AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

FASA: FREQUENCY-AWARE SPARSE ATTENTION

Yifei Wang, Yueqi Wang, Zhenrui Yue, Huimin Zeng, Yong Wang, Ismini Lourentzou, Zhengzhong Tu, Xiangxiang Chu, Julian McAuley

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 11 citations (OpenAlex)

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
use FASA frequency-chunk token selection instead of SnapKV at 256 budgetKV-cache eviction (Llama3.2-3B) · LongBench-V1average score: 37.0 -> 41.5Improved
use FASA instead of QuestKV-cache eviction (Qwen2.5-7B) · LongBench-V1average score: 31.4 -> 47.9Improved
use FASA instead of full KV cacheQwen2.5-7B / Mistral-7B · LongBench-V1average score: 47.8 -> 47.9 (Qwen), 47.4 -> 47.8 (Mistral)No clear effect
use FASA instead of full KV cacheLlama3.2-3B / Llama3.1-8B / Qwen2.5-14B-1M · LongBench-V1average score: 42.2 -> 41.5, 48.7 -> 48.2, 50.3 -> 49.2Hurt
use FASA instead of RKV on long-CoT reasoningDeepSeek-R1-Distill-Llama-8B · MATH500 (budget 300)accuracy: 24.0 -> 62.2Improved
use FASA at 10% budget vs FKVDeepSeek-R1-Distill-Qwen-32B · MATH500 (budget 300)accuracy: 92.6 -> 86.4Hurt
vary calibration window K (ablation)FASA · LongBench Qasperaverage score: 44.5 (K=128) vs 43.9 (K=512)No clear effect
vary offline calibration dataset (ablation)FASA · LongBench 2Wiki/Musi/Hqa/Qasp/MF-en/Nqascore: CV .007 to .014No clear effect
add PyramidKV layer-wise budget allocation (ablation)FASA · LongBench Qasperscore: 43.7 -> 44.4 (budget 256)Improved
apply FASA to ALiBi modelBaichuan-13B-Chat · LongBench Qasperscore: 9.11 -> 7.80Hurt
apply FASA to Partial-RoPE modelDeepSeek-V2-Lite-Chat · LongBenchscore: Qasper 33.18 -> 33.46, Lcc 63.40 -> 62.49Mixed
use Triton FASA-C kernel at 64K contextFASA · decoding latencyspeedup: up to 2.56xImproved

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