Skip to content- Sources: primary
- Summary: The authors define entropy shape, the sequence of next-token entropies, and an Entropy Shape Correlation between student and teacher. Interventions on teacher entropy that preserve token preferences show that where to branch causally drives OPD gains, and pre-OPD ESC picks the best teacher better than existing predictors. Code is released.
- Why it matters: Teams post-training with on-policy distillation get a cheap pre-training check for choosing a teacher instead of assuming the strongest model helps most.