John Kos 

Ph.D. Student in Human-centered Computing  

School of Interactive Computing  

Georgia Institute of Technology  

Date: May 6th 2026 

Time: 2:00pm-4:00pm 

Location: CODA 1203 Five Points 

Teams Link 

Committee: 

Ashok Goel (Advisor) - School of Interactive Computing, Georgia Tech 

Chris Dede - Graduate School of Education, Harvard University 

Christopher MacLellan - School of Interactive Computing, Georgia Tech 

Christina Schwarz - College of Education, Michigan State University 

Sashank Varma - School of Interactive Computing, Georgia Tech  

Emily Weigel – School of Biological Sciences, Georgia Tech 

Abstract  

Adult learners are often asked to address open-ended problems that require critical thinking such as scientific inquiry. In scientific inquiry, scientists use inquiry-based modeling to develop hypotheses, evaluate the model based on data, revise the model, and evaluate the outcome of their hypotheses. Supporting inquiry education is challenging because it requires the learners to engage in both cognitive (thinking about a problem) and metacognitive (thinking about one’s thinking, high level planning, and reflection) processes. 

AI and education tools built to assist these processes have been an open line of research in the literature for two decades. A series of obstacles including modeling the learner’s thought process, establishing the necessary cognitive understanding, and properly encouraging active engagement with metacognitive thinking have presented challenges to tools seeking to support metacognitive thinking using traditional machine learning methods. Traditional machine learning techniques have only been modestly successful in inferring intent from learner modeling actions, and as a result AI agents have struggled with supporting metacognitive thinking. 

I propose that the advent of Generative AI, paired with learner self-explanations, allows for evaluation and guidance of the learner's metacognitive thinking. My research explores the use of this new paradigm to support the learner through two different metacognitive processes: inquiry and metamodeling. Using the interactive learning environment VERA as a platform, I design specialized LLM agents to enable these metacognitive processes in the context of modeling ecological systems. To additionally support these processes, an additional AI agent orchestration system is designed to evaluate user self-explanation and cognitively scaffold user inquiry using a handful of other AI agents. I plan to evaluate the above Agentic AI system with a set of pilot studies designed to measure the gains in learning outcomes from the metacognitive agents. Additionally, I will do an analysis of differing architectures for AI agent orchestration and selection, to demonstrate proper deployment of the correct agent based on the learner’s self-explanation.