pith. sign in

arxiv: 2606.22349 · v1 · pith:NJDB5RY2new · submitted 2026-06-21 · 💻 cs.CL · cs.HC

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

Gevindu Ganganath , Pasindu Bolonghege , Qianru Lyu , Pradeep Varakantham , Thivya Kandappu This is my paper

Pith reviewed 2026-06-26 10:57 UTC · model grok-4.3

classification 💻 cs.CL cs.HC
keywords curiositylinguistic interventionsLLM tutoringexploratory learningBerlyne collative variablesconversational agencyproductive struggle
0
0 comments X

The pith

Curiosity-oriented linguistic interventions in LLM tutoring increase exploratory learner behaviors up to 2.4 times.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether language in tutoring dialogues can be deliberately shaped to boost curiosity-driven exploration. It does so by embedding interventions drawn from Berlyne's four collative variables into LLM tutor responses and measuring the downstream effects on learners across 270 conversations. The interventions produced substantially more learner turns under fixed time limits, and these gains appeared on the learner side even when the tutor's own instructional quality stayed the same. The work also supplies a learner-centered measurement scheme for exploratory questioning, agency, productive struggle, and observable curiosity, positioning LLM conversations as a controllable testbed for studying how language influences learning behavior.

Core claim

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 tu

What carries the argument

CURIOBOT, the framework that turns Berlyne's collative variables into real-time linguistic interventions inside LLM tutor utterances.

If this is right

  • Exploratory questioning, conversational agency, productive struggle, and observable curiosity all rise in the learner.
  • The increase in learner activity occurs independently of any measured improvement in tutor instructional quality.
  • LLM-mediated dialogue supplies a scalable, controllable setting for testing how specific language choices affect exploratory learning.
  • The effects appear across different model families and domains, indicating the interventions are not tied to one particular LLM.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same style of interventions might be tested outside tutoring, for example in open-ended question-answering or collaborative problem-solving sessions.
  • If the mechanism is truly independent of content quality, it could be combined with other tutoring strategies without trade-offs.
  • Longer-term studies could check whether the increased exploratory turns during a session lead to better retention or transfer on later tasks.

Load-bearing premise

The chosen linguistic patterns successfully isolate the four collative variables without being altered by the underlying model's own tendencies, the topic choices, or the way the new evaluation metrics are scored.

What would settle it

A controlled replication in which the same interventions produce no increase in learner conversational turns relative to baseline under identical time budgets and topic sets.

Figures

Figures reproduced from arXiv: 2606.22349 by Gevindu Ganganath, Pasindu Bolonghege, Pradeep Varakantham, Qianru Lyu, Thivya Kandappu.

Figure 1
Figure 1. Figure 1: CURIOBOT Architecture. Curiosity Operators The five operators define distinct conversational strategies for introducing epistemic tension during tutoring dialogue (Fig￾ure 2 and 7). NOVELTY reframes concepts through unexpected analogies or perspectives; COMPLEX￾ITY expands the surrounding conceptual space by exposing additional mechanisms and dependencies; CONFLICT introduces contradictions or competing ex… view at source ↗
Figure 2
Figure 2. Figure 2: Example turns representing how CURIOBOT modulates each operator. topic complexity. We study three frontier LLM families commonly used in conversational learn￾ing settings: Claude (claude-opus-4-6), Gemini (gemini-3.1-pro-preview), and GPT (gpt-5.4), all accessed through their respective APIs using de￾fault decoding parameters. For each model family, we evaluate three tu￾toring variants: (i) the unmodulated… view at source ↗
Figure 3
Figure 3. Figure 3: Example interaction trajectories under Baseline and C [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average curiosity score changes across operator [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pearson correlations between learner-side and tutor-side evaluation dimensions under (a) Baseline and (b) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Student participants demographic data distri [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example turns representing how CURIOBOT modulates each operator. I Conversation Examples [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Example conversation on the topic Tragedy of the Commons. [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example conversation on the topic Political Polarization. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Aggregated results over LLMs; columns represents subject areas while rows represents complexity levels. [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
read the original abstract

Large Language Models (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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces CURIOBOT, a framework operationalizing Berlyne's collative variables (novelty, complexity, conflict, uncertainty) as adaptive linguistic interventions in LLM-based tutoring dialogues. Across 270 conversations spanning model families, domains, and topic complexities, it reports that curiosity-oriented interventions consistently increase exploratory learner behaviors, yielding up to 2.4x more conversational turns under fixed time budgets. Learner gains persist even when tutor instructional quality is unchanged, and the work introduces a learner-centered evaluation framework measuring exploratory questioning, conversational agency, productive struggle, and observable curiosity. The results position LLM-mediated dialogue as a scalable experimental tool for studying language's role in exploratory learning.

Significance. If the central independence claim holds after verification of controls, the work offers a new experimental paradigm for isolating linguistic effects on exploratory cognition at scale. The CURIOBOT framework and learner-centered metrics could enable reproducible studies of interaction-level mechanisms in educational AI, extending beyond traditional tutoring research.

major comments (2)
  1. [Abstract] Abstract: The claim that 'learner-side gains persisted even when tutor-side instructional quality remained unchanged' is load-bearing for the independence conclusion, yet the abstract provides no evidence that interventions were checked for systematic effects on response length, coherence, or topic coverage that could mechanically increase turn counts.
  2. [Abstract] Abstract: The reported 2.4x increase in conversational turns requires statistical details, per-condition sample sizes, and validation that the learner-centered metrics are insensitive to generation artifacts; none of these are supplied, undermining assessment of whether the effect is robust to the noted confounds.
minor comments (1)
  1. [Abstract] The acronym CURIOBOT is used without an explicit expansion or definition on first use in the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these targeted comments on the abstract. They correctly identify that the abstract must better support its key claims about independence and effect robustness. We will revise the abstract accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that 'learner-side gains persisted even when tutor-side instructional quality remained unchanged' is load-bearing for the independence conclusion, yet the abstract provides no evidence that interventions were checked for systematic effects on response length, coherence, or topic coverage that could mechanically increase turn counts.

    Authors: The full manuscript reports post-hoc analyses confirming that curiosity interventions produced no significant differences in tutor response length, coherence scores, or topic coverage relative to control conditions. These checks are described in the experimental design and results sections. We agree the abstract should reference this verification to strengthen the independence claim and will add a concise clause to that effect in the revision. revision: yes

  2. Referee: [Abstract] Abstract: The reported 2.4x increase in conversational turns requires statistical details, per-condition sample sizes, and validation that the learner-centered metrics are insensitive to generation artifacts; none of these are supplied, undermining assessment of whether the effect is robust to the noted confounds.

    Authors: The abstract summarizes the maximum observed effect size across the 270 conversations (balanced across conditions and models). The learner-centered metrics were validated via human annotation with reported inter-rater agreement to mitigate generation artifacts. Full per-condition sample sizes, confidence intervals, and statistical tests appear in the results section. Due to abstract length limits we cannot include all details, but we will revise to state the total sample size and note the artifact validation. We view this as a partial revision. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical study is self-contained

full rationale

The paper reports results from an empirical user study across 270 tutoring conversations, measuring behavioral outcomes (conversational turns, exploratory questioning, agency) under curiosity-oriented interventions operationalized from Berlyne's variables. No derivation chain, equations, fitted parameters renamed as predictions, or self-citation load-bearing steps exist that reduce claims to inputs by construction. Central findings rest on observed data from multiple models and domains rather than definitional equivalence or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that Berlyne's psychological variables can be translated into controllable linguistic interventions and that the new evaluation framework measures the intended constructs; no free parameters or invented physical entities are evident from the abstract.

axioms (1)
  • domain assumption Berlyne's collative variables (novelty, complexity, conflict, uncertainty) influence curiosity and can be operationalized as adaptive linguistic interventions in dialogue.
    The CURIOBOT framework is built directly on this psychological theory to design the tutor responses.
invented entities (1)
  • CURIOBOT no independent evidence
    purpose: Framework that operationalizes collative variables as linguistic interventions for conversational tutoring.
    New named system introduced to implement and test the interventions.

pith-pipeline@v0.9.1-grok · 5713 in / 1447 out tokens · 34754 ms · 2026-06-26T10:57:07.035343+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    Neuron , volume=

    States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit , author=. Neuron , volume=. 2014 , publisher=

    2014

  2. [2]

    Nature communications , volume=

    Humans monitor learning progress in curiosity-driven exploration , author=. Nature communications , volume=. 2021 , publisher=

    2021

  3. [3]

    , author=

    Conflict, arousal, and curiosity. , author=. 1960 , publisher=

    1960

  4. [4]

    , author=

    Motivational problems raised by exploratory and epistemic behavior. , author=. 1962 , publisher=

    1962

  5. [5]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  6. [6]

    Text summarization branches out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text summarization branches out , pages=

  7. [7]

    BERTScore: Evaluating Text Generation with BERT

    Bertscore: Evaluating text generation with bert , author=. arXiv preprint arXiv:1904.09675 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  8. [8]

    Unifying AI tutor evaluation: An evaluation taxonomy for pedagogical ability assessment of LLM-powered AI tutors , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  9. [9]

    Ceur Workshop Proceedings , volume=

    Automating pedagogical evaluation of LLM-based conversational agents , author=. Ceur Workshop Proceedings , volume=. 2025 , organization=

    2025

  10. [10]

    arXiv preprint arXiv:2205.07540 , year=

    The AI teacher test: Measuring the pedagogical ability of blender and GPT-3 in educational dialogues , author=. arXiv preprint arXiv:2205.07540 , year=

    work page arXiv

  11. [11]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Mathdial: A dialogue tutoring dataset with rich pedagogical properties grounded in math reasoning problems , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  12. [12]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Stepwise verification and remediation of student reasoning errors with large language model tutors , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  13. [13]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Bridging the novice-expert gap via models of decision-making: A case study on remediating math mistakes , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  14. [14]

    arXiv preprint arXiv:2507.10579 , year=

    Findings of the bea 2025 shared task on pedagogical ability assessment of ai-powered tutors , author=. arXiv preprint arXiv:2507.10579 , year=

    work page arXiv 2025

  15. [15]

    Dean of LLM Tutors: A Framework for Automated Quality Review of AI-generated Feedback

    Dean of llm tutors: exploring comprehensive and automated evaluation of llm-generated educational feedback via llm feedback evaluators , author=. arXiv preprint arXiv:2508.05952 , year=

    work page internal anchor Pith review arXiv

  16. [16]

    Advances in neural information processing systems , volume=

    Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Advances in neural information processing systems , volume=

  17. [17]

    The Innovation , year=

    A survey on llm-as-a-judge , author=. The Innovation , year=

  18. [18]

    arXiv preprint arXiv:2407.12687 , year=

    Towards responsible development of generative AI for education: An evaluation-driven approach , author=. arXiv preprint arXiv:2407.12687 , year=

    work page arXiv

  19. [19]

    Proceedings of the Eleventh ACM Conference on Learning@ Scale , pages=

    Autotutor meets large language models: A language model tutor with rich pedagogy and guardrails , author=. Proceedings of the Eleventh ACM Conference on Learning@ Scale , pages=

  20. [20]

    arXiv preprint arXiv:2306.10052 , year=

    Assigning AI: Seven approaches for students, with prompts , author=. arXiv preprint arXiv:2306.10052 , year=

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2407.05181 , year=

    Instructors as innovators: A future-focused approach to new AI learning opportunities, with prompts , author=. arXiv preprint arXiv:2407.05181 , year=

    work page arXiv

  22. [22]

    Communications of the ACM , volume=

    Computing education in the era of generative AI , author=. Communications of the ACM , volume=. 2024 , publisher=

    2024

  23. [23]

    International workshop on AI in education and educational research , pages=

    Enhancing critical thinking in education by means of a Socratic chatbot , author=. International workshop on AI in education and educational research , pages=. 2024 , organization=

    2024

  24. [24]

    , author=

    EULER: Fine-Tuning a Large Language Model for Socratic Interactions. , author=. AIxEDU@ AI* IA , volume=

  25. [25]

    Proceedings of the 23rd Koli calling international conference on computing education research , pages=

    Codehelp: Using large language models with guardrails for scalable support in programming classes , author=. Proceedings of the 23rd Koli calling international conference on computing education research , pages=

  26. [26]

    Proceedings of the 54th ACM Technical Symposium on Computer Science Education V

    Using large language models to enhance programming error messages , author=. Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 , pages=

  27. [27]

    Proceedings of the 2023 ACM Conference on International Computing Education Research-Volume 1 , pages=

    Exploring the responses of large language models to beginner programmers’ help requests , author=. Proceedings of the 2023 ACM Conference on International Computing Education Research-Volume 1 , pages=

    2023

  28. [28]

    arXiv preprint arXiv:2302.04662 , year=

    Generating high-precision feedback for programming syntax errors using large language models , author=. arXiv preprint arXiv:2302.04662 , year=

    work page arXiv

  29. [29]

    Proceedings of the 2023 ACM Conference on International Computing Education Research-Volume 2 , pages=

    Generative AI for programming education: Benchmarking ChatGPT, GPT-4, and human tutors , author=. Proceedings of the 2023 ACM Conference on International Computing Education Research-Volume 2 , pages=

    2023

  30. [30]

    arXiv preprint arXiv:2412.16429 , year=

    Learnlm: Improving gemini for learning , author=. arXiv preprint arXiv:2412.16429 , year=

    work page arXiv

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    SocraticLM: Exploring socratic personalized teaching with large language models , author=. Advances in Neural Information Processing Systems , volume=

  32. [32]

    Khanmigo: AI-powered tutor , year =

  33. [33]

    Learning and individual differences , volume=

    ChatGPT for good? On opportunities and challenges of large language models for education , author=. Learning and individual differences , volume=. 2023 , publisher=

    2023

  34. [34]

    Medical education , volume=

    Resolving the 50-year debate around using and misusing Likert scales , author=. Medical education , volume=. 2008 , publisher=

    2008

  35. [35]

    Advances in health sciences education , volume=

    Likert scales, levels of measurement and the “laws” of statistics , author=. Advances in health sciences education , volume=. 2010 , publisher=

    2010

  36. [36]

    Neuron , volume=

    The psychology and neuroscience of curiosity , author=. Neuron , volume=. 2015 , publisher=

    2015

  37. [37]

    Annual Review of Psychology , volume=

    Cognitive modeling using artificial intelligence , author=. Annual Review of Psychology , volume=. 2025 , publisher=

    2025

  38. [38]

    Proceedings of the National Academy of Sciences , volume=

    The neural architecture of language: Integrative modeling converges on predictive processing , author=. Proceedings of the National Academy of Sciences , volume=. 2021 , publisher=

    2021

  39. [39]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Rarely a problem? Language models exhibit inverse scaling in their predictions following few-type quantifiers , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  40. [40]

    Proceedings of the National Academy of Sciences , volume=

    Using cognitive psychology to understand GPT-3 , author=. Proceedings of the National Academy of Sciences , volume=. 2023 , publisher=

    2023

  41. [41]

    Nature Computational Science, 3 (10), 833-838 , author=

    Human-like intuitive behavior and reasoning biases emerged in large language models but disappeared in ChatGPT. Nature Computational Science, 3 (10), 833-838 , author=

  42. [42]

    Political Analysis , volume=

    Out of one, many: Using language models to simulate human samples , author=. Political Analysis , volume=. 2023 , publisher=

    2023

  43. [43]

    Behavior Research Methods , volume=

    Can large language models help augment English psycholinguistic datasets? , author=. Behavior Research Methods , volume=. 2024 , publisher=

    2024

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Enhancing personalized multi-turn dialogue with curiosity reward , author=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    International Conference on Human-Computer Interaction , pages=

    Generating Neurolinguistic Stimuli Using LLM Prompting , author=. International Conference on Human-Computer Interaction , pages=. 2025 , organization=

    2025

  46. [46]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Mind the gap: The divergence between human and llm-generated tasks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  47. [47]

    Proceedings of the 18th workshop on innovative use of NLP for building educational applications (BEA 2023) , pages=

    Generating better items for cognitive assessments using large language models , author=. Proceedings of the 18th workshop on innovative use of NLP for building educational applications (BEA 2023) , pages=

    2023

  48. [48]

    Computers & Education , volume=

    Do curious students learn more science in an immersive virtual reality environment? Exploring the impact of advance organizers and epistemic curiosity , author=. Computers & Education , volume=. 2022 , publisher=

    2022

  49. [49]

    , author=

    Examining the Relationship between Epistemic Curiosity and Achievement Goals. , author=. Eurasian Journal of Educational Research (EJER) , number=

  50. [50]

    Journal of Qualitative Research in Education , number=

    Does E-learning Trigger Epistemic Curiosity? , author=. Journal of Qualitative Research in Education , number=

  51. [51]

    Cognition & Emotion , volume=

    Epistemic curiosity, feeling-of-knowing, and exploratory behaviour , author=. Cognition & Emotion , volume=. 2005 , publisher=

    2005

  52. [52]

    Current opinion in behavioral sciences , volume=

    Epistemic curiosity and the region of proximal learning , author=. Current opinion in behavioral sciences , volume=. 2020 , publisher=

    2020

  53. [53]

    Symbol Emergence Systems: An Interdisciplinary Discussion about Cognition, Language and Society , pages=

    Curiosity and Exploration: Why Do We Want to Learn? , author=. Symbol Emergence Systems: An Interdisciplinary Discussion about Cognition, Language and Society , pages=. 2026 , publisher=

    2026

  54. [54]

    Learning and Individual Differences , volume=

    Achievement motivation and knowledge development during exploratory learning , author=. Learning and Individual Differences , volume=. 2015 , publisher=

    2015

  55. [55]

    Cognition and instruction , volume=

    Productive failure , author=. Cognition and instruction , volume=. 2008 , publisher=

    2008

  56. [56]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Towards human-centered proactive conversational agents , author=. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=