TL;DR of the paper, generated automatically.

On-policy distillation looks near-perfect in-distribution but misses generalization.

• Gap: out-of-distribution scores sit 2.45-16.25% below the base model.
• Compositional: composed tasks land 18.19-24.34% below the best teachers.
• Fixes: the authors propose geometry-based patching and base-preserving replay.