Towards Comprehensive Basketball Understanding
BasketballBench tests whether models can connect clips, players, events, and game records in one basketball reasoning task.
The paper introduces a 7,980-question multimodal benchmark built from the 2025-2026 NBA season. It combines text, image, and video tasks with official play-by-play data, rosters, player profiles, and 2,501 possession-level broadcast clips. The authors also propose BasketballSkills, an agent that sequences eight basketball-specific perception and retrieval tools. In their experiments, current multimodal LLMs struggle most when several abilities have to be integrated, while BasketballSkills performs better. ArXiv · AI/CL/LG's note
The paper introduces a 7,980-question multimodal benchmark built from the 2025-2026 NBA season. It combines text, image, and video tasks with official play-by-play data, rosters, player profiles, and 2,501 possession-level broadcast clips. The authors also propose BasketballSkills, an agent that sequences eight basketball-specific perception and retrieval tools. In their experiments, current multimodal LLMs struggle most when several abilities have to be integrated, while BasketballSkills performs better. ArXiv · AI/CL/LG's note
score 4