Researchers from Tsinghua University have introduced Cache-to-Cache (C2C), a framework that enables large language models to exchange internal memory states directly. By bypassing text-based communication, the system uses key-value caches as a medium for sharing semantic information between a 'Sharer' model and a 'Receiver' model.
The framework utilizes a lightweight component called the 'Cache Fuser' to project and integrate the Sharer's cache into the Receiver's space. This design avoids increasing the overall cache size, allowing for efficient data transfer without additional storage overhead.
Testing across Qwen, Llama, and Gemma model families showed accuracy improvements of 6.4% to 14.2% compared to individual models, and 3.1% to 5.4% over traditional text-based collaboration. Communication latency was reduced by 2 to 2.5 times in standard setups, reaching up to 14 times faster in specific configurations.
The research, detailed in the paper 'Cache-to-Cache: Direct Semantic Communication Between Large Language Models,' was published as an arXiv preprint on October 3, 2025. It has been accepted for presentation at ICLR 2026, and the team has made the code and a demo available on GitHub.
A current limitation is the need to train the Cache Fuser for each specific pair of models. Scaling this approach to pipelines involving numerous heterogeneous models remains an area for further engineering development.
