On October 24, 2026, a new study titled “TutorCaseBench” was officially published on the arXiv platform (paper ID 2610.30828v1). This is the first large-scale benchmark for evaluating the tutoring capabilities of large language models (LLMs), specifically designed to assess students with learning and behavioral challenges, such as attention deficits, emotional regulation difficulties, and social communication challenges.
TutorCaseBench departs from the conventional educational assessment paradigm and instead grounds itself in qualitative case studies drawn from pedagogy, educational psychology, and special education. Through a systematic process, it translates literature‑based categories of student behavior into concrete teaching scenarios, samples of teacher–student dialogue, and empirically grounded scoring rubrics. The benchmark comprises 923 structured teacher–student dialogues, covering five typical types of challenging behaviors, and incorporates 1,560 open-ended response‑generation items along with 2,537 strategy‑selection questions for dual‑level evaluation.
In tests of 11 of the most advanced LLMs currently available, the top-performing model met its predefined evaluation criteria in only 55.8% of open-ended generation tasks; by contrast, its ability to recognize strategies in multiple-choice questions was significantly stronger. This gap highlights the widespread inability of LLMs to translate intervention principles into concrete, appropriate, and context‑specific responses, a limitation that is particularly pronounced in models with open-source weights. All data has been made publicly available on GitHub (https://github.com/kixlab/TutorCaseBench）。)