PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov, Ilya Makarov

arXiv:2610.28905v1cs.AI

TL;DR

The paper asks whether independent diffusion-LM agents can communicate through hidden continuous states and whether receivers use the messages’ example-specific content. It introduces a supervised message channel and finds that message content and the timing of transfer matter, with timing effects depending on task structure.

  • Problem

    Multi-agent systems usually discard diffusion models’ revisable continuous states, leaving limited evidence on whether independent diffusion-LM agents can transfer useful hidden content.

    Source: §1 · pp. 1–2

  • Method

    A supervised signed dense message channel transfers compressed denoising states between agents while preserving their independent contexts, canvases, and trajectories.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Wrong-example messages and targeted fact substitutions predictably change fixed-receiver decisions, while the better delivery schedule varies with task dependency structure.

    Source: §6 · p. 6

  • Takeaways & Limitations

    Hidden communication is a temporal coordination problem: transfer timing can matter more than exchange count, depending on when information becomes useful.

    Source: §8 · p. 6, §6 · p. 6

    Evidence is limited to Dream-7B on natural-question QASC, and the supervised task-specific adapters do not establish emergent language or task-independent protocols.

    Source: §7 · p. 6

Abstract

from arXiv · show

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. In Hidden Messages, a sender compresses its denoising state into four signed dense vectors; the receiver consumes them while retaining its own context, canvas, and trajectory. Removing a message tests whether the channel matters; replacing it with a message from another example tests whether its content matters. Across controlled tasks, derangement removes a +63.6-point advantage on Tiny-A2D and a +100.0-point advantage on Dream-7B, while targeted substitutions steer the receiver toward the substituted fact. On QASC with the two supporting facts split across agents, matched communication reaches 80.8%, but a wrong-example message lowers accuracy to 46.8%, a content effect of +34.0 points. The surprising result is temporal: one late transfer reaches 87.8%, exceeding four recurrent exchanges by +7.0 points. These experiments provide causal evidence for continuous hidden communication between diffusion-LM agents and show that when information is transferred can matter more than how many times agents exchange it.

1 Introduction

The paper studies whether independent diffusion-LM agents can exchange compact continuous messages while retaining separate contexts and trajectories. It tests whether the messages carry example-specific content and finds that useful exchange depends on when a task can use the information.

1 Introduction

  • Independent diffusion-LM agents exchange compact hidden messages while each retains its own context, canvas, and trajectory.Unlike text-only protocols, the message can be used while answers are still being revised.Source: §1 · p. 1
  • Matched-versus-deranged replacement tests whether the channel matters and whether the receiver uses the correct example-specific content.The receiver stays fixed while message origin changes, distinguishing content effects from generic activation or extra-compute benefits.Source: §1 · p. 2
  • The supervised signed dense message channel is tested on Tiny-A2D and Dream-7B with independent contexts, canvases, random-number streams, and denoising trajectories.The channel is task-specific; supervision trains the communication interface rather than asking a protocol to emerge without supervision.Source: §1 · p. 2
  • The temporal principle is task-dependent: recurrent exchange helps sequential dependencies, while late fusion can be simpler and more accurate when evidence is independently computable.The introduction contrasts an out-and-back dependency with distributed QASC, where one late transfer beats four recurrent exchanges.Source: §1 · p. 2

2 Related work

Prior latent multi-agent work shows that text is not the only viable agent interface, while other diffusion work transfers state within or across models. This paper focuses on denoising-time communication between separate diffusion-LM agents and tests message identity and delivery time with receiver inputs held fixed.

2 Related work

  • HIDDEN MESSAGES transfers compressed hidden messages between separate diffusion-LM agents, unlike methods that preserve one model's state or transfer partially denoised sequences.The receiving canvas is never replaced, helping distinguish communication effects from trajectory selection.Source: §2 · p. 2
  • The receiver is scored rather than a vote, with message identity and delivery time intervened on while local inputs remain fixed.Matched-versus-deranged replacement tests content; recurrent-versus-late delivery tests timing.Source: §2 · p. 3
  • Figure 1 depicts four-slot continuous-state exchange with independent agent canvases and contrasts matched-versus-deranged content with recurrent-versus-late delivery.The receiver is designated by a star.Source: Figure 1 · p. 3

3 Hidden communication between diffusion agents

The method transfers compact hidden states between independent diffusion agents while preserving each agent’s private context, canvas, and trajectory. A supervised channel is tested with message-replacement and timing interventions.

3.1 Independent trajectories and a continuous interface

  • Agents retain private contexts, canvases, and denoising trajectories while an encoder transfers a compact message that the receiver fuses into its own conditioning state.On Tiny-A2D the decoded message enters the answer state through a gated residual; on Dream-7B it becomes virtual prefix embeddings for the receiver’s next pass.Source: §3.1 · p. 3

3.2 Supervised task-specific channel adaptation

  • The task-specific communication adapter trains the interface around frozen backbones, without giving the receiver teacher context, labels, or text messages at inference.The reported settings train separate task-specific channels rather than transferring one universal adapter across all tasks.Source: §3.2 · p. 3

3.3 Causal and temporal interventions

  • The causal audit compares matched messages with no-message and deranged-message conditions while holding the receiver input, local trajectory, and evaluation item fixed.It also compares recurrent delivery with a late-only handoff after agents process private evidence independently.Source: §3.3 · p. 4

4 Experimental setup

The experiments progress from controlled synthetic exchanges to split natural evidence, using fixed receiver settings and paired evaluation conditions. Table 1 summarizes the task and schedule comparisons.

4.1 Two backbones, three views of the channel

  • HM-Synth tests whether a compact channel transfers sender-specific facts across fixed receiver contexts using 1,000 counterfactual pairs.The frozen Tiny-A2D backbone sees no union context at inference, and only a 6.1 MiB communication module is learned.Source: §4.1 · p. 4
  • Dream-7B requires an out-and-back exchange: one agent sends the requested slot, the other returns its table code, and the first commits its answer.The evaluation uses 512 counterfactual pairs with disjoint private contexts and includes a table-entry substitution intervention.Source: §4.1 · p. 4
  • Distributed QASC splits two supporting facts across agents, with a fixed receiver combining evidence absent from its own text context.The channel trains on 7,134 derived examples and is evaluated on all 1,000 derived development examples.Source: §4.1 · p. 4
  • QASC compares four synchronous recurrent exchanges with one late exchange after independent agent passes, using identical item order and receiver inputs.Accuracy contrasts use paired bootstrap resampling; the Dream-7B row aggregates locked evaluations.Source: §4.1 · p. 4
  • Table 1 compares channel presence, message content, and schedule across Tiny-A2D, sequential Dream-7B, and split-evidence QASC, with late fusion reversing direction across the latter tasks.Tiny-A2D content CI: [+61.5, +65.8]; QASC content CI: [+30.7, +37.2]; QASC late−matched CI: [+4.4, +9.6].Source: Table 1 · p. 5, §4.1 · p. 5

5 Results

Matched hidden messages improve performance across synthetic, dependency-spanning, and natural-question tasks, while derangement or fact substitution shows sensitivity to message content. The best exchange schedule depends on whether agents need another agent’s result to compute their own evidence.

5.1 The receiver follows the message, not merely the channel

  • +63.6 points: Tiny-A2D’s matched-minus-deranged HM-Synth effect, with matched accuracy 72.9% and deranged accuracy 9.3%.Replacing the sender’s fact changes the receiver’s answer in 89.4% of examples, which selects the substituted target in 72.9%.Source: §5.1 · p. 5
  • 100.0%: Dream-7B matched two-stage communication accuracy, versus 9.8% without a message and 0.0% under derangement.Swapping the table fact steers the receiver to the replacement target in 100.0% of examples.Source: §5.1 · p. 5
  • +34.0 points: QASC’s matched-minus-deranged content effect, with matched accuracy 80.8% and wrong-example-message accuracy 46.8%.Accuracy without communication is 72.6%, compared with 80.8% with matched recurrent messages.Source: §5.1 · p. 5

5.2 The right moment beats more rounds

  • 100.0%: recurrent Dream-7B communication accuracy, while one late message reaches only 10.8% because the requested slot must redirect the other agent’s computation.The first transfer is useful because it changes what the other agent computes next, not because it contains the final answer.Source: §5.2 · p. 5
  • 87.8%: QASC late-only accuracy, outperforming the four-round recurrent schedule by +7.0 points when each agent can interpret its evidence before exchange.Late-only uses one transmission and three agent forward passes, versus four transmissions and nine forward passes for recurrent communication.Source: §5.2 · p. 5
  • The useful exchange point aligns with the task’s latent dependency: early messages redirect conditional computation, while independently formed evidence can favor late fusion.Round count alone is a poor proxy for collaboration depth.Source: §5.2 · p. 5

6 Discussion

Replacement experiments indicate that hidden messages carry decision-relevant content, while the value of communication depends on when information becomes actionable. The useful schedule differs between dependency-spanning computation and independently formed evidence.

What travels through the channel

  • Wrong-example messages hurt, while targeted fact substitutions redirect the receiver toward the substituted answer.Changing message origin while holding the receiver fixed supports content-sensitive hidden communication, without establishing a human-readable or emergent language.Source: §6 · p. 6

Communication has a clock

  • In Dream-7B, delaying the first message removes the opportunity for the other agent to perform the computation it enables.In QASC, early repeated fusion can disturb agents’ independent evidence representations; both settings use the same Dream-7B backbone.Source: §6 · p. 6

7 Limitations and future work

The evidence is limited in task and model coverage, and the channels are supervised, task-specific adapters. Zero-shot reuse across heterogeneous tasks did not show a detectable benefit, leaving transfer and broader applicability unestablished.

7 Limitations and future work

  • Natural-question evidence comes only from Dream-7B on QASC; larger teams, other diffusion backbones, and long-form generation may have different channel geometries.The experiments therefore do not establish how the channel behaves in those settings.Source: §7 · p. 6
  • The channels are supervised, task-specific adapters, and initial zero-shot reuse across heterogeneous tasks yielded no detectable benefit.New tasks may require new supervision and adapter retraining or explicit cross-domain alignment; successful zero-shot transfer is not established.Source: §7 · p. 6
  • Matched text-channel and wall-clock comparisons remain future work, and dense latent states may encode sensitive information despite not being human-readable.The paper identifies context-partitioned software engineering with content-dependent message timing as a primary direction for future work.Source: §7 · p. 6

8 Conclusion

The conclusion presents supervised hidden-message exchange as useful across the controlled tasks and distributed QASC. Its central finding is that timing can matter more than round count, with the useful schedule depending on whether one agent’s computation depends on another’s earlier message.

8 Conclusion

  • Removing, deranging, or surgically replacing messages changes a fixed receiver in the predicted direction across Tiny-A2D, Dream-7B, and distributed QASC.The conclusion treats this intervention sequence as evidence that example-specific content crosses the latent boundary.Source: §8 · p. 6
  • Two-stage exchange is essential when one agent’s computation depends on another’s earlier message, while a single late transfer is superior when private evidence can mature independently.The conclusion characterizes hidden communication as a temporal coordination problem in which transfer timing can matter more than the number of rounds.Source: §8 · p. 6

A Implementation and evaluation details

The implementation uses Tiny-A2D and Dream-7B backbones with task-specific channel training and evaluation. The paper records dataset splits, model and adapter configurations, released artifacts, and reported compute time.

A Implementation and evaluation details

  • The derived QASC split has 7,134 training, 1,000 development, and 926 held-out examples, and the paper reports the full development set.Tiny-A2D and Dream-v0-Instruct-7B use BF16 weights; QASC is used under CC-BY-4.0.Source: A Implementation and evaluation details · p. 8
  • HM-Synth contains 5,000 training, 500 validation, and 1,000 test counterfactual pairs, with disjoint generator ranges and two to four hops.The Tiny-A2D channel has four 128-dimensional slots and width 256; only split private contexts are available at evaluation.Source: A Implementation and evaluation details · p. 8
  • The Dream encoder uses LayerNorm–Linear–SiLU–Linear with per-slot LayerNorm, and its prefix projector is RMS-calibrated to input embeddings.Representation learning combines answer-class cross-entropy and a feature-variance penalty; prefix alignment uses cosine distance plus 0.25 relative MSE.Source: A Implementation and evaluation details · p. 8
  • The implementation, exact configurations and revisions, learned channel weights, per-example predictions, and table-generation code are available in the project repository.Exact pinned model and dataset revisions are recorded with the released configurations.Source: A Implementation and evaluation details · p. 8
  • Tiny-A2D used 1,293.3 seconds (0.36 GPU-hours), while Dream QASC evaluation took 754.6 seconds (0.21 GPU-hours).Dream controlled channel training plus evaluation took 8,988.7 seconds (2.50 GPU-hours); totals exclude downloads, artifact verification, and other runs.Source: A Implementation and evaluation details · p. 8

B Intervention details

The intervention details test message content against alternative message conditions while preserving each agent’s independent trajectory and the receiver’s unchanged backbone path.

B Intervention details

  • Deranged messages are permuted across evaluation rows while the receiver’s prompt, private context, and target stay fixed.QASC permutations use a different answer label; controlled-task counterfactual pairs and different-slot batches define fact and slot replacements.Source: B Intervention details · p. 8
  • Zero, random, and self controls distinguish task content from vector presence, magnitude, and receiver-side activation.Zero replaces each message element with zero; random samples a tensor matching matched messages’ empirical first and second moments; self routes an agent’s own message through the same projection.Source: B Intervention details · pp. 8–9
  • Each agent has a separate token row and diffusion state, and communication-off predictions exactly match the unmodified backbone path.Agent 0 is always the primary output; the evaluation uses no vote, canvas replacement, logit averaging, or union-context oracle, and paired bootstrap intervals use 10,000 row-resampled replicates.Source: B Intervention details · p. 9
Loading 2610.28905v1…