Reproduce “One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-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

Khouloud Saadi, Di Wang

MCPD trains a reasoning model by averaging a frozen teacher’s predictions across independently masked partial views of a verified solution at student-generated prefixes. This replaces supervision from one complete reference with a probability-space marginal over views. In the reported evaluation, it lowers reference sensitivity on 80.4% of 500 problems; on Qwen3-8B AIME25, Avg@12 is 72.8 versus OPSD’s 70.8.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.