Abstract
Collaborative problem-solving (CPS) is a critical competency in modern science, technology, engineering, and mathematics (STEM) education, yet orchestrating effective peer collaboration in complex domains like physics remains a significant pedagogical challenge. This quasi-experimental study evaluated the efficacy of an artificial intelligence (AI)-driven formative scaffolding system, "PhysCollab-AI," designed to support high school students during collaborative thermodynamics problem-solving. A sample of 142 eleventh-grade physics students was assigned to either an experimental condition (n = 72), which received real-time, AI-driven cognitive, metacognitive, and social prompts, or a control condition (n = 70), which received traditional, teacher-led scaffolding. Quantitative analyses revealed that students in the experimental condition achieved significantly higher conceptual learning gains on a physics post-test compared to the control group, controlling for pre-test scores. Furthermore, discourse analysis of group interactions indicated that AI-driven scaffolding significantly increased the frequency of high-level cognitive talk, such as explanation construction and negotiation, and promoted more equitable participation patterns among group members. These findings demonstrate the potential of AI-driven formative scaffolding to enhance domain-specific learning and collaborative dynamics, offering key insights for the design of adaptive learning environments in the learning sciences.
