Unveiling the Capabilities of Large Language Models in Simulating Student Behavioral Dynamics and Supporting Peer Feedback to Augment Task Performance

Large language models (LLMs) show promise as student simulators for learning research. Yet prior work offers only a coarse view, focusing mainly on predicting answer accuracy while overlooking finer-grained factors such as prior knowledge, contextual experience, and broader behavioral signals like sensory actions. Explanations for when and why LLMs succeed or fail in these simulations also remain limited. To address this gap, we conduct large-scale, fine-grained simulation experiments to examine LLMs’ capabilities and challenges in capturing such behavioral dynamics. We further explore explanations and probe underlying mechanisms through ablation studies on input information and embedding space analyses. Furthermore, we highlight new HCI opportunities by presenting a case study that illustrates how insights above can inform example simulator design and practical applications. For example, in an N = 188 study, LLMs mimicked real peers to deliver peer-pressure feedback, accelerating students’ cognitive problem solving and matching real peer effects.