From EduVisBench to EduVisAgent: A Benchmark and Multi-Agent Framework for Reasoning-Driven Pedagogical Visualization

Foundation models (FMs) have shown promise in educational contexts, yet struggle to generate pedagogically effective visual explanations for complex reasoning tasks. Current approaches focus primarily on textual outputs, overlooking the critical role of structured visualizations in supporting conceptual understanding. To address this gap, we introduce EduVisBench, a multi-domain benchmark featuring 1154 STEM problems requiring visually grounded solutions, evaluated using a fine-grained rubric informed by pedagogical theory. Our analysis reveals that existing models frequently fail to decompose complex reasoning into interpretable visual representations aligned with human cognitive processes. To overcome these limitations, we propose EduVisAgent, a multi-agent collaborative framework coordinating specialized agents for instructional planning, reasoning decomposition, metacognitive prompting, and visualization design. Experimental results demonstrate that EduVisAgent substantially outperforms all baselines, achieving a 40.2\% improvement in generating educationally effective visualizations.