Research map / Large language models / Retrieval and knowledge

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

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds K12-KGraph, a curriculum knowledge graph from Chinese K-12 textbooks, and uses it to derive K12-Bench (23,640 multi-select items for curriculum cognition) and K12-Train (7,335 KG-guided SFT samples).

Problem

Existing K-12 LLM benchmarks measure factual answering but not curriculum cognition (prerequisite chains, concept taxonomies, experiment-concept links, textbook location), and training data does not teach this structure.

Approach

Extract a heterogeneous textual and multimodal knowledge graph from official PEP textbooks with LLM schema-guided extraction and DAG validation, then generate graph-derived multi-select benchmark items and KG-grounded QA training pairs from the same graph.

Key results

• Best overall exact match on K12-Bench is 57.1% (Gemini-3-Flash); best open-source Gemma-4-31B-IT reaches 46.4%.
• On GaokaoBench with Qwen3-4B-Base, K12-Train-Text reaches 1009.96 total vs 985.91 for the strongest baseline DataFlow (+24.1).
• On Gaokao-MM, K12-Train-Full reaches 39.9% overall vs 36.1% for Qwen3.5-2B (Instruct).

Surprising findings

• LLaMA-3-8B-Instruct (EM 7.2%) is near random guessing (6.7%) on K12-Bench.
• K12-Train-Text improves Chinese and humanities subjects it has no in-domain data for, suggesting transfer of answer style.

Limitations

• Benchmark and training data depend on the correctness of LLM-extracted KG; relies on GPT-5.2 extraction and filtering.
• Coverage limited to PEP textbooks and four subjects (math, physics, chemistry, biology).

Experimental setup

Datasets: K12-KGraph (PEP textbooks) OpenHermes-2.5 Infinity-Instruct UltraChat WizardLM_evol_instruct_V2_196k DataFlow-10K-Instruct LMSYS SmolTalk Tulu-3-SFT

Benchmarks: K12-Bench GaokaoBench EduEval Gaokao-MM MDK12-Bench K12Vista

Metrics: exact match instance-level macro F1 accuracy score totals

Baselines: OpenHermes Infinity-Instruct UltraChat WizardLM DataFlow LMSYS SmolTalk Tulu-3

Builds on: MinerU LLaMA-Factory C-Eval CMMLU GaokaoBench EduEval

Closest papers by meaning

• CK12: A Rounded K12 Knowledge Graph Based Benchmark for Chinese Holistic Cognition Evaluation (AAAI 2024 · similarity 0.88)
• The Signal is in the Steps: Local Scoring for Reasoning Data Selection (ICML 2026 · similarity 0.86)
• Improving Complex Reasoning over Knowledge Graph with Logic-Aware Curriculum Tuning (AAAI 2025 · similarity 0.86)
• CurBench: Curriculum Learning Benchmark (ICML 2024 · similarity 0.86)
• Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.85)
• SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models (ICML 2024 · similarity 0.85)