Less is More: Selective Reflection for Compatible and Efficient Knowledge Distillation in Large Language Models | Synapse