Research map / Architectures and efficiency / Attention and transformers
Sparse Attention Adaptation for Long Reasoning
Yizhao Gao, Shuming Guo, Shijie Cao, Yuqing Xia, Yu Cheng, Lei Wang, Lingxiao Ma, Yutao Sun, Tianzhu Ye, Li Dong, Hayden Kwok-Hay So, Yu Hua and 3 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 |
|---|---|---|---|
| replace Quest training-free selection with learned self-distilled AttnGate | sparse decoding, Qwen3-8B · AIME24, 4k budget | accuracy: 44.2 -> 72.3 | Improved |
| replace Quest with SeerAttention-R at 2k budget | sparse decoding, Qwen3-8B · AIME24, 2k budget | accuracy: 13.3 -> 56.6 | Improved |
| replace Quest with SeerAttention-R at 8k budget | sparse decoding, Qwen3-8B · AIME24, 8k budget | accuracy: 59.6 -> 75.1 | Improved |
| replace Quest with SeerAttention-R | sparse decoding, Qwen3-8B · AIME24, 2k budget | generation length (k): 30.0 -> 19.8 | Improved |
| increase token budget from 2k to 8k for SeerAttention-R | SeerAttention-R, Qwen3-8B · AIME24 | generation length (k): 19.8 -> 15.1 | Improved |
| increase token budget from 2k to 8k for Quest | Quest, Qwen3-8B · AIME24 | accuracy: 13.3 -> 59.6 | Improved |
| use larger block size (16 to 128) for Quest | Quest, Qwen3-4B/8B · AIME24, 4k budget | accuracy | Hurt |
| use larger block size (32 to 128) for SeerAttention-R | SeerAttention-R, Qwen3-4B/8B · AIME24, 4k budget | accuracy | No clear effect |
| add hybrid dense attention in the first two layers to Quest (ablation) | Quest, Qwen3-4B · AIME24, block 64 | accuracy | Improved |
| add hybrid dense attention in the first two layers to SeerAttention-R (ablation) | SeerAttention-R, Qwen3-4B · AIME24, block 64 | accuracy | No clear effect |
| use threshold sparsification instead of token budget (ablation) | SeerAttention-R, Qwen3-4B · AIME24 | accuracy at high sparsity | Improved |
| add positional embedding (RoPE) in AttnGate (ablation) | SeerAttention-R · reasoning benchmarks | accuracy | Improved |
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 — 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
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.88)
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.88)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.87)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.87)
- LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss (NeurIPS 2026 · similarity 0.87)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
