Reproduce “Unpacking the Compute Efficiency of Strong-to-Weak OPD against GRPO” 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
Zeyu Leo Liu, Yang Li, Bo Pang, Emre Can Acikgoz, Xuan-Phi Nguyen, Anurag Koul, Srijan Bansal, Erik Nijkamp and 2 more
The authors compare on-policy distillation (OPD) with group-relative policy optimization (GRPO) using held-out accuracy against training FLOPs. External-teacher OPD generally trails GRPO, but a Qwen3-4B-Base student distilled from its own GRPO checkpoint recovers the teacher’s OOD math score with about 50× fewer distillation-stage FLOPs. Consolidating math and code specialists reaches joint-GRPO performance with 42.0% fewer total FLOPs in the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.