A Simple Plug-in for Improving Eviction-Based KV Cache Compression

Abstract:KV cache growth is a major bottleneck for long-context inference in large language models. Existing methods are often dominated by binary eviction or representation approximation, which may underutilize tokens that are not critical for exact retention but are still reconstructable. We present VECTOR, a plug-and-play augmentation for eviction-based pipelines that introduces three-way token routing: retention, approximation, and eviction. VECTOR combines an importance signal from the base scorer with a reconstructability signal from an offline-calibrated regression-based value estimation. By leveraging reconstructability, VECTOR recovers useful value information that would otherwise be irreversibly lost under binary eviction, while preserving key vectors for attention routing stability. On LongBench, RULER with contexts up to 128K tokens, and multi-turn SCBench, VECTOR yields significant gains over its base eviction methods under high compression. At a 90% compression ratio, SnapKV augmented with VECTOR outperforms six competing compression methods, including D2O and RocketKV, on all three evaluated models. Since VECTOR makes all routing decisions at prefill, decoding runs on a standard compressed cache with no measurable overhead.

Machine Learning (cs.LG)

arXiv:2605.23258 [cs.LG]

(or arXiv:2605.23258v2 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2605.23258 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.