Research map / Architectures and efficiency / Attention and transformers
ProtoKV: Long-context Knowledges Are Already Well-Organized Before Your Query
Zhiyuan Yu, Shijian Xiao, Zhangyue Yin, Xiaoran Liu, Lekai Xing, Wenzhong Li, Nguyen Cam-Tu, Sanglu Lu
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 |
|---|---|---|---|
| retain KV pairs by outlier degree only (SATKV) instead of SnapKV selection | Llama3-8B KV retention · NarrativeQA (budget 256) | score: 23.32 -> 21.12 | Hurt |
| retain KV pairs by outlier degree only (SATKV) instead of SnapKV selection | Llama3-8B KV retention · LCC (budget 256) | score: 60.27 -> 58.46 | Hurt |
| add semantic-prototype clustering (ProtoKV) over SnapKV | Llama3-8B · RULER NIAH (budget 128) | accuracy: 97.2 -> 98.2 | Improved |
| add semantic-prototype clustering (ProtoKV) over ChunkKV | Llama3-8B · RULER QA (budget 512) | accuracy: 42.2 -> 43.6 | Improved |
| add prototype clustering (ProtoKV) over SnapKV | Llama3-8B · RULER QA (budget 512) | accuracy: 43.8 -> 43.6 | No clear effect |
| use ProtoKV instead of K-means cluster methods (SqueezeAtt) with decoding-stage recall | Llama3-8B · LongBench (budget 512) | average score: 41.47 -> 42.59 | Improved |
| use ProtoKV instead of SqueezeAtt with prefilling-only retention | Llama3-8B · LongBench (budget 512) | average score: 40.42 -> 40.76 | Improved |
| ProtoKV instead of SnapKV on Needle In A Haystack | Llama3-8B-Instruct · Needle In A Haystack (budget 128) | retrieval accuracy: 94.2% -> 97.3% | Improved |
| combine ProtoKV with layer-wise budget allocation | Mistral-7B-Instruct · LongBench SDQA (budget 256) | score: 32.87 -> 33.12 | Improved |
| combine ProtoKV with head-wise budget allocation | Mistral-7B-Instruct · LongBench Code (budget 256) | score: 59.38 -> 60.16 | Improved |
| replace hybrid prototypes with chunk-only aggregation (remove SAT prototypes) (ablation) | ProtoKV · LongBench average, budgets 128-512 | average score | Improved |
| replace hybrid prototypes with LSH bucketing of all tokens (remove PDT chunks) (ablation) | ProtoKV · LongBench average, budgets 128-512 | average score | Improved |
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.
- ChunkKV — ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025) · better on 3 of 4 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
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.94)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.91)
- LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models (ICML 2025 · similarity 0.90)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.90)
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.90)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.90)
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.
