---
id: 20260910-T0-12
title: "CONDUIT框架：使视觉语言模型更高效复用KV缓存"
title_en: "CONDUIT Framework Restores KV Cache for VLMs"
url: https://ai.daily.yangsir.net/daily/20260910-T0-12
issue_date: 2026-09-10
publish_date: 2026-09-09T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.05821
---

# CONDUIT框架：使视觉语言模型更高效复用KV缓存

来自 arXiv 的研究提出 CONDUIT 框架，用于解决视觉语言模型（VLM）在重复视觉内容场景下缓存复用失效的问题。该框架通过统一残差流恢复机制，在图像内容有细微变化时依然可以高效复用历史 KV 缓存，避免重复计算昂贵的视觉前缀，从而显著提升模型在视频理解、多轮对话等场景中的推理效率。

## English Version

**CONDUIT Framework Restores KV Cache for VLMs**

A new arXiv paper introduces CONDUIT, a unified residual-stream restoration framework designed to address the failure of KV cache reuse in vision-language models (VLMs) when recurring visual content appears with slight variations. The framework enables efficient reuse of historical KV caches by restoring the residual stream, avoiding expensive re-encoding of visual prefixes. This improves inference efficiency in applications like video understanding and multi-turn dialogue.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.05821)

**详情页**：https://ai.daily.yangsir.net/daily/20260910-T0-12

---

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