Multimodal Large Language Models as a Catalyst for Advancing Learning Analytics in Early Childhood Education

Abstract

Early childhood education (ECE) involves rich, informal, and multimodal learning processes that are difficult to assess with traditional methods. This paper presents a novel framework that integrates Multimodal Large Language Models (MLLMs) into ECE learning analytics to capture children’s spontaneous expressions—such as drawings, speech, gestures, and social interactions—in a scalable, child-centered manner. The system includes multimodal data collection, MLLM-based feature extraction, automated developmental analytics, and educator-in-the-loop feedback. A real-world case from a rural kindergarten illustrates the framework’s ability to generate interpretable indicators and actionable insights. We discuss opportunities for individualized assessment and developmentally appropriate practices, as well as challenges related to interpretability, privacy, and equity. This work demonstrates the potential of MLLMs to support holistic, play-based learning analytics in early childhood settings.