Tsinghua Paper Details Cache-to-Cache Method That Lets AI Models Swap Internal Memory Without Text

The method targets what the paper describes as a specific inefficiency in pipelines where several language models work together. When two models cooperate, one currently has to write its output as sentences before the second can read it. That writing step consumes computing time and discards detail contained in the first model's raw internal state. Every model maintains a working memory of what it has processed, known technically as a cache, and C2C passes that memory straight into a second model's memory bank rather than converting it to language.

A small program called a Fuser handles the transfer, reshaping and rotating the information so the receiving model can use it. The problem is that different models store memory in different internal layouts, sizes and structures, so dumping one model's raw memory into another would likely confuse it and degrade its answers. C2C therefore includes a selective gating filter that decides which pieces of incoming memory should be absorbed immediately. Some layers accept the new information at once, while others continue reasoning independently without outside interference.

The paper reports speed gains of between 100% and 150% during shared collaborative tasks. The upper figure corresponds to roughly two and a half times faster than the usual back-and-forth process of writing and reading text. The team also reports accuracy gains as high as 14.2% when models work together rather than operating alone, and improvements of 3.1% to 5.4% compared with older setups in which models still communicate through typed text.

The approach carries a practical restriction. It currently works only with open-weight models, because it requires direct access to a model's internal cache and layer structure. Most widely used AI tools, including the ones ordinary users chat with online, hide those internal details from outside access. That means such applications cannot adopt the shortcut unless their own developers build it in privately. It is not publicly known whether any company has begun using a similar method internally.

The research team argues that typed language has always slowed machines down because it forces them to think the way humans do. That claim comes with a caveat: the same team that built the system also ran every test cited as evidence that it works. TechRadar reported that whether the method speeds up collaboration as much as claimed will depend on external testing by other groups.

Editor's Summary

Tsinghua University researchers have published a paper on Cache-to-Cache, a method that lets one AI model hand its internal cache directly to another instead of writing text first, with a learned Fuser and selective gating to make the transfer usable. The authors report 100% to 150% faster collaborative inference and accuracy gains of up to 14.2% over independent operation. The technique currently applies only to open-weight models whose internal structures are accessible, and its performance claims have so far been validated only by the team that developed it.