Tag
This paper introduces DRPO, which replaces the hard mask in DPPO with a smooth advantage-weighted quadratic regularizer to improve stability and efficiency in LLM reinforcement learning by providing continuous gradient corrections beyond trust-region boundaries.