A new benchmark called BasketballBench has been introduced to evaluate a team's ability to understand a basketball game. The benchmark includes 7,980 questions across ten tasks in text, image, and video formats, using data from the 2025-2026 NBA season. It tests various skills such as recognizing events, localizing actions, identifying players, and relating these to structured game knowledge. A system called BasketballSkills has been proposed that can compose perception and retrieval tools to answer questions requiring multiple capabilities. Experiments show that current large language models struggle with these types of questions, but BasketballSkills outperforms them. --- Why it matters: This matters because it highlights the limitations of current AI systems in understanding complex tasks like basketball analysis, and demonstrates a more effective approach to integrating multiple skills for comprehensive understanding. Source: https://arxiv.org/abs/2608.23435