metadata
license: mit
language:
- en
tags:
- DLLM
- diffusion-language-model
- on-policy-distillation
- post-training
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen3-4B
datasets:
- divelab/opdlm_train_data
arxiv: 2606.06712
OPDLM-4B
OPDLM-4B is a block diffusion language model (DLM) obtained by post-training an autoregressive language model (ARLM) into a diffusion language model via on-policy distillation. arXiv report: arxiv.org/abs/2606.06712
Highlights
- Converted, not pretrained from scratch: built from a strong ARLM, reusing its prior.
- Training-efficient: ~0.076B tokens of conversion vs. ~50B tokens for from-scratch DLM training (same base ARLM).
- Inference-efficient: parallel token decoding via block diffusion.
Model Details
- Developed by: DIVE Lab, Texas A&M University
- Base model: Qwen3-4B
- Model type: Block diffusion language model (decoder-based)
- Block size: 4
- Parameters: ~4B
- Language: English
- License: MIT
Training
- Method: On-policy distillation from a frozen ARLM teacher into a block DLM student.
- Conversion budget: ~0.076B tokens
- Data: opdlm_train_data
Results
For detailed results and benchmarks, please refer to our paper: arxiv.org/abs/2606.06712
Citation
@misc{su2026dataefficientautoregressivetodiffusionlanguagemodels,
title={Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation},
author={Xingyu Su and Jacob Helwig and Shubham Parashar and Atharv Chagi and Lakshmi Jotsna and Degui Zhi and James Caverlee and Dileep Kalathil and Shuiwang Ji},
year={2026},
eprint={2606.06712},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.06712},
}