Imagine an AI tackling complex math problems, not just equations, but word problems with graphs and diagrams—all in Chinese. Researchers have created CMM-Math, a massive dataset of over 28,000 Chinese math problems spanning elementary to high school levels, to evaluate and push the boundaries of AI's mathematical reasoning. Unlike traditional text-based math datasets, CMM-Math incorporates multiple images within questions and answers, mirroring the visual challenges of real-world tests. This multimodal approach demands AI models not just to understand language but also to interpret visual information and logically connect concepts, like geometry or graph theory. Initial tests with state-of-the-art multimodal AI models, including the likes of GPT-4V, show a significant struggle with the complexities of CMM-Math. While these AI excel at elementary arithmetic and basic statistics, they stumble when faced with geometry and higher-level math concepts. Even few-shot prompting, a technique to guide AI with example solutions, offers limited improvement. The findings highlight a glaring gap in existing AI capabilities, particularly in handling visually rich, complex mathematical reasoning. To tackle this challenge, the researchers developed Math-LMM, a specifically designed LLM trained in three stages: foundational pre-training (aligning visual data with language), foundational fine-tuning (learning general problem-solving), and mathematical fine-tuning (focused training on complex math problems). The results are promising, with Math-LMM outperforming existing open-source models, showcasing its improved ability to reason mathematically with multiple visual inputs. However, more research is needed to bridge the gap between human-like mathematical thinking and current AI capabilities. The CMM-Math dataset provides an important benchmark for future development, pushing towards AI that can not only ‘see' but also ‘understand’ and ‘solve' complex math problems in diverse languages and contexts.