As learning moves to digital and AI-supported platforms, we risk losing the rich social dynamics that drive curiosity, deep engagement, and social development in collaborative learning. MoSaIC (Methods for Social and Individual Curiosity) addresses this challenge with a research tool that detects and visualizes how curiosity and social learning features emerge and spread in learning conversations.
The framework classifies four curiosity types (specific, diversive, epistemic, social) alongside interaction features (questioning, explanation, uncertainty, agreement, and disagreement). These are presented through interactive timeline and network visualizations that enable researchers to connect automated analysis with subjective insights into collaboration. MoSaIC evolved from an NLP-researcher hybrid pipeline into a hybrid NLP-LLM framework. High-confidence discourse cases are handled locally by NLP, while low-confidence segments are routed to an LLM. This improves detection of subtle, context-dependent phenomena while prioritizing privacy, cost efficiency, and scalability. Validation shows stronger alignment with expert judgment (κ ≈ 0.66 for curiosity presence and interaction type classification).
Our collaborative-task studies with peer learners (n=22) demonstrate how questioning, explaining, and turn-taking behaviors connect to learners' curiosity, interest, and social perception. Ongoing work includes comparative analyses revealing differences in how these dynamics unfold in peer-peer versus student-AI study sessions, offering insight into what may be lost when learners collaborate with AI assistants instead of peers. By making social learning processes like curiosity spread visible, MoSaIC contributes both methodological tools for researchers and insights for designing AI that supports the collaborative essence of learning.