Researchers from Tsinghua University developed a method for AI models to skip text entirely and share knowledge through internal memory states
What if AI models could skip the middleman of language and just beam their understanding directly into another model’s brain? That’s essentially what a new framework called Cache-to-Cache, or C2C, accomplishes. Instead of forcing large language models to write out their thoughts in text and then having another model read them back in, C2C lets models exchange compressed internal representations directly.
The result: accuracy gains of up to 14.2% over single models, and communication speeds up to 14 times faster than traditional text-based approaches.
How C2C actually works #
C2C, introduced in a paper titled “Cache-to-Cache: Direct Semantic Communication Between Large Language Models,” takes a radically different approach to model collaboration. It uses the key-value caches, the internal memory states that LLMs build up while processing information, as the communication medium itself.
The system designates one model as the “Sharer” and another as the “Receiver.” A lightweight neural component called the “Cache Fuser” projects and fuses the Sharer’s KV-cache directly into the Receiver’s representation space. The Receiver then processes this fused information as if it had built up the context on its own.
The Cache Fuser is deliberately designed to be small. It doesn’t increase the size of the KV-cache, which means you’re getting richer semantic information without the storage overhead.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The numbers tell a compelling story #
The research team, primarily from Tsinghua University’s THU-NICS group, tested C2C across multiple model families including Qwen, Llama, and Gemma. Compared to individual models working alone, C2C delivered accuracy improvements between 6.4% and 14.2%. When stacked against traditional text-to-text communication between models, accuracy still improved by 3.1% to 5.4%.
On the latency front, C2C reduced communication time by a factor of 2 to 2.5 times in standard configurations. Certain setups pushed that speedup all the way to 14 times faster. The gains become even more pronounced with larger Sharer models and longer context inputs.
Why multi-model communication is the next frontier #
The paper was published as an arXiv preprint on October 3, 2025, and has already been accepted at ICLR 2026, one of the top machine learning conferences globally. The team has released both code and a demo on GitHub.
The key limitation worth watching is interoperability. C2C currently works across model families like Qwen, Llama, and Gemma, but the Cache Fuser needs to be trained for each specific pairing. Whether this approach can scale gracefully to dozens of heterogeneous models in a single pipeline remains an open engineering question.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our