The dream of personalized AI tutors has captivated educators and tech enthusiasts alike. Imagine a world where every student has access to a tireless, infinitely patient virtual guide, tailored to their individual learning style and pace. But how close are we to realizing this vision? A new research paper, "Unifying AI Tutor Evaluation," delves into this question by examining the pedagogical abilities of state-of-the-art large language models (LLMs) when they're tasked with helping students overcome mistakes in mathematics. The researchers have developed a comprehensive evaluation framework based on core learning science principles to assess how well LLMs can truly tutor. They’ve tested powerful models like GPT-4, Gemini, and others, prompting them to act as expert tutors in conversations where students exhibit confusion or make errors. The results reveal a complex picture. While some LLMs excel at identifying mistakes and pinpointing their location, they often fall short in providing effective guidance without simply revealing the answer. This is a critical distinction between a question-answering system and a true tutor. A good tutor scaffolds learning, offering hints and explanations that encourage students to actively participate in the process of understanding. The study shows that even the most advanced LLMs sometimes struggle with this nuanced pedagogical approach. Interestingly, the research also highlights the shortcomings of human novice tutors, whose feedback often lacks the clarity and actionability needed to guide students effectively. This underscores the difficulty of teaching itself, regardless of whether the tutor is human or AI. While the vision of universally accessible, effective AI tutors remains a work in progress, this research provides a valuable roadmap for future development. By identifying the specific areas where LLMs need to improve, it paves the way for training AI tutors that can truly understand, guide, and inspire students.