How GenAI Mentor Configurations Shape Early Collaborative Dynamics: A Classroom Comparison of Individual and Shared Agents

A research team from institutions including Tsinghua University and Beijing Normal University, led by Siyu Zha, conducted a novel classroom experiment to understand how the structure of AI assistance impacts group learning. The study, published on arXiv, involved two eighth-grade classes engaging in creative problem-solving. One class used a 'shared-AI' configuration where a single GenAI mentor served an entire small group, while the other used an 'individual-AI' setup where each student had a personal AI instance. The researchers employed advanced multi-layer discourse coding combined with lag sequential analysis (LSA) and ordered network analysis (ONA) to map the complex interaction patterns.

Results revealed starkly different collaborative dynamics. The shared AI configuration acted as a central hub, fostering 'convergence-oriented collaboration' with stronger alignment in the group's shared regulatory states and more coordinated reasoning. Conversely, the individual AI configuration distributed support, leading to more cycles of individual exploration and evaluation. However, this also resulted in more fragmented group interaction patterns, which were accompanied by a measurable increase in teacher intervention needed to manage the resulting divergence. The findings position AI configuration not just as a tool choice, but as a critical 'structural design variable' that fundamentally reorganizes the regulatory ecology of classroom collaboration, with clear trade-offs between group cohesion and individual exploration.