When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning

Xiaogeng Liu, Xinyan Wang, Yingzi Ma, Yechao Zhang, Chaowei Xiao, [0.6em] Johns Hopkins University · 2026

Read When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning by Xiaogeng Liu, Xinyan Wang, Yingzi Ma, Yechao Zhang, Chaowei Xiao, [0.6em] Johns Hopkins University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-21606.txt.

Description

Xiaogeng Liu and co-authors ask when teacher tokens are reliable, giving place-weighted training for reasoning. This trains a student on its own rollouts with a special teacher. But its goal weights all made tokens equally, treating the teacher target as reliable at every student prefix. Entropy methods relax this sameness by tuning token guidance with teacher entropy. But high entropy from the teacher in reasoning has unclear meaning, showing either non-viable doubt or benign solution mix. They add a branch health check to spot this effect.

Catalog record

CREATORXiaogeng Liu, Xinyan Wang, Yingzi Ma, Yechao Zhang, Chaowei Xiao, [0.6em] Johns Hopkins University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65974
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.21606

Browse the full library · About the Tower of Babel