K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
Hao Liang, Qihan Lin, Zhaoyang Han, Xiaochen Ma et al. · arXiv · 2026
Abstract
Most education benchmarks ask whether a model can answer exam questions. This one asks whether it grasps how a curriculum is structured — prerequisite chains, concept hierarchies, the link between an experiment and the idea it demonstrates, and where a concept appears in a textbook diagram. The authors extract a knowledge graph from official school textbooks, derive a 23,640-question benchmark from it, and find leading models scoring near or below sixty percent.
Why it matters
A good demonstration that a benchmark score answers only the question the benchmark asks. Models that ace exam items still fumble the structure underneath them — a real gap for anyone building tutoring products.
https://arxiv.org/abs/2605.09635