AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

LycheeDecode: Accelerating Long-Context LLM Inference via Hybrid-Head Sparse Decoding

Gang Lin, Dongfang Li, Zhuoen Chen, Yukun Shi, Xuhui Chen, Baotian Hu, Min Zhang

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 4 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 head-level sharing (LycheeDecode) instead of layer-level (TidalDecode), budget 1024Qwen3-8B · LongBenchaverage score: 29.70 -> 32.38Improved
use LycheeDecode instead of TidalDecode, budget 4096Llama-3-8B-Gradient · LongBenchaverage score: 32.86 -> 33.07Improved
use LycheeDecode instead of TidalDecodeDeepSeek-R1-Distill-Qwen-7B · Math reasoningaverage accuracy: 30.2 -> 44.2Improved
add cache correctionLycheeDecode · Math reasoning (Llama-8B distill)average accuracy: 36.8 -> 40.3Improved
add cache correctionLycheeDecode · Math reasoning (Qwen-7B distill)average accuracy: 44.2 -> 44.9Improved
add cache correctionTidalDecode · Math reasoning (Llama-8B distill)average accuracy: 31.6 -> 35.7Improved
increase top-k token budget from 1024 to 4096 (ablation)LycheeDecode · LongBenchaverage score: 31.02 -> 33.07Improved
use Top-p p=0.99 instead of Top-k 1024 at similar sparsity (ablation)LycheeDecode · LongBenchaverage score: 31.02 -> 27.26Hurt
use Ratio theta=70% instead of Top-k 4096 (ablation)LycheeDecode · LongBenchaverage score: 33.07 -> 32.41Hurt
use Threshold tau=1e-4 instead of Top-k 1024 (ablation)LycheeDecode · LongBenchaverage score: 31.02 -> 21.18Hurt
add sparse decoding with 4096 budget at longer contextLycheeDecode vs full attention · RULERaverage score: 63.30 -> 62.79Hurt
use HardKuma instead of HardConcrete or direct optimization for head identification (ablation)LycheeDecode · LongBenchaverage scoreImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 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