AI Paper Atlas

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

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs

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 changedWhereMeasuredEffect
replace Quest training-free selection with learned self-distilled AttnGatesparse decoding, Qwen3-8B · AIME24, 4k budgetaccuracy: 44.2 -> 72.3Improved
replace Quest with SeerAttention-R at 2k budgetsparse decoding, Qwen3-8B · AIME24, 2k budgetaccuracy: 13.3 -> 56.6Improved
replace Quest with SeerAttention-R at 8k budgetsparse decoding, Qwen3-8B · AIME24, 8k budgetaccuracy: 59.6 -> 75.1Improved
replace Quest with SeerAttention-Rsparse decoding, Qwen3-8B · AIME24, 2k budgetgeneration length (k): 30.0 -> 19.8Improved
increase token budget from 2k to 8k for SeerAttention-RSeerAttention-R, Qwen3-8B · AIME24generation length (k): 19.8 -> 15.1Improved
increase token budget from 2k to 8k for QuestQuest, Qwen3-8B · AIME24accuracy: 13.3 -> 59.6Improved
use larger block size (16 to 128) for QuestQuest, Qwen3-4B/8B · AIME24, 4k budgetaccuracyHurt
use larger block size (32 to 128) for SeerAttention-RSeerAttention-R, Qwen3-4B/8B · AIME24, 4k budgetaccuracyNo clear effect
add hybrid dense attention in the first two layers to Quest (ablation)Quest, Qwen3-4B · AIME24, block 64accuracyImproved
add hybrid dense attention in the first two layers to SeerAttention-R (ablation)SeerAttention-R, Qwen3-4B · AIME24, block 64accuracyNo clear effect
use threshold sparsification instead of token budget (ablation)SeerAttention-R, Qwen3-4B · AIME24accuracy at high sparsityImproved
add positional embedding (RoPE) in AttnGate (ablation)SeerAttention-R · reasoning benchmarksaccuracyImproved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy