Research map / Architectures and efficiency / Attention and transformers
A Unified Sparse Attention via Multi-Granularity Compression
Siran Liu, Zane Cao, Yongchao He
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 average pooling instead of max pooling for composite tokens (ablation) | UniSparse (0.95) · RULER, Qwen2.5-7B-Instruct | overall accuracy: 81.01 -> 84.26 | Improved |
| use average pooling instead of stochastic pooling (ablation) | UniSparse (0.95) · HELMET, Qwen2.5-7B-Instruct | overall accuracy: 44.05 -> 47.03 | Improved |
| use average pooling instead of max pooling (ablation) | UniSparse (0.95) · HELMET, Llama-3.1-8B-Instruct | overall accuracy: 54.83 -> 56.21 | Improved |
| use average pooling instead of stochastic pooling (ablation) | UniSparse (0.95) · RULER, Llama-3.1-8B-Instruct | overall accuracy: 92.21 -> 92.56 | Improved |
| balanced Q-K compression (cq=ck=8) instead of query-heavy cq=32, ck=2 (ablation) | UniSparse (0.95) · RULER, Llama-3.1-8B-Instruct | overall accuracy: 91.37 -> 92.56 | Improved |
| balanced Q-K compression instead of cq=32, ck=2 (ablation) | UniSparse (0.95) · RULER, Qwen2.5-7B-Instruct | overall accuracy: 81.93 -> 84.26 | Improved |
| key-heavy compression cq=4, ck=16 instead of balanced (ablation) | UniSparse (0.95) · HELMET, Qwen2.5-7B-Instruct | overall accuracy: 47.03 -> 47.27 | No clear effect |
| compression factor c=16 instead of c=8 (ablation) | UniSparse (0.95) · RULER, Qwen2.5-7B-Instruct | overall accuracy: 84.26 -> 83.43 | Hurt |
| compression factor c=4 instead of c=8 (ablation) | UniSparse (0.95) · RULER, Qwen2.5-7B-Instruct | overall accuracy: 84.26 -> 84.51 | No clear effect |
| add head compression ch=2 at Top-P 0.95 (ablation) | UniSparse (0.95) · RULER, Llama-3.1-8B-Instruct | sparsity: 46.33% -> 45.13% | Hurt |
| add head compression ch=4 at Top-P 0.9 (ablation) | UniSparse (0.95) · RULER, Qwen2.5-7B-Instruct | sparsity: 60.12% -> 46.82% | Hurt |
| add head compression ch=2 (ablation) | UniSparse (0.95) · block selection overhead, 128K tokens | speedup over XAttention-0.95: 1.62x -> 2.64x | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- XAttention (0.9) — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 9 of 9 shared results
- FlexPrefill (0.95) — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 8 of 9 shared results
- FlashAttention — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022) · better on 2 of 5 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
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.90)
- 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)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.89)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.89)
- UniGist: Towards General and Hardware-aligned Sequence-level Long Context Compression (NeurIPS 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.
