| title | Learning from the Self-future: On-policy Self-distillation for dLLMs |
|---|---|
| authors | Yifu Luo, Zeyu Chen, Haoyu Wang, Xinhao Hu, Yuxuan Zhang, Zhizhou Sha, Shiwei Liu |
| date | 2026-06-16 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2606.18195 |
| affiliation | Max Planck Institute for Intelligent Systems |
| upvotes | 38 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | On-policy self-distillation (OPSD) has proven effective for post-training large language models (LLMs), yet its application to diffusion LLMs (dLLMs) remains unexplored. Existing OPSD methods are inherently autoregressive-centric. They inject privileged information via left-to-right prefix conditioning with token-level divergence supervision, a design that fundamentally conflicts with the arbitraryorder generation of dLLMs. We introduce d-OPSD, the first OPSD framework tailored for dLLMs. Our approach makes two core contributions. First, we reframe self-teacher construction by using self-generated answers as suffix conditioning, enabling the student model to learn from "self future-experience" rather than privileged prefixes. Second, we shift supervision from token-level to step-level, aligning training with the iterative denoising process of dLLMs. Experiments across four reasoning benchmarks show that d-OPSD consistently outperforms RLVR and SFT baselines with superior sample efficiency, requiring only around 10% of the optimization steps by RLVR and opening a promising pathway for dLLM posttraining. The code is available at https://github.com/xingzhejun/d-OPSD. |
cite
@misc{luo2026learning,
title={Learning from the Self-future: On-policy Self-distillation for dLLMs},
author={Yifu Luo and Zeyu Chen and Haoyu Wang and Xinhao Hu and Yuxuan Zhang and Zhizhou Sha and Shiwei Liu},
year={2026},
eprint={2606.18195},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2606.18195}
}