pith. sign in

Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it
abstract

On-policy distillation (OPD) trains a student on its own rollouts with token-level teacher supervision. Recent selective OPD methods exploit the non-uniformity of OPD signals by prioritizing high-entropy or high-disagreement tokens. We revisit this principle and ask: which token-level teacher signals are actually learnable? Using a fixed-context diagnostic that measures same-context teacher-student KL reduction, we show that raw KL disagreement is a coarse proxy for learning value. It conflates learnable disagreement, where the teacher assigns corrective mass to the student's top-K candidates, with incompatible disagreement, where the teacher places mass mostly off the student's current support. We formalize this local compatibility as token teachability and show that it better predicts fixed-context improvement than raw KL alone. Motivated by this finding, we propose Teachability-Aware OPD (TA-OPD), a lightweight token-position selection method that applies OPD loss to high-teachability positions without reward models or verifiers. Across Qwen2.5 and Qwen 3 teacher-student settings, TA-OPD often surpasses full-token OPD with only 5% retained tokens and improves over entropy- and divergence-based baselines. Our results reframe selective OPD as selecting learnable teacher signals rather than merely salient tokens.

fields

cs.AI 1

years

2026 1

verdicts

UNVERDICTED 1

representative citing papers

DOPD: Dual On-policy Distillation

cs.AI · 2026-06-29 · unverdicted · novelty 5.0

DOPD is an advantage-aware dual distillation method that dynamically assigns token supervision from either privileged teacher or student to transfer capability while mitigating non-replicable information asymmetry in on-policy distillation.

citing papers explorer

Showing 1 of 1 citing paper.

  • DOPD: Dual On-policy Distillation cs.AI · 2026-06-29 · unverdicted · none · ref 38 · internal anchor

    DOPD is an advantage-aware dual distillation method that dynamically assigns token supervision from either privileged teacher or student to transfer capability while mitigating non-replicable information asymmetry in on-policy distillation.