AI models can now talk without words: What is Cache-to-Cache communication?

Researchers have developed Cache-to-Cache (C2C), a method that allows AI models to share information directly through their internal memory. The technique delivered faster responses and improved accuracy in controlled tests, but several limitations remain.

Researchers have developed a method that allows large language models (LLMs) to exchange information directly without generating written messages for each other. Called Cache-to-Cache (C2C), the technique could help AI systems work together faster and produce more accurate answers. The research was published at the International Conference on Learning Representations (ICLR) 2026.

The study, led by researchers from Tsinghua University and collaborating institutions, explores a problem in AI systems that use multiple models. Currently, one model typically generates a text response, which another model must read and process. C2C allows information to move between their internal memory structures, reducing the need for this intermediate........

© News9Live