Most existing educational benchmarks evaluate exam-style Q&A but fail to measure how models understand curriculum structure and visual grounding — what this paper calls "curriculum cognition." That gap matters because real classroom tasks require chaining prerequisites, navigating concept taxonomies, linking experiments to concepts, sequencing pedagogy, and grounding content in textbook visuals. K12-KGraph attacks that problem by extracting a nine-node-type, fourteen-relation knowledge graph from official People's Education Press mathematics, physics, chemistry, and biology textbooks across primary, middle, and high school and deriving both benchmarks and training data from it.
Key Findings
- The project produces K12-Bench (23,640 multi-select questions across five task families: Ground, Prereq, Neighbor, Evidence, Locate) and K12-Train (7,335 supervised samples: 2,267 text QA + 5,068 multimodal VQA). This couples structured curriculum knowledge with visual grounding at scale.
- Off-the-shelf large models struggle: Gemini-3-Flash achieves 57% exact match and Gemma-4-31B-IT reaches 46% on K12-Bench, with Prereq and Neighbor tasks being the hardest, indicating weakness in modeling dependency chains and local graph neighborhoods.
- Domain-specific supervision helps: under a matched 2,300-sample budget, K12-Train-Text outperforms equal-sized subsets of eight mainstream instruction-tuning corpora on GaokaoBench and EduEval. For vision-language models, K12-Train-Full yields top results on Gaokao-MM, MDK12-medium, and K12Vista versus larger baselines, showing text+visual supervision is complementary.
- The authors release the knowledge graph, benchmark, training corpora, multimodal pairs, and the full pipeline to enable reproducible research and dataset-driven fine-tuning.
Who it's for and trade-offs
Great fit if you are building or evaluating LLMs for K–12 educational applications, studying curriculum-aware reasoning, or training multimodal models that must ground textbook visuals. It is also useful for researchers comparing curriculum-structure capabilities across models. Look elsewhere if your focus is generic open-domain QA, conversational tutoring at scale, or non-Chinese curricula: the graph is extracted from People's Education Press textbooks and reflects Chinese K–12 sequencing and visual styles, so transfer to other curricula will require adaptation. The dataset emphasizes structured curriculum relations and textbook visuals rather than long-form pedagogical dialogue or classroom interaction logs.