---
id: 20260526-T0-01
title: "AI模型之间不用文字也能对话：Latent Cache Flow绕过文本解码"
title_en: "Latent Cache Flow: AI Models Communicate Without Text"
url: https://ai.daily.yangsir.net/daily/20260526-T0-01
issue_date: 2026-05-26
publish_date: 2026-05-25T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2605.22863
---

# AI模型之间不用文字也能对话：Latent Cache Flow绕过文本解码

当前LLM多智能体协作依赖文本通信，需经过自回归解码和重新编码，带来显著延迟和信息损失。新方法Latent Cache Flow（LCF）让模型之间直接通过隐层缓存共享内部状态，跳过文本生成环节，降低了通信延迟并减少信息损失。该方法在多智能体协作场景下可提升系统级通信效率，对需要多轮交互的AI Agent框架具有实际应用价值。

## English Version

**Latent Cache Flow: AI Models Communicate Without Text**

Current LLM agent communication relies on text, causing latency and information loss from autoregressive decoding. Latent Cache Flow (LCF) lets models share internal states directly via latent caches, bypassing text generation entirely. This reduces communication latency and information loss. It can improve efficiency in multi-agent AI frameworks requiring frequent interactions.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2605.22863)

**详情页**：https://ai.daily.yangsir.net/daily/20260526-T0-01

---

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