Learning from the Self-future: On-policy Self-distillation for dLLMs
d-OPSD applies on-policy self-distillation to diffusion LLMs via suffix conditioning and step-level supervision, cutting optimization steps by ~90% versus RLVR while outperforming baselines on reasoning benchmarks.
Published 2026Sydney Poster Session 3 · Wed, Dec 9, 10:00 AM–1:00 PM local time · Hall 1-4▲ 174 on Hugging FaceCode ★ 18arXiv ↗OpenReview ↗
Only vote on papers you've read. Sign in with GitHub to vote.
d-OPSD delivers a structurally sound, highly efficient self-distillation fix for diffusion LLMs via suffix conditioning and step-level supervision, though the 10% step claim and benchmark choices need verification.
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.