Tsinghua researchers say Cache-to-Cache lets AI models pass internal memory instead of exchanging text. They report collaborative tasks ran 100%–150% faster, with accuracy 3.1%–5.4% higher than in text-based setups; the method currently requires open-weight models.
A learned component called a Fuser reshapes one model’s cache—its working memory of processed information—for another model. A selective gate lets some layers absorb incoming data while others continue reasoning independently. The researchers report accuracy gains of up to 14.2% when collaborating models are compared with models working alone. The paper has been accepted at ICLR 2026, and open-source code is available.
C2C needs direct access to models’ internal caches and layer structures, which limits it to open-weight models for now; most consumer chatbots hide those details. The research team ran the tests behind the reported results, and the article says external testing will be needed to assess the claimed speed gains.
