A new research paper introduces K12-KGraph, a curriculum-aligned knowledge graph designed to benchmark and train large language models (LLMs) for K-12 education. The work addresses a gap in existing benchmarks, which primarily test exam question answering rather than what the authors call curriculum cognition—the ability to understand how knowledge is structured and visually presented.
K12-KGraph is extracted from official People's Education Press textbooks covering mathematics, physics, chemistry, and biology across primary, middle, and high school. It includes nine node types and fourteen relation types that capture prerequisite chains, concept taxonomies, experiment-concept links, pedagogical sequencing, and visual grounding.
From this graph, the team derived K12-Bench, a multi-select benchmark with 23,640 questions across five task families: Ground, Prereq, Neighbor, Evidence, and Locate. They also built K12-Train, a graph-guided supervised fine-tuning corpus of 7,335 samples, including 2,267 text-only QA pairs and 5,068 multimodal VQA pairs.
Evaluation results show that even advanced models struggle: Gemini-3-Flash achieves only 57% exact match, and Gemma-4-31B-IT reaches 46%, with Prereq and Neighbor being the hardest tasks. Training experiments demonstrate that domain-specific supervision can reduce this gap. Under a matched budget of 2,300 samples, K12-Train-Text consistently outperforms equally sized subsets of eight mainstream instruction-tuning corpora on GaokaoBench and EduEval.
For vision-language models, K12-Train-Full achieves the best overall results on Gaokao-MM, MDK12-medium, and K12Vista among all compared training configurations, despite using fewer samples than full DataFlow and WizardLM baselines. It also surpasses both text-only and multimodal-only variants, showing that textual and visual supervision are complementary.
The graph, benchmark, training data, and construction pipeline are all open-sourced. The updated v2 includes 9 node types, 14 relation types, 23,640 benchmark questions, and 7,335 training samples (including 5,068 multimodal VQA samples). The authors hope K12-KGraph will support further research on educational and multimodal LLMs.