AI Paper Atlas

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

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 7 citations (OpenAlex)

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 changedWhereMeasuredEffect
use ProxyAttn instead of full attention at gamma=0.95Llama3.1-8B-Instruct · RULERaverage accuracy: 89.46 -> 90.18Improved
use ProxyAttn instead of XAttention on QwenQwen2.5-7B-Instruct-1M · RULERaverage accuracy: 83.90 -> 87.30Improved
use ProxyAttn instead of full attentionLlama3.1-8B-Instruct · InfiniteBench + LongBench-v2overall score: 35.38 -> 36.06Improved
use ProxyAttn instead of full attentionQwen2.5-7B-Instruct-1M · InfiniteBench + LongBench-v2overall score: 38.22 -> 38.33No clear effect
use ProxyAttn instead of MInference/XAttentionLlama3.1-70B-Instruct · RULERaccuracy and TTFT: 62.23 vs 60.33/58.59; TTFT 2.08xMixed
raise min budget from 0 to 2048 (ablation)ProxyAttn · RULER (Llama3.1-8B, gamma=0.95)average accuracy: 90.18 -> 89.33Hurt
raise min budget from 0 to 2048 (ablation)ProxyAttn · RULER (Qwen, gamma=0.95)average accuracy: 87.11 -> 87.30Improved
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 -> 0.81Mixed
increase stride from 1 to 4 (ablation)ProxyAttn · RULER (Llama3.1-8B)average accuracy: 87.87 -> 88.42No clear effect
use dynamic per-head budget instead of static top-K (ablation)ProxyAttn · RULER 128Kperformance under sparsityImproved
use more proxy heads on Llama (ablation)ProxyAttn · RULER 128KaccuracyNo clear effect
use about 4 proxy heads on Qwen instead of 1 (ablation)ProxyAttn · RULER 128KaccuracyImproved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy