Research map / Architectures and efficiency / Attention and transformers
DSSA: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation
Weilin Zhao, Zihan Zhou, Zhou su, Chaojun Xiao, Yuxuan Li, Yanghao Li, Yudi Zhang, Weilun Zhao, Zhen Li, Yuxiang Huang, Ao Sun, Xu Han and 1 more
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 InfLLM-V2 shared-parameter sparse attention instead of NSA for short-to-long adaptation | 8B model · RULER 32k | average accuracy: 59.92 -> 82.62 | Improved |
| use InfLLM-V2 instead of NSA | 8B model · LongPPL | perplexity: 4.24 -> 2.12 | Improved |
| use InfLLM-V2 instead of NSA | 8B model · long reasoning average | average: 37.28 -> 42.66 | Improved |
| switch InfLLM-V2 to dense mode at inference (ablation) | InfLLM-V2 · RULER 32k | average accuracy: 82.62 -> 88.32 | Improved |
| add LSE Approximation to block selection (ablation) | InfLLM-V2 (Sparse) · RULER 32k | average accuracy: 82.09 -> 82.62 | No clear effect |
| add LSE Approximation to block selection (ablation) | InfLLM-V2 kernel · Block selection time, A100 128k | time (ms): 75.36 -> 56.59 | Improved |
| return to dense attention after long finetuning with InfLLM-V2 | InfLLM-V2 (Dense) · general short tasks average | average: 67.41 -> 66.76 | No clear effect |
| use NSA instead of full attention during long finetuning | 8B model · general short tasks average | average: 67.41 -> 60.63 | Hurt |
| InfLLM-V2 Sparse vs FullAttn on long reasoning | 8B model · AIME 25 | accuracy: 30.63 -> 29.38 | No clear effect |
Closest papers by meaning
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.89)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.89)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.89)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.88)
- Curse of High Dimensionality Issue in Transformer for Long Context Modeling (ICML 2025 · similarity 0.88)
- Hardware-aligned Hierarchical Sparse Attention for Efficient Long-term Memory Access (NeurIPS 2025 · similarity 0.88)
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.
