K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

K12-KGraph is a new curriculum-aligned knowledge graph extracted from official Chinese K–12 textbooks, covering mathematics, physics, chemistry, and biology. It features seven node types and nine relation types, designed to capture structured curriculum understanding beyond factual recall. From this graph, researchers derived K12-Bench, a 23,640-question multi-select benchmark probing five task families of curriculum cognition, and K12-Train, a ~2,300 QA pair supervised fine-tuning corpus. Experiments on K12-Bench revealed that even strong models like Gemini-3-Flash achieved only 57% exact match, and Gemma-4-31B-IT reached 46%, with prerequisite and neighbor tasks being the hardest. Conversely, K12-Train, despite its modest size, consistently outperformed eight mainstream instruction-tuning corpora, improving Qwen3-4B-Base by +24.1 and Llama3.1-8B-Base by +32.4 on GaokaoBench, demonstrating its sample-efficiency for educational LLM training.

Key takeaway

For AI scientists and ML engineers developing educational LLMs, you should prioritize integrating curriculum-aligned knowledge graphs. Current models lack structural understanding, as shown by low K12-Bench scores. By utilizing resources like K12-KGraph and its derived training data, you can significantly improve model performance on complex educational tasks, even with modest data budgets. Focus on teaching relational knowledge to build more effective and pedagogically coherent AI tutors.

Key insights

Curriculum-aligned knowledge graphs are crucial for developing LLMs with structural educational understanding beyond factual recall.

Principles

• LLMs struggle with structural curriculum cognition.
• KG-guided SFT is highly sample-efficient.
• Educational AI needs curriculum cognition.

Method

A five-stage pipeline extracts K–12 textbook content via OCR, segments it, uses LLMs for schema-guided node/edge extraction, merges graphs hierarchically, and validates DAGs.

In practice

• Use K12-KGraph for curriculum-aware LLM training.
• Benchmark LLMs on structural understanding tasks.
• Synthesize QA pairs from KG relations.

Topics

• Knowledge Graphs
• Educational LLMs
• Curriculum Cognition
• Supervised Fine-tuning
• K12-KGraph
• K12-Bench

Code references

• haolpku/K12-Dataset
• TapXWorld/ChinaTextbook

Best for: AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• What Makes Knowledge Graphs AI-Ready — Knowledge Graphs for AI bridge siloed data semantics with organizational ontology for machine-readable reasoning.
• Search-on-Graph-R1: Training Large Language Models to Search Knowledge Graphs with Reinforcement Learning — Internalizing KG navigation into compact LLMs via SFT and RL outperforms frontier models, reducing cost and complexity.
• KGCQual: An Interpretable Framework for Evaluating the Knowledge Graph Construction Quality from Text — KGCQual offers an interpretable, model-agnostic framework for intrinsically evaluating automated Knowledge Graph construction quality by comparing extracted graphs to an "ideal" text representation.
• Forget GraphRAG: A 4B AI does the work NOW — Agent-native knowledge orchestration via K1's 4B LLM and multimodal graph overcomes GraphRAG limitations for scientific reasoning.
• BLINKG: A Benchmark for LLM-Integrated Knowledge Graph Generation — BLINKG benchmarks LLM capabilities in mapping diverse data to ontologies for Knowledge Graph generation.
• Sharing all KGC 2026 decks. More production-grade KG systems than I've seen at any conference. [D] — Knowledge graphs are maturing into critical production infrastructure for reasoning and compliance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.