Cache-to-Cache: Direct Semantic Communication Between LLMs (2025)

AI 导读

文章提出Cache-to-Cache技术，通过KV缓存直接传输语义信息，解决LLM间文本通信的信息丢失与延迟问题，提升多模型协作效率。

This paper introduces Cache-to-Cache, a method enabling direct semantic communication between LLMs via KV-cache, improving collaboration efficiency by avoiding text-based communication limitations.

重点速览

• 现有LLM通信依赖文本转换导致信息丢失与延迟 Text-based LLM communication causes information loss and latency
• C2C通过KV缓存直接传输语义信息 C2C transfers semantics directly via KV-cache
• 引入可学习门控机制选择目标层进行通信 Learnable gating selects beneficial target layers
• 实验显示C2C准确率提升6.4-14.2% Experiments show 6.4-14.2% accuracy improvement
• 实现2.5倍延迟降低与3.1-5.4%性能超越 Achieves 2.5x latency reduction and 3.1-5.4% performance gain

一句话：C2C通过深度语义直接传输突破文本通信瓶颈，显著提升多模型协作效率。 / C2C breaks text communication bottlenecks by enabling direct semantic transfer between LLMs.

推荐理由 该方法创新性地利用缓存进行语义传输，为多模型系统优化提供可落地的技术路径。