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
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:7941-7960, 2026.

Abstract

Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank–aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75% KV cache compression and up to 20$\times$ overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9$\times$ speedup for the attention module and 3.1$\times$ end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-bhatnagar26a, title = {{STAR}-{KV}: Low-Rank {KV} Cache Compression via Soft Thresholding for Adaptive Rank Control}, author = {Bhatnagar, Priyansh and Moradifirouzabadi, Ashkan and Yang, Se-Hyun and Lee, Seungjae and Choi, Jungwook and Kang, Mingu}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {7941--7960}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/bhatnagar26a/bhatnagar26a.pdf}, url = {https://proceedings.mlr.press/v306/bhatnagar26a.html}, abstract = {Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank–aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75% KV cache compression and up to 20$\times$ overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9$\times$ speedup for the attention module and 3.1$\times$ end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.} }
Endnote
%0 Conference Paper %T STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control %A Priyansh Bhatnagar %A Ashkan Moradifirouzabadi %A Se-Hyun Yang %A Seungjae Lee %A Jungwook Choi %A Mingu Kang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-bhatnagar26a %I PMLR %P 7941--7960 %U https://proceedings.mlr.press/v306/bhatnagar26a.html %V 306 %X Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank–aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75% KV cache compression and up to 20$\times$ overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9$\times$ speedup for the attention module and 3.1$\times$ end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.
APA
Bhatnagar, P., Moradifirouzabadi, A., Yang, S., Lee, S., Choi, J. & Kang, M.. (2026). STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:7941-7960 Available from https://proceedings.mlr.press/v306/bhatnagar26a.html.

Related Material