Learning from the Self-future: On-policy Self-distillation for dLLMs

Research question and approach

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.

When this work is relevant

Use this work when discussing application of on-policy self-distillation to diffusion language models and its proposed self-future learning approach.

Author-written 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.

Abstract source: https://arxiv.org/abs/2606.18195. Checked 2026-09-14. Bibliographic metadata uses the linked paper record or author-maintained catalog. Results, limitations and experimental settings remain defined by the original source.

Citation

BibTeX · CITATION.cff

BibTeX is preserved from the author-maintained citation repository. The CFF uses the source metadata shown on this page. Version titles or author lists can differ; choose the version you used.

@article{luo2026learning,
  title={Learning from the Self-future: On-policy Self-distillation for dLLMs},
  author={Luo, Yifu and Chen, Zeyu and Wang, Haoyu and Hu, Xinhao and Zhang, Yuxuan and Sha, Zhizhou and Liu, Shiwei},
  journal={arXiv preprint arXiv:2606.18195},
  year={2026}
}