Reproduce “Proposing Better Rollouts for On-Policy Distillation” 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
Sirui Chen, Yunzhe Qi, Haobo Xu, Jiaru Zou, Mengting Ai, Zihao Li, Jingrui He
MPOPD changes on-policy distillation rollouts, shifting student sampling toward tokens the teacher supports more while keeping the policy close to the student. A learned proposal approximates this target without running teacher and student at every decoding step. On Qwen3-4B-Base, average accuracy across eight math benchmarks is 35.37%, versus 28.46% for OPD, and generation length falls 41% in this evaluated setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.