Runyang You, Yongqi Li, Tiezheng Yu, Wenjie Li · 2026 · arXiv (Cornell University)
DOI: 10.48550/arxiv.2610.28048 · cited by 0 works (OpenAlex) · open-access PDF
OPD(On-Policy Distillation)는 학생이 생성한 로루팅에서 교사와 학생의 최종 계층 배포를 일치함으로써 전문가 능력을 단일 학생으로 통합합니다. 우리는 깊이있는 예측 전환, 계층 간 예측의 변화가 추가 작업 관련 감독을 제공할 수 있는지 여부를 조사합니다. 우리의 진단 연구는 중간 계층 전환에서 강력한 교사-학생 합의가 질문 내에서 더 높은 응답 정확성과 관련이 있으며, 정확한 응답은 측정된 중간 계층 쌍에 걸쳐 더 낮은 평균 전환 KL를 나타냅니다. 이러한 관찰은 정책 전환 분포 (OPTD)를 동기화하여 최종 계층 배포를 증가시키는 것이며, 학생이 생성한 로루팅에서 교사와 학생의 계층-계
Machine translation (Meta M2M100); the English original follows.
On-policy distillation (OPD) consolidates expert capabilities into a single student by matching teacher and student final-layer distributions on student-generated rollouts. We investigate whether depthwise prediction transitions, the changes in predictions between layers, can provide additional task-relevant supervision. Our diagnostic study finds that stronger teacher--student agreement in intermediate transitions is associated with higher answer correctness within questions, with correct responses exhibiting lower mean transition KL across all measured intermediate layer pairs. These observations motivate On-Policy Transition Distillation (OPTD), which augments final-layer distribution matching by aligning teacher and student layer-to-layer log-probability increments on student-generated rollouts. Evaluations across three backbones and eight benchmarks spanning mathematics, coding, and instruction following show that OPTD achieves the best or tied-best overall performance in both dual- and three-domain distillation. In the three-domain setting, OPTD recovers $\approx 90\%$ of the overall teacher--student capability gap, compared with $\approx 75\%$ for OPD.
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
Topics: Distillation, Matching (statistics), Correctness, Transition (genetics), Computer science, Distribution (mathematics)
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.28048, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange