A recent paper on arXiv introduces a method called dense latent communication, in which language-model agents share their internal representations rather than relying solely on text. The authors argue that these representations capture not only what an agent observes but also the reasoning it performs, so transmitting them can convey both kinds of information at once.
The proposal is motivated by the fact that language-model agents already build rich internal states as they process input and generate output. Instead of compressing that state into natural language, agents could exchange the latent vectors directly. This would let a receiving agent see what the sender saw and know what the sender thought, without the lossy step of verbalization.
The paper emphasizes that this works across heterogeneous agents—different models or architectures—since the representations are dense and learned rather than tied to a specific vocabulary. While the abstract does not provide experimental details, it frames the idea as a promising direction for more efficient and faithful agent-to-agent communication. As this is a single source, the claims are preliminary and have not yet been independently verified.