A Distilled Model for Collaborative Problem Solving Skill Classification on Resource-Limited Devices

Abstract

Collaborative Problem Solving (CPS), a critical 21st-century skill, requires real-time classification to optimize classroom instruction. While mobile technologies offer unprecedented opportunities for such evaluation, existing methods fail to meet real-time demands in dynamic mobile environments due to their requirement for heavy computational resources. This paper introduces a novel knowledge distillation framework to enable lightweight, accurate CPS classification on resource-limited devices. Our approach distills knowledge from large-scale teacher models into compact student networks, reducing model size by 32% compared to baseline model while maintaining 82.6% accuracy on the target dataset. This work bridges the gap between AI-driven classification and mobile educational ecosystems, thereby offering educators a scalable tool for continuous formative evaluation and adaptive pedagogy.