Reproduce “OC-OPD: Observation-coordinated Supervision 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
Ruoran Xu, Yilun Zhao, Ziyu Chen, Arman Cohan
OC-OPD transfers a teacher’s next-token knowledge to smaller models using student-generated prefixes. It compares student and teacher candidate distributions, derives reverse- and forward-KL corrections, and coordinates them with entropy-based preferences and a discrepancy gate. On Qwen3-1.7B, it exceeds the strongest compared baseline by 6.88 Avg@32 and 11.11 Pass@32 points across AIME24–26; the guarantee is local to student-logit space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.