Found 1 document(s)
Page 1 of 1
-
Reverse Kullback-Leibler (KL) distillation has shown empirical promise for training language models, but its statistical guarantees and limits remain poorly understood. In particular, the effects of two design choices re...