Knowledge-Augmented Reasoning Distillation for Small Language Models in Knowledge-Intensive Tasks