Diffusion LM agents pass usable hidden messages in latent states
Original title: Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication
On QASC, matched messages reach 80.8%, while a wrong-example message lowers accuracy to 46.8%.
Hidden Messages tests whether one diffusion language model can pass four signed dense vectors from its denoising state to another agent that uses their content. Derangement removes a 63.6-point advantage on Tiny-A2D and a 100.0-point advantage on Dream-7B, and substitutions steer the receiver toward the substituted fact. On split-fact QASC, matched communication reaches 80.8% versus 46.8% with a wrong-example message. One late transfer reaches 87.8%, 7.0 points above four recurrent exchanges.
- The sender compresses its denoising state into four signed dense vectors.
- The QASC content effect is reported as +34.0 points.
- A single late transfer beats four recurrent exchanges by 7.0 points.
- Removing a message tests the channel; swapping examples tests content.
From the original article
1 min read at huggingface.co
Diffusion language models form answers through a sequence of revisable continuous states, but multi-agent systems usually discard those states and make agents communicate only after decoding text. We ask whether one independent diffusion trajectory can pass a compact hidden message into another and, crucially, whether the receiver uses what that message means.
By Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.