Research map / Architectures and efficiency / Attention and transformers
ProxyAttn: Guided Sparse Attention via Representative Heads
Yixuan Wang, Huang He, Siqi Bao, Hua Wu, Haifeng Wang, Qingfu Zhu, Wanxiang Che
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 |
|---|---|---|---|
| use ProxyAttn instead of full attention at gamma=0.95 | Llama3.1-8B-Instruct · RULER | average accuracy: 89.46 -> 90.18 | Improved |
| use ProxyAttn instead of XAttention on Qwen | Qwen2.5-7B-Instruct-1M · RULER | average accuracy: 83.90 -> 87.30 | Improved |
| use ProxyAttn instead of full attention | Llama3.1-8B-Instruct · InfiniteBench + LongBench-v2 | overall score: 35.38 -> 36.06 | Improved |
| use ProxyAttn instead of full attention | Qwen2.5-7B-Instruct-1M · InfiniteBench + LongBench-v2 | overall score: 38.22 -> 38.33 | No clear effect |
| use ProxyAttn instead of MInference/XAttention | Llama3.1-70B-Instruct · RULER | accuracy and TTFT: 62.23 vs 60.33/58.59; TTFT 2.08x | Mixed |
| raise min budget from 0 to 2048 (ablation) | ProxyAttn · RULER (Llama3.1-8B, gamma=0.95) | average accuracy: 90.18 -> 89.33 | Hurt |
| raise min budget from 0 to 2048 (ablation) | ProxyAttn · RULER (Qwen, gamma=0.95) | average accuracy: 87.11 -> 87.30 | Improved |
| lower gamma from 0.95 to 0.90 (ablation) | ProxyAttn · RULER (Llama3.1-8B) | average accuracy and sparsity: 90.18 -> 89.32; sparsity 0.69 -> 0.81 | Mixed |
| increase stride from 1 to 4 (ablation) | ProxyAttn · RULER (Llama3.1-8B) | average accuracy: 87.87 -> 88.42 | No clear effect |
| use dynamic per-head budget instead of static top-K (ablation) | ProxyAttn · RULER 128K | performance under sparsity | Improved |
| use more proxy heads on Llama (ablation) | ProxyAttn · RULER 128K | accuracy | No clear effect |
| use about 4 proxy heads on Qwen instead of 1 (ablation) | ProxyAttn · RULER 128K | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- XAttention — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 5 of 5 shared results
- FlexPrefill — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 2 of 2 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
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.90)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.90)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.89)
- Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning (NeurIPS 2025 · similarity 0.88)
- Block Sparse Flash Attention (NeurIPS 2026 · similarity 0.88)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.88)
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.
