Inferring Self-Efficacy and Growth Mindset from Student-Agent Interaction to Tailor Teacher Feedback

Abstract

Sustaining student motivation and engagement remains a difficult challenge in digital learning environments. To effectively address this, Intelligent Virtual Agents (IVAs) designed for education must move beyond generic pedagogical support and adapt to learners’ dynamic socio-cognitive states. Recent advances in Large Language Models (LLMs) offer unprecedented potential for IVAs to achieve the nuanced social perception required to foster meaningful human-AI cooperation. This study investigates how an LLM-powered conversational agent infers socio-cognitive states of self-efficacy (SE) and growth mindset (GM) from two complementary data sources: what students say (dialogue traces) and what they do (interaction logs). Using data from high school students receiving motivational support provided by the "Grow-Together" IVA during math problem-solving, we evaluate the comparative regression and correlation metrics of ten state-of-the-art LLMs. By comparing models’ inferences with students’ self-reported baselines, we reveal an important finding: while LLMs can infer the highly reactive state of SE through surface-level linguistic tone interpretation, they struggle to infer the deep-seated state of GM without exhibiting a systematic pessimistic bias. These findings advocate for a Teacher-in-the-Loop paradigm within IVAs, evidencing that human oversight remains essential to interpret psychological factors and provide targeted pedagogical support in digital learning environments.