# Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why

arXiv: https://arxiv.org/abs/2605.10889 · HTML: https://mlsanity.com/paper/2605.10889/

## Key facts

- Cited by 25 papers in the map (data through 14 Oct 2026).
- Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning.
- Builds on 27 of its 31 references that are in the map.
- First posted on arXiv 11 May 2026.

## Placed in

- LLM reasoning / [Prompting and chain-of-thought reasoning](https://mlsanity.com/lr/prompting-and-chain-of-thought-reasoning/)
- Knowledge distillation / [On-policy and self-distillation for LLMs](https://mlsanity.com/distill/on-policy-and-self-distillation-for-llms/)

## Builds on (most cited first, 27 of 27)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://mlsanity.com/paper/2503.14476/) (2025)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [Understanding R1-Zero-Like Training: A Critical Perspective](https://mlsanity.com/paper/2503.20783/) (2025)
- [On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes](https://mlsanity.com/paper/2306.13649/) (2023)
- [Solving math word problems with process- and outcome-based feedback](https://mlsanity.com/paper/2211.14275/) (2022)
- [Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks](https://mlsanity.com/paper/1506.03099/) (2015)
- [Sequence-Level Knowledge Distillation](https://mlsanity.com/paper/1606.07947/) (2016)
- [GLM-5: from Vibe Coding to Agentic Engineering](https://mlsanity.com/paper/2602.15763/) (2026)
- [Improved Knowledge Distillation via Teacher Assistant](https://mlsanity.com/paper/1902.03393/) (2019)
- [Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models](https://mlsanity.com/paper/2601.18734/) (2026)
- [On the Efficacy of Knowledge Distillation](https://mlsanity.com/paper/1910.01348/) (2019)
- [Reinforcement Learning via Self-Distillation](https://mlsanity.com/paper/2601.20802/) (2026)
- [Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe](https://mlsanity.com/paper/2604.13016/) (2026)
- [Self-Distillation Enables Continual Learning](https://mlsanity.com/paper/2601.19897/) (2026)
- [MiMo-V2-Flash Technical Report](https://mlsanity.com/paper/2601.02780/) (2026)
- [On-Policy Context Distillation for Language Models](https://mlsanity.com/paper/2602.12275/) (2026)
- [Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation](https://mlsanity.com/paper/2602.12125/) (2026)
- [Learning by Distilling Context](https://mlsanity.com/paper/2209.15189/) (2022)
- [Small Models Struggle to Learn from Strong Reasoners](https://mlsanity.com/paper/2502.12143/) (2025)
- [Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?](https://mlsanity.com/paper/2603.24472/) (2026)
- [Privileged Information Distillation for Language Models](https://mlsanity.com/paper/2602.04942/) (2026)
- [Distillation Scaling Laws](https://mlsanity.com/paper/2502.08606/) (2025)

## Built on by (most cited first, 25 of 25)

- [Rethinking On-Policy Distillation of Large Language Models II: One Training Example](https://mlsanity.com/paper/2609.04172/) (2026)
- [A Formula-Driven Survey and Research Agenda for On-Policy Distillation](https://mlsanity.com/paper/2606.22793/) (2026)
- [Lightning OPD 2.0: Mitigating Style Bias in Cross-Teacher On-Policy Distillation for Large Reasoning Models](https://mlsanity.com/paper/2607.28449/) (2026)
- [β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation](https://mlsanity.com/paper/2607.28582/) (2026)
- [1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation](https://mlsanity.com/paper/2609.24432/) (2026)
- [What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation](https://mlsanity.com/paper/2609.25623/) (2026)
- [A Good Self-Teacher Meets the Student Where They Are: Joint On-Policy Learning and Teaching](https://mlsanity.com/paper/2610.10447/) (2026)
- [AlphaG-OPD: Reliability-Gated Sibling Counterfactuals for On-Policy Distillation in Symbolic Alpha Factor Discovery](https://mlsanity.com/paper/2608.01303/) (2026)
- [One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation](https://mlsanity.com/paper/2608.25936/) (2026)
- [VERPO: Verified Evidence Regularized Policy Optimization](https://mlsanity.com/paper/2609.06100/) (2026)
- [Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition](https://mlsanity.com/paper/2609.14708/) (2026)
- [Teaching a Moving Student: Rethinking the Curriculum of On-Policy Distillation](https://mlsanity.com/paper/2609.25048/) (2026)
- [What Shared Prefixes Hide: Trajectory Dropout for On-Policy Distillation](https://mlsanity.com/paper/2609.33455/) (2026)
- [SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models](https://mlsanity.com/paper/2609.34044/) (2026)
- [Look Before You Select: Rethinking Vocabulary Sparsification in On-Policy Distillation](https://mlsanity.com/paper/2609.34386/) (2026)
- [On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics](https://mlsanity.com/paper/2609.35259/) (2026)
- [Interactive-Policy Distillation with Bidirectional Propose-and-Verify](https://mlsanity.com/paper/2609.36546/) (2026)
- [Dr. OPD: Learning What to Follow for Optimal On-Policy Distillation of Large Language Models](https://mlsanity.com/paper/2609.38025/) (2026)
- [RISED: RubrIcs for agentic multi-environment Selection and sElf-Distillation](https://mlsanity.com/paper/2610.00979/) (2026)
- [DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation](https://mlsanity.com/paper/2610.04596/) (2026)
- [UP-MOPD: Update Projection in Multi-Teacher On-Policy Distillation](https://mlsanity.com/paper/2610.08398/) (2026)
- [Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals](https://mlsanity.com/paper/2610.11247/) (2026)
- [When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better](https://mlsanity.com/paper/2610.11291/) (2026)
- [Branching On-Policy Distillation](https://mlsanity.com/paper/2610.13481/) (2026)
- [Reward-Aligned Distillation](https://mlsanity.com/paper/2610.14572/) (2026)

## Abstract (arXiv)

On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions this signal is beneficial and under which it is detrimental. Which teacher model should be used, and in the case of self-distillation, which specific context should serve as the supervisory signal? Does the optimal choice vary from one token to the next? At present, addressing these questions typically requires costly training runs whose aggregate performance metrics obscure the dynamics at the level of individual tokens. We introduce a training-free diagnostic framework that operates at the highest resolution: per token, per question, and per teacher. We derive an ideal per-node gradient defined as the parameter update that maximally increases the student's probability of success. We then develop a scalable targeted-rollout algorithm to estimate this gradient efficiently, even for long chains of intermediate thoughts. The gradient alignment score, defined as the cosine similarity between this ideal gradient and any given distillation gradient, quantifies the extent to which a particular configuration approximates the ideal signal. Across a range of self-distillation settings and external teacher models, we observe that distillation guidance exhibits substantially higher alignment with the ideal on incorrect rollouts than on correct ones, where the student already performs well and the teacher's signal tends to become noisy. Furthermore, we find that the optimal distillation context depends jointly on the student model's capacity and the target task, and that no single universally effective configuration emerges. These findings motivate the use of per-task, per-token diagnostic analyses for distillation.