Imagine asking an AI assistant a seemingly simple question: "If most people in a town own a car, and a small number take the bus, how many people use both?" Sounds easy enough, right? But for large language models (LLMs), these “fuzzy” concepts like "most" and "few" pose a surprisingly tough challenge. A new benchmark called FRoG (Fuzzy Reasoning of Generalized Quantifiers) is putting LLMs to the test, revealing just how much they struggle with this type of reasoning. FRoG uses real-world math problems, tweaking them to include generalized quantifiers instead of exact numbers. So, instead of saying "20% of people," the problem might say "a small number of people." This forces the LLM to not only perform the math but also interpret the meaning of vague quantifiers. The results? Even the most advanced LLMs are stumbling. Many show an "inverse scaling" effect, where bigger models, surprisingly, perform worse than smaller ones. Traditional methods for boosting reasoning, like code or math-specific training, don't seem to help much either. The FRoG research shows a fascinating gap in current AI capabilities. While LLMs excel at precise calculations and complex language tasks, they're still learning to navigate the ambiguity of human language. This has real-world implications. Think of AI assistants handling nuanced customer requests or medical diagnoses dealing with uncertain symptoms—fuzzy reasoning is critical. The FRoG benchmark highlights a key area for future AI development: building models that understand not only what we say but also what we mean.