From Turns to Trajectories: Sequential Analytics of AI‑Generated Student Discourse and Teacher Feedback in a 3D Teacher Simulation

• In Event: AERA Poster Session Sunday 11:45 am
  In Poster Session: Making Learning Work: Design and Technology

Abstract

Teacher simulations increasingly incorporate generative AI agents to mimic students, yet the impact of AI architectures on human-teacher feedback remains underexplored. Analyzing 251 conversation sessions with 1,618 coded dialogue turns across GPT and rule-based conditions, this study employed sequential data analytics, including pattern mining, Markov modeling, and lag sequential analysis. Findings indicate GPT agents elicit more diverse student responses, such as emotional expressions (S:EE → T:Q, 0.672 probability) and complex multi-turn patterns (30 unique). Conversely, rule-based agents foster structured questioning (T:Q → S:R, 28.3%) and sustained inquiry (29 patterns). Student difficulty expressions (S:ED) prompt scaffolded responses predictably (GPT: 0.701; Rule: 0.873). Teacher querying persists across lags (Lag-2: 47.9%), while directives signal termination.

Authors

• Jewoong Moon, University of Alabama

• Jieun Lim, Daegu National University of Education

• Yunseo Lee, University of Wisconsin - Madison

• Gyuri Byun, Seoul National University

• Yeil Jeong, Seoul National University

• Unggi Lee, Korea University