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
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 head-level sharing (LycheeDecode) instead of layer-level (TidalDecode), budget 1024 | Qwen3-8B · LongBench | average score: 29.70 -> 32.38 | Improved |
| use LycheeDecode instead of TidalDecode, budget 4096 | Llama-3-8B-Gradient · LongBench | average score: 32.86 -> 33.07 | Improved |
| use LycheeDecode instead of TidalDecode | DeepSeek-R1-Distill-Qwen-7B · Math reasoning | average accuracy: 30.2 -> 44.2 | Improved |
| add cache correction | LycheeDecode · Math reasoning (Llama-8B distill) | average accuracy: 36.8 -> 40.3 | Improved |
| add cache correction | LycheeDecode · Math reasoning (Qwen-7B distill) | average accuracy: 44.2 -> 44.9 | Improved |
| add cache correction | TidalDecode · Math reasoning (Llama-8B distill) | average accuracy: 31.6 -> 35.7 | Improved |
| increase top-k token budget from 1024 to 4096 (ablation) | LycheeDecode · LongBench | average score: 31.02 -> 33.07 | Improved |
| use Top-p p=0.99 instead of Top-k 1024 at similar sparsity (ablation) | LycheeDecode · LongBench | average score: 31.02 -> 27.26 | Hurt |
| use Ratio theta=70% instead of Top-k 4096 (ablation) | LycheeDecode · LongBench | average score: 33.07 -> 32.41 | Hurt |
| use Threshold tau=1e-4 instead of Top-k 1024 (ablation) | LycheeDecode · LongBench | average score: 31.02 -> 21.18 | Hurt |
| add sparse decoding with 4096 budget at longer context | LycheeDecode vs full attention · RULER | average score: 63.30 -> 62.79 | Hurt |
| use HardKuma instead of HardConcrete or direct optimization for head identification (ablation) | LycheeDecode · LongBench | average score | Improved |
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.
- TidalDecode — TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention (ICLR 2025) · better on 6 of 6 shared results
- DuoAttention — DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025) · better on 2 of 2 shared results
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 2 of 2 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
- DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025 · similarity 0.92)
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.92)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.92)
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.91)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.91)
- ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference (ICML 2025 · similarity 0.91)
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.
