Tag
Tsinghua University and WuWenXinQiong teams have open-sourced a paper introducing C2C, a method for direct communication between large language models via cache fusion, eliminating text intermediaries to boost inference speed and accuracy significantly.
Chinese researchers open-sourced Cache-to-Cache (C2C), a new paradigm enabling LLMs to communicate directly through KV-cache fusion, bypassing text generation to reduce latency and improve accuracy by up to 14.2%.