Research map / Architectures and efficiency / Quantization
STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control
Priyansh Bhatnagar, Ashkan Moradifirouzabadi, Se-Hyun Yang, SeungJae Lee, Jungwook Choi, Mingu Kang
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 learned adaptive rank selection instead of uniform static rank (ablation) | STAR-KV at 75% compression · Commonsense zero-shot (6 tasks avg) | accuracy: 38.88 -> 60.49 | Improved |
| add block-wise Hadamard to mixed-precision quantization (ablation) | STAR-KV 75% + 3.2-bit quant · Commonsense zero-shot (5 tasks avg) | accuracy: 57.27 -> 59.18 | Improved |
| use block-wise Hadamard with mixed precision instead of global Hadamard at uniform 3-bit (ablation) | STAR-KV 75% + quant · Commonsense zero-shot (5 tasks avg) | accuracy: 58.23 -> 59.18 | Improved |
| hybrid decomposition: head-wise keys, joint values instead of joint for both (ablation) | STAR-KV · Reconstruction latency vs accuracy | latency | Improved |
| apply soft-threshold STAR-KV at 60% instead of Palu at 50% compression | LLaMA-3.1-8B-Instruct · LongBench (7 tasks avg) | score: 32.21 -> 39.58 | Improved |
| STAR-KV 60% instead of Palu 60% on RULER 16K | LongChat-v1.5-32K · RULER 16K avg | accuracy: 62.94 -> 84.16 | Improved |
| compression-loss weight gamma 1.0 / 0.1 / 0.01 (ablation) | STAR-KV 60% · Commonsense zero-shot (6 tasks avg) | accuracy: 61.28 / 61.53 / 61.33 | No clear effect |
| change calibration dataset from FineWeb-Edu to C4 or RedPajama (ablation) | STAR-KV 60% on LongChat-v1.5-32K · Commonsense zero-shot (6 tasks avg) | accuracy: 61.53 -> 61.39 / 61.18 | No clear effect |
| change calibration dataset from FineWeb-Edu to C4 or RedPajama (ablation) | STAR-KV 60% on LLaMA-3.1-8B-Inst · Commonsense zero-shot (6 tasks avg) | accuracy: 66.43 -> 66.08 / 65.75 | No clear effect |
| outlier:inlier split 20:80 vs 10:90 (ablation) | STAR-KV 75% + quant · Commonsense zero-shot (5 tasks avg) | accuracy: 58.80 -> 59.18 | Improved |
| outlier:inlier split 40:60 vs 20:80 (ablation) | STAR-KV 75% + quant · Commonsense zero-shot (5 tasks avg) | accuracy: 59.18 -> 59.52; compression 20.00 -> 18.82 | Mixed |
| combine 60% low-rank STAR-KV with 60% H2O token pruning | LongChat-v1.5-32K · Commonsense zero-shot (6 tasks avg) | accuracy: 59.95 -> 59.99 | Improved |
Closest papers by meaning
- KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation (NeurIPS 2026 · similarity 0.93)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.90)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.88)
- Lexico: Extreme KV Cache Compression via Sparse Coding over Universal Dictionaries (ICML 2025 · similarity 0.88)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.88)
- CompilerKV: Risk-Adaptive KV Cache Compression via Offline Experience Compilation (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.
