Suen, H.-Y., & Hung, K.-E. (2026) — National Taiwan Normal University. IEEE Transactions on Learning Technologies.
Key Finding
Closed-loop ITS with multimodal affective scoring (facial, vocal, textual, oculomotor) produced significant presentation skill gains (Cohen's d = 0.39-0.90, N=204) over 30 days.Synthesis
This paper presents one of the most comprehensive closed-loop intelligent-tutoring-systems for soft-skill training. The system operationalizes a seven-dimension BARS across facial, vocal, textual, and oculomotor inputs, using an XGBoost backbone for interpretable scoring that approaches expert-rater reliability (Spearman's rho 0.69-0.78). The three-layer feedback architecture — rubric-aligned scoring, audience-expressive diagnostics, and retrieval-augmented-generation conversational coaching — creates a complete deliberate practice loop. With 204 adult learners and Cohen's d of 0.39-0.90 across all seven dimensions, this is among the stronger efficacy signals in ITS research. The interpretability requirement (feedback traceable to observable cues) directly addresses concerns raised in educational-llm-alignment about opaque AI feedback, while the multimodal approach extends beyond text-only systems like cyberscholar-genai-writing-feedback. The closed-loop architecture shares philosophical ground with ai-tutor-behavioral-evaluation's call for behavioral feedback loops, and the retrieval-augmented coaching component parallels retrieval-augmented-tutoring-algorithm-kite's approach to grounded tutoring.Related Pages
- ai-enabled-serious-games — Complements multimodal ITS architectures with game-based training contexts
- epistemic-emotions-collaborative-problem-solving — Ordered Network Analysis reveals structured persistence and transition patterns of confusion and fru