NVIDIA Research reports that Semi-OPD outperformed OPD in 14 of 17 tested teacher–student pairs, across models from 1.5B to 235B parameters, and reports up to 13.6% higher accuracy and 11.4x faster training.
The mechanism is a mismatch between where examples come from and what the teacher can effectively supervise. For misaligned pairs, student rollouts can become increasingly off-policy relative to the teacher as context grows; Semi-OPD uses offline initial-student rollouts and shorter contexts while covering full trajectories. The authors identify output-token overlap as a measure of initial alignment and say OPD is beneficial only when overlap is high.
Why now: the paper supplies a concrete selection criterion for a training choice often framed as a general preference for on-policy data. Measure teacher–student output-token overlap before committing to OPD. The evidence boundary is that the reported comparison covers 17 pairs and the study's own training setups; it does not establish a universal winner across tasks, datasets, or stronger OPD variants. Whether follow-up work validates overlap as a reliable selector on broader teacher–student pairs and settings is the key watch condition. The source is NVIDIA Research: https://research.nvidia.com/publication/2026-12_when-do-we-need-policy-distillation-distilling-offline-student-rollouts-often