Mathematical reasoning is a cornerstone of human intelligence, and a crucial frontier for AI. While large language models (LLMs) excel at language tasks, math presents a unique challenge. A new research survey explores the burgeoning field of mathematical reasoning in multimodal LLMs (MLLMs), revealing both exciting progress and persistent hurdles. The survey, covering over 200 studies, dissects how MLLMs tackle math problems. It examines benchmarks used to test their abilities, the various methodologies employed to enhance their mathematical prowess, and the key challenges that still stand in the way of AI achieving true mathematical understanding. The study categorizes how MLLMs are used in mathematical reasoning into three main paradigms: as Reasoners, directly tackling problems; as Enhancers, improving data and refining solutions; and as Planners, coordinating the problem-solving process. While LLMs as Reasoners are the most common, the Planner role, though less explored, holds significant promise, particularly as multi-agent AI systems advance. Recent models have begun incorporating multimodal approaches, tackling geometry problems with diagrams and even incorporating visual elements like sketches, demonstrating a shift toward more holistic mathematical reasoning. However, significant challenges remain. MLLMs still struggle with complex visual reasoning, such as interpreting 3D shapes or understanding the nuances of diagrams. They often lack the ability to generalize across different mathematical domains, excelling in algebra, for example, but faltering in geometry. Furthermore, providing effective error feedback remains a major obstacle, hindering the models' learning process. And importantly, current approaches often don't reflect real-world educational practices, like the use of draft work and handwritten notes, limiting their practical application in educational settings. The survey highlights the need for richer, more diverse datasets that incorporate various modalities beyond just text and images. Future research must address the limitations in visual reasoning and domain generalization, and develop better methods for providing feedback and integrating with real-world educational needs. As AI continues to evolve, the ability to reason mathematically will be a key indicator of its true potential, unlocking new possibilities in education, science, and beyond.