Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation
About
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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Scientific Question Answering | GPQA Diamond | Accuracy47.47 | 123 | |
| Mathematical Reasoning | MATH-500 1 (test) | Accuracy87.4 | 38 | |
| Code Generation | HumanEval v1 (test) | Accuracy79.88 | 37 | |
| Instruction Following | IFEval v1 (test) | Accuracy78.93 | 28 | |
| Mathematical Reasoning | AIME 2024 1 | Accuracy30 | 24 | |
| Mathematical Reasoning | AIME 2025 1 | Accuracy25 | 24 |