Designing LLM-Powered Conversational Agents for STEM+C Learning and Assessment

Abstract

Large language models (LLMs) are increasingly being integrated into STEM+C learning environments, yet their educational value remains uncertain. Too often, they are deployed as generic helpers rather than pedagogically grounded systems. This creates a central design problem: how to build LLM-powered agents that make learning more visible and better supported without bypassing the reasoning, struggle, and collaboration through which students learn. This dissertation addresses that problem through several years of participatory design with teachers and students, resulting in three LLM-based systems: CoTAL, a prompt engineering method for aligning LLM scoring and feedback with teacher rubrics; Inquizzitor, a mentor agent that turns formative assessment evidence into reflective dialogue; and Copa, a peer agent that uses multimodal learner evidence to scaffold collaborative computational modeling. Findings show that these systems can provide timely feedback, adapt scaffolding, adhere to pedagogical principles, and promote active discussion between collaborating students. This dissertation contributes design principles, alignment methods, agent architectures, and evaluation approaches for LLM-powered systems in STEM+C education. More broadly, it argues that educational LLMs should strengthen teacher judgment, student reasoning, collaboration, trust, and growth rather than make learning effortless.