Rethinking OPD Exposes Distillation’s Hidden Failure Modes

The key shift is from ranking teachers by benchmark score to measuring whether their token-level behavior is learnable at states the student actually visits. OPD is less a magic post-training recipe than a compatibility-and-coverage engineering problem.

• –Successful runs progressively align high-probability tokens, with a small shared token set capturing 97–99% of probability mass.
• –A stronger teacher can fail when its reasoning distribution is incompatible with the student’s; same-family models may appear nearly indistinguishable from the student’s perspective.
• –Teacher-aligned prompt selection gives practitioners a practical way to expose the student to learnable, high-value states.
• –The repository includes implementations built around verl and LlamaFactory, plus diagnostics for measuring token overlap.
• –The dense token-level signal is powerful but may become expensive or ineffective for long-horizon reasoning traces.