Imagine a world where AI can seamlessly navigate complex datasets, effortlessly connecting the dots between different types of information. This is the vision behind CMDBench, a groundbreaking benchmark designed to evaluate how AI systems discover and utilize data from diverse sources like text, tables, and knowledge graphs. Why is this so important? Because real-world data isn't neatly organized. It lives in silos, scattered across different formats and systems. Think about a sports analytics company. Teams focusing on basketball, football, or soccer each have their own data sources, from player stats to historical articles and even relationship graphs of players and teams. CMDBench mimics this complexity, using the world of basketball as a testing ground. Researchers have created a dataset combining Wikipedia articles, statistical tables, and a knowledge graph extracted from Wikidata, all related to the NBA. This allows them to test how well AI systems can find the right data to answer complex questions like, "Which team drafted the tallest players on average for the guard position over the years?" This requires the AI to not only understand the question but also figure out which data sources are relevant and how to combine information from them. Initial experiments using CMDBench have revealed a significant challenge: even the most advanced AI models struggle when faced with this real-world complexity. There's a noticeable drop in accuracy when AI has to discover the data compared to when it's simply given the relevant information. This highlights the need for better data discovery methods. CMDBench is a crucial step towards unlocking the full potential of AI in data-rich environments. By providing a standardized way to evaluate these systems, it paves the way for more sophisticated AI assistants that can truly understand and analyze the complex world around us.