Off-Policy Distillation
A model distillation approach in which training data is generated offline by the teacher model before training begins. The student model learns from a fixed dataset. This approach is contrasted with on-policy distillation.
Timeline (last 90 days)
Off-policy distillation directly uses trajectories generated by the teacher; the student is exposed to the teacher's distribution, which deviates from its own test-time behavior and can easily cause forgetting or degraded generalization.
On held-out test sets, SSD consistently outperforms off-policy distillation.
All Facts (2)
Off-policy distillation directly uses trajectories generated by the teacher; the student is exposed to the teacher's distribution, which deviates from its own test-time behavior and can easily cause forgetting or degraded generalization.
50%UnverifiedOn held-out test sets, SSD consistently outperforms off-policy distillation.
50%