Privileged-context on-policy self-distillation degrades thinking models' long-budget accuracy by suppressing forking and self-correction behaviors, while helping instruction-tuned models.
hub Canonical reference
Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
Canonical reference. 80% of citing Pith papers cite this work as background.
34
Pith papers citing it
Background
80% of classified citations
abstract
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.
hub tools
citation-role summary
background 4
other 1
citation-polarity summary
years
2026 34representative citing papers
Purified OPSD subtracts a reference-only teacher's signal from standard OPSD supervision and applies PMI to create a cleaner distillation target, yielding gains on long-CoT models while preserving epistemic behavior.
SC-GRPO improves RL with verifiable rewards by multiplying GRPO gradients with self-induced per-token KL divergence, outperforming GRPO by 8.1% and DAPO by 5.9% on math, code, and agent benchmarks.
ZPPO improves distillation to small vision-language models by using binary and negative candidate prompts plus a replay buffer for hard questions, outperforming standard distillation and GRPO on a 31-benchmark suite with largest gains at the 0.8B scale.
OPRD performs distillation in hidden-state space on on-policy data for deterministic gradients and better math benchmark performance, plus OPRD-Bridge for cross-architecture transfer via low-rank projectors.
DistIL applies distributional DAgger with forward cross-entropy to achieve monotonic policy improvement and better Pass@N from rich feedback in RL for reasoning tasks.
VPD frames language feedback learning as variational EM so the teacher policy refines itself via trust-region updates on outcomes while the student learns dense token distributions on its own rollouts, outperforming fixed-teacher baselines on reasoning and code tasks.
EGRSD and CL-EGRSD advance the accuracy-length frontier in LLM reasoning by entropy-guided weighting of token-level distillation signals from the teacher.
RLRT augments GRPO by reinforcing tokens on correct student rollouts that the teacher would not have predicted, outperforming standard self-distillation and exploration baselines on Qwen3 models.
TRACE improves math reasoning by distilling only on annotator-marked critical spans with forward KL on correct key spans, optional reverse KL on errors, and GRPO elsewhere, gaining 2.76 points over GRPO while preserving OOD performance.
vOPD stabilizes on-policy distillation gradients by subtracting a closed-form per-token negative reverse KL baseline as a detached control variate, preserving unbiasedness while lowering variance and matching expensive full-vocabulary methods.
PBSD derives a reward-reweighted teacher distribution as the analytic optimum of a reward-regularized objective, yielding better stability and performance than KL-based self-distillation on math reasoning and tool-use tasks.
OPD for LLMs suffers length inflation and repetition collapse; StableOPD uses reference divergence and rollout mixing to prevent it and improve math reasoning performance by 7.2% on average.
TESSY creates stylistically consistent synthetic data via teacher-student token interleaving, yielding 11.25% and 6.68% gains on code benchmarks where pure teacher data causes 3.25% and 10.02% drops.
Disagreement-modulated reverse-KL barycenter targets let on-policy self-distillation attenuate privileged leakage while preserving exploration, beating SDPO and GRPO on SciKnowEval and GPQA.
A geometric self-distillation objective using Hellinger loss and Fisher-Rao proximal regularization prevents predictive drift in LLM post-training, improving out-of-distribution reasoning by 5.7-8.6 points.
Transferring a weak model’s RL-induced log-ratio policy shift on a strong student’s own rollouts raises AIME accuracy more cheaply than imitating the weak teacher or running matched-step RL on the student.
Sleep-time Knowledge Seeding plus Dreaming lets LLMs expand capacity, distill fragile in-context memories into stable parameters, and self-improve without human labels.
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
MixSD uses dynamic mixing of the model's expert and naive conditionals to create distribution-aligned supervision that improves the memorization-retention tradeoff over standard SFT.
CIPO jointly optimizes standard RLVR rewards with correction samples derived from the model's own failed attempts, yielding better reasoning and self-correction on math and code benchmarks.
RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
Anti-Self-Distillation reverses self-distillation signals via PMI to fix overconfidence on structural tokens, matching GRPO baseline accuracy 2-10x faster with up to 11.5 point gains across 4B-30B models.
citing papers explorer
-
Rethinking On-Policy Self-Distillation for Thinking Models
Privileged-context on-policy self-distillation degrades thinking models' long-budget accuracy by suppressing forking and self-correction behaviors, while helping instruction-tuned models.
-
Purified OPSD: On-Policy Self-Distillation Without Losing How to Think
Purified OPSD subtracts a reference-only teacher's signal from standard OPSD supervision and applies PMI to create a cleaner distillation target, yielding gains on long-CoT models while preserving epistemic behavior.
-
Learning from Own Solutions: Self-Conditioned Credit Assignment for Reinforcement Learning with Verifiable Rewards
SC-GRPO improves RL with verifiable rewards by multiplying GRPO gradients with self-induced per-token KL divergence, outperforming GRPO by 8.1% and DAPO by 5.9% on math, code, and agent benchmarks.
-
Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients
ZPPO improves distillation to small vision-language models by using binary and negative candidate prompts plus a replay buffer for hard questions, outperforming standard distillation and GRPO on a 31-benchmark suite with largest gains at the 0.8B scale.
-
OPRD: On-Policy Representation Distillation
OPRD performs distillation in hidden-state space on on-policy data for deterministic gradients and better math benchmark performance, plus OPRD-Bridge for cross-architecture transfer via low-rank projectors.
-
Reinforcement Learning from Rich Feedback with Distributional DAgger
DistIL applies distributional DAgger with forward cross-entropy to achieve monotonic policy improvement and better Pass@N from rich feedback in RL for reasoning tasks.
-
Learning from Language Feedback via Variational Policy Distillation
VPD frames language feedback learning as variational EM so the teacher policy refines itself via trust-region updates on outcomes while the student learns dense token distributions on its own rollouts, outperforming fixed-teacher baselines on reasoning and code tasks.
-
Respecting Self-Uncertainty in On-Policy Self-Distillation for Efficient LLM Reasoning
EGRSD and CL-EGRSD advance the accuracy-length frontier in LLM reasoning by entropy-guided weighting of token-level distillation signals from the teacher.
-
Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVR
RLRT augments GRPO by reinforcing tokens on correct student rollouts that the teacher would not have predicted, outperforming standard self-distillation and exploration baselines on Qwen3 models.
-
TRACE: Distilling Where It Matters via Token-Routed Self On-Policy Alignment
TRACE improves math reasoning by distilling only on annotator-marked critical spans with forward KL on correct key spans, optional reverse KL on errors, and GRPO elsewhere, gaining 2.76 points over GRPO while preserving OOD performance.
-
KL for a KL: On-Policy Distillation with Control Variate Baseline
vOPD stabilizes on-policy distillation gradients by subtracting a closed-form per-token negative reverse KL baseline as a detached control variate, preserving unbiasedness while lowering variance and matching expensive full-vocabulary methods.
-
Preference-Based Self-Distillation: Beyond KL Matching via Reward Regularization
PBSD derives a reward-reweighted teacher distribution as the analytic optimum of a reward-regularized objective, yielding better stability and performance than KL-based self-distillation on math reasoning and tool-use tasks.
-
Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Models
OPD for LLMs suffers length inflation and repetition collapse; StableOPD uses reference divergence and rollout mixing to prevent it and improve math reasoning performance by 7.2% on average.
-
How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data
TESSY creates stylistically consistent synthetic data via teacher-student token interleaving, yielding 11.25% and 6.68% gains on code benchmarks where pure teacher data causes 3.25% and 10.02% drops.
-
DemoPSD: Disagreement-Modulated Policy Self-Distillation
Disagreement-modulated reverse-KL barycenter targets let on-policy self-distillation attenuate privileged leakage while preserving exploration, beating SDPO and GRPO on SciKnowEval and GPQA.
-
Geometric Self-Distillation for Reasoning Generalization
A geometric self-distillation objective using Hellinger loss and Fisher-Rao proximal regularization prevents predictive drift in LLM post-training, improving out-of-distribution reasoning by 5.7-8.6 points.
-
Weak-to-Strong Generalization via Direct On-Policy Distillation
Transferring a weak model’s RL-induced log-ratio policy shift on a strong student’s own rollouts raises AIME accuracy more cheaply than imitating the weak teacher or running matched-step RL on the student.
-
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
Sleep-time Knowledge Seeding plus Dreaming lets LLMs expand capacity, distill fragile in-context memories into stable parameters, and self-improve without human labels.
-
Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoning
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
-
MixSD: Mixed Contextual Self-Distillation for Knowledge Injection
MixSD uses dynamic mixing of the model's expert and naive conditionals to create distribution-aligned supervision that improves the memorization-retention tradeoff over standard SFT.
-
Learning from Failures: Correction-Oriented Policy Optimization with Verifiable Rewards
CIPO jointly optimizes standard RLVR rewards with correction samples derived from the model's own failed attempts, yielding better reasoning and self-correction on math and code benchmarks.
-
Learning with Rare Success but Rich Feedback via Reflection-Enhanced Self-Distillation
RESD turns failure trajectories into token-level supervision via retrospective reflections and a persistent global playbook, enabling faster improvement than standard self-distillation or GRPO with only one rollout per prompt.
-
Multi-Rollout On-Policy Distillation via Peer Successes and Failures
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
-
Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information
Anti-Self-Distillation reverses self-distillation signals via PMI to fix overconfidence on structural tokens, matching GRPO baseline accuracy 2-10x faster with up to 11.5 point gains across 4B-30B models.
-
Selective Off-Policy Reference Tuning with Plan Guidance
SORT turns all-wrong prompts into selective learning signals by weighting tokens more predictable under plan guidance from reference solutions, improving over GRPO on reasoning benchmarks especially for weaker models.
-
Multilingual Safety Alignment via Self-Distillation
MSD enables cross-lingual safety transfer in LLMs via self-distillation with Dual-Perspective Safety Weighting, improving safety in low-resource languages without target response data.
-
AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency
AtManRL learns an additive attention mask on CoT traces to produce a saliency reward that, when combined with outcome rewards in GRPO, trains LLMs to generate reasoning that genuinely influences final predictions.
-
Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
On-policy distillation works when student and teacher models share thinking patterns and the teacher adds new capabilities, with success tied to alignment on a small set of high-probability tokens.
-
Procedural Memory Distillation: Online Reflection for Self-Improving Language Models
PMD extracts and distills cross-episode procedural knowledge from RL rollouts into LLM policies at three abstraction levels, yielding 3.8-13.6% gains over SDPO on SCIKNOWEVAL and LIVECODEBENCH via co-evolution.
-
Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation
FiRe-OPD introduces a two-stage filter-then-soft-reweight procedure for trajectory- and token-level supervision in on-policy distillation, claiming gains over prior token-level methods.
-
Physics-Guided Policy Optimization with Self-Distillation
PGPO modulates per-step trust in self-distilled updates via a mutual-information estimate derived from a viscous-fluid analogy, preserves SGD weak-approximation order, and reports gains of up to 4.5 points on Science-QA while avoiding late-training collapse.
-
A Brief Overview: On-Policy Self-Distillation In Large Language Models
This overview paper explains the conceptual foundations and design principles of On-Policy Self-Distillation for large language models from a beginner's perspective.
- Constitutional On-Policy Safe Distillation
- On-Policy Distillation with Best-of-N Teacher Rollout Selection