Cache-to-Cache reportedly lets AI models communicate without generating words

A post describes Cache-to-Cache (C2C) as an open-source method for language models to communicate through cached internal data, known as the KV cache, rather than generated words. It says a neural network transfers and combines the caches, while a learned gate selects which model layers benefit most. The post claims the approach eliminates intermediate text-generation latency, improves benchmark accuracy by up to 14.2% over individual models and delivers a 2.5× overall speedup.