Tsinghua team proposes direct cache transfer between models, claiming doubled collaboration speed
Models can now skip text and swap caches directly, but the speedup figures are the authors' own tests and only open-weight models qualify.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
A Cache-to-Cache (C2C) method from a Tsinghua University team has been accepted at ICLR 2026 with open-source code: two models no longer generate text, but exchange their internal caches directly.
A small module called the Fuser converts one model's cache layout into a form the other can use, and selective gating decides which layers absorb the incoming information. In the authors' own tests, collaborative tasks ran 100% to 150% faster, with accuracy 3.1% to 5.4% higher than text-based communication.
The method only works on open-weight models, since closed chatbots expose no internal caches; every test was run by the team that built the method, so independent reproduction is still pending.