Research map / Architectures and efficiency / Quantization
PatternKV: Flattening KV Representation Expands Quantization Headroom
Ji Zhang, Yiwei Li, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Yueqi Zhang, Jiayi Shi, Chuyi Tan, Boyuan Pan, Yao Hu, Kan Li
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 |
|---|---|---|---|
| PatternKV pattern-aligned residual quantization instead of KIVI | KIVI, Llama3.1-8B-Instruct · LongBench (2-bit) | average score: 44.33 -> 45.33 | Improved |
| PatternKV instead of KIVI on 70B | KIVI, Llama3.1-70B-Instruct · LongBench (2-bit) | average score: 51.48 -> 51.61 | Improved |
| PatternKV instead of KIVI at 4-bit | KIVI, Llama3.1-8B-Instruct · LongBench (4-bit) | average score: 46.34 -> 46.41 | Improved |
| PatternKV instead of KIVI on long-CoT AIME 25 | KIVI, Llama-8B · AIME 25 (2-bit) | Avg@8: 12.50 -> 17.50 | Improved |
| PatternKV instead of KIVI on Qwen-14B AIME 25 | KIVI, Qwen-14B · AIME 25 (2-bit) | Avg@8: 37.08 -> 35.42 | Hurt |
| add K pattern alignment (ablation) | PatternKV · LongBench (2-bit) | average score: 44.53 -> 45.33 | Improved |
| add V pattern alignment (ablation) | PatternKV · LongBench (2-bit) | average score: 44.96 -> 45.33 | Improved |
| add new pattern vectors during decoding (ablation) | PatternKV · LongBench (2-bit) | average score: 45.37 -> 45.33 | No clear effect |
| add new pattern vectors during decoding (ablation) | PatternKV · GSM8K (2-bit, zero-shot CoT) | accuracy: 75.49 -> 75.58 | Improved |
| add adaptive z-test threshold for V patterns (ablation) | PatternKV · LongBench (2-bit) | average score: 24.67 -> 45.33 | Improved |
| add adaptive z-test threshold for V patterns (ablation) | PatternKV · GSM8K (2-bit, zero-shot CoT) | accuracy: 0.30 -> 75.58 | Improved |
| increase number of patterns from 2 to 32 (ablation) | PatternKV · LongBench (2-bit) | average score: 44.57 -> 45.33 | 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.
- ZipCache — ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 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
- KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference (ICML 2025 · similarity 0.90)
- NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache (NeurIPS 2025 · similarity 0.89)
- KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization (NeurIPS 2024 · similarity 0.89)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.89)
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · similarity 0.88)
- Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation (NeurIPS 2026 · similarity 0.88)
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.
