English

Beyond Token Eviction: Mixed-Dimension Budget Allocation for Efficient KV Cache Compression

Machine Learning 2026-03-24 v1

Abstract

Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.

Keywords

Cite

@article{arxiv.2603.20616,
  title  = {Beyond Token Eviction: Mixed-Dimension Budget Allocation for Efficient KV Cache Compression},
  author = {Ruijie Miao and Zhiming Wang and Wang Li and Shiwei Wu and Shufan Liu and Yanbing Jiang and Tong Yang},
  journal= {arXiv preprint arXiv:2603.20616},
  year   = {2026}
}
R2 v1 2026-07-01T11:30:58.004Z