AURAL-RL

AURAL performs adaptive reasoning in continuous latent space before generating an answer. This checkpoint uses Qwen2.5-Omni-7B with compression factor c = 4, chunk size m = 10, four Gaussian components, and a covariance rank of 16.

Paper · Code and training · AuralReason

Quick start

Use Python 3.12, Linux, and an NVIDIA GPU. Install the AURAL package:

git clone https://github.com/JensenYX/AURAL.git
cd AURAL
python -m pip install -e .
python -m pip install flash-attn==2.8.3 --no-build-isolation
python -m aural.infer --model YuxiangW/AURAL --audio speech.wav

An audio clip can be accompanied by an instruction:

python -m aural.infer --audio speech.wav --text "Describe the emotion in the speaker's voice."

For text input:

python -m aural.infer --text "What is 17 + 25?"

The AURAL loader reads the backbone, joint latent head, processor, and reasoning configuration from this repository. Generation uses latent and EOL temperatures of 1, a maximum of 100 latent states, and greedy answer decoding. The output is text.

Files

• model-*.safetensors: Qwen2.5-Omni backbone weights.
• latent_policy.safetensors: the joint low-rank Gaussian mixture head.
• aural_config.json: latent dimensions, compression, chunk size, and prompt.
• Tokenizer and processor files, including the speaker dictionary required by the Qwen2.5-Omni loader.

The code repository includes SFT, RL, custom-data formats, and training commands.

Citation

@article{wang2026aural,
  title={AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models},
  author={Wang, Yuxiang and others},
  journal={arXiv preprint arXiv:2610.01560},
  year={2026}
}