When Clipping Reverses Correction: Failure Dynamics of Pointwise Forward-KL On-Policy Self-Distillation

arXiv:2609.38995 · cs.CL · Submitted 2026-09-30 · Read on arXiv

cs.CL

Submitted: 2026-09-30

Updated: 2026-09-30

Code: https://github.com/huggingface/math-verify

Terminology

• Distillation — 29× in this paper · explained in 2 episodes
• Self-distillation — 22× in this paper · explained in 6 episodes
• On-Policy Self-Distillation — 14× in this paper
• Accuracy — 9× in this paper · explained in 1 episode
• Benchmark — 4× in this paper · explained in 2 episodes
• On-Policy Distillation — 4× in this paper · explained in 3 episodes
• Trace — 4× in this paper · explained in 2 episodes
• Alignment — 3× in this paper · explained in 1 episode
• Calibration — 2× in this paper · explained in 7 episodes
• On-policy Self-Distillation (OPSD) — 2× in this paper · explained in 2 episodes
• Reinforcement Learning — 2× in this paper · explained in 4 episodes
• Adaptation — 1× in this paper · explained in 1 episode
• Steering — 1× in this paper · explained in 2 episodes
• Transformer — 1× in this paper · explained in 2 episodes
• Verification — 1× in this paper · explained in 2 episodes

Sources

• SCOPE-OPSD: Fisher-Conditioned Privileged Subspaces for On-Policy Self-Distillation
• DualOPSD: Adaptive Privileged Teachers for On-Policy Self-Distillation
• PAST: Privileged Adaptation from Complete Student Trajectories for On-Policy Self-Distillation
• Rethinking Reward Supervision: Rubric-Conditioned Self-Distillation
• DASH: Divergence-Adaptive Supervision Horizons for On-Policy Self-Distillation of Reasoning Models
• Reinforcement Learning via Self-Distillation
• Privileged Solutions or Context-Induced Teacher Behavior? Dissecting On-Policy Self-Distillation
• On-Policy Self-Distillation without Any Supervision
• Search-E1: Self-Distillation Drives Self-Evolution in Search-Augmented Reasoning
• When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning
• RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation
• HybridFlow: A Flexible and Efficient RLHF Framework
• PAINT: Partial-Solution Adaptive Interpolated Training for Self-Distilled Reasoners
• Self-Supervised On-Policy Distillation for Reasoning Language Models
• TRACE: Distilling Where It Matters via Token-Routed Self On-Policy Alignment
• Qwen3 Technical Report
• OGLS-SD: On-Policy Self-Distillation with Outcome-Guided Logit Steering for LLM Reasoning
• What Does Privileged Information Add to On-Policy Self-Distillation?
• Is More Privileged Information Better? From Solution Traces to Problem-Solving Structure in Self-Distilled Reasoning

Related papers

• Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
• Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
• Subliminal Steering: Stronger Encoding of Hidden Signals
• MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
• The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
• Untangling the Mechanisms of Misleading Context in Medical Question Answering