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
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: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
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
By Junyoung Park, Jungwook Choi, Mingu Lee
The paper investigates how to recover language model quality lost when using low‑bit key–value caches for autoregressive decoding. By keeping the quantizer fixed and distilling the full‑precision cache behavior into low‑rank Q/K/V projection updates, the authors demonstrate that 4‑bit affine‑cache adapters recover roughly 54 % of the perplexity gap on TinyLlama‑1.1B and 76 % on Gemma‑4‑12B, while preserving most long‑context retrieval. Even a 2‑bit rank–token sweep can dramatically reduce TinyLlama’s perplexity, though it only partially restores retrieval performance.
By Seifeldin Abdellatif
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 introduces Random Attention, a method that evicts KV cache entries uniformly at random within each attention head while preserving the prompt. Experiments on four models and six reasoning tasks show that this simple strategy matches the performance of the best existing eviction methods and achieves 32‑43% higher throughput in vLLM deployments. The authors explain that the prompt is the most fragile cache component and that reasoning traces are redundantly stored across text and attention heads, making a selection score unnecessary.
By Heng Wang, Jielin Qiu, Wenting Zhao, Cheng Qian, Liangwei Yang, Jiawei Han, Heng Ji, Silvio Savarese, Shelby Heinecke, Huan Wang
Virtual Head Attention (VHA) is a lightweight attention mechanism that improves long‑context large language models by adding two linear operations—Q Premix and Linear Postmix—to recover head diversity and fuse inter‑head features without nonlinearities. VHA reduces KV cache size, training and prefill FLOPs, and maintains or surpasses model quality compared to existing methods such as GQA, MLA, and MFA across multiple scales and benchmarks. The design allows VHA to be folded into standard GQA inference pipelines, enabling efficient deployment with only two KV heads and full exploitation of Tensor Core parallelism.
By Xibo Ding, Guoxia Wang, Jinle Zeng, Jiabin Yang, Sirui Tao, Chao Yang, Dianhai Yu
arXiv:2607. 11942v1 Announce Type: cross Abstract: KV-cache compression methods are predominantly evaluated with the query appended to the context before compression -- a query-aware protocol.
By Daming Luo, Christy Liang, Junyu Xuan
arXiv:2607. 24555v1 Announce Type: cross Abstract: Serving large language models at long context is bottlenecked by the key-value (KV) cache, which is read in full at every decode step.
By Junsung Hwang
Page-EntroKV is a new KV‑cache eviction framework designed for grouped‑query attention (GQA) in long‑context autoregressive language models. It pools token importance scores across query heads within each physical group using sink‑isolated Renyi‑2 entropy, then projects these pooled scores onto page frames and performs eviction at the hardware tuple level. The authors prove theoretical guarantees on union overhead, budget preservation, and needle‑retention, and demonstrate on a Qwen2.5‑1.5B‑Instruct pilot that Page‑EntroKV achieves exact union overhead control, eliminates sink masquerade, and retains 100% of critical tokens compared to mean‑pooling baselines.
By Inbasekaran S
arXiv:2609.36722v1 Announce Type: new
Abstract: Large language model (LLM) agents repeatedly load reusable content, such as skills, documents, and memory entries, into the current context. Re-encodin...
By Xinghao Chen, Junnan Dong, Cai Ke, Chak Tou Leong, Haocheng Sun, Keyu Chen, Siyu An, Ruizhi Qiao, Xing Sun, Wenjie Li, Xiaoyu Shen