Reproduce “Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models” 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
Ruize Zheng, Xuyang Liu, Siyu Ren, Xiaobing Tu, Yinggui Wang, Jinkui Ren, Xiantao Zhang, Linfeng Zhang
The authors introduce On-Policy Mid-Training Distillation (OPMD), which selects corpus contexts, lets a student generate continuations, and trains it using feedback from a frozen teacher on the student’s generated states. On two mathematics benchmarks, OPMD improved both pass@16 and avg@16 over the Qwen3-1.7B-Base model and avoided the decline observed with next-token prediction; broader-domain gains are unestablished.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.