Superficial Self-Improved Reasoners Benefit from Model Merging