Research map / Architectures and efficiency / Attention and transformers
Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers
Zecheng Tang, Quantong Qiu, Yi Yang, Zhiyi Hong, Haiya Xiang, Kebin Liu, Qingqing Dang, Juntao Li, Min Zhang
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 |
|---|---|---|---|
| enlarge router MLP hidden size from 2x to 4x d' (ablation) | Elastic Attention (Qwen3-4B) · Avg of LongBench, RULER, LongBench-V2 | score: 45.22 -> 45.92 | Improved |
| enlarge router MLP hidden size from 4x to 8x d' (ablation) | Elastic Attention (Qwen3-4B) · Avg of LongBench, RULER, LongBench-V2 | score: 45.92 -> 46.40 | Mixed |
| use XAttention for retrieval heads too (XA-SSA, full sparse regime) (ablation) | Qwen3-4B · Avg of LongBench, RULER, LongBench-V2 | score: 46.80 -> 45.93 | Hurt |
| use XAttention for retrieval heads too (XA-SSA) (ablation) | Qwen3-8B · Avg of LongBench, RULER, LongBench-V2 | score: 53.26 -> 49.42 | Hurt |
| use XAttention for retrieval heads too (XA-SSA) (ablation) | Llama-3.1-8B-Instruct · Avg of LongBench, RULER, LongBench-V2 | score: 56.48 -> 50.02 | Hurt |
| use XAttention instead of full attention for retrieval heads (FA-XA vs FA-SSA sparse heads) | Elastic Attention on Llama-3.1-8B-Instruct · RULER average | accuracy: 72.85 -> 81.82 | Improved |
| use XAttention sparse heads instead of streaming sparse heads | Elastic Attention on Qwen3-8B · LongBench-V2 | accuracy: 33.41 -> 30.74 | Hurt |
| lower target sparsity t_sen for sensitive tasks from 0.7 to 0.4 (ablation) | Elastic Attention · overall long-context performance | performance: t_sen 0.7 -> 0.4 | Improved |
| replace retrieval heads with sparse heads, Omega_MSR 0.0 to 0.2 (ablation) | Llama-3.1-8B-Instruct (static head replacement) · LongBench Single-Doc QA | retention %: 100.00 -> 71.32 | Hurt |
| replace retrieval heads with sparse heads, Omega_MSR 0.0 to 0.2 (ablation) | Llama-3.1-8B-Instruct (static head replacement) · LongBench Summarization | retention %: 100.00 -> 98.00 | No clear effect |
| raise Omega_MSR from 0.0 to 1.0 by replacing retrieval heads (ablation) | Llama-3.1-8B-Instruct (static head replacement) · LongBench Code | retention %: 100.00 -> 99.46 | No clear effect |
| Elastic Attention vs PruLong at fixed 0.70 sparsity | Llama-3.1-8B-Instruct · RULER average | accuracy: 48.82 -> 72.85 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- DuoAttention — DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025) · better on 4 of 4 shared results
- XAttention — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 3 of 3 shared results
- MoBA — MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 2025) · 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
- Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity (NeurIPS 2025 · similarity 0.90)
- AdaSplash: Adaptive Sparse Flash Attention (ICML 2025 · similarity 0.90)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.90)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.89)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix (ICLR 2025 · 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.
