Tag
This arXiv paper introduces CVPO, a reinforcement learning method for LLMs that adapts value-variance for advantage estimation and uses dynamic curriculum learning to match question difficulty, achieving better reasoning performance than VAPO on math tasks.