Cache-to-Cache: Direct Semantic Communication Between Large Language Models

by Doris Putra 18 September 2026 2 min read 34x views

[Submitted on 3 Oct 2025 (v1), final revised 2 Mar 2026 (this version, v2)]

Abstract:Multi-LLM systems harness the complementary strengths of varied Large Language Models, achieving achievement and effectiveness gains that are not attainable by a sole model. In existing designs, LLMs communicate through text, forcing inner representations to be transformed into output token sequences. This procedure the two loses affluent semantic data and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments display that enriching the KV-Cache semantics can enhance reply norm without expanding cache size, supporting KV-Cache as an productive average for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for straightforward semantic communication between LLMs. C2C uses a neural network to project and fuse the origin model's KV-cache alongside that of the mark example to allow straightforward semantic transfer. A learnable gating scheme selects the mark layers that advantage from cache communication. Compared alongside content communication, C2C utilizes the deep, specialized semantics from the two models, during avoiding definitive intermediate content generation. Experiments display that C2C achieves 6.4-14.2% higher average accuracy than idiosyncratic models. It additional outperforms the content communication paradigm by about 3.1-5.4%, during delivering an average 2.5x speedup in latency. Our code is accessible at this https URL.