Researchers at Tsinghua University have published a paper detailing a technique that allows separate AI models to exchange information directly through their internal memory systems, eliminating the need to generate human-readable text. The method, called Cache-to-Cache (C2C), has been accepted at ICLR 2026 and is accompanied by open-source code for developers.

What You Need to Know

C2C targets a bottleneck that occurs when multiple language models collaborate: the overhead of converting internal states into written text. By passing raw cache data between models, C2C can double or more the inference speed and improve accuracy. The approach, however, works only with open-weight models that expose their internal structures. That limitation means popular closed-source AI tools cannot use C2C unless their developers implement it privately.

Bypassing the Text Bottleneck

When two AI models work together today, one must convert its internal reasoning into sentences before the other can process them. That writing step consumes computing time and discards subtle details hidden in the first model's raw thinking process. C2C eliminates typed language by letting one model transfer its working memory, technically called a Cache, directly into a second model's memory bank. A small helper program called a Fuser handles this handoff, reshaping the data so the receiving model can interpret it correctly. Different AI models store information using entirely different internal layouts, so C2C includes a selective gating mechanism that controls which layers receive the incoming data. That prevents confusion and preserves independent reasoning in layers that do not need outside input.

  • Cache-to-Cache (C2C): Lets separate AI models exchange internal states without generating text.
  • Fuser: Converts one model's cache data into a format another model can use.
  • Selective gating: Controls which layers accept the incoming information to avoid interference.

Performance Gains and Limitations

The research team reports that their setup makes AI models run between 100% and 150% faster during collaborative tasks. That upper figure translates to roughly two and a half times quicker than the conventional back-and-forth text process. Accuracy improvements reach as high as 14.2% when models work together instead of alone. Compared with older systems that still communicate through text, accuracy improved by 3.1% to 5.4%. The technique, however, currently applies only to open-weight models because it requires direct access to a model's internal cache and layer structure. Popular AI tools that ordinary people interact with online hide these details. The researchers note that typed language has always slowed machines by forcing them to mimic human communication. The same team that built C2C also ran every test validating its performance, so independent verification remains necessary.

Why This Matters

This development marks a step toward more efficient multi-model systems, particularly for agentic AI pipelines where multiple models must cooperate rapidly. If the gains hold under broader testing, C2C could reduce inference costs for companies running large-scale AI workflows that rely on open-weight models. The speed improvement from removing the text generation step directly lowers the latency and compute resources required for tasks like real-time translation or reasoning chains. That is economically significant for startups and research labs that cannot afford massive GPU clusters. The primary hurdle is the restriction to open-weight models. The industry's major players using closed models, such as OpenAI and Google, would need to implement similar methods internally. Whether they are already doing so remains unknown. The approach underscores a wider trend of optimizing AI systems by stripping away human-centric processes that add overhead but not value.