Large Language Models and Mathematical Reasoning Failures