Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains unattainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), 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 that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 8.5-10.5% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.0-5.0%, while delivering an average 2.0× speedup in latency.
Example: In T2T, ambiguous text fails to convey structural semantics. C2C directly transfers precise semantic understanding without intermediate text generation.
High-dimensional representations compressed into linear text
Natural language is inherently vague and can be misinterpreted
Token-by-token generation for every communication
Two oracle experiments validate that KV-Cache enables effective inter-model communication
Can enriched KV-Cache improve performance without extending sequence length?
| Method | Cache Len | Accuracy |
|---|---|---|
| Direct | |X| | 58.42% |
| Few-shot | |E|+|X| | 63.39% |
| Oracle | |X| | 62.34% |
✓ Enriched cache improves quality at same length
Can one model's KV-Cache be used by another?
✓ Cache is generally convertible between LLMs
C2C Fuser: Projects and fuses KV-Caches from Sharer into Receiver through (1) Projection, (2) Dynamic Weighting, and (3) Learnable Gating modules.
Concatenates and processes KV-Caches from both models
Input-aware modulation for adaptive information flow
Per-layer gates select which layers benefit from fusion
Training: Freeze both models, train only C2C module with next-token prediction loss
Select model combinations to explore C2C performance across all benchmarks
The model that provides contextual knowledge through its KV-Cache representations
The model that receives and utilizes knowledge from the Sharer via KV-Cache projection
| Receiver | Sharer | Benchmark | Receiver Only | Sharer Only | Routing | Text-to-Text | Cache-to-Cache |
|---|
Generalization:
First direct semantic communication between LLMs beyond text
8.5-10.5% higher accuracy than individual models; 3.0-5.0% better than text communication; 2× faster
Works across models, families, sizes, and specializations
@article{fu2025c2c,
title={Cache-to-Cache: Direct Semantic Communication Between Large Language Models},
author={Tianyu Fu and Zihan Min and Hanling Zhang and Jichao Yan and Guohao Dai and Wanli Ouyang and Yu Wang},
journal={arXiv preprint arXiv:2510.03215},
year={2025},
}