Design and Evaluation of a Photorealistic AI Virtual Peer in Elementary Collaborative Classroom

Paper Title

Design and Evaluation of a Photorealistic AI Virtual Peer in Elementary Collaborative Classroom

Publication Info

• Topic area: AI-driven educational tools for enhancing collaborative learning in elementary classrooms.
• Keywords: AI virtual peer, collaborative learning, GPT-4o-mini, elementary education, classroom participation, photorealistic agents, teacher control, small class challenges, educational technology, peer influence.

Background and Problem

• Problem / challenge: Elementary classrooms, especially small ones, face challenges in fostering diverse perspectives and active participation due to limited student numbers and psychological barriers to speaking. Existing educational technologies have not adequately addressed these issues in group settings.
• Significance: Enhancing student participation and collaborative learning is critical for effective knowledge acquisition, particularly in small or rural classrooms where natural peer interaction is limited.
• Motivation and related work: Prior research highlights the importance of peer influence and active learning but has focused on ideal conditions or university-level settings. Few studies have explored AI agents as peer substitutes in elementary classrooms, particularly for real-time group dialogue generation aligned with teacher intentions.

Solution

• Proposed approach: The study introduces "Saya," a photorealistic AI virtual peer powered by GPT-4o-mini, designed to act as an additional student in classrooms. Saya generates teacher-controlled speech acts to enhance classroom discussions.
• Novelty:
  1. Development of a photorealistic AI virtual peer with five teacher-controlled speech acts (expand, probe, summarize, lighten, incorrect answer).
  2. Integration of GPT-4o-mini for real-time dialogue generation in elementary classrooms.
  3. Empirical evaluation of Saya's impact in contrasting class sizes (large vs. small) in authentic educational settings.
• Procedure and key techniques:
  1. Conducted formative studies with 12 teachers to identify design requirements.
  2. Developed Saya with a teacher-controlled interface for selecting speech acts.
  3. Implemented Saya in field studies in Japanese elementary schools, comparing lessons with and without Saya integration.
  4. Evaluated system performance, student speaking time, teacher feedback, and student acceptance.

Results

• Concrete findings:
  • Student speaking time increased by 1.28 times in a large class (27 students) and 2.07 times in a small class (2 students).
  • Saya achieved practical response times (3.48–3.54 seconds) and was positively received by 95.6% of students in the large class and 100% in the small class.
• Advantage over baselines:
  • In the small class, Saya enabled critical opinion changes and collaborative discussion previously unattainable with only two students.
  • In the large class, Saya helped synthesize and organize discussions, reducing teacher dominance in speaking time.
• Experiments / evaluation:
  • Conducted field studies in two classrooms (large and small) using Japanese language arts lessons.
  • Metrics included student speaking time, system performance, and qualitative feedback from teachers and students.
• Limitations and future work:
  • Limited to single-lesson evaluations, which may not account for novelty effects or long-term adaptation.
  • Lessons differed between conditions, potentially influencing results.
  • Future work should include longitudinal studies, cross-cultural validation, and exploration of alternative agent appearances (e.g., robots, anime-style characters).

Summary

The study presents "Saya," a photorealistic AI virtual peer designed to enhance collaborative learning in elementary classrooms. Saya uses GPT-4o-mini to generate teacher-controlled speech acts, addressing challenges in student participation and opinion diversity, particularly in small classes. Field studies demonstrated that Saya increased student speaking time (1.28x in large classes, 2.07x in small classes) and was highly accepted by students (95.6%-100%). Teachers noted improved listening behaviors and critical engagement, especially in small classes where Saya enabled meaningful discussions. Future work should explore sustained effectiveness, cross-cultural applicability, and alternative agent designs.