English

Stable On-Policy Distillation through Adaptive Target Reformulation

Machine Learning 2026-04-21 v2 Artificial Intelligence

Abstract

Knowledge distillation (KD) is a widely adopted technique for transferring knowledge from large language models to smaller student models; however, conventional supervised KD often suffers from a distribution mismatch between training and inference. While on-policy KD approaches attempt to mitigate this issue by learning directly from student-generated outputs, they frequently encounter training instabilities because the distributional gap between the novice student and the expert teacher is often too wide to bridge directly. These challenges manifest as pathological gradients in forward KL objectives or diversity collapse in reverse KL regimes. To address these limitations, we propose Veto, an objective-level reformulation that constructs a geometric bridge in the logit space. Unlike prior methods that mix data samples, Veto creates an intermediate target distribution that promotes alignment between the teacher and the student. By introducing a tunable parameter beta, Veto serves as an Adaptive Gradient Veto that stabilizes optimization by suppressing harmful gradients on low-confidence tokens, while simultaneously acting as a Decisiveness Knob to balance reward-driven performance with output diversity. Extensive experiments across various reasoning and generation tasks demonstrate that Veto consistently outperforms supervised fine-tuning and existing on-policy baselines.

Keywords

Cite

@article{arxiv.2601.07155,
  title  = {Stable On-Policy Distillation through Adaptive Target Reformulation},
  author = {Ijun Jang and Jewon Yeom and Juan Yeo and Hyunggu Lim and Taesup Kim},
  journal= {arXiv preprint arXiv:2601.07155},
  year   = {2026}
}

Comments

10 pages, 5 figures, Accepted to Findings of ACL 2026

R2 v1 2026-07-01T08:59:58.609Z