Reproduce “Bootstrapped 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
Yue Wu, Xueyu Zhou, Yangrong Hu, Jian Huang
BOPD trains a fresh student through two-stage on-policy distillation. First, reference-guided training constructs a teacher; then that teacher guides the student without reference access. In Qwen3-4B experiments, BOPD reaches 23.98% mean Avg@12 accuracy, above the initial model and OPSD teacher. This aggregate gain is not universal: on HMMT 2025, BOPD scores below both comparators.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.