Beyond Labels: Aligning Large Language Models with Human-like Reasoning