AI summaryⓘ
The authors created ELBench, a new test to check how well language models work as educational tools by measuring accuracy, safety, usefulness, and alignment with teaching goals all at once. They tested nine models and found that no single overall score tells the full story; different models excel in different areas, and being very safe can sometimes mean being less practical for teaching. Chinese models stood out in safety, especially with local content, while education-specific models didn't lead in educational tasks. All models struggled with a high-level task that requires judgment, often prioritizing teaching style over actual goals. This suggests that improving education-focused models is still a challenge compared to general ones.
large language modelsbenchmarkeducation technologymodel safetyinstructional alignmentevaluation metricspedagogical goalsmodel assessmenthigh-level cultivationdomain post-training
Authors
Yilin Jiang, Xiaorong Zhu, Fei Tan, Zicheng Zhang, Kaiyi Huang, Yang Yu, Zexuan Fei, Yiming Luo, Keqian Li, Hao Hao, Aimin Zhou, Guangtao Zhai
Abstract
Large language models are increasingly deployed in education as tutors, teaching assistants, and content generators. These roles place demands that ordinary question answering does not: a usable education-facing model is supposed to be accurate, safe under sensitive prompts, instructionally useful, and aligned with pedagogical goals at the same time. Existing benchmarks evaluate these requirements largely in isolation, so none assesses education-facing suitability as an integrated profile. We introduce ELBench, the first benchmark to evaluate all four requirements (General Capability, Safety and Trustworthiness, Basic Education, and High-Level Cultivation) on the same models under a common protocol, combining curated public sources with newly synthesized safety and cultivation data. We evaluate nine models, seven frontier general-purpose systems and two education-specialized variants, and report three findings. First, module-level profiles are more informative than a single aggregate: the top six models are statistically indistinguishable on overall score, yet their module leaders differ substantially, and safety is anti-correlated with practical teaching (r = -0.83). Second, the Chinese-developed models lead the safety module, the most discriminative in the suite; this advantage is largest on region-specific normative content and narrows, but does not vanish, on universal-harm content. Third, the two education-specialized models lead neither education module, and on High-Level Cultivation all models share a systematic blind spot: on the structured judgment task they converge on the same non-reference option, favoring pedagogical style over fit to the stated goal, so the module scores uniformly low and does not separate models. This raises, but does not resolve, whether domain post-training keeps pace with frontier systems on education tasks.