Novobo: Supporting Teachers' Peer Learning of Instructional Gestures by Teaching a Mentee AI-Agent Together

Paper Title

Novobo: Supporting Teachers' Peer Learning of Instructional Gestures by Teaching a Mentee AI-Agent Together

Publication Info

• Topic area: Teacher professional development through AI-assisted peer learning of instructional gestures.
• Keywords: Instructional gestures, teacher professional development, teachable agents, peer learning, embodied practice, Retrieval-Augmented Generation (RAG), large language models (LLMs), tacit knowledge, collaborative learning, human-computer interaction (HCI).

Background and Problem

• Problem / challenge: Existing methods for training teachers in instructional gestures are time-intensive, overly prescriptive, and fail to address the tacit, experiential nature of these skills. Peer learning is underutilized in this context, and current AI systems lack the ability to support collaborative, reflective learning of nonverbal teaching behaviors.
• Significance: Instructional gestures are crucial for enhancing communication, student comprehension, and engagement. Developing effective training methods can improve teaching quality and student outcomes.
• Motivation and related work: Prior approaches, such as video analysis and behavior modification systems, are limited by high resource demands, lack of contextual adaptability, and overly simplified feedback mechanisms. Research on teachable agents (TAs) has focused on students, leaving a gap in exploring their potential for teacher professionalization.

Solution

• Proposed approach: Novobo, a teachable AI agent designed to act as a mentee, facilitates teachers’ peer learning of instructional gestures through verbal and bodily interactions.
• Novelty:
  1. Introduction of a teachable AI agent for teacher professional development, leveraging the "learning by teaching" approach.
  2. Integration of Retrieval-Augmented Generation (RAG) to provide well-referenced, theoretical, and practical knowledge on instructional gestures.
  3. Use of a skeletal mirror for embodied practice, reducing self-consciousness and fostering reflection.
  4. Design of mentor-mentee dynamics between teachers and the AI to mitigate social pressures and encourage inclusive collaboration.
• Procedure and key techniques:
  • Teachers interact with Novobo in a four-stage loop: posing questions, providing commentary, demonstrating gestures, and explaining their rationale.
  • Novobo generates gesture suggestions using a multi-agent LLM pipeline with RAG, while teachers evaluate, refine, and demonstrate gestures using a skeletal mirror.
  • The system synthesizes teachers’ inputs into summarized principles, fostering knowledge co-construction.

Results

• Concrete findings:
  • Teachers externalized tacit knowledge, socialized insights, and co-constructed understanding of instructional gestures during interactions with Novobo.
  • Novobo’s pipeline outperformed a baseline architecture in pedagogical meaningfulness (4.10 vs. 3.36), appropriateness (4.19 vs. 3.43), naturalness (4.16 vs. 3.44), and understandability (4.08 vs. 3.33), with significant p-values (< 0.001).
  • Teachers appreciated the system’s theoretical grounding, skeletal mirror design, and mentor-mentee dynamics.
• Advantage over baselines: Novobo’s multi-agent architecture with RAG significantly improved the quality of gesture suggestions compared to a single-LLM baseline, achieving higher ratings across all evaluation metrics.
• Experiments / evaluation:
  • Conducted with 30 teachers across 10 collaborative sessions.
  • Teachers interacted with Novobo in real-world study group settings, followed by focus group discussions.
  • Evaluations included vignette analysis and thematic analysis of teacher feedback.
• Limitations and future work:
  • Text-based gesture descriptions may lead to interpretation variability; future work could explore multimodal outputs (e.g., animations).
  • Skeletal mirror view was sensitive to camera positioning, limiting gesture expansiveness.
  • Study participants were from high socio-economic regions; future research should include diverse demographics.
  • Asynchronous interaction and quantification of learning outcomes remain unexplored.

Summary

This study introduces Novobo, a teachable AI agent designed to support teachers in collaboratively learning instructional gestures. By leveraging Retrieval-Augmented Generation (RAG) and a skeletal mirror for embodied practice, Novobo facilitates the externalization, socialization, and co-construction of tacit knowledge. Teachers appreciated its theoretical grounding, reduced social pressures, and reflective learning opportunities. Empirical results demonstrated significant improvements in gesture generation quality and teacher engagement. Future work could explore multimodal outputs, broader demographic inclusion, and asynchronous learning environments. Novobo highlights the potential of teachable agents to enhance teacher professional development and peer learning.