Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models

Hugging Face Daily Papers Papers

Summary

This paper investigates generalization in on-policy distillation for large language models, showing that it transfers reasoning behaviors and that teacher-student origin alignment is crucial, with multi-teacher combinations causing capability trade-offs.

On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.
Original Article
View Cached Full Text

Cached at: 08/24/26, 08:28 AM

Paper page - Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models

Source: https://huggingface.co/papers/2608.16647 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

On-policy distillation transfers reasoning behaviors rather than specific answers, with generalization strongly tied to teacher-student origin alignment and multi-teacher combinations causing capability trade-offs.

On-policy distillation(OPD) transfers teacher capabilities by supervising trajectories sampled from the student’s own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domaindistribution shifts tocross-domain transferand themulti-teachersetting. We find that OPD transfers a teacher’sreasoning behaviorrather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student:same-originpairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereascross-originpairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher’s influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosingmulti-teacherOPD.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2608\.16647

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.16647 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.16647 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.16647 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes

Hugging Face Daily Papers

This paper presents a comprehensive empirical study on on-policy distillation for large language models, identifying failure mechanisms like distribution mismatch and optimization instability, and proposing fixes such as stop-gradient objectives and RLVR-adapted teachers.

Behavior Leverage Imbalance in Multi-Teacher On-Policy Distillation

arXiv cs.CL

This paper identifies a failure mode in multi-teacher on-policy distillation for agentic language models where the student becomes over-reliant on tool-calling, and proposes Soft Clamp, a per-token divergence calibration method that reduces over-calling without sacrificing accuracy.