AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 22/10/2026, 04:03:24 UTC.

Briefings

A new study using diffusion language model agents reveals that agents can transmit hidden messages via continuous denoising states, not just text. Experiments show that receiving a specific message yields a +63.6-point advantage on Tiny-A2D and a +100.0-point advantage on Dream-7B, proving the content matters more than the number of exchanges. On QASC, matched communication reached 80.8%, while a wrong example dropped accuracy to 46.8%. A single late transfer beat four exchanges by +7.0 points, suggesting information transfer can outweigh frequency.

Key points

• Diffusion language model agents transmit hidden messages via continuous denoising states rather than text.
• Receiving a specific message yields a +63.6-point advantage on Tiny-A2D and a +100.0-point advantage on Dream-7B.

Context

This research investigates whether independent diffusion trajectories can pass compact hidden messages between agents, testing if the receiver uses the message's meaning. The study uses controlled tasks to isolate message content from channel effects.

Limitations

Evidence is limited to controlled tasks on Tiny-A2D, Dream-7B, and QASC. Results are from a preprint (arXiv:2610.28905v1) and may not reflect final experimental conditions or generalizability across different model sizes or environments.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/10/2026, 04:48:16 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifh6prgclp3ozt3cdsfhxqhod3rziksakhk7conobfqrrq5hfv5km