Reflective Self-Distillation

A technique combining knowledge distillation with self-training. Through explicit optimization, the model autonomously produces improved versions of its outputs, thereby internalizing reflective signals and continuously improving generation quality.