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

This research introduces Cache-to-Cache (C2C), a new way for large language models to exchange information directly through KV-cache rather than generating intermediate text, improving both quality and speed. For CIOs, the business implication is clearer: multi-model AI systems can become more accurate, lower-latency, and potentially cheaper to operate, which strengthens the case for deploying orchestration-heavy AI workflows in customer support, knowledge work, and automation. Strategically, this points to a shift from text-based model integration toward semantic interoperability between models, requiring IT teams to rethink how they design, govern, and optimize AI pipelines.