Rethinking On-Policy Self-Distillation for Thinking Models

• URL: http://arxiv.org/abs/2607.05184v1
• Date: Mon, 06 Jul 2026 15:01:35 GMT
• Title: Rethinking On-Policy Self-Distillation for Thinking Models
• Abstract summary: Self-distillation is a promising recipe for self-improvement in language models.<n>We show that privileged self-distillation degrades thinking models on long reasoning traces.<n>Our diagnostics link this failure mode to how privileged teacher context reshapes learning at high-entropy forking positions.
• Score: 32.29580112160733
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Self-distillation is a promising recipe for self-improvement in language models. In this setting, a model can serve as its own teacher when given privileged information, such as a solution to a math problem. This seems especially appealing for thinking models, which can use test-time reasoning to absorb the privileged information. Surprisingly, we show that privileged self-distillation degrades thinking models on long reasoning traces: across five Qwen3 and OLMo thinking models evaluated on AIME24, AIME25, and HMMT25, privileged-context distillation causes a relative drop of up to 17% in avg@16 accuracy. The degradation scales with the amount of privileged context withheld from the student and is most pronounced at long rollout budgets, where thinking models otherwise obtain their largest gains. This failure mode is not specific to self-distillation: on-policy distillation (OPD) improves thinking models, but privileged OPD reverses these gains. Our diagnostics link this failure mode to how privileged teacher context reshapes learning at high-entropy forking positions, where multiple continuations remain plausible and may lead to different reasoning paths. Privileged context lowers fork rates in thinking-model rollouts but not in instruction-model rollouts. This leads to an interesting dichotomy, where privileged context can help instruction-tuned models but hurts stronger thinking models. The effect is visible when the student begins a self-correction branch, where privileged OPD penalizes sampled reconsideration tokens that vanilla OPD supports. Thinking models trained with a privileged teacher produce fewer verification, backtracking, and hedging markers, even after length normalization. These findings indicate that self-distillation for strong thinking models requires attention to token-level signal, especially around correction and reasoning steps.

Related papers

• Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models [38.84286306535778]
  We show that stopping at the first correct prefix improves accuracy over standard reasoning up to 21%.<n>While common efficiency strategies like early stopping substantially reduce verbose overthinking, they fail to mitigate harmful overthinking.<n>Our findings generalize to language-only reasoning benchmarks, highlighting harmful overthinking as a broader reliability risk.
  arXiv  Detail & Related papers  (2026-06-01T19:59:27Z)
• Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information [22.436966302243565]
  We propose Anti-Self-Distillation, which ascends a divergence between student and teacher rather than descending it.<n>AntiSD reaches the GRPO baseline's accuracy in 2 to 10x fewer training steps and improves final accuracy by up to 11.5 points.
  arXiv  Detail & Related papers  (2026-05-12T06:40:43Z)
• Pando: Do Interpretability Methods Work When Models Won't Explain Themselves? [53.07826484214082]
  We introduce Pando, a model-organism benchmark that breaks this confound via an explanation axis.<n>Pando predicts held-out model decisions from $10$ labeled query-response pairs.<n>When explanations are faithful, black-box elicitation matches or exceeds all white-box methods.
  arXiv  Detail & Related papers  (2026-04-13T06:42:24Z)
• Your Models Have Thought Enough: Training Large Reasoning Models to Stop Overthinking [50.97239453902612]
  Large Reasoning Models (LRMs) have achieved impressive performance on challenging tasks, yet their deep reasoning often incurs substantial computational costs.<n>Inspired by Evidence Accumulation Models, we find that LRMs have accumulated sufficient information early in reasoning, making further reasoning steps redundant.<n>We propose Just-Enough Thinking (JET), which trains models to proactively terminate unnecessary reasoning.
  arXiv  Detail & Related papers  (2025-09-27T16:25:06Z)
• Reasoning about Uncertainty: Do Reasoning Models Know When They Don't Know? [7.423494663010787]
  Reasoning language models have set state-of-the-art (SOTA) records on many challenging benchmarks.<n>Like previous language models, reasoning models are prone to generating confident, plausible responses that are incorrect.<n>Knowing when and how much to trust these models is critical to the safe deployment of reasoning models in real-world applications.
  arXiv  Detail & Related papers  (2025-06-22T21:46:42Z)
• How Well Can Reasoning Models Identify and Recover from Unhelpful Thoughts? [35.00173704760055]
  We investigate how well reasoning models identify and recover from four types of unhelpful thoughts.<n>We show that models are effective at identifying most unhelpful thoughts but struggle to recover from the same thoughts when these are injected into their thinking process.
  arXiv  Detail & Related papers  (2025-06-12T17:59:53Z)
• Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models [130.5487886246353]
  Extending thinking traces using prompts like "Wait" or "Let me rethink" can improve performance.<n>This raises a natural question: Does thinking more at test-time truly lead to better reasoning?<n>We show a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to "overthinking"
  arXiv  Detail & Related papers  (2025-06-04T17:55:09Z)
• The Price of a Second Thought: On the Evaluation of Reasoning Efficiency in Large Language Models [54.88805865447848]
  We show that instruct models achieve higher efficiency overall, and problem difficulty affects efficiency.<n>We propose COTHINK, a simple two-stage pipeline: an instruct model drafts a brief outline, and a thinking model expands it.<n>On GSM8K, MATH500, and AIME24, COTHINK cuts token usage by 21.1% while keeping accuracy on four thinking models, and remains competitive with strong efficiency baselines.
  arXiv  Detail & Related papers  (2025-05-28T06:24:45Z)
• SuperCorrect: Advancing Small LLM Reasoning with Thought Template Distillation and Self-Correction [89.56181323849512]
  SuperCorrect is a novel two-stage framework that uses a large teacher model to supervise and correct both the reasoning and reflection processes of a smaller student model.<n>In the first stage, we extract hierarchical high-level and detailed thought templates from the teacher model to guide the student model in eliciting more fine-grained reasoning thoughts.<n>In the second stage, we introduce cross-model collaborative direct preference optimization (DPO) to enhance the self-correction abilities of the student model.
  arXiv  Detail & Related papers  (2024-10-11T17:25:52Z)
• Symbolic Chain-of-Thought Distillation: Small Models Can Also "Think" Step-by-Step [133.60124577507727]
  Chain-of-thought prompting primes large language models to verbalize rationalization for their predictions. We show that orders-of-magnitude smaller models can still benefit from chain-of-thought prompting. We introduce Chain-of-Thought Distillation (SCoTD), a method to train a smaller student model on rationalizations sampled from a significantly larger teacher model.
  arXiv  Detail & Related papers  (2023-06-24T20:15:07Z)
• Explain, Edit, and Understand: Rethinking User Study Design for Evaluating Model Explanations [97.91630330328815]
  We conduct a crowdsourcing study, where participants interact with deception detection models that have been trained to distinguish between genuine and fake hotel reviews. We observe that for a linear bag-of-words model, participants with access to the feature coefficients during training are able to cause a larger reduction in model confidence in the testing phase when compared to the no-explanation control.
  arXiv  Detail & Related papers  (2021-12-17T18:29:56Z)

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.