In elementary education, students struggle to articulate uncertainties, limiting diverse perspectives in classroom discussions, particularly in small schools where limited participants constrain collaborative learning. This study designed and evaluated ``Saya,'' a photorealistic AI virtual peer functioning as an additional student. We implemented five teacher-controlled speech acts (expand, probe, summarize, lighten, and incorrect answer) through dynamic classroom dialogue generation using GPT-4o-mini. Field studies in Japanese elementary schools (large class: 27 students, small class: 2 students) demonstrated that Saya integration increased the proportion of student speaking time by 1.28 times and 2.07 times respectively, with 95.6% and 100% of students expressing desire for future Saya-integrated lessons. Teachers reported enhanced student concentration and listening behaviors, noting that interactions with Saya prompted students to reconstruct their own understanding of the learning material. This research provides new insights into design principles for collaborative learning agents in elementary education settings, effective implementation scenarios based on class size, and the future potential of AI-enhanced collaborative learning.
ACM CHI Conference on Human Factors in Computing Systems