Research map / Architectures and efficiency / Attention and transformers
Tactic: Adaptive Sparse Attention with Clustering and Distribution Fitting for Long-Context LLMs
Kan Zhu, Tian Tang, Qinyu Xu, Zhan Jin, Yile Gu, Zhichen Zeng, Rohan Kadekodi, Liangyu Zhao, Ang Li, Arvind Krishnamurthy, Baris Kasikci
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 |
|---|---|---|---|
| add union of selected tokens across grouped query heads (ablation) | Tactic on GQA model · attention latency across context lengths and P | attention speedup: up to 1.65x | Improved |
| use cumulative-attention-score thresholds instead of fixed token budgets (Tactic vs Quest at 75%) | Llama-3.1-8B-Instruct · RULER average | accuracy: 69.2 -> 84.7 | Improved |
| use Tactic instead of PyramidKV at 90% budget | Llama-3.1-8B-Instruct · RULER average | accuracy: 73.0 -> 84.6 | Improved |
| use Tactic instead of Quest at 75% budget | MegaBeam-Mistral-7B-512k · RULER average | accuracy: 75.4 -> 85.2 | Improved |
| raise threshold from 75% to 90% | Tactic Llama-3.1-8B-Instruct · RULER average | accuracy: 84.7 -> 84.6 | No clear effect |
| raise threshold from 75% to 90% | Tactic MegaBeam-Mistral-7B-512k · RULER average | accuracy: 85.2 -> 85.4 | No clear effect |
| raise threshold from 70% to 90% | Tactic Llama-3.1-8B-Instruct · LongBench Qasper | score: 43.30 -> 44.20 | Improved |
| raise threshold from 70% to 90% | Tactic Llama-3.1-8B-Instruct · LongBench Musique | score: 28.57 -> 30.71 | Improved |
| use Tactic instead of Quest at 90% budget | MegaBeam-Mistral-7B-512k · LongBench HotpotQA | score: 51.69 -> 49.59 | Hurt |
| use Tactic instead of PyramidKV at 70% budget | Llama-3.1-8B-Instruct · LongBench Qasper | score: 27.50 -> 43.30 | Improved |
| use Tactic instead of PyramidKV at 70% budget | MegaBeam-Mistral-7B-512k · LongBench MultiFieldQA | score: 36.74 -> 50.50 | Improved |
| use Tactic at 90% instead of Ada-SnapKV | MegaBeam-Mistral-7B-512k · RULER 96K | accuracy: 87.1 -> 82.6 | Hurt |
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.
- Quest 75% — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 8 of 8 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
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.91)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.91)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.90)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.90)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.90)
- Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · similarity 0.90)
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.
