K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
The paper says curriculum cognition is still a weak spot for education-focused LLMs.
K12-KGraph is built from official People’s Education Press textbooks across math, physics, chemistry, and biology. The authors derive a 23,640-question benchmark and a 7,335-sample training set from the graph, including multimodal VQA. On K12-Bench, Gemini-3-Flash reaches 57% exact match and Gemma-4-31B-IT reaches 46%, with prerequisite and neighbor tasks hardest. Their fine-tuning results suggest curriculum-specific text and visual supervision improves educational model performance. HF Daily Papers' note
K12-KGraph is built from official People’s Education Press textbooks across math, physics, chemistry, and biology. The authors derive a 23,640-question benchmark and a 7,335-sample training set from the graph, including multimodal VQA. On K12-Bench, Gemini-3-Flash reaches 57% exact match and Gemma-4-31B-IT reaches 46%, with prerequisite and neighbor tasks hardest. Their fine-tuning results suggest curriculum-specific text and visual supervision improves educational model performance. HF Daily Papers' note
score 4