The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes

Siqi Zhu, Xuyan Ye, Hongyu Lu, Weiye Shi, Ge Liu

arXiv:2605.11182 · 2026-05-13 공개 · arXiv · PDF

language-models on-policy-distillation rlvr teacher-student system-prompt distribution-mismatch loss-formulation topk-gradients

Abstract

On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensitive to teacher choice and loss formulation, whereas OPSD fails in our tested settings due to test-time absence of instance-specific privileged information (PI). In contrast, OPSD is effective when PI represents a shared latent rule, such as a system prompt or alignment preference. We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.

한국어 요약

📋 한 줄 요약

**[LLM 사후 학습 / 증류]** On-Policy Distillation(OPD)과 On-Policy Self-Distillation(OPSD)이 언제·왜 실패하는지 진단하고 세 가지 실패 메커니즘과 대응 처방을 제시.

🎯 핵심 기여도

💡 핵심 아이디어

"학생이 자기 자신을 가르치라"는 OPSD가 직관적으로 보이지만, PI가 instance-specific하면 학생은 결국 PI 없이 PI-conditioned 교사들을 평균낸 정책을 학습할 뿐이며 이는 본질적으로 부족하다. 즉 OP(S)D는 PI 구조에 따라 작동 여부가 갈리는 조건적 도구이다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: 사후 학습 분야에서 "OP(S)D가 작동한다/안 한다" 논쟁을 PI 구조와 그래디언트 편향 관점에서 정리해 실무 채택 지침을 제공. **한계**: 분석이 주로 수학·정렬 도메인에 집중되며 도구 사용·멀티모달 등 다른 도메인 일반화는 추가 검증 과제.

🚀 실용적 활용