Tsinghua Research Unveils Inter-AI 'Modem': Models Can Exchange Thoughts Without Writing Text

A research team from Tsinghua University has released a paper introducing a new technique called Cache-to-Cache (C2C), a protocol that allows two artificial intelligence models to exchange internal information without generating intermediate text. The method has been accepted at the ICLR 2026 conference and its source code is openly available to developers.

Until now, when several large language models (LLMs) work together in a pipeline, one model must convert its reasoning output into written sentences before another model can read it. That rewriting process not only consumes computing time but also sacrifices fine-grained details stored in the first model's internal representations. C2C offers a shortcut: instead of typing, the first model sends a 'cache' — the working memory containing its processing traces — directly into the second model's memory bank.

The main obstacle to this approach is the difference in internal architecture between models. Each model stores memory with a non-uniform layout, size, and structure, so transferring a raw cache risks confusing the receiving model. To address this, C2C includes a helper program called Fuser, tasked with reshaping and rotating the information so it can be digested. In addition, there is a 'selective gating' mechanism that filters which parts of the memory are worth absorbing immediately, while other layers continue to reason independently without external interference.

According to the researchers, this acceleration could change how multi-agent AI systems are designed. Until now, architectures relying on text communication between models have often become a bottleneck, especially in applications demanding fast responses such as virtual assistants or real-time decision-support systems. With C2C, the latency arising from the 'write and read' process can be significantly reduced.

"Written language always slows machines down because it forces them to think like humans," the research team argued, adding that a cache-based approach is better suited to how parallel computing works.

However, that claim warrants scrutiny. The team that developed the system also ran all the internal tests, so independent validation is still awaited. In addition, C2C is currently compatible only with open-weight models. Popular commercial models used by the wider public — such as chatbots owned by large technology companies — hide the details of their caches and internal layers, so the technique cannot be applied as-is without private modifications by their respective developers.

Questions about industry adoption have also emerged. There is no public information on whether AI companies have adopted a similar method internally. If C2C proves effective in external testing, it is not impossible that the approach becomes a new standard in multi-model orchestration, driving energy and computing-cost efficiency that has long been a bane for large-scale AI service providers.

For Indonesia, this development offers both opportunity and challenge. Local researchers and open-source model developers can use the C2C code to build more efficient AI systems, especially in sectors with limited computing resources such as precision agriculture, regional health services, or public administration. However, the reliance on open-weight models means adoption in enterprise environments using closed APIs remains impossible in the near term. Regulators also need to examine the security implications of memory exchange between models, since the information exchanged may contain sensitive data without passing through the text encryption that is normally audited.

Going forward, the question is not only how quickly C2C will be adopted, but whether this 'talking through cache' approach will spur the birth of a safer and more transparent inter-AI standard protocol. If so, the AI development landscape could shift from merely chasing large parameter counts toward smarter architectural efficiency.