Research map / Architectures and efficiency / Attention and transformers
xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction
Chi-Chih Chang, Wei-Cheng Lin, Chien-Yu Lin, Hung-Yueh Chiang, Yash Akhauri, Xilai Dai, Huiqiang Jiang, Yucheng Li, Luis Ceze, Kai-Chiang Wu, Mohamed S. Abdelfattah
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 |
|---|---|---|---|
| share a cross-layer SVD basis instead of per-layer SVD | Single SVD · RULER 64K (Llama-3.1-8B-Instruct) | avg accuracy: 45.71 -> 88.50 | Improved |
| use cross-layer SVD instead of cosine-similarity SLERP merging | MiniCache · RULER 64K (Llama-3.1-8B-Instruct) | avg accuracy: 45.04 -> 88.50 | Improved |
| use cross-layer SVD instead of cosine-similarity merging | MiniCache · LongBench (Llama-3.1-8B-Instruct) | avg score: 30.63 -> 42.27 | Improved |
| use cross-layer SVD instead of KIVI-2 quantization | KIVI-2 · RULER 64K (Llama-3.1-8B-Instruct) | avg accuracy: 86.87 -> 88.50 | Improved |
| use cross-layer SVD instead of token eviction (SnapKV) | SnapKV · RULER 64K (Llama-3.1-8B-Instruct) | avg accuracy: 89.68 -> 88.50 | No clear effect |
| use cross-layer SVD instead of token eviction on a compact-KV model | SnapKV/PyramidKV · RULER 64K (Qwen2.5-7B-Instruct-1M) | avg accuracy: 89.22 | Improved |
| replace single-layer SVD key compression with cross-layer (xK-SR) in ShadowKV | ShadowKV · RULER 64K (selective reconstruction) | avg accuracy: 87.17 -> 89.70 | Improved |
| use cross-layer compression at extreme rate (1.71x, 11.59x GPU memory) | ShadowKV · RULER 64K (selective reconstruction) | avg accuracy: 43.59 -> 79.18 | Improved |
| compress both keys and values with cross-layer SVD (xKV-SR) vs ShadowKV value compression | ShadowKV-V · RULER 64K (selective reconstruction) | avg accuracy: 70.94 -> 89.69 | Improved |
| use xK-SR/xKV-SR instead of Quest | Quest · RULER 64K (selective reconstruction) | avg accuracy: 84.87 -> 89.70 | Improved |
| use selective reconstruction instead of dense reconstruction | xKV · Efficiency (A100) | kernel speedup: below 1.0x -> 3.5x | Improved |
| keep compressed cache on GPU (xKV-SR) rather than offloading values (ShadowKV) | ShadowKV · Efficiency (A100) | throughput: approx. 30% higher | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · 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
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.93)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.92)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.91)
- RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression (ICML 2025 · similarity 0.91)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.91)
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · 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.
