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

Synthesis: Ganganath et al. (2026) introduce CURIOBOT, a framework that operationalizes Berlyne's four collative variables (novelty, complexity, conflict, uncertainty) as adaptive linguistic interventions in conversational tutoring. Across 270 tutoring conversations spanning multiple Large Language Models (LLMs) model families, domains, and topic complexity levels, curiosity-oriented interventions consistently increased exploratory learner behaviors, producing up to 2.4x more conversational turns under fixed time budgets. A learner-centered evaluation framework captured exploratory questioning, conversational agency, productive struggle, and observable curiosity. Critically, learner-side gains persisted even when tutor-side instructional quality remained unchanged, suggesting that curiosity functions as a partially independent interaction-level mechanism — not merely a byproduct of instruction quality. This work demonstrates that LLM-mediated dialogue can serve as a scalable experimental framework for studying how language shapes exploratory learning behavior, with direct implications for Metacognition and Self-Regulated Learning research, Intelligent Tutoring design, and Scaffolding strategies.

What this means for practice

• Designers. Treat curiosity as a sequencing problem, not a standing instruction. The strongest curiosity score gains followed transitions out of uncertainty — especially toward novelty and conflict — while repeated application of the same operator trended neutral or negative.
• Designers. Instrument the learner side of a tutor separately from tutor-side quality. CURIOBOT improved exploratory questioning (L1: +21–32%), learner curiosity (L3: +26–35%), and conversational agency (L4: +30–34%) even where tutor-side instructional quality and cognitive load management declined significantly, so a tutor-side rubric alone will miss real learner gains.
• Researchers. Use LLM-mediated dialogue as an experimental setting: 270 conversations across three model families, three domains, and three complexity levels let the design isolate conversational framing from content while holding time-on-task fixed.
• Designers. Equalize time-on-task before comparing dialogue designs. Interaction duration scaled with complexity — 10 minutes at low, 20 at medium, 30 at high — with 120 minutes of total interaction per participant, which is what makes the roughly 2.4× increase in conversational turns interpretable.

Limitations

• All 45 participants were students recruited from the authors' own institution (64.10% male, 35.90% female, mostly aged 24–26 and enrolled in postgraduate programs), so the sample is one site and not representative of broader learner populations.
• Learner-side constructs such as productive struggle and curiosity were scored by an LLM-as-a-judge pipeline across three frontier models, not by independent human raters, and the authors state that text-only evaluation cannot capture the physiological arousal central to Berlyne's framework.
• Every interaction ran through a text-only interface with self-reported moderate English proficiency (reading 3.26±0.12, writing 3.16±0.13 on a 5-point scale), so learners who communicate in technical rather than verbose markers may be systematically underscored.
• The operator-selection policy is a fixed prompted decision rule mapping engagement signals to operators, and the authors describe it as unoptimized, leaving learned controllers such as a contextual bandit over the five-operator action space as future work.

Citation

Gevindu Ganganath, Pasindu Bolonghege, Qianru Lyu, Pradeep Varakantham, Thivya Kandappu (2026). Curiosity as Linguistic Intervention: Using LLM Tutoring Dialogues to Influence Exploratory Learning Behavior. submitted to EMNLP 2026.