Reproduce “On the Off-Policy Teacher in 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
Langlin Huang, Hao Liu, Mononito Goswami, Xinyu Li, Prithwith Jana, Nikos Kanakaris, Patrick Blöbaum, Purak Jain
SCOUT adapts a teacher for on-policy distillation by training it on student-generated prefixes with verifiable outcome rewards. In AIME 2025, continuation accuracy declined with longer student prefixes. For a Qwen3-4B-Instruct-2507 teacher and Qwen3-1.7B student, accuracy rose from 49.2±1.5 to 51.4±0.6 across six math benchmarks. Results cover math and function-level code generation; teacher rollouts and RL updates add cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.