Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication
One late hidden-message transfer outperformed four recurrent exchanges by 7 percentage points.
The sender compresses its internal denoising state into four signed dense vectors, which the receiver uses while keeping its own context and trajectory. On QASC, with supporting facts split between agents, matched messages achieved 80.8% accuracy; messages from the wrong example cut that to 46.8%. Targeted substitutions also steered receivers toward the substituted fact, supporting the claim that message content affects their answers. A single late transfer reached 87.8%, showing that timing mattered more than exchange count in this experiment.
HF Daily Papers' note
The sender compresses its internal denoising state into four signed dense vectors, which the receiver uses while keeping its own context and trajectory. On QASC, with supporting facts split between agents, matched messages achieved 80.8% accuracy; messages from the wrong example cut that to 46.8%. Targeted substitutions also steered receivers toward the substituted fact, supporting the claim that message content affects their answers. A single late transfer reached 87.8%, showing that timing mattered more than exchange count in this experiment.
HF Daily Papers' note
score 5