Large Language Models (LLMs) are powerful tools for natural language processing tasks, but their capabilities in formal reasoning, especially mathematical logic, are still limited. This study focuses on the phenomenon of hallucinations in LLMs—errors where the model produces illogical or fabricated outputs—in the context of mathematical reasoning. By systematically categorizing and analyzing errors across multiple mathematical domains, we propose a framework for detecting and mitigating such hallucinations. Experimental results reveal common patterns of failure, while suggested interventions, such as integrating external theorem provers and refining datasets, demonstrate potential for improving logical accuracy.