Imagine an AI that could effortlessly summarize dense, equation-filled scientific papers. That's the ambitious goal researchers tackled by creating "derivation graphs." These graphs map the relationships between equations in a paper, showing how one equation leads to another, like a roadmap of mathematical reasoning. To test this idea, they hand-labeled the equation dependencies in 107 STEM papers from arXiv, creating a dataset of these derivation graphs. They then challenged several algorithms—from simple text analysis to cutting-edge Large Language Models (LLMs) like Google's Gemini—to reconstruct these graphs automatically. The results? While both LLMs and a basic "brute force" search for explicit textual references between equations achieved decent accuracy (around 90%), they struggled with precision. This means they often identified spurious connections where no real mathematical derivation existed. The best performing models only correctly identified about half of the actual derivation relationships (F1 score of ~48%). This study reveals that even the most advanced AI still has a long way to go in truly understanding the complex web of reasoning in scientific literature. It highlights the need for better methods to represent mathematical knowledge and the unique challenges of deciphering the language of mathematics. Could future AI tools act as automated research assistants, quickly summarizing key findings and suggesting new avenues of inquiry? This research takes a promising step towards that exciting possibility, while also revealing how much we have yet to learn about bridging the gap between human and machine understanding of math.