The paper investigates KV‑cache eviction strategies for sparse‑attention models, showing that selecting the largest attention weights is nearly optimal—closing only a median 2–5 % of the gap to full attention. It further demonstrates that differences in performance between eviction methods largely stem from memory usage, with the new training‑free ContourKV allocator outperforming state‑of‑the‑art methods in most pairwise comparisons while matching their byte‑efficiency.
By Jack Shi, Jerry Gu
arXiv:2608. 01247v1 Announce Type: cross Abstract: Query-agnostic KV cache eviction compresses a context once and reuses the resulting cache for arbitrary future queries, but performance can collapse under tight budgets.
By Changwoo Baek, Seungjun Shin, Kyeongbo Kong
Query-agnostic KV cache eviction compresses a context once and reuses the resulting cache for arbitrary future queries, but performance can collapse under tight budgets. Existing methods primarily improve which original KV pairs are retained.
arXiv:2609.03949v2 Announce Type: replace-cross
Abstract: A long-lived KV cache must be compressed before the queries that will read it exist. Selection by observed attention collapses there: on a No...
By WenJie Fan
VestigeKV is a new KV‑cache technique that uses a 64‑dimensional vestigial branch—originally a RoPE component repurposed during NoPE training—as a query‑independent eviction signal. By reading only 11 % of each cache row, the method partitions the cache into an attended tier (top‑m rows) and an archive tier (all other rows), which is GPU‑resident and never deleted. The approach achieves near‑perfect retrieval (1.00 at 8×, 0.92 at 32×) without any training, quantization, or changes to weights or kernels.
By WenJie Fan
arXiv:2607. 15498v1 Announce Type: cross Abstract: The key-value (KV) cache is the main memory bottleneck in long-context large language model (LLM) inference.
By Shahrzad Esmat, Dhawal Shah, Ali Jannesari
arXiv:2609.05637v2 Announce Type: replace
Abstract: A popular way to improve Retrieval-Augmented Generation (RAG) is to rewrite the user's question into several variants and search with all of them....
By Sara Shanian, Xiaoqin Yi, Pavlo Ruban, Kurt MacDonald
arXiv:2602. 08686v3 Announce Type: replace-cross Abstract: Prefill-only KV compression freezes a token subset at the end of prefill and decodes from it without further eviction.
By Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang
arXiv:2609.10266v1 Announce Type: new
Abstract: LLM serving systems already reuse KV caches, but only when the reused text sits at the very start of the prompt. Two growing workloads break this condi...
By Xi Shi, Qian Lou
arXiv:2608. 16309v1 Announce Type: cross Abstract: Static pruning is widely used to accelerate sparse neural retrieval, yet existing studies each validate their conclusions within a single custom pipeline, leaving it unclear which findings transfer to modern engines with different index organizations and dynamic pruning mechanisms.
By Zirui Song, Yuye Zhu, Yang Yang
arXiv:2606. 09916v1 Announce Type: cross Abstract: Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning.
By Junjie Li, Jiong Lou, Jie Li
The paper investigates how temporal aggregation of token scores during decoding-time KV cache compression affects eviction decisions. It shows that using an exponential‑moving‑average (EMA) aggregation keeps ranking changes minimal for many scorer variants, while others like KeyDiff or recency significantly alter rankings and degrade performance. Building on this, the authors propose InertiaKV and its lazy variant, which achieve 1.34–1.46× faster decoding than full refresh, and also evaluate a score‑free approach that slightly improves quality while eliminating further scoring.