The paper argues OPD learns where reasoning paths stay open

The paper argues OPD learns where reasoning paths stay open

Authors Shiping Yang, Chuanqi Tang, Hossein Hajimirsadeghi, Layla El Asri, Mohammad Amin Shabani, and Angel X. Chang argue that on-policy distillation (OPD) learns more than which candidate tokens a student considers. It also learns where the student keeps reasoning paths open, which the authors call the “shape” of uncertainty. The authors introduce entropy shape—the sequence of next-token entropies—and Entropy Shape Correlation (ESC), a measure of student-teacher agreement on that sequence. They report that intervening on a teacher’s entropy while preserving its token preferences indicates that where to branch causally drives OPD gains, independently of how to branch. They also frame predicting OPD outcomes as a teacher-selection task. In their evaluation protocol, pre-OPD ESC selects the best teacher for both students and outperforms existing predictors. If this measure helps identify a useful teacher before distillation, it could make teacher selection more informed; the abstract does not establish how well the result transfers beyond that protocol. What results would show whether pre-OPD ESC predicts teacher quality for students and settings outside the authors’ evaluation protocol?

Publication: daadfca5-7868-4fbd-8892-e560800eca6c

attribution — pass: The six listed authors match the arXiv page. The claims are framed as what the authors argue, introduce or report, not as established fact or EM opinion.

Authors:Shiping Yang, Chuanqi Tang, Hossein Hajimirsadeghi, Layla El Asri, Mohammad Amin Shabani, Angel X. Chang

summary — uncertain: The reading tracks the abstract faithfully, but a non-specialist may not follow it. It never says that OPD is a post-training method in which a student language model learns from a teacher model. It also leaves 'tokens' and 'next-token entropies' unexplained (add tooltips). It omits the motivating puzzle that a stronger teacher can fail to improve the student. 'Both students' appears without saying two student models were tested. Smallest repair: in the first paragraph, add the post-training context and the stronger-teacher puzzle. Tooltips should cover OPD, token and entropy.

On-policy distillation (OPD) is increasingly adopted in the post-training of large language models. However, a stronger teacher may fail to improve the student.

date — pass: originalPublished 2026-10-28 matches the v1 submission date. The later feed date (Oct 31) was correctly not used.

[Submitted on 28 Oct 2026]

support — pass: Each claim is backed by an exact abstract excerpt. The causal finding is attributed as the authors' empirical claim. The teacher-selection result is scoped to their evaluation protocol. The benefit statement is explicitly conditional, and the limitations note that the evidence is abstract-only. The question follows from that unresolved scope.

we formulate OPD outcome prediction as a teacher selection task and establish an evaluation protocol for it, under which pre-OPD ESC selects the best teacher for both students and outperforms existing predictors.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.