---
id: 20260826-T0-02
title: "KVBoost：按块缓存KV，偏差指导重算，推理提速"
title_en: "KVBoost: Chunk-Level KV Cache Reuse Cuts LLM Prefill Latency"
url: https://ai.daily.yangsir.net/daily/20260826-T0-02
issue_date: 2026-08-26
publish_date: 2026-08-25T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2608.21362
---

# KVBoost：按块缓存KV，偏差指导重算，推理提速

arXiv新论文提出KVBoost，一种针对Transformer大模型推理的优化方案。它通过按块复用键值（KV）缓存，并结合偏差指导的重算策略，解决了传统前缀缓存要求提示词共享连续前缀的限制。该方法可显著降低预填充（prefill）延迟，提升吞吐量。对于处理长上下文或高并发请求的推理服务，KVBoost提供了一种更灵活的缓存复用方案，有望降低运营成本。

## English Version

**KVBoost: Chunk-Level KV Cache Reuse Cuts LLM Prefill Latency**

A new arXiv paper introduces KVBoost, an optimization for Transformer LLM inference. It enables chunk-level KV cache reuse with deviation-guided recomputation, overcoming the limitation of traditional prefix caching that requires shared leading contiguous prefixes. This method significantly reduces prefill latency and improves throughput. For inference services handling long contexts or high concurrency, KVBoost offers a more flexible caching strategy that could lower operational costs.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2608.21362)

**详情页**：https://ai.daily.yangsir.net/daily/20260826-T0-02

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*