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

Abstract: LLMs provide a new opportunity to study how language shapes exploratory cognition because conversational strategies can be systematically manipulated at inference time. We introduce CURIOBOT, a framework that operationalizes Berlyne's collative variables, novelty, complexity, conflict, and uncertainty, as adaptive linguistic interventions for conversational tutoring. Across 270 tutoring conversations spanning multiple 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. To measure these effects, we further introduce a learner-centered evaluation framework capturing exploratory questioning, conversational agency, productive struggle, and observable curiosity. Learner-side gains persisted even when tutor-side instructional quality remained unchanged, suggesting that curiosity functions as a partially independent interaction-level mechanism. More broadly, our results demonstrate that LLM-mediated dialogue can serve as a scalable experimental framework for studying how language shapes exploratory learning behavior.

• The paper introduces CurioBot, an inference-time prompting framework that adaptively applies novelty, complexity, conflict, uncertainty, or neutral framing across 270 tutoring conversations with 45 participants.
• CurioBot increased dialogue depth by about 2.4× and significantly improved exploratory questions, productive struggle, curiosity, and conversational agency across three model families, academic domains, and complexity levels.
• The findings suggest curiosity can operate partly independently of instructional quality, while conflict and transitions from uncertainty to novelty or conflict appear especially promising for personalized tutoring policies.

Overview

This paper introduces CurioBot, a framework that operationalizes Berlyne's collative variables—novelty, complexity, conflict, and uncertainty—as adaptive, inference-time linguistic interventions for LLM-based tutoring. The central claim is that curiosity can be treated as a controllable interaction-level mechanism: by systematically manipulating how a tutor frames information through prompting alone, without fine-tuning or parameter updates, one can reliably increase exploratory learner behavior. The authors support this claim with a controlled $3 \times 3 \times 3$ factorial study (model family × academic domain × topic complexity) involving 45 human participants and 270 tutoring conversations, evaluated via a learner-centered rubric scored by a multi-judge LLM-as-a-judge pipeline. Beyond its educational contribution, the paper positions conversational tutoring as a scalable experimental instrument for studying how language shapes exploratory cognition.

Framework design

CurioBot decomposes curiosity modulation into five operators: Novelty (unexpected analogies and perspectives), Complexity (exposing additional mechanisms), Conflict (introducing contradictions within the learner's reasoning), Uncertainty (foregrounding unresolved aspects), and Neutral (no modulation). At each tutor turn, a Modulator agent selects an operator conditioned on conversational signals such as response depth, exploratory questioning, confidence cues, and initiative; a Generator then produces the response conditioned on both dialogue context and the selected operator. Both stages are prompted black-box LLMs, making the intervention portable across model families.

The prompting strategy also enforces discourse-level constraints designed to sustain exploration: tutors avoid direct quizzing, maintain open informational gaps, ground terminology before use, and end turns with knowledge gaps that invite learner questions. Crucially, the underlying topic and learning objective are held fixed across conditions; only conversational framing varies. This separation of operator selection from response generation additionally enables post-hoc analysis of operator transitions on downstream learner behavior.

Experimental design

The study addresses four research questions: whether curiosity interventions increase exploratory learner behaviors (RQ1); how individual operators and their sequencing affect exploration dynamics (RQ2); whether learner-side gains are explainable purely by tutor-side instructional quality (RQ3); and whether effects generalize across models, domains, and complexity levels (RQ4).

Three frontier model families were tested (Claude, Gemini, GPT), alongside provider-native study-oriented modes (Gemini Guided Learning, Claude learning style) as approximate upper-bound references rather than formal baselines, since these lack API-level controllability. Topics spanned STEM, social sciences, and arts/humanities at low, medium, and high complexity. Each participant completed two blocks totaling 120 minutes of interaction time, with variant allocation counterbalanced and all conversations initialized from a standardized learner prompt.

Evaluation uses a ten-dimensional rubric: four learner-side dimensions (exploratory questions $L_1$, productive struggle $L_2$, observable curiosity $L_3$, conversational agency $L_4$) and six tutor-side dimensions (instructional quality $T_1$ through engagement facilitation $T_6$), each on 7-point behaviorally anchored scales. Because preliminary analysis showed non-trivial judge variance, each conversation was scored by three judge models with three runs each, yielding nine evaluations per conversation.

Effects of curiosity modulation

The headline result is engagement depth: under fixed time budgets, CurioBot produced roughly 2.4× more conversational turns than baseline (16.12 vs. 6.66 overall), with increases of approximately 2–3× consistent across all domains and complexity levels. Qualitatively, baseline interactions converge toward short acknowledgment-style exchanges, whereas CurioBot interactions sustain self-initiated questioning and follow-up exploration.

All four learner-side dimensions improve significantly ($p < 0.005$) across all three model families. The largest relative gains appear in productive struggle (+44–47%), conversational agency (+30–34%), learner curiosity (+26–35%), and exploratory questioning (+21–32%). Tutor-side effects are less uniform: Claude and GPT show broad instructional improvements, but Gemini exhibits significant degradation in instructional quality ($T_1$: 5.30 → 3.66) and cognitive load management ($T_3$: 4.42 → 3.25). This asymmetry is itself informative—it directly motivates the paper's argument that learner-side gains cannot be reduced to tutor-side quality improvements.

Against the commercial reference conditions, CurioBot matches or exceeds study-oriented modes on exploratory questioning and agency despite relying solely on inference-time prompting, though study modes retain clear advantages on productive struggle and most tutor-side dimensions.

Operator effectiveness and sequencing

Operator-level analysis reveals substantial heterogeneity. Conflict achieves the highest downstream success rate (79.03%) despite being selected least often among non-neutral operators (72 appearances), while Novelty is the most frequently applied operator (258 appearances) but succeeds only 61.42% of the time; Uncertainty performs worst (44.44%). This ordering is consistent with Berlyne's theoretical hierarchy, in which cognitive conflict generates stronger epistemic tension than informational expansion alone.

Transition analysis adds a temporal dimension: the largest curiosity gains follow transitions out of Uncertainty toward Novelty or Conflict, whereas repeated application of the same operator tends toward neutral or negative effects. The implication is that curiosity induction depends jointly on operator type, magnitude of the induced information gap, and sequencing—an observation the authors identify as a tractable target for personalized dialogue policies, though they do not optimize the selection policy themselves.

Curiosity as a partially independent mechanism

The strongest conceptual claim in the paper is that curiosity operates as a partially independent interaction-level mechanism rather than a byproduct of better instruction. Two lines of evidence support this. First, the dissociation is clearest under Gemini, where learner curiosity and agency improve significantly even as instructional quality and load management degrade. Second, Pearson correlations between learner-side and tutor-side dimensions weaken systematically under CurioBot across nearly all dimension pairs, indicating that exploratory behavior becomes decoupled from tutor-side pedagogical evaluations when curiosity modulation is active. If accepted, this implies that evaluation frameworks for conversational tutoring systems should include learner-centered dimensions; tutor-side metrics alone would miss the primary effect the intervention produces.

Generalization

Learner-side gains replicate across all three model families, all three domains, and most complexity levels: CurioBot improves learner-side scores in 8 of 9 domain-complexity settings, with 7 reaching significance, including all high-complexity conditions. Tutor-side gains, however, become constrained under STEM and high-complexity conditions, where differences between variants shrink despite continued learner-side improvement. This further reinforces the learner–tutor dissociation. The consistency across heterogeneous base models suggests the effect stems from the linguistic intervention itself rather than idiosyncratic model behavior, though per-judge analyses reveal meaningful cross-family variation in how judges score tutor-side dimensions, including a self-preference artifact on instructional quality where each judge rates its own family's baseline higher.

Limitations

The paper concedes two principal limitations. First, the Modulator is a prompted agent following a fixed decision rule; the operator-selection policy is unoptimized, and the authors explicitly note that a learned controller—a contextual bandit over the five-operator action space or an RL policy trained on offline trajectories—could discover transitions their rule-based scheme misses. Second, and more fundamentally, all psychological constructs (productive struggle, epistemic curiosity) are inferred from text transcripts via LLM judges. The text-only protocol cannot capture the physiological arousal central to Berlyne's framework, and learners who express engagement tersely may be systematically underscored. The authors propose triangulation with eye-tracking, skin conductance, or EEG as necessary future validation. A related open question concerns durability: the study measures within-session exploratory behavior only, leaving unresolved whether induced curiosity translates into retained learning outcomes beyond the interaction itself.

Conclusion

This work demonstrates that Berlyne-inspired curiosity variables can be operationalized as adaptive inference-time prompting operators that consistently increase exploratory learner behavior across model families, domains, and complexity levels, with up to 2.4× more conversational turns under fixed time budgets. The finding that learner-side gains persist independently of tutor-side instructional quality supports treating curiosity as a partially autonomous interaction-level mechanism and motivates learner-centered evaluation of tutoring systems. More broadly, the results establish LLM-mediated dialogue as a controllable experimental setting for studying how specific linguistic strategies shape exploratory cognition during learning interactions.