Diffusion LLM Agents Learn to Whisper in Vectors: +34 Points on QASC

Diffusion language models build an answer through a series of continuous states that they keep revising. In multi-agent systems, though, those states are usually thrown away: agents communicate only after decoding into text. The authors of the Hidden Messages paper tested a different setup. One agent compresses its denoising state into four dense symbolic vectors and passes them to a second agent. On QASC, where two supporting facts are split between the agents, working together yields 80.8% accuracy. If the message is swapped for one from another example, the result falls to 46.8%. The 34.0-point gap is the content effect.

Four vectors instead of text

The channel is simple. The sender follows its own diffusion trajectory and, at a chosen moment, compresses its current state into four vectors. The receiver accepts them as an additional input but does not abandon its own: it keeps its own context, its own canvas, and its own denoising trajectory. The message does not replace the receiver's reasoning; it is blended into it.

This design matters for experimental cleanliness. The agents are independent and share no state, so everything one knows about the other passes through the four vectors.

How they showed the receiver reads meaning

The main weakness of many papers on hidden communication is that a gain in metrics can be attributed to the mere presence of an extra signal. Here the authors set two controls. First, the message is removed entirely to see whether the channel is needed at all. Second, the message is replaced with one from a different example (a derangement). The format and statistics of the signal stay the same; only the content changes. If accuracy drops, the receiver is using meaning, not just the presence of a vector.

It drops sharply:

• Tiny-A2D: the derangement removes an advantage of +63.6 points.
• Dream-7B: the derangement removes an advantage of +100.0 points.
• QASC with two facts split between agents: 80.8% with the correct message versus 46.8% with a message from an unrelated example, a content effect of +34.0 points.
• Targeted substitutions: when a specific different fact is planted in the message, the receiver shifts its answer toward it.

The last item is arguably the most telling. The degradation shows that the channel matters. The controllable shift toward the planted fact shows that the four vectors carry readable information.

A late handoff beats frequent exchanges

The most surprising result concerns timing. Intuition suggests that more rounds of exchange should let agents synchronize better. The experiment shows the opposite: a single late handoff yields 87.8%, which is 7.0 points higher than four recurrent exchanges. By simple arithmetic, four exchanges stay at around the 80.8% level.

A plausible explanation lies in the nature of diffusion. In early steps the trajectory is still blurry, and the sender's state contains a lot of noise and unstable hypotheses. By the late steps the answer has taken shape, and the compressed state carries what the sender has actually "decided." This explanation remains a hypothesis, but the result itself breaks the common assumption that more frequent communication is automatically useful.

Where this can be applied in practice

For developers of multi-agent systems, the takeaway is direct: treat the timing of information transfer as a hyperparameter, on par with the number of rounds. One cheap exchange at the right point in the trajectory can outperform a four-exchange scheme, which means saving compute and latency.

The second implication concerns debugging. The test of swapping in a message from another example is easy to reproduce in any system with a hidden channel, and it quickly separates real communication from a chance effect of an extra input.

The limitations are also clear. The tests were run on controlled tasks and on models such as Tiny-A2D and Dream-7B, not on long agentic scenarios with tools. Still, the work provides a methodology and the first causal evidence that diffusion agents can pass meaning to each other without text. For architectures where decoding into tokens is considered a mandatory step of communication, this is a serious reason to rethink the design.