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

The paper introduces Token Teachability to separate learnable teacher corrections from incompatible disagreements in on-policy distillation.

It proposes TA-OPD, which selects only high-teachability token positions instead of training on all token-level differences.

On Qwen2.5 and Qwen3, TA-OPD often matched or beat full-token OPD while keeping only about 5% of tokens.

The method improves efficiency without using reward models or verifiers, suggesting better student learning from fewer teacher signals.