Reproduce “O$^2$PD: Overlap-Gated On-Policy Distillation on Reliable Prefix” 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
Shuaishuai Zu, Jihao Zhao, Pengfei Li, Guanghuan Fang, Ben Feng, Yuhan Liu, Biao Qin, Baochuan Yang
O²PD trains only the first 20% of each student-generated response and adjusts token alignment according to overlap between teacher and student top-k predictions. On the Qwen3-1.7B student and Qwen3-4B-NTRL teacher pair, its seven-benchmark average is 50.58, versus 45.29 for top-k OPD. O²PD has a higher average than OPD across all three tested student–teacher pairs; the evidence is limited to math benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.