Bridging the Early Science Gap with Artificial Intelligence: Evaluating Large Language Models as Tools for Early Childhood Science Education

2025-01-02

AI Summary

This study addresses the challenge of adapting large language models (LLMs) for preschool science education. It presents the first interdisciplinary (biology, chemistry, physics) empirical evaluation targeting children aged 3–6. Grounded in developmental psychology and early science education standards, we developed a multidimensional evaluation framework assessing accuracy, comprehensibility, engagement, and developmental appropriateness. Thirty early childhood educators conducted mixed qualitative and quantitative assessments of age-appropriate explanations generated by GPT-4, Claude, Gemini, and Llama. Results reveal that Claude significantly outperforms others in biology explanations, and model performance strongly correlates with disciplinary abstraction level; systematic cross-model and cross-dimension variations are observed. This work establishes the first empirically grounded benchmark for LLMs in preschool science education and provides actionable guidance on prompt engineering and pedagogical integration.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Early childhood science education is crucial for developing scientific literacy, yet translating complex scientific concepts into age-appropriate content remains challenging for educators. Our study evaluates four leading Large Language Models (LLMs) - GPT-4, Claude, Gemini, and Llama - on their ability to generate preschool-appropriate scientific explanations across biology, chemistry, and physics. Through systematic evaluation by 30 nursery teachers using established pedagogical criteria, we identify significant differences in the models' capabilities to create engaging, accurate, and developmentally appropriate content. Unexpectedly, Claude outperformed other models, particularly in biological topics, while all LLMs struggled with abstract chemical concepts. Our findings provide practical insights for educators leveraging AI in early science education and offer guidance for developers working to enhance LLMs' educational applications. The results highlight the potential and current limitations of using LLMs to bridge the early science literacy gap.

Problem

• Large Language Models
• Preschool Science Education
• Complex Concept Simplification

Innovation

• Language Models
• AI in Education
• Science Learning