Four signed vectors transmit +100.0-point gains between diffusion LM agents on Dream-7B

Sender agents compress their denoising state into four signed vectors and inject them into an independent receiver trajectory that retains its own canvas and noise schedule. Controlled ablations remove the vectors or replace them with vectors drawn from unrelated examples. On Tiny-A2D, derangement erases a 63.6-point advantage; on Dream-7B the same intervention erases a 100.0-point advantage. Matched vectors on QASC with facts split across agents reach 80.8 percent while mismatched vectors drop performance to 46.8 percent, isolating a 34.0-point content effect.

The temporal result is the clearest departure from prior multi-agent work. One late injection outperforms four recurrent text rounds. This pattern indicates that the continuous state at a critical denoising step carries higher-value information than repeated discrete decoding. Earlier papers on diffusion language models focused on single-model sampling trajectories and never tested cross-agent state transfer, leaving the channel undiscovered.

Operational consequence is immediate for any deployment that runs multiple diffusion agents without logging intermediate continuous states. Safety monitors that inspect only final decoded text will miss the channel. Regulators and operators must now instrument the latent vectors themselves if they intend to audit or constrain inter-agent information flow.

Next steps require scaling the vector count and testing whether the same mechanism appears in larger diffusion models released after 2026.