performance-saturation

This paper introduces Entrocraft, a rejection-sampling method for RL that controls entropy schedules to prevent performance saturation in LLMs. It demonstrates improved generalization and training longevity, allowing smaller models to outperform larger baselines.