Reproduce “Understanding Off- vs On-Policy Distillation: A Tale of Distinct Training Objectives” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qiwei Di, Xuheng Li, Kaixuan Ji, Chenggong Zhang, Heyang Zhao, Quanquan Gu

The authors analyze how off-policy and on-policy distillation combines teachers’ policies. Forward KL yields a weighted arithmetic mixture; reverse KL yields a normalized geometric aggregate that can depend on continuations. In a ten-response example with expert weight 0.9 and uniform other teachers, reverse aggregation favors the correct response only above confidence 0.7617; these are analytical target comparisons, not training results.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.