Researchers Show AI Models From Different Families Passing Internal State Instead of Text

The method, called SwaT or Swarm Telepathy, is set out in a preprint posted to arXiv on October 10 by Abdulrahman Alabdulkareem, Irene Tenison, Yang Zhang, Yada Zhu, Boris Katz and Dan Gutfreund. It has since been covered by South Korea's Chosun Ilbo. Secondary coverage links the authors to MIT and the MIT-IBM research collaboration.

The work targets the key-value cache, the store of intermediate calculations a model builds as it reads its input. In multi-agent systems, one model normally hands its output to another as text, and the receiving model then reprocesses it from scratch, a step known as prefill. SwaT uses a small learned translator to convert one model's cache into a form another model can use, even when the two come from different families. The authors call it the first cross-family method of this kind that allows prefill-free communication.

According to the paper, combined with cache compression the approach was on average 40% more accurate and four times faster in wall-clock time than models exchanging text. The authors say they tested it on 17 benchmarks across 12 models from six families. Both the sending and receiving models stay frozen, and training the translator took under 50 million tokens and less than one GPU-day. The paper also reports that the visual understanding of a vision-language model could be passed to a text-only model.

The results have not been peer reviewed. They point to cheaper and faster coordination between AI agents built on models from different vendors, which matters to companies designing agent pipelines. They also raise an oversight question that commentators flagged quickly: much of today's AI safety monitoring depends on reading the text that models produce and exchange. If agents swap internal states that humans cannot read, that window closes.