mlsanityData to 7 Oct 2026

Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?

arXiv 2603.24472 ↗First posted 25 Mar 2026cs.CL cs.LGcited by 87

On the map: LLM reasoning → Prompting and chain-of-thought reasoning

Abstract (arXiv)

Self-distillation has emerged as an effective post-training paradigm for LLMs, often improving performance while shortening reasoning traces. However, in mathematical reasoning, we find that it can reduce response length while degrading performance. We trace this degradation to the suppression of epistemic verbalization - the model's expression of uncertainty during reasoning. Through controlled experiments varying conditioning context richness and task coverage, we show that conditioning the teacher on rich information suppresses uncertainty expression, enabling rapid in-domain optimization with limited task coverage but harming OOD performance, where unseen problems benefit from expressing uncertainty and adjusting accordingly. Across Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, and Olmo3-7B-Instruct, we observe performance drops of up to 40%. Our findings highlight that exposing appropriate levels of uncertainty is crucial for robust reasoning and underscore the importance of optimizing reasoning behavior beyond merely reinforcing correct answer traces.

Key facts. Cited by 87 papers in the map (data through 7 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 20 of its 29 references that are in the map. First posted on arXiv 25 Mar 2026.

Builds on

20 of 29 references.

Closest ancestors · 1

Same direction, newest first

  1. 2026Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models26 Jan

Other references · 19

  1. 2025Qwen3 Technical Reportcited by 8,195
  2. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
  3. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
  4. 2024LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Codecited by 1,906
  5. 2025OpenAI GPT-5 System Cardcited by 910
  6. 2023ToolAlpaca: Generalized Tool Learning for Language Models with 3000 Simulated Casescited by 383
  7. 2026Reinforcement Learning via Self-Distillationcited by 285
  8. 2026Self-Distillation Enables Continual Learningcited by 212
  9. 2026On-Policy Context Distillation for Language Modelscited by 161
  10. 2022Learning by Distilling Contextcited by 128
  11. 2026Privileged Information Distillation for Language Modelscited by 79
  12. 2025Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoningcited by 75
7 more
  1. 2026Expanding the Capabilities of Reinforcement Learning via Text Feedbackcited by 49
  2. 2026Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimizationcited by 17
  3. 2026Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertaintycited by 8
  4. 2026ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressurecited by 4
  5. 2026Accordion-Thinking: Self-Regulated Step Summaries for Efficient and Readable LLM Reasoningcited by 4
  6. 2026Compress the Easy, Explore the Hard: Difficulty-Aware Entropy Regularization for Efficient LLM Reasoningcited by 4
  7. 2025In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedbackcited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 87.

87 papers.

Sort:

By area:

  1. 2026Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipecited by 262
  2. 2026Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routingcited by 65
  3. 2026The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixescited by 37
  4. 2026Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillationcited by 31
  5. 2026Rethinking On-Policy Self-Distillation for Thinking Modelscited by 28
  6. 2026Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Modelscited by 24
  7. 2026KL for a KL: On-Policy Distillation with Control Variate Baselinecited by 24
  8. 2026Prune-OPD: Efficient and Reliable On-Policy Distillation for Long-Horizon Reasoningcited by 23
  9. 2026Purified OPSD: On-Policy Self-Distillation Without Losing How to Thinkcited by 23
  10. 2026Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Whycited by 21
  11. 2026Respecting Self-Uncertainty in On-Policy Self-Distillation for Efficient LLM Reasoningcited by 20
  12. 2026STEPS: Selective On-Policy Self-Distillation for Reasoningcited by 15
38 more
  1. 2026OGLS-SD: On-Policy Self-Distillation with Outcome-Guided Logit Steering for LLM Reasoningcited by 14
  2. 2026Trajectory-Refined Distillationcited by 14
  3. 2026Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVRcited by 13
  4. 2026Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Informationcited by 12
  5. 2026Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillationcited by 12
  6. 2026Preference-Based Self-Distillation: Beyond KL Matching via Reward Regularizationcited by 10
  7. 2026Crosslingual On-Policy Self-Distillation for Multilingual Reasoningcited by 10
  8. 2026Reinforcement Learning from Rich Feedback with Distributional DAggercited by 9
  9. 2026OPRD: On-Policy Representation Distillationcited by 9
  10. 2026Multi-Rollout On-Policy Distillation via Peer Successes and Failurescited by 8
  11. 2026On-Policy Delta Distillationcited by 8
  12. 2026On-Policy Distillation with Best-of-N Teacher Rollout Selectioncited by 7
  13. 2026Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoningcited by 7
  14. 2026The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillationcited by 6
  15. 2026Learning with Rare Success but Rich Feedback via Reflection-Enhanced Self-Distillationcited by 6
  16. 2026Learning from Language Feedback via Variational Policy Distillationcited by 6
  17. 2026Reducing the Safety Tax in LLM Safety Alignment with On-Policy Self-Distillationcited by 6
  18. 2026OmniOPSD: Rationale-Privileged On-Policy Self-Distillation for Affective Computingcited by 6
  19. 2026Denser neq Better: Limits of On-Policy Self-Distillation for Continual Post-Trainingcited by 6
  20. 2026DemoPSD: Disagreement-Modulated Policy Self-Distillationcited by 6
  21. 2026Agentic Reinforcement Learning with Observation-Calibrated Self-Distillationcited by 6
  22. 2026Rethinking Privileged Information in On-Policy Self-Distillationcited by 6
  23. 2026Breaking the Tokenizer Barrier: On-Policy Distillation across Model Familiescited by 5
  24. 2026A Formula-Driven Survey and Research Agenda for On-Policy Distillationcited by 5
  25. 2026ReNIO: Reweighting Negative Trajectory Importance for LLM On-Policy Distillationcited by 5
  26. 2026Diagnosing and Mitigating Thinking Collapse in On-Policy Self-Distillationcited by 5
  27. 2026Towards Understanding Modality Interaction in Multimodal Language Models via Partial Information Decompositioncited by 4
  28. 2026CriPO: Enhancing Rubric-based RL via Self-Distillationcited by 4
  29. 2026AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signalscited by 3
  30. 2026Procedural Memory Distillation: Online Reflection for Self-Improving Language Modelscited by 3
  31. 2026Simple-OPD: Demystifying Warm-up for On-policy Distillationcited by 3
  32. 2026Matching Supervision to the Student's Learning Capacity: A Unified Framework for On-Policy Self-Distillationcited by 3
  33. 2026Privileged Likelihood Is Not Automatically Value: Three Checks for Token Credit in On-Policy Self-Distillationcited by 3
  34. 2026Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agentscited by 3
  35. 2026RISE: Recursive Improvement via Self-Extrapolating Policy Distillationcited by 3
  36. 2026MixSD: Mixed Contextual Self-Distillation for Knowledge Injectioncited by 2
  37. 2026DRIFT: Difficulty Routing Self-DIstillation with Rhythm-Gated Exploration and Success BuFfer Trainingcited by 2
  38. 2026Geometric Self-Distillation for Reasoning Generalizationcited by 2

50 most cited of 87 All 87 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2026Skill-Conditioned Gated Self-Distillation for LLM Reasoning
  2. 2026The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning
  3. 2025Harnessing Negative Signals: Reinforcement Distillation from Teacher Data for LLM Reasoning
  4. 2025UNDO: Understanding Distillation as Optimization
  5. 2025Beyond Scaling Law: A Data-Efficient Distillation Framework for Reasoning
  6. 2025Exploring LLM Reasoning Through Controlled Prompt Variations
  7. 2025Distilling the Essence: Efficient Reasoning Distillation via Sequence Truncation
  8. 2025Thinking Out Loud: Do Reasoning Models Know When They're Right?