Reproduce “Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability” 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

Bingxi Hou, Guochao Jiang, Guofeng Quan, Weiqing Li, Wenfeng Feng, Guohua Liu, Yuewei Zhang

The method compares teacher and student predictions at positions where one token on each side spans the same text, restricting reverse-KL supervision to shared vocabulary choices. Across three model pairs, tested span log-probability mean squared error additions lowered Full accuracy in all 18 positive-weight settings. Strict top-16 reached 32.64% versus 31.59% for SimCT and 26.96% for Base, on math and code evaluations.

Summary by alphaXiv.

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