention | OLMoE 0.25B-1.75B · MMLU SocSci and Humanities (5-shot) | accuracy: 0.2640 -> 0.2552; 0.2520 -> 0.2468 | Hurt |
| replace dense attention with SPAttention | 8-layer Transformer · training throughput | tokens/second: 3,214 -> 6,167 | Improved |
| use SPAttention instead of Longformer | 8-layer Transformer · Average | accuracy: 0.3967 -> 0.4048 | Improved |
| use SPAttention instead of BigBird | 8-layer Transformer · HellaSwag | accuracy: 0.2772 -> 0.2856 | Improved |
| use balanced full-coverage bands instead of sliding window (ablation) | SPAttention · Average (0.25B-1.75B) | accuracy: 0.3376 -> 0.3434 | Improved |
| use complete bands instead of gapped bands (first half only) (ablation) | SPAttention · Average (0.25B-1.75B) | accuracy: 0.3315 -> 0.3434 | Improved |
| use balanced partitioning instead of exclusive unbalanced bands (ablation) | SPAttention · Average (0.25B-1.75B) | accuracy: 0.3289 -> 0.3434 | Improved |
| use distance-band partitioning instead of dense attention | simulated 4-layer 8-head Transformer · head diversity | std of head behaviour: 0.0000-0.0005 -> 0.1845-0.1847 | Improved |
| use distance-band partitioning instead of dense attention | simulated 4-layer 8-head Transformer · attention entropy | entropy: 4.5461 -> 3.6276 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BigBird — Big Bird: Transformers for Longer Sequences (NeurIPS 2020) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.91)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.90)
- Virtual Head Attention (NeurIPS 2026 · similarity 0.90)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.89)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.89)
- HyperAttention: Long-context Attention in Near-Linear Time (ICLR 2024 · 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.
