Copa LLM agent boosts STEM learning without causing student over-reliance

A team of 16 researchers from multiple institutions (including Vanderbilt University) has introduced Copa, a Collaborative Peer Agent built as a multi-agent LLM system for STEM+C education. Unlike typical LLM tutors that encourage answer-seeking and cognitive offloading, Copa is grounded in the Evidence-Decision-Feedback (EDF) framework, which is itself rooted in Social Cognitive Theory and Social Constructivism. The agent uses adaptive, dialogic scaffolding to promote sense-making, intervening only when students need help building understanding rather than just providing answers.

In a controlled study with 33 dyads from an authentic high school computational modeling course, Copa demonstrated two key outcomes. First, it supported students' confidence and ability to verbalize conceptual understanding without creating dependence — a common failure of naive LLM tutors. Second, it provided personalized, adaptive feedback interpretable from multimodal student input data (e.g., text, code interactions). The results position theory-guided multi-modal LLM agents as a viable path for classroom AI that amplifies reasoning instead of replacing it, addressing long-standing concerns about over-reliance and "gaming" behaviors in AI-assisted learning. The paper is currently under review at Computers & Education.