Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication

Hidden Messages Between Diffusion Language Model Agents. Causal Evidence for Latent Communication. By Alan-Barsag Gazzaev and three others, from ITMO University and H S E University. Published on the twentieth of October. Its code repository is worked out: about forty source files, with a readme.

Diffusion language models form their answers through a sequence of revisable continuous states. Multi-agent systems usually throw those states away and let agents talk only after the text has been decoded. The claim here is that one independent diffusion trajectory can pass a compact hidden message into another, and, the harder part, that the receiver actually uses what the message means. On top of that sits a temporal claim: when information is transferred can matter more than how many times agents exchange it.

The setup is called Hidden Messages. A sender compresses its denoising state into four signed dense vectors. The receiver consumes them while keeping its own context, canvas, and trajectory. Two interventions do the causal work. Removing a message tests whether the channel matters at all. Replacing it with a message taken from a different example tests whether the content matters.

Across controlled tasks, derangement, the swap of messages between examples, removes an advantage of sixty-three point six points on Tiny-A2D, and of a full one hundred points on Dream-7B. Targeted substitutions go further and steer the receiver toward the substituted fact.

On QASC, the two supporting facts are split between agents. Matched communication reaches eighty point eight percent. A message from the wrong example drops accuracy to forty-six point eight percent, so the content alone is worth thirty-four points. Then the timing result. A single late transfer reaches eighty-seven point eight percent, which beats four recurrent exchanges by seven points.

The evidence is from controlled tasks, with QASC as the setting where facts are deliberately divided between two agents. The channel studied is specific: four vectors drawn from a denoising state, passed between diffusion-LM agents, with the receiver keeping its own trajectory. The timing comparison is one late transfer against four recurrent exchanges. What the abstract establishes is causal evidence that this channel carries usable content, measured by removal and substitution, rather than a general account of agent communication.

The design of the test is what makes the claim more than correlation. Taking the message away shows the channel is used. Swapping in someone else's message shows the meaning is used, because the receiver then follows the wrong fact. And the late single transfer outscoring repeated exchange shifts the question from how much agents talk to when.