Reproduce “$D^2$-OPD: Separating Direction and Distance in Extrapolative 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
Bowen Fang, Dingmin Wang, Hengzhi Pei
The authors present D²-OPD for distilling a smaller language model from a larger model’s guidance on its own generated answers. At each prefix, it combines two signals to choose the target’s direction and uses a per-prefix KL bound to limit its distance from the teacher. Across three math benchmarks, it averages 47.19, versus 37.74 for TOP-D; experiments cover three Qwen3 pairs, with one run per configuration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.