A new preprint on arXiv looks at how multi-agent systems built from frozen language models talk to each other. Normally these agents communicate through text, but that text leaves out the internal state the sender already computed. As a result, the receiving agent has to spend extra computation to rebuild that state on its own.

The authors position this as a cost problem: text-based communication makes the receiver pay for work the sender has already done. They point to prior work on internal channels as a way to avoid that cost, and they propose a 2x2 study of internal-state sharing, with the title suggesting two levers: augmenting the cache and replacing the activation.

Because this is a single-source item, there are no independent findings to compare or contrast. The abstract is brief, so the details of the proposed methods and experimental results are not yet available from the source material.