Cache-to-Cache (C2C) enables Large Language Models to communicate directly through their KV-Caches, bypassing text generation. By projecting and fusing KV-Caches between models, C2C achieves 8.5–10.5% ...