Curiosity as Linguistic Intervention: Using LLM Tutoring Dialogues to Influence Exploratory Learning Behavior

This paper introduces CURIOBOT, a framework designed to influence exploratory learning by embedding Berlyne’s collative variables—novelty, complexity, conflict, and uncertainty—into LLM-based tutoring dialogues. Rather than fine-tuning models, the authors use these variables as inference-time linguistic interventions. The system dynamically selects an 'operator' based on the learner's current interaction state, such as their level of initiative or response depth, to frame the tutor's responses in a way that stimulates epistemic curiosity.

The researchers conducted a 3x3x3 factorial study involving 270 tutoring conversations across three model families, three academic domains, and three levels of topic complexity. They compared a baseline model, the CURIOBOT-modulated model, and commercially available 'study-oriented' modes. To assess the impact, they developed a learner-centered evaluation framework that tracks dimensions like exploratory questioning, conversational agency, and productive struggle, using an LLM-as-a-judge pipeline to score interactions.

The results indicate that curiosity-oriented interventions consistently boost exploratory behavior, leading to up to 2.4 times more conversational turns within fixed time budgets compared to baseline models. Notably, these gains in learner engagement persisted even when the tutor's instructional quality remained static or degraded. This suggests that curiosity acts as an independent interaction-level mechanism, and that LLM-mediated dialogue can serve as a powerful, scalable experimental tool for studying how language shapes cognitive processes during learning.

Most current educational LLMs prioritize instructional correctness and pedagogical fluency. This research shifts the focus toward the learner's internal state, demonstrating that specific linguistic framing can actively foster curiosity and agency. By treating LLMs as experimental instruments, this work provides a scalable way to study human cognition and learning behavior in a controlled, interactive environment.