Multi-LLM systems combine diverse models for better performance, but existing designs force communication through text, losing rich semantics and incurring token-by-token latency. Can LLMs communicate beyond text? Oracle experiments show enriching KV-Cache semantics improves response quality without increasing cache size.

Thus, Cache-to-Cache (C2C) is proposed: a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with the target model's, enabling direct semantic transfer. A learnable gating mechanism selects target layers that benefit from cache communication.

Compared with text communication, C2C utilizes deep, specialized semantics from both models while avoiding explicit intermediate text generation. Experiments show C2C achieves 6.4-14.2% higher average accuracy than individual models, outperforms text communication by approximately 3.1-5.4%, and delivers an average 2.5x speedup in latency. Code is available at the provided URL.