The paper investigates how the positional geometry of cached key‑value (KV) states affects inference quality in sparse KV compression. It introduces RoPERelay, a method that reassigns compact, order‑preserving RoPE coordinates and re‑rotates cached key phases to better align with current queries, without requiring new selection or extra cache budget. Experiments on two models and multiple selectors show that RoPERelay improves answer quality in most cases, notably raising Llama‑3.1‑8B‑Instruct accuracy from 23.6% to 35.4% on LongMemEval‑S with only a 1–3% runtime overhead.
arXiv:2610.08811v1 Announce Type: new Abstract: As context windows scale to tens or hundreds of thousands of tokens, KV cache compression has become essential for efficient LLM inference. Existing me...
arXiv:2609.36835v1 Announce Type: new Abstract: Long-context large language model inference is bottlenecked by KV caches that grow linearly with sequence length. This burden is especially severe for...
arXiv:2607. 06519v1 Announce Type: new Abstract: Long-context LLM inference is increasingly limited by the memory and bandwidth cost of KV caches, yet aggressive compression can remove the layer-specific evidence needed for retrieval and multi-step reasoning.
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.
arXiv:2608.21362v1 Announce Type: new Abstract: Transformer-based large language models (LLMs) incur high prefill latency because key-value (KV) tensors must be recomputed for each request. Existing...
arXiv:2608.23843v1 Announce Type: new Abstract: Long-context inference in large language models (LLMs) is increasingly limited by the memory required for the key-value (KV) cache. KV cache compressio...
arXiv:2609.37988v1 Announce Type: new Abstract: As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This i...
KV$^2$ is a query‑agnostic key‑value cache compression technique that selectively reconstructs only informative in‑context tokens using a lightweight proxy scorer before final eviction scoring. On benchmarks such as RULER, Needle‑in‑a‑Haystack, and LongBench, KV$^2$ outperforms baseline methods, especially under tight memory budgets, achieving higher scores with lower runtime and peak memory than full‑context reconstruction. The approach demonstrates that reusable KV‑cache compression can avoid reprocessing the entire prompt while maintaining quality.
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.
The paper introduces iS-KV, an online low‑rank KV‑cache compression technique that uses block‑incremental SVD to manage memory during long‑horizon autoregressive decoding. Unlike token‑eviction methods, iS‑KV retains all positions in a compact representation by keeping a recent window exact and incrementally folding older states into bounded‑rank bases, synchronizing coordinates as the basis evolves. Experiments on DeepSeek‑R1‑Distill‑Llama‑8B and Qwen3‑8B show that iS‑KV achieves high accuracy (82.6% and 89.2% respectively) while providing 4.06‑fold and 5.64‑fold compression, outperforming token‑eviction baselines under matched memory budgets.
The paper addresses the challenge of long input contexts in Retrieval-Augmented Generation (RAG) systems, where concatenating many retrieved chunks increases prefill workload and time to first token (TTFT). It proposes a dual strategy: fine‑tuning the model to be aware of KV cache concatenation and selectively recomputing only part of the KV caches. Experiments on the RULER benchmark show that for a 124k‑token input, this combined method boosts the RULER score by 9.7 points over a baseline that recomputes caches only, while cutting TTFT by 80% compared with full attention.
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.