Imagine a world where grading math homework isn't a tedious chore but a seamless, accurate process, even for those tricky 'edge cases' where students get the right answer in unexpected ways. That's the promise of new research using large language models (LLMs) to revolutionize how we assess student learning. Researchers explored the challenges of grading open-response math questions, particularly those that stump traditional rule-based systems. Using a new dataset called AMMORE, gathered from a real-world math tutoring platform used by students in several African countries, they put LLMs to the test. AMMORE offers a unique peek into how students learn math in diverse contexts, providing valuable insights into the nuances of math acquisition. The researchers experimented with various LLM-driven grading methods, including zero-shot, few-shot, and the innovative 'chain-of-thought' prompting. This last technique proved particularly effective, achieving remarkable accuracy even with those tricky edge cases. Why does this matter? Because misgrading a correct answer can frustrate students and hinder their progress. These LLMs, by understanding the logic behind a student's solution, even if expressed unconventionally, offer a more nuanced and encouraging approach to grading. The impact goes beyond individual questions. When integrated into intelligent tutoring systems, these accurate LLM graders can significantly improve how we estimate student mastery of concepts, allowing for more personalized learning experiences. This research is a step toward more widespread use of open-ended questions in math assessments. By automating the grading process, we can unlock the rich insights these questions provide without overburdening educators, ultimately leading to more effective learning and a better understanding of how students grasp mathematical concepts.