Tag
This paper proposes SDPG, a self-distilled policy-gradient framework that combines on-policy self-distillation with verifier advantages and KL regularization to improve reinforcement learning stability and performance.