China publishes 'landmark paper' on AI-to-AI technique that kicks human 'bottleneck' out of the loop and replaces us with an AI 'modem' — C2C brainwave direct connection achieves 150% boost in inference speed
Tsinghua researchers developed C2C, allowing AI models to exchange internal information directly while reportedly improving collaborative inference speed and accuracy.
Chinese researchers have unveiled a groundbreaking technique called Cache-to-Cache (C2C) that enables separate artificial intelligence models to exchange internal information without producing any text. This method, published by Tsinghua University and accepted at the upcoming ICLR 2026 conference, promises significant improvements in AI collaboration.
By bypassing the need for models to generate text, C2C streamlines the information exchange process, potentially boosting inference speed by up to 150%. The key to this innovation lies in a learned Fuser, which transforms one model's internal data into a format suitable for another model. C2C's selective gating mechanism controls which layers receive this information, allowing models to retain their independent reasoning capabilities.
While the technique could result in accuracy gains of up to 14.2%, it is currently limited to open-weight models, as it requires direct access to a model's internal cache and layer structure. Most popular AI applications, however, do not expose this level of detail, meaning everyday users may not immediately benefit from this advancement.
Nonetheless, the researchers argue that C2C represents a step towards more efficient AI collaboration, potentially outperforming older methods where models still communicate through typed text.
Written by urgent.news from TechRadar's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
