On-policy distillation gains don't come from distillation, study finds
A new analysis challenges the theoretical foundation of on-policy distillation, a popular technique for training language models with dense supervision. Researchers discovered that teacher models generate substantial noise when scoring student-generated trajectories, yet students remain surprisingly robust to this corruption. The finding suggests OPD's gains may stem from mechanisms other than faithful knowledge transfer, raising questions about whether the technique actually distills teacher expertise or simply provides a regularization signal. This matters for practitioners choosing between distillation and reinforcement learning approaches, and for understanding what actually drives improvement in modern LLM training pipelines.
Modelwire context
Skeptical readThe critical finding isn't that teacher noise exists (expected), but that students improve despite it. The paper stops short of explaining what mechanism actually drives gains if not knowledge transfer, leaving open whether OPD is a misnomer for a regularization technique or whether the noise measurement itself is flawed.
This connects directly to the self-improvement work from late August. S3Gym and ASPIRE both probe whether models can improve autonomously, but they assume some form of reliable feedback signal. This paper suggests that signal may be noisier than practitioners believe, which has immediate implications for how to design self-judging loops (as tested in S3Gym) and whether vague feedback degrades self-evolution (the core question in ASPIRE). If teachers are poor scorers, self-improvement frameworks that rely on model-generated supervision may need explicit noise-handling mechanisms.
If follow-up work shows OPD gains persist when teacher scoring is replaced with random labels of similar magnitude, that confirms OPD is primarily a regularizer, not distillation. Conversely, if practitioners report OPD failures correlate with teacher model quality in production settings over the next 6 months, the noise finding becomes actionable rather than theoretical.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsOn-policy distillation · Reinforcement learning with verifiable rewards
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.