RobotLearningLab/hidden-messages-tiny-a2d

Hidden Messages: Tiny-A2D HM-Synth 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 Tiny-A2D backbone is downloaded separately. The adapter consists of a sender-state compressor and a gated residual receiver fuser, with four 128-dimensional message slots. It was trained on HM-Synth distributed chain-lookup tasks.

Files and provenance

• tiny_a2d_channel.pt: original PyTorch state dictionary.
• config.yaml: original backbone revision, synthetic-task settings, channel architecture, training settings, and evaluation settings.
• 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/tiny_a2d_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-tiny-a2d",
    filename="tiny_a2d_channel.pt",
)
config_path = hf_hub_download(
    repo_id="RobotLearningLab/hidden-messages-tiny-a2d",
    filename="config.yaml",
)

The download example does not run inference. The original ReceiverChannel class is defined in phase5_hm_synth_tiny.py, with compressor and fusion modules in code/src/hidden_messages/communication/. Use that architecture and the accompanying config to load the trusted state dictionary with torch.load(..., weights_only=True). The original driver includes training and evaluation, not just a checkpoint-only inference CLI. This repository does not implement Transformers AutoModel.from_pretrained or a hosted text-generation endpoint.

Required backbone: dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1, revision c8d24a3f4adaeef46881b450e1bf7d1005203bd7. Use the pinned dependencies in the code repository and a CUDA environment for the reported model workflow. The full backbone weights are not included here.

Scope and limitations

Training uses union-context teacher states and paired matched-versus-deranged supervision. Evaluation uses split private contexts. The adapter is task-specific; it is not an emergent language, a privacy mechanism, or a demonstrated universal cross-task protocol. It should not be assumed interchangeable with the Dream QASC adapter. Independent agent trajectories, the designated receiver, and causal message interventions are part of the evaluated setup.

For reported results and experimental scope, see the paper and evidence/tiny_controlled_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}
}