Research map / Architectures and efficiency / Attention and transformers
Long-Context Modeling with Dynamic Hierarchical Sparse Attention for Memory-Constrained LLM Inference
Siheng Xiong, Joe Zou, Faramarz Fekri, Yae Jee Cho
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 dynamic chunking (learned boundary predictor) over fixed chunks (ablation) | DHSA · LongBench (density 12.5%), Llama-3.1-8B (4-bit) | avg score: 28.0 -> 31.8 | Improved |
| add length-normalized robust chunk representation (ablation) | DHSA · LongBench (density 12.5%), Llama-3.1-8B (4-bit) | avg score: 30.7 -> 31.8 | Improved |
| remove both dynamic chunking and robust chunk representation (ablation) | DHSA · LongBench (density 12.5%), Llama-3.1-8B (4-bit) | avg score: 31.8 -> 27.9 | Hurt |
| increase token density from 6.25% to 25.0% (ablation) | DHSA · LongBench, Llama-3.1-8B (4-bit) | avg score: 27.7 -> 34.4 | Improved |
| use DHSA instead of Block Sparse Attention at 12.5% density | DHSA · LongBench, Llama-3.1-8B (4-bit) | avg score: 27.9 -> 31.8 | Improved |
| use DHSA instead of Block Sparse Attention | DHSA · RULER 48K, Llama-3.1-8B (4-bit) | avg accuracy: 43.1 -> 71.5 | Improved |
| use DHSA instead of Quest at 12.5% density | DHSA · LongBench, Llama-3.1-8B (4-bit) | avg score: 30.9 -> 31.8 | Improved |
| sparse DHSA at 12.5% density versus dense attention | DHSA · LongBench, Llama-3.1-8B (4-bit) | avg score: 32.7 -> 31.8 | No clear effect |
| Block Sparse Attention fixed-size chunks on needle retrieval | Block Sparse · Needle-in-a-Haystack | retrieval accuracy | Hurt |
| add NMS to boundary selection (window 8) (ablation) | DHSA boundary predictor · boundary inference | performance | Improved |
| fuse multiple features (diff, product, cosine) instead of cosine similarity only (ablation) | DHSA boundary predictor · boundary detection | boundary stability | Improved |
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.
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 3 of 3 shared results
- DuoAttention — DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025) · better on 1 of 1 shared result
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
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.93)
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.92)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.92)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.91)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.91)
- Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity (NeurIPS 2025 · similarity 0.91)
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.
