NestedKV: Nested Memory Routing for Long-Context KV Cache Compression

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Chen, Hong, Liu, Xiang, Gao, Yubo, Fan, Yuxuan, Wang, Bo, Chu, Yuanlin, Lin, Yuanguo, Hu, Xuming
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911718497583104
author Chen, Hong
Liu, Xiang
Gao, Yubo
Fan, Yuxuan
Wang, Bo
Chu, Yuanlin
Lin, Yuanguo
Hu, Xuming
author_facet Chen, Hong
Liu, Xiang
Gao, Yubo
Fan, Yuxuan
Wang, Bo
Chu, Yuanlin
Lin, Yuanguo
Hu, Xuming
contents Long-context language models are limited by the memory footprint of the key-value (KV) cache. Existing training-free KV compression methods usually rank tokens by one importance signal -- attention, recency, layer-wise allocation, or key distinctiveness -- which becomes brittle when useful context is globally distinctive, locally episodic, or immediately relevant. We introduce NestedKV, a key-only KV cache compression method inspired by the Continuum Memory System in Nested Learning. NestedKV maintains global, block-level, and sliding-window key anchors, scores tokens by multi-time-scale cosine anomaly, and combines the resulting rankings with a training-free outer learner using head-adaptive mixing and surprise-gated token routing. The score is paired with adaptive per-head budgets and requires no training or LLM modification. Across RULER (4k--32k), LooGLE, LongBench, LongBench-E, InfiniteBench, and MMLU-Pro on Qwen3 and Llama-3.2 models, NestedKV is strongest when the retained cache is small. On Qwen3-4B, it improves over KeyDiff by up to 19.10 points on RULER and 19.29 on LongBench at $r=0.75$; at $r=0.95$, it retains 37.32 on LongBench versus 17.55 for KeyDiff.
format Preprint
id arxiv_https___arxiv_org_abs_2605_26678
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle NestedKV: Nested Memory Routing for Long-Context KV Cache Compression
Chen, Hong
Liu, Xiang
Gao, Yubo
Fan, Yuxuan
Wang, Bo
Chu, Yuanlin
Lin, Yuanguo
Hu, Xuming
Computation and Language
Long-context language models are limited by the memory footprint of the key-value (KV) cache. Existing training-free KV compression methods usually rank tokens by one importance signal -- attention, recency, layer-wise allocation, or key distinctiveness -- which becomes brittle when useful context is globally distinctive, locally episodic, or immediately relevant. We introduce NestedKV, a key-only KV cache compression method inspired by the Continuum Memory System in Nested Learning. NestedKV maintains global, block-level, and sliding-window key anchors, scores tokens by multi-time-scale cosine anomaly, and combines the resulting rankings with a training-free outer learner using head-adaptive mixing and surprise-gated token routing. The score is paired with adaptive per-head budgets and requires no training or LLM modification. Across RULER (4k--32k), LooGLE, LongBench, LongBench-E, InfiniteBench, and MMLU-Pro on Qwen3 and Llama-3.2 models, NestedKV is strongest when the retained cache is small. On Qwen3-4B, it improves over KeyDiff by up to 19.10 points on RULER and 19.29 on LongBench at $r=0.75$; at $r=0.95$, it retains 37.32 on LongBench versus 17.55 for KeyDiff.
title NestedKV: Nested Memory Routing for Long-Context KV Cache Compression
topic Computation and Language
url https://arxiv.org/abs/2605.26678