Endpoint Distillation
A policy distillation method that uses only initial and terminal checkpoints for finite-difference approximation, and thus cannot capture the non-monotonic changes in policy shift during training.
Timeline (last 90 days)
The endpoint-style policy distillation method uses the difference in log probabilities of the teacher model before and after RL training as the supervision signal, applied to trajectories generated by the student model itself.
Endpoint-style distillation cannot distinguish between "a sustained consistent positive policy shift" and "a sharp early increase followed by a later decline," because their endpoint differences may be exactly the same.
All Facts (2)
The endpoint-style policy distillation method uses the difference in log probabilities of the teacher model before and after RL training as the supervision signal, applied to trajectories generated by the student model itself.
50%UnverifiedEndpoint-style distillation cannot distinguish between "a sustained consistent positive policy shift" and "a sharp early increase followed by a later decline," because their endpoint differences may be exactly the same.
50%