Self-Training Large Language Models with Confident Reasoning