Self-Distilled Policy Gradient

In three linesSDPG combines policy self-distillation with group-relative verifier advantages and KL regularization. The method uses full-vocabulary reverse KL divergence to supervise language model generations. Code available on GitHub.

Summary generated by Claude — human-verified