Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes

• URL: http://arxiv.org/abs/2603.25562v1
• Date: Thu, 26 Mar 2026 15:35:59 GMT
• Title: Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes
• Abstract summary: On-policy distillation (OPD) is appealing for large language model (LLM) post-training because it evaluates teacher feedback on student-generated rollouts rather than fixed teacher traces.<n>We revisit OPD from the estimator and implementation sides.<n>We identify three failure modes of sampled-token OPD: an imbalanced one-token signal, unreliable teacher guidance on student-generated prefixes, and distortions caused by tokenizer or special-token mismatch.
• Score: 31.95045602299568
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: On-policy distillation (OPD) is appealing for large language model (LLM) post-training because it evaluates teacher feedback on student-generated rollouts rather than fixed teacher traces. In long-horizon settings, however, the common sampled-token variant is fragile: it reduces distribution matching to a one-token signal and becomes increasingly unreliable as rollouts drift away from prefixes the teacher commonly visits. We revisit OPD from the estimator and implementation sides. Theoretically, token-level OPD is biased relative to sequence-level reverse-KL, but it has a much tighter worst-case variance bound; our toy study shows the same tradeoff empirically, with stronger future-reward coupling producing higher gradient variance and less stable learning. Empirically, we identify three failure modes of sampled-token OPD: an imbalanced one-token signal, unreliable teacher guidance on student-generated prefixes, and distortions caused by tokenizer or special-token mismatch. We address these issues with teacher top-K local support matching, implemented as truncated reverse-KL with top-p rollout sampling and special-token masking. Across single-task math reasoning and multi-task agentic-plus-math training, this objective yields more stable optimization and better downstream performance than sampled-token OPD.

Related papers

• A Token-Level Analysis of Sampled-Token Reverse-KL On-Policy Distillation [59.43605629311855]
  We analyze the gradient of the per-token K2 estimator of reverse KL with respect to the student logits.<n>Our primary contribution is therefore a gradient-level characterization of reverse-KL OPD trained with the K2 estimator, with SuRe as one empirical instantiation.
  arXiv  Detail & Related papers  (2026-08-26T11:14:42Z)
• Mismatch Matters: On-Policy Distillation Beyond Token Agreement [27.396104103072844]
  We show that mismatch tokens can be mainly categorized into two types: student-excess tokens and student-deficit tokens.<n>We propose TIDE, which applies bounded Hellinger shaping to suppress the most severe sampled excesses and an analytic teacher top-$K$ injection to restore deficient probability mass.
  arXiv  Detail & Related papers  (2026-08-10T16:53:14Z)
• Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance [50.77757355236912]
  We propose RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), which applies distillation selectively and precisely where it matters most.<n> Experiments demonstrate that RSTG substantially outperforms naive GRPO+OPD by +4.02% on math and+3.05% on code.
  arXiv  Detail & Related papers  (2026-08-01T17:29:14Z)
• dOPSD: On-Policy Self-Distillation for Diffusion Language Models [52.60302464420127]
  Diffusion large language models generate text by iteratively denoising a masked sequence.<n>Supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards.<n>On-policy self-distillation (OPSD) offers a promising alternative, using one model as both student and teacher.
  arXiv  Detail & Related papers  (2026-07-05T17:47:29Z)
• Trajectory-Refined Distillation [25.346810464266497]
  Trajectory-Refined Distillation (TRD) is a trajectory-level correction method that revises the student's rollout under the teacher guidance while within on-policy support.<n> TRD can also be applied to on-policy self-distillation (OPSD), a parameter-sharing variant that uses the student model conditioned on privileged informations as the teacher.
  arXiv  Detail & Related papers  (2026-06-07T03:17:25Z)
• OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification [21.095025092772257]
  On-Policy Distillation (OPD) trains a student model on its own generative trajectories under dense token-level feedback from a stronger teacher.<n>We introduce OmniOPD, a novel framework that addresses both limitations through a logit-free, chunk-level supervision signal.
  arXiv  Detail & Related papers  (2026-05-31T22:31:15Z)
• Your Teacher Can't Help You Here: Combating Supervision Fidelity Decay in On-Policy Distillation [81.10000755917712]
  On-policy distillation transfers reasoning capabilities by training a student model on its own generated trajectories using token-level feedback from a teacher.<n>As student-generated prefixes lengthen, the teacher's next-token distribution becomes less confident and less discriminative.<n>To mitigate SFD, we introduce textbfLookahead Group Reward (ours).
  arXiv  Detail & Related papers  (2026-05-29T04:39:20Z)
• Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation [28.572361799234784]
  We show that raw KL disagreement is a coarse proxy for learning value.<n>We formalize this local compatibility as token teachability.<n>We propose Teachability-Aware OPD, a lightweight token-position selection method.
  arXiv  Detail & Related papers  (2026-05-26T10:56:46Z)
• Prefix Teach, Suffix Fade: Local Teachability Collapse in Strong-to-Weak On-Policy Distillation [49.117085054884676]
  On-policy distillation trains a student model on its own rollouts using dense feedback from a stronger teacher.<n>We operationalize this principle through a trajectory-specific release rule.<n> Experimental results across strong-to-weak distillation tasks indicate that this release rule consistently outperforms standard full-trajectory OPD.
  arXiv  Detail & Related papers  (2026-05-13T15:05:30Z)
• Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning [17.384089089363382]
  We identify a root cause that existing methods overlook: the uniform penalization of errors.<n>Current approaches treat all incorrect rollouts within a group identically.<n>We propose the Asymmetric Confidence-aware Error Penalty (ACE)
  arXiv  Detail & Related papers  (2026-02-24T22:46:43Z)
• Probability-Entropy Calibration: An Elastic Indicator for Adaptive Fine-tuning [55.2818264614932]
  RankTuner introduces a probability--entropy calibration signal, the Relative Rank Indicator, which compares the rank of the ground-truth token with its expected rank under the prediction distribution.<n>The inverse indicator is used as a token-wise Relative Scale to reweight the fine-tuning objective, focusing updates on truly under-learned tokens.
  arXiv  Detail & Related papers  (2026-02-02T07:27:19Z)
• Long-Chain Reasoning Distillation via Adaptive Prefix Alignment [57.130176131042965]
  We propose a framework that exploits teacher CoTs for distillation through adaptive prefix alignment.<n>P-ALIGN adaptively truncates teacher-generated reasoning trajectories by determining whether the remaining suffix is concise.<n>Experiments on multiple mathematical reasoning benchmarks demonstrate that P-ALIGN outperforms all baselines by over 3%.
  arXiv  Detail & Related papers  (2026-01-15T04:40:45Z)
• Stable On-Policy Distillation through Adaptive Target Reformulation [7.361248172930405]
  Veto is an objective-level reformulation that constructs a geometric bridge in the logit space.<n>Veto consistently outperforms supervised fine-tuning and existing on-policy baselines.
  arXiv  Detail & Related papers  (2026-01-12T02:57:39Z)
• Taming the Tail: Stable LLM Reinforcement Learning via Dynamic Vocabulary Pruning [35.41241409574854]
  We show that inference engines and numerically-precise training systems produce different probability distributions from the same parameters, creating a training-inference mismatch.<n>By pruning such tokens, we trade large, systematically biased mismatches for a small, bounded optimization bias.
  arXiv  Detail & Related papers  (2025-12-28T21:44:07Z)
• Sebra: Debiasing Through Self-Guided Bias Ranking [54.09529903433859]
  Ranking samples by fine-grained estimates of spuriosity has recently been shown to significantly benefit bias mitigation.<n>We propose a debiasing framework based on our novel ulSelf-Guided ulBias ulRanking (emphSebra)<n>Sebra mitigates biases via an automatic ranking of data points by spuriosity within their respective classes.
  arXiv  Detail & Related papers  (2025-01-30T11:31:38Z)
• Noisy Correspondence Learning with Self-Reinforcing Errors Mitigation [63.180725016463974]
  Cross-modal retrieval relies on well-matched large-scale datasets that are laborious in practice. We introduce a novel noisy correspondence learning framework, namely textbfSelf-textbfReinforcing textbfErrors textbfMitigation (SREM)
  arXiv  Detail & Related papers  (2023-12-27T09:03:43Z)
• Prompt Tuning Pushes Farther, Contrastive Learning Pulls Closer: A Two-Stage Approach to Mitigate Social Biases [13.837927115198308]
  We propose an adversarial training-inspired two-stage debiasing model using Contrastive learning and Continuous Prompt Augmentation. Our approach guides the model to achieve stronger debiasing performance by adding difficulty to the training process.
  arXiv  Detail & Related papers  (2023-07-04T09:35:03Z)
• Relieving Long-tailed Instance Segmentation via Pairwise Class Balance [85.53585498649252]
  Long-tailed instance segmentation is a challenging task due to the extreme imbalance of training samples among classes. It causes severe biases of the head classes (with majority samples) against the tailed ones. We propose a novel Pairwise Class Balance (PCB) method, built upon a confusion matrix which is updated during training to accumulate the ongoing prediction preferences.
  arXiv  Detail & Related papers  (2022-01-08T07:48:36Z)
• Accelerated Convergence for Counterfactual Learning to Rank [65.63997193915257]
  We show that convergence rate of SGD approaches with IPS-weighted gradients suffers from the large variance introduced by the IPS weights. We propose a novel learning algorithm, called CounterSample, that has provably better convergence than standard IPS-weighted gradient descent methods. We prove that CounterSample converges faster and complement our theoretical findings with empirical results.
  arXiv  Detail & Related papers  (2020-05-21T12:53:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.