Research map / Architectures and efficiency / Attention and transformers
Self-Indexing KVCache: Predicting Sparse Attention from Compressed Keys
Xu Yang, Jiapeng Zhang, Dongyang Zhao, Guo Chen, Zhuo Tang
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 |
|---|---|---|---|
| add sign bit to the 2-bit key quantization (ablation) | Self-Indexing KVCache · LongBench RB-P (Llama3.1-8B) | score: 50.9 -> 54.4 | Improved |
| add magnitude-based VQ to sign-only retrieval (ablation) | Self-Indexing KVCache · LongBench MF-en (Llama3.1-8B) | score: 52.8 -> 54.6 | Improved |
| add 64 full-precision sink tokens (ablation) | Self-Indexing KVCache · LongBench MF-en (Llama3.1-8B) | score: 52.6 -> 54.6 | Improved |
| add sink tokens (ablation) | Self-Indexing KVCache · LongBench HPQA and RB-P (Llama3.1-8B) | score: 55.2 -> 55.7 (HPQA) | No clear effect |
| quantize K and V to 2 bits instead of 16 bits | Self-Indexing KVCache · LongBench avg (Llama3.1-8B) | average score: 58.4 -> 58.2 | No clear effect |
| replace KMeans (20 iterations) with one-pass sign-based clustering | codebook construction · 16K input, batch 10 | time (ms): 2113.9 -> 88.85 | Improved |
| retrieve via compressed-domain LUT-GEMV instead of full dot product | top-k retrieval · 16K input, batch 10 | time (ms): 0.166 -> 0.039 | Improved |
| use custom sparse FlashAttention at 7.5% tokens vs full FlashAttention2 | attention stage · 16K input, batch 10 | time (ms): 0.776 -> 0.116 | Improved |
| sparse FlashAttention kernel vs Page Attention | attention stage · 16K input, batch 10 | time (ms): 0.101 -> 0.116 | No clear effect |
| add compressed-domain sparse attention on top of FlashAttention2 | prefill · TT2T 32K prompt | TT2T (s): 6.18 -> 6.51 | Hurt |
| use dynamic compressed-key retrieval instead of static SnapKV pruning | KV cache sparsification · RULER 32K (Llama3.1-8B) | average score: 70.8 -> 89.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 4 of 4 shared results
- FlashAttention2 — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022) · better on 0 of 1 shared result
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
- Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · similarity 0.91)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.90)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.89)
- Value-Guided KV Compression for LLMs via Approximated CUR Decomposition (NeurIPS 2025 · similarity 0.89)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.89)
- SALS: Sparse Attention in Latent Space for KV Cache Compression (NeurIPS 2025 · similarity 0.89)
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.
