Preprint
Aug 2026
A Token-Level Analysis of Sampled-Token Reverse-KL On-Policy Distillation
A gradient-level characterization of reverse-KL OPD trained with the K2 estimator, with Surprise-aware Reweighting (SuRe), a detached, bounded weighting rule that further amplifies this existing allocation.
Bing Shao, Jia-Zheng Zhang, Long Ma et al.
· 8 citations