Research map / Architectures and efficiency / Attention and transformers
Token Sparse Attention: Efficient Long-Context Inference with Interleaved Token Selection
Dongwon Jo, Beomseok Kang, Jiwon Song, Jae-Joon Kim
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 Token Sparse Attention to FlexPrefill | FlexPrefill, LLaMA-3.1-8B-Instruct · RULER 128K attention speedup | speedup: ×2.44 -> ×2.76 | Improved |
| add Token Sparse Attention to MInference | MInference, LLaMA-3.1-8B-Instruct · RULER 128K attention speedup | speedup: ×1.12 -> ×1.38 | Improved |
| add Token Sparse Attention to MInference | MInference, LLaMA-3.1-8B-Instruct · RULER avg | accuracy: 86.49 -> 86.05 | Hurt |
| add Token Sparse Attention to FlashAttention | Mistral-Nemo-12B-Instruct · RULER avg | accuracy: 67.60 -> 67.37 | No clear effect |
| add Token Sparse Attention to FlexPrefill | FlexPrefill, LLaMA-3.1-8B-Instruct · InfiniteBench avg | accuracy: 49.53 -> 49.23 | No clear effect |
| use dynamic token coverage instead of fixed sparsity ratio (matched ~1.5x speedup) (ablation) | Token Sparse Attention · RULER avg | accuracy: 85.43% (s=0.5, ×1.57) -> 86.84% (tau=0.010, ×1.51) | Improved |
| replace token eviction (FastKV) with interleaved Token Sparse Attention at matched speedup | LLaMA-3.1-8B-Instruct prefill · RULER avg | accuracy: 85.64 -> 86.84 | Improved |
| apply token-level sparsity to all layers instead of selected stable layers (ablation) | Token Sparse Attention · RULER | accuracy | Hurt |
| select layers with low inter-layer representation drift (ablation) | Token Sparse Attention · RULER 4K, random 3-layer sparsification | accuracy | Improved |
| add Token Sparse Attention on top of FlexPrefill vs tuning FlexPrefill gamma | FlexPrefill · RULER accuracy-speedup Pareto | accuracy at matched speedup | Improved |
| increase token coverage tau from 0.005 to 0.010 (ablation) | Token Sparse Attention on FlashAttention · RULER avg, LLaMA-3.1-8B | speedup and accuracy: ×1.36 -> ×1.51; 87.02% -> 86.84% | Mixed |
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.
- FlexPrefill — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 3 of 4 shared results
- FlashAttention — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022) · better on 4 of 4 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
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.93)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.91)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.91)
- OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference (ICML 2025 · similarity 0.90)
- Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning (NeurIPS 2025 · similarity 0.90)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic 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.
