LLMs are bad at counting things just in general. It’s hard to say whether the failures here are vision based or just an inherent weakness of the language model.