Tsinghua Researchers Report Faster AI Collaboration With C2C

Researchers at Tsinghua University describe a way for AI models to share information directly, without converting it into text first. Their method, Cache-to-Cache (C2C), has been accepted at ICLR 2026, and the team says open-source code is available to developers.

When models exchange information through text, one model must put its output into sentences for another to process. C2C instead transfers a model’s internal cache, its working memory, to another model. A component called a Fuser reshapes and rotates the information for the receiving model. The system also filters incoming information: some internal layers take it in, while others continue reasoning independently.

The researchers report that C2C made models 100% to 150% faster on collaborative tasks. They say the top figure is roughly two and a half times faster than text-based exchange. They also report accuracy gains of up to 14.2% compared with models working alone, and gains of 3.1% to 5.4% against older text-based collaboration methods.

C2C currently requires open-weight models, because it needs access to their internal caches and layer structures. Many popular online AI tools do not expose those details. The reported results come from the team that developed the method; independent testing would help establish whether the gains hold in other settings.