EduIllustrate: Towards Scalable Automated Generation Of Multimodal Educational Content

A research team led by Shuzhen Bi, Mingzi Zhang, and Keqian Li has introduced EduIllustrate, a new benchmark designed to rigorously evaluate how well large language models (LLMs) can generate multimodal educational content. The benchmark focuses on a critical gap in AI-assisted education: the ability to produce coherent, diagram-rich explanations for K-12 STEM subjects. It comprises 230 problems across five subjects and three grade levels, challenging models to interleave step-by-step reasoning with geometrically accurate visuals. The evaluation of ten leading LLMs revealed a significant performance spread, with Google's Gemini 3.0 Pro Preview achieving the top score of 87.8% across an 8-dimension rubric grounded in multimedia learning theory. Notably, Kimi-K2.5 emerged as the most cost-efficient model, delivering 80.8% accuracy at just $0.12 per problem.

The study also introduced and validated a novel 'sequential anchoring' protocol, a generation method that enforces visual consistency across multiple diagrams in a single explanation. Workflow ablation tests confirmed this technique improves the Visual Consistency dimension by 13% while reducing costs by a remarkable 94%. Human evaluation with 20 expert raters validated that LLM-as-judge scoring is reliable for objective quality dimensions (correlation ρ ≥ 0.83), though it highlighted limitations for more subjective visual assessments. This benchmark shifts the focus from simple question-answering to the complex task of instructional content creation, providing a standardized tool to measure progress in a key application area for generative AI.