Hidden Messages: Dream-7B QASC communication channel

Task-specific, supervised communication-channel weights accompanying Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication, accepted at the DiffuLM Workshop at NeurIPS 2026.

This is a communication adapter, not a standalone language model. The frozen Dream backbone is downloaded separately. The channel compresses a sender hidden state into four 128-dimensional continuous vectors and projects received messages into virtual prefix embeddings for the receiving diffusion agent.

Files and provenance

• qasc_dense_latent_prefix_channel.pt: original PyTorch state dictionary.
• config.yaml: original model/data revisions, architecture, training settings, and evaluation settings for distributed QASC.
• LICENSE: Apache-2.0 license for this release's channel weights and documentation.
• THIRD_PARTY_LICENSES.md: attribution for external dependencies and datasets.

These are the existing paper artifacts, mirrored without retraining or checkpoint conversion from the public code release. The checkpoint source is checkpoints/qasc_dense_latent_prefix_channel.pt. No new inference or benchmark validation was performed for this Hub mirror.

Download

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="RobotLearningLab/hidden-messages-dream-qasc",
    filename="qasc_dense_latent_prefix_channel.pt",
)
config_path = hf_hub_download(
    repo_id="RobotLearningLab/hidden-messages-dream-qasc",
    filename="config.yaml",
)

The download example does not run inference. Use the original DenseLatentPrefixChannel and QASC driver for the architecture and denoising-time integration. Read the driver's CLI options before running it: it also contains training and evaluation workflows. Use torch.load(..., weights_only=True) when loading the trusted state dictionary. This repository does not implement Transformers AutoModel.from_pretrained or a hosted text-generation endpoint.

Required backbone: Dream-org/Dream-v0-Instruct-7B, revision 05334cb9faaf763692dcf9d8737c642be2b2a6ae. Use the pinned dependencies in the code repository and a CUDA environment for the reported model workflow. The full backbone weights and QASC source data are not included here.

Scope and limitations

The adapter was trained with supervision for split-evidence QASC. It is not an emergent language or a universal cross-task communication protocol. Successful zero-shot transfer to other tasks or backbones is not established. A dense latent message is not a privacy mechanism. Independent agent contexts and trajectories, the designated receiver, message interventions, and delivery schedule are part of the evaluated setup; downloading the adapter alone does not reproduce the results.

For reported results and experimental scope, see the paper and evidence/qasc_results.json.

Citation

@misc{gazzaev2026hiddenmessages,
  title={Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication},
  author={Alan-Barsag Gazzaev and Alexey Gavrilov and Sergey Muravyov and Ilya Makarov},
  year={2026},
  eprint={2610.28905},
  archivePrefix={arXiv}
}