Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Investigating Co-Constructive Behavior of Large Language Models in Explanation Dialogues

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Prompting an LLM to explain co-constructively produces more interactive, engaging dialogues but leaves its monitoring and scaffolding too inconsistent to raise average objective understanding.

desk verdict A solid, transparent user study showing prompt-level co-constructive instructions shift LLM explainer behavior but do not reliably improve objective understanding; the behavioral result is credible, but the abstract's causal 'can improve' claim over-reaches the design. read the letter →

arxiv 2504.18483 v2 pith:Y2JM2ZO4 submitted 2025-04-25 cs.CL

Leandra Fichtel , Maximilian Spliethöver , Eyke Hüllermeier , Patricia Jimenez , Nils Klowait , Stefan Kopp , Axel-Cyrille Ngonga Ngomo , Amelie Robrecht
show 4 more authors
Ingrid Scharlau Lutz Terfloth Anna-Lisa Vollmer Henning Wachsmuth
This is my paper · ORCID
classification cs.CL
keywords co-constructiveexplanationsexplanationdialoguesmonitoringandscaffoldinglargelanguagemodelsuserstudyexplainableAIdialogueactsLlama3.1
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether an out-of-the-box instruction-tuned LLM can lead a co-constructive explanation dialogue, in which the explainer continuously monitors what the explainee understands and adapts explanations to close emerging gaps. In a user study with 277 participants and three topics, the authors compare a minimal 'explainer' prompt with an enhanced prompt that defines monitoring and scaffolding and instructs the model to apply them. The enhanced model asks more verification questions, tests prior knowledge more often, gives shorter responses, and is rated more co-constructive, and the dialogues become more interactive. Yet average objective understanding is no better than in the base setting, and readability-based scaffolding only succeeds for some topics. The paper concludes that prompting can induce some co-constructive surface behavior, but consistent monitoring and scaffolding that reliably helps every explainee remain out of reach for current LLMs.

What carries the argument

The central objects are the two processes that define co-constructive explanation: monitoring, the explainer's ongoing use of diagnostic and verification questions to identify the explainee's knowledge gap, and scaffolding, adjusting the explanation's content and complexity to that gap. The argument is carried by comparing two system prompts, a minimal base prompt and an enhanced prompt that defines these processes, and by measuring behavior in three ways: automatic annotation of dialogue acts and explanation moves, readability metrics (Gunning Fog Index, type-token ratio, Shannon entropy) as a proxy for scaffolding before and after the explainee signals non-understanding, and pre/post questionnaires that separate subjective understanding, objective comprehension, enabledness, and perceived co-constructiveness.

What would settle it

Hand-annotate a random sample of the 277 dialogues and compare the rates of check questions and prior-knowledge tests between settings; if human annotation finds no significant increase, the monitoring result collapses. A cheaper calculation: re-run the quantitative comparison of Section 6.2 with a classifier whose agreement with human labels is above 0.7 and check whether the significant differences survive.

Watch

Extended reading notes

Core claim

The paper's central claim is that LLMs are capable of partial, surface-level co-construction in explanation dialogues: explicit prompting produces measurable co-constructive behavior, mainly in the form of monitoring questions, and this behavior can engage explainees and improve understanding in selected cases, but it is applied inconsistently and does not translate into better average understanding. The study's numbers show the split clearly: participants in the enhanced setting interacted longer, sent more queries, and rated the LLM more co-constructive (4.1 vs. 3.7), and objective-comprehension scores spread into wider tails with more participants doing better or worse. But overall objective comprehension (74.6% vs. 75.1%) and enabledness (69.5% vs. 67.0%) are statistically similar between settings, and after an explainee signals non-understanding the enhanced model lowers explanation complexity for Sleep and Black holes but sharply increases it for Quarto. The authors interpret the missing average gain not as evidence against co-constructive explanations, but as evidence that the LLM's monitoring and scaffolding capabilities are the bottleneck.

Load-bearing premise

The claim that the enhanced LLM really monitors more rests on an automatic labeler that agrees with human annotations only 37 to 48 percent of the time, and the study treats those labels as reliable evidence for Llama 3.1 explanation dialogues.

Editorial extensions

If this is right

  • Explicit prompting alone can shift an LLM explainer from monologue toward dialogue: shorter turns, more verification questions, and more frequent tests of prior knowledge, with no fine-tuning.
  • Because average understanding does not improve, users should not yet expect prompt-instructed LLMs to reliably personalize explanations to their level.
  • Objective comprehension in the enhanced setting splits into wider tails, so co-constructive prompting helps some explainees while leaving others worse off; mean-based evaluation hides this.
  • Participants who were led to provide explanations themselves achieved higher objective comprehension, pointing to self-explanation as a concrete mechanism worth eliciting deliberately.
  • Scaffolding success varies by topic, so conclusions about an LLM's adaptivity drawn from a single topic may mislead.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An untested extension would keep the enhanced prompt but add an explicit rule that the explainer must elicit a brief self-explanation after each chunk; the paper's own correlation suggests this should raise objective comprehension, but the paper does not test it.
  • The low agreement of the automatic dialogue-act labeler suggests the monitoring differences in Section 6.2 may be partly an artifact; a hand-annotated replication is the natural next check.
  • Since the enhanced prompt produced wider outcome tails, stratifying participants by prior motivation and topic interest would likely show that co-constructive prompting amplifies motivation rather than compensating for it; the paper reports a motivation correlation but does not pursue this causal reading.
  • A multimodal version of the same study, with voice, face, or backchannel signals, could find stronger monitoring because the explainer would receive implicit cues that text-only chat withholds; the paper notes the unimodal limitation but does not quantify how much it costs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper studies whether an instruction-tuned LLM (Llama 3.1 70B) can act as a co-constructive explainer in explanation dialogues. In a user study with 277 participants after filtering, each participant chatted with the LLM under one of two system prompts: a minimal 'base' explainer prompt and an 'enhanced' prompt instructing the LLM to apply monitoring and scaffolding. The authors measure subjective and objective comprehension, enabledness, motivation, and perceived co-constructiveness, and they automatically annotate dialogue acts and explanation moves, compute readability metrics, and analyze nine dialogues qualitatively. They report that the enhanced prompt leads the LLM to ask more verification questions and to test prior knowledge more often, and that participants in the enhanced setting interact longer and produce more diverse dialogue acts. Subjective comprehension and motivation increase in both settings, objective comprehension is similar on average, but the distribution of objective scores has heavier tails in the enhanced setting, and scaffolding, measured via readability adjustment after signals of non-understanding, is inconsistent. The paper's central claim is that LLMs exhibit some co-constructive behaviors that foster engagement and can improve understanding, while their monitoring and scaffolding abilities remain limited.

Significance. If the central claim were fully supported, the paper would make a useful contribution to XAI and human-LLM interaction by showing that prompt-level instructions can shift an LLM's explanation behavior toward more interactive, verification-oriented dialogue, and by documenting the limits of that shift. The study is transparent: the code and data are released, the setup uses a fixed seed, the questionnaires were validated by domain experts, and the paper carefully discusses limitations. The behavioral findings about increased interaction, more check questions, and higher perceived co-constructiveness are plausible and interesting. However, the causal claim that co-constructive behavior 'can improve understanding' is not supported by the study design, as the authors themselves concede in Appendix A.5. The quantitative monitoring and scaffolding evidence also relies on an auto-annotator with low macro F1 and on a very small scaffolding sample. The paper is therefore best viewed as an exploratory evaluation of prompt-induced co-constructive surface behavior, with the understanding-improvement claim needing either a different experimental design or a substantial softening.

major comments (4)
  1. [Abstract and §5.3/Appendix A.5] The abstract's claim that co-constructive behaviors 'can improve understanding of a topic' is not testable from the reported design. Objective comprehension and enabledness are measured only after the interaction, and there is no non-interaction control condition. The authors state in Appendix A.5 that 'to draw conclusions about the understanding gained during the interaction, two different questionnaires of equal complexity would be required' and that this is 'out of scope for our study.' Consequently, the improvement component of the central claim is unsupported, and the subjective comprehension gains reported in Table 3 are self-reported and increase in both settings. Please either add a no-interaction control or a pre-test with matched forms, or revise the abstract and conclusion to say that co-constructive prompting is associated with particular interaction behaviors and may support understanding in selected cases rather than that it improves understanding.
  2. [§6.2 and Appendix B.4] The quantitative evidence that the enhanced setting produces more co-constructive behavior rests on an automatically annotated dialogue-act and explanation-move classifier with macro F1 scores between 0.37 and 0.48 on its training corpora (Table 6). No validation on Llama 3.1 explanation dialogues is reported, and the domain shift from ELI-5 and Wired 5-Levels to this study's conversations could be substantial. The differences highlighted in Section 6.2 (e.g., Ask a check question 28.4 vs. 38.4, Test prior knowledge 15.8 vs. 19.3) may therefore be artifacts of annotation error rather than true behavioral differences. This is load-bearing for the monitoring claim. Please report human agreement on a sample of the study's dialogues or explicitly treat the automatic annotation results as exploratory, and adjust the strength of the wording accordingly.
  3. [§6.3 and Table 2] The scaffolding analysis is based on only 14 interactions (Quarto: 2, Sleep: 4, Black holes: 8), with extremely high variance, such as the Quarto readability increase of +109.4% ± 100.9. The statement that the enhanced LLM 'successfully reduces its explanation complexity for two topics' is not supported robustly by these data, and no significance test is reported. The authors do note the small sample, but the quantitative scaffolding claim still overstates the evidence. Please report per-dialogue values, provide nonparametric tests or confidence intervals, or recast this section as a qualitative observation about variability rather than a quantitative finding.
  4. [§6.5 and Figure 4] The correlations used to suggest a link between co-constructive behavior and objective understanding (Kendall's τ = 0.15 to 0.17 for informing statements, provide explanation, and participant self-explanations) are observational and cannot establish causal direction. The 'heavier tails' interpretation of Figure 4 is also purely descriptive, with no statistical test for differences in distribution shape. Because these correlational results are used to support the claim that co-constructive behavior can improve understanding, the causal language should be removed or supplemented with a proper causal or longitudinal design. In addition, the labels feeding these correlations come from the low-F1 classifier discussed above, which further weakens the inference.
minor comments (6)
  1. [Table 1] The column header 'Sent's' appears to be a typo for 'Sentences'; please correct it.
  2. [Figure 3 and Appendix B.4] The definition of the 'Rest' category is inconsistent: Figure 3 says labels with proportion smaller than 5% are grouped into Rest, while Appendix B.4 says the threshold is 2%. Please unify the definitions across the main text and appendix.
  3. [Table 3 and Figure 11] The composite scores for motivation, subjective comprehension, and co-constructiveness are averaged over Likert items, but no reliability statistics (e.g., Cronbach's alpha) are reported. Given that one co-constructiveness item was excluded after the fact, reporting scale reliability would strengthen the interpretation.
  4. [§6.1] The 'processing time' metric is described as the time to respond to the LLM's answer, which conflates reading time, thinking time, and typing time. Please clarify this in the text so readers do not interpret it as pure reading time.
  5. [Appendix B.5] The qualitative selection procedure says the authors 'filter the 25% best and worst participants' but does not specify how the composite criterion of subjective gain, objective comprehension, and co-constructiveness score was combined. Please describe the selection rule precisely.
  6. [Appendix A.2] The hyperparameters include temperature 1.0 and top-p 1.0, which makes exact reproduction unlikely despite the fixed seed. Please either state that the seed was verified to yield identical outputs or note that nondeterminism remains.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the study is an empirical prompt comparison whose self-referential measurement instruments are tools, not fitted predictions, and the acknowledged baseline gap is an evidentiary limitation rather than a circular reduction.

full rationale

The paper contains no derivation chain in which an output is equivalent to an input by construction. The central comparison is empirical: two system prompts (base vs. enhanced) are administered to Llama 3.1, and the resulting dialogue behavior and questionnaire outcomes are measured. The enhanced prompt explicitly instructs monitoring and scaffolding, but whether the model follows that instruction is an empirical question, and the paper's finding is that it does so partially and inconsistently; the behavior is therefore not defined into existence. The dialogue-act and explanation-move classifier (Alshomary et al., 2024, with one overlapping author) and the questionnaires (Buhl et al., 2025; Terfloth and Schaffer, 2025, also with an overlapping author) are measurement instruments, not fitted parameters or renamed predictions; their low macro-F1 (0.37-0.48) is a measurement-validity concern, not circularity. The abstract's clause that co-constructive behaviors 'can improve understanding' is not supported by the design, because objective understanding is measured only post-interaction and the paper itself states in Appendix A.5 that 'two different questionnaires of equal complexity would be required' to draw conclusions about understanding gained during the interaction, but this is an acknowledged evidentiary gap and overclaim rather than a reduction of the claim to its inputs. Accordingly, no circular step is identified.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

This is an empirical user study, not a derivation. There are no fitted numeric parameters and no invented entities. The load-bearing assumptions are measurement-level: the validity of the co-construction framework, the transferability of the dialogue-act classifier with macro F1 0.37-0.48, the validity of the questionnaires, and the use of readability indices as proxies for scaffolding. These assumptions determine whether the observed group differences should be read as evidence about co-constructive behavior.

assumptions (5)
  • domain assumption The co-constructive explanation framework of Rohlfing et al. (2021) and Wachsmuth and Alshomary (2022), including the categories monitoring, scaffolding, and explanation moves, is a valid operationalization for evaluating explanation quality.
    Used throughout Sections 3 and 5 to define what counts as co-constructive behavior and to choose annotation labels.
  • domain assumption The re-trained dialogue-act and explanation-move classifier by Alshomary et al. (2024) transfers to Llama 3.1 explanation dialogues with acceptable accuracy.
    Macro F1 is 0.37-0.48 (Appendix B.4), but the quantitative monitoring results in Section 6.2 rely on these labels.
  • domain assumption The questionnaire items from Buhl et al. (2025) and Terfloth and Schaffer (2025) validly measure subjective comprehension, objective comprehension, enabledness, and co-constructive behavior.
    Section 5.3 and Appendix A.5 use these scales without independent validation in this setting.
  • domain assumption Readability indices (Gunning Fog, Type-Token Ratio, Shannon Entropy) are valid proxies for scaffolding and adaptation of explanation complexity.
    Section 6.3 interprets decreases in these metrics as successful scaffolding.
  • standard math Statistical tests assume independent observations across participants and no unmodeled confounds from topic assignment.
    Mann-Whitney U and Kendall's tau tests are applied in Sections 5.4 and 6 without covariate adjustment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigating Co-Constructive Behavior of Large Language Models in Explanation Dialogues." pith.science (2026). https://pith.science/paper/Y2JM2ZO4

@misc{pith2026250418483,
  author       = {Pith},
  title        = {Pith review of: Investigating Co-Constructive Behavior of Large Language Models in Explanation Dialogues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y2JM2ZO4}},
  note         = {Machine review of arXiv:2504.18483}
}
read the original abstract

The ability to generate explanations that are understood by explainees is the quintessence of explainable artificial intelligence. Since understanding depends on the explainee's background and needs, recent research focused on co-constructive explanation dialogues, where an explainer continuously monitors the explainee's understanding and adapts their explanations dynamically. We investigate the ability of large language models (LLMs) to engage as explainers in co-constructive explanation dialogues. In particular, we present a user study in which explainees interact with an LLM in two settings, one of which involves the LLM being instructed to explain a topic co-constructively. We evaluate the explainees' understanding before and after the dialogue, as well as their perception of the LLMs' co-constructive behavior. Our results suggest that LLMs show some co-constructive behaviors, such as asking verification questions, that foster the explainees' engagement and can improve understanding of a topic. However, their ability to effectively monitor the current understanding and scaffold the explanations accordingly remains limited.

Figures

Figures reproduced from arXiv: 2504.18483 by the authors.

Figure 1
Figure 1. An explanation dialogue about the sleep cycle [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. System prompts used for the LLM in the base and enhanced settings. They include different instructions related to the LLM’s desired behavior. 4 Co-Constructive LLMs Conceptually, the introduced process of co￾constructing explanations appears to fit well the interaction capabilities of instruction-tuned LLMs: Although such LLMs are not inherently designed for the role of a co-constructive explainer, their pre￾trainin… view at source ↗
Figure 3
Figure 3. Proportions of annotated dialogue acts and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Results of the post objective comprehension [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The study application interface for answering [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: The study application interface for interacting [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The questionnaire to assess the participants’ [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The questionnaire to assess the participants’ [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: The questionnaire to assess the participants’ [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: The questionnaire to assess the participants’ [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: The questionnaire to assess the co￾constructive behavior of the LLM. The items are adopted from Buhl et al. (2025). The participants rate the statements on a five-point Likert scale. unchangeable, rather than telling the LLM to avoid long answers. Another participant …
Figure 12
Figure 12. Figure 12: Proportions of annotated dialogue acts and [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Vibe Coding vs. Agentic Coding: Fundamentals and Practical Implications of Agentic AI

    cs.SE 2025-05 conditional novelty 3.0 of 10

    A qualitative taxonomy positions vibe coding and agentic coding as complementary paradigms rather than rivals in AI-assisted software development.

Reference graph

Works this paper leans on

65 extracted references · 64 canonical work pages · cited by 1 Pith paper

  1. [1]

    I have prior knowledge of sleep cycles

    work page

  2. [3]

    At some point in the past, I did study sleep cycles

    work page

  3. [9]

    Figure 7: The questionnaire to assess the participants’ prior subjective comprehension of sleep cylces and the motivation to learn about it before interacting with the LLM

    I want to be acknowledged by others through my knowledge about sleep cycles. Figure 7: The questionnaire to assess the participants’ prior subjective comprehension of sleep cylces and the motivation to learn about it before interacting with the LLM. The statements are based on Buhl and Herrmann (2025) and are rated on a five-point Likert scale. Figure 10 ...

    work page 2025

  4. [10]

    I want to deepen my understanding of sleep cycles to be the best in the field

    work page

  5. [11]

    I want to be acknowledged by others through my knowledge about sleep cycles. Figure 8: The questionnaire to assess the participants’ post subjective comprehension of the explanandum re- lated to sleep and the motivation to learn about it after interacting with the LLM. The statements are based on Buhl and Herrmann (2025). The participants rate the stateme...

    work page 2025

  6. [12]

    I’m motivated to learn something new about sleep cycles

    work page

  7. [21]

    I’m a real expert on the topic of sleep cycles

    work page

  8. [22]

    I know a lot about sleep cycles

    work page

Show all 65 references
  1. [23]

    I understand what the sleep stages of a sleep cycle are

  2. [24]

    I can explain the sleep stages of a sleep cycle

  3. [25]

    Motivation

    I know the sleep stages of a sleep cycle. Motivation

  4. [26]

    I was motivated by the explanation to discover new connections between the concepts related to sleep cycles and its stages

  5. [27]

    I found the explanation to be very interesting

  6. [28]

    I had fun understanding the explanations

  7. [29]

    I find the topic of sleep cycles very exciting

  8. [30]

    I would also engage in the topic of sleep cycles in my free time

  9. [31]

    In general, I find the topic of sleep cycles boring

  10. [32]

    I have fun engaging in the topic of sleep cycles

  11. [33]

    I want to deepen my understanding of sleep cycles to know more about it

  12. [34]

    I want to deepen my understanding of sleep cycles to know more about it than others

  13. [37]

    The sleep cycle always runs in fixed phases

  14. [38]

    Sleep cycles can be altered by external influences such as light and temperature

  15. [39]

    Slow-wave sleep is crucial for physical recovery

  16. [40]

    Everyone needs the same amount of sleep

  17. [41]

    A complete sleep cycle lasts about 60 minutes

  18. [42]

    There are five sleep stages

  19. [43]

    Melatonin production is highest during REM sleep

  20. [44]

    We only dream during REM sleep

  21. [45]

    Our body temperature and heart rate drop during deep sleep

  22. [46]

    It’s harmful to wake up during REM sleep

  23. [47]

    Slow-wave sleep is dominated by delta waves with frequencies below 4 Hz

  24. [48]

    Melatonin and Cortisol are primarily involved in regulating the sleep-wake cycle

  25. [49]

    Serotonin is the primary neurotransmitter responsible for REM sleep

  26. [50]

    Figure 9: The questionnaire to assess the participants’ objective comprehension of the explanandum related to sleep after interacting with the LLM

    Alpha waves play a role in the initial sleep onset. Figure 9: The questionnaire to assess the participants’ objective comprehension of the explanandum related to sleep after interacting with the LLM. The statements are based on Terfloth and Schaffer (2025). The participants va...

    2025

  27. [51]

    They are likely experiencing restorative processes

  28. [52]

    They may be in a state of wakefulness or light sleep

  29. [53]

    A person's sleep was monitored

    They are in a slow-wave sleep with minimal brain activity. A person's sleep was monitored. The result is plotted in the sleep graph seen below. What sleep stage does the yellow segment (marked with an arrow) represent?

  30. [54]

    What does this pattern most likely indicate about their sleep architecture?

    Slow-wave sleep A person exhibits prolonged periods of slow-wave sleep during the first half of the night, with a gradual increase in REM duration towards the morning. What does this pattern most likely indicate about their sleep architecture?

  31. [55]

    A typical and healthy sleep cycle

  32. [56]

    Unusually low amount of slow-wave sleep

  33. [57]

    A lack of REM rebound If someone only gets about 5 hours of sleep per night, which sleep stages are most likely to be impacted?

  34. [58]

    REM sleep and slow-wave sleep

  35. [59]

    Light and moderate sleep

  36. [60]

    Only slow-wave sleep

  37. [61]

    All sleep stages equally Which of the following is a correct sequence of sleep stages in one complete sleep cycle?

  38. [62]

    Light sleep → Moderate sleep → REM → Slow-wave sleep

  39. [63]

    REM → Light sleep → Moderate sleep → Slow-wave sleep

  40. [64]

    Light sleep → Moderate sleep → Slow-wave sleep → REM

  41. [65]

    The questions are based on Terfloth and Schaffer (2025)

    Slow-wave sleep → Moderate sleep → Light sleep → REM Figure 10: The questionnaire to assess the participants’ enabledness of the explanandum related to sleep after interacting with the LLM. The questions are based on Terfloth and Schaffer (2025). For every question, the partic...

    2025

  42. [66]

    My dialogue partner carefully adapted their utterances to my utterances

  43. [67]

    My dialogue partner considered my understanding

  44. [68]

    While explaining, my dialogue partner met me halfway

  45. [69]

    My dialogue partner took my statements into account

  46. [70]

    The explanation was meant to encourage me to question my understanding

  47. [71]

    My dialogue partner intended to provide me with the opportunity to build an understanding of the topic by asking questions

  48. [72]

    My dialogue partner encouraged me to continuously think about further details of the topic

  49. [73]

    When learning about a new topic, it’s better to think about details yourself, rather than having everything fully explained

  50. [74]

    My dialogue partner encouraged me to visualize the different processes of the topic

  51. [75]

    While explaining, it was important to my dialogue partner to continuously consider if I understood the explanation

  52. [76]

    While explaining, it was important to my dialogue partner to monitor whether I understood everything

  53. [77]

    Figure 11: The questionnaire to assess the co- constructive behavior of the LLM

    My dialogue partner responded, when I signaled non- understanding. Figure 11: The questionnaire to assess the co- constructive behavior of the LLM. The items are adopted from Buhl et al. (2025). The participants rate the statements on a five-point Likert scale. unchangeable, r...

    2025

  54. [2001]

    Mama always had a way of explaining things so I could understand

    Fam: Ein fragebogen zur erfassung aktuller motivation in lern- und leistungssituationen. Diag- nostica, 47:57–66. Amelie S. Robrecht, Christoph R. Kowalski, and Stefan Kopp. 2025. SNAPE-PM: Building and Utilizing Dynamic Partner Models for Adaptive Explanation Generation. arXi...

    2025 arXiv

  55. [2010]

    In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Valletta, Malta

    Towards an ISO standard for dialogue act an- notation. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Valletta, Malta. European Language Re- sources Association (ELRA). Hendrik Buschmeier, Heike M. Buhl, Friederike Kern, ...

    2023

  56. [2011]

    In Towards Ubiqui- tous Learning, pages 461–466, Berlin, Heidelberg

    The effect of dynamic computerized scaffold- ing on collaborative discourse. In Towards Ubiqui- tous Learning, pages 461–466, Berlin, Heidelberg. Springer Berlin Heidelberg. Robert Nimmo, Marios Constantinides, Ke Zhou, Daniele Quercia, and Simone Stumpf. 2024. User Characteri...

    2024

  57. [2020]

    arXiv preprint

    Longformer: The Long-Document Trans- former. arXiv preprint. Meisam Booshehri, Hendrik Buschmeier, Philipp Cimi- ano, Stefan Kopp, Jaroslaw Kornowicz, Olesja Lam- mert, Marco Matarese, Dimitry Mindlin, Amelie So- phie Robrecht, Anna-Lisa V ollmer, Petra Wagner, and Britta Wred...

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.