大型语言模型实现直接语义通信新方法

研究人员提出了Cache-to-Cache（C2C）范式，使大型语言模型能够实现直接的语义通信1。该方法通过神经网络投影和融合源模型的KV-Cache与目标模型的KV-Cache，避免了基于文本通信的信息损失和生成延迟1。

根据实验结果，C2C范式相比单个模型的准确率平均提升6.4-14.2%，相比文本通信方式提升3.1-5.4%，同时实现约2.5倍的延迟加速1。该方法采用可学习的门控机制，能够自动选择受益于缓存通信的目标层1。论文于2025年10月3日首次提交，并在2026年3月2日进行了修订1。

Researchers have introduced a novel Cache-to-Cache (C2C) paradigm that enables direct semantic communication between large language models, bypassing the inefficiencies of text-based interaction 1. The method employs neural network projection and fusion techniques to combine the KV-Cache from a source model with that of a target model, circumventing the information loss and generation latency inherent in traditional text-based communication 1.

Experimental results demonstrate substantial performance improvements across multiple dimensions 1. The C2C approach achieves accuracy gains of 6.4 to 14.2 percent compared to individual models operating independently, and 3.1 to 5.4 percent improvement over text-based communication paradigms 1. Additionally, the method delivers approximately 2.5 times speedup in latency 1. The framework incorporates a learnable gating mechanism that selectively determines which target layers benefit most from cache-based communication 1. The initial version of the research was submitted on October 3, 2025, with a revised version submitted on March 2, 2026 1.