🧠 AI Ed Wiki

📄 Research Article

ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models

● high · 2026-08-13benchmark, llm-evaluation, ai-ed-evaluation, pedagogical-safety, llm, safety, generative-ai

Synthesis: Jiang et al. (2026) introduce ELBench, the first benchmark to evaluate education-facing LLMs on all four required dimensions — General Capability, Safety and Trustworthiness, Basic Education, and High-Level Cultivation — under a common protocol, combining curated public sources with newly synthesized safety and cultivation data. Testing nine models, they find module-level profiles are more informative than a single aggregate: the top six models are statistically indistinguishable overall yet differ substantially by module leader, and safety is anti-correlated with practical teaching (r = −0.83). The two education-specialized models lead neither education module, and all models share a systematic blind spot on High-Level Cultivation's structured-judgment task. The work connects to Benchmark, AI Ed Evaluation, and AI Ed Evaluation frameworks.

An Integrated Profile, Not a Single Score

A usable education-facing model must 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; ELBench is the first to assess all four as an integrated profile under a common protocol.

Three Findings

  • Module-level profiles beat a single aggregate: the top six models are statistically indistinguishable on overall score, yet their module leaders differ substantially; safety is anti-correlated with practical teaching (r = −0.83).
  • Chinese-developed models lead the safety module, the most discriminative in the suite; the advantage is largest on region-specific normative content and narrows (but does not vanish) on universal-harm content.
  • Education-specialized models lead neither education module, and on High-Level Cultivation all models converge on the same non-reference option on the structured-judgment task, favoring pedagogical style over fit to the stated goal — a systematic blind spot.

    • Connected Concepts

  • Benchmark
  • AI Ed Evaluation
  • AI Ed Evaluation
  • Pedagogical Safety
  • Pedagogical Safety
  • Generative AI
  • LLM
  • AI Tutoring
  • Research Methods AIED
  • Educational Measurement

    • Connected Articles

  • Teachbench LLM Teaching Evaluation
  • Eduagentbench Agent Teaching Benchmark
  • AI Tutor Behavioral Evaluation
  • Solving Vs Evaluating GenAI Solutions
  • Nsmq Riddles Science Math Benchmark
  • Academiclaw Student Agent Benchmark
  • Knowledge Distillation AI Tutor Evaluation
  • Eduframetrap LLM Sycophancy Educational Safety

    • Citation

    Jiang, Y., Zhu, X., Tan, F., Zhang, Z., Huang, K., Yu, Y., Fei, Z., Luo, Y., Li, K., Hao, H., Zhai, G., & Zhou, A. (2026). ELBench: A multi-dimensional benchmark for education-facing large language models. arXiv:2608.09548.