dOPSD: On-Policy Self-Distillation for Diffusion Language Models

Hugging Face Daily Papers Papers

Summary

This paper introduces dOPSD, an on-policy self-distillation method for diffusion language models that leverages internal denoising trajectories to improve mathematical reasoning and code generation.

Diffusion large language models (dLLMs) generate text by iteratively denoising a masked sequence, offering a parallel alternative to autoregressive models, but eliciting strong reasoning through post-training remains difficult: supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards and is hard to apply without tractable sequence likelihoods. On-policy self-distillation (OPSD) offers a promising alternative, using one model as both student and teacher to provide dense, token-level, on-policy supervision, but its effectiveness hinges on giving the teacher privileged information (PI) - typically an instance-specific ground-truth reference unavailable at inference - so the student ends up distilling a weak PI-free consensus policy that yields little improvement on dLLM reasoning. We introduce dOPSD, which instead derives the teacher's privilege directly from the student's own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher's advantage emerges from the model's own decoding process; on Dream and LLaDA, dOPSD improves both in-domain math reasoning and out-of-domain code generation, outperforming supervised and on-policy baselines.
Original Article
View Cached Full Text

Cached at: 07/07/26, 06:42 AM

Paper page - dOPSD: On-Policy Self-Distillation for Diffusion Language Models

Source: https://huggingface.co/papers/2607.04428

Abstract

Diffusion large language models face challenges in reasoning enhancement through post-training, but a novel on-policy self-distillation method using internal denoising trajectories improves mathematical reasoning and code generation performance.

Diffusion large language models(dLLMs) generate text by iteratively denoising a masked sequence, offering a parallel alternative toautoregressive models, but eliciting strong reasoning through post-training remains difficult:supervised fine-tuningis off-policy and suffers fromexposure bias, whilereinforcement learninggives only sparse, sequence-level rewards and is hard to apply without tractable sequence likelihoods.On-policy self-distillation(OPSD) offers a promising alternative, using one model as both student and teacher to provide dense, token-level, on-policy supervision, but its effectiveness hinges on giving the teacherprivileged information(PI) - typically an instance-specific ground-truth reference unavailable at inference - so the student ends up distilling a weak PI-free consensus policy that yields little improvement on dLLM reasoning. We introduce dOPSD, which instead derives the teacher’s privilege directly from the student’s owndenoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher’s advantage emerges from the model’s own decoding process; on Dream and LLaDA, dOPSD improves bothin-domain math reasoningandout-of-domain code generation, outperforming supervised and on-policy baselines.

View arXiv pageView PDFGitHub2Add to collection

Get this paper in your agent:

hf papers read 2607\.04428

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.04428 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.04428 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.04428 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles