Pith. sign in

REVIEW 3 major objections 6 minor 31 references

LLM tutor response styles track whether students continue productively, with only small global differences but much larger gaps under high cognitive load and in debugging.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 14:35 UTC pith:3RQ5YZCU

load-bearing objection Solid observational map of which LLM response styles track next-turn student engagement in real programming help dialogues; small global effects, clearer context variation, and one real but not fatal labeling caveat. the 3 major comments →

arxiv 2607.09919 v1 pith:3RQ5YZCU submitted 2026-07-10 cs.HC

When LLM Tutoring Responses Work: Evidence from Student Programming Conversations

Mohammad Fahim Abrar , Shayla Sharmin , Roghayeh Leila Barmaki This is my paper
classification cs.HC
keywords LLM TutoringComputer Science EducationProgramming EducationStudent Help-SeekingResponse StrategiesConversational AnalysisHuman-AI InteractionLearning Analytics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks what kind of ChatGPT reply helps a programming student keep going productively after asking for help. From a course dataset it builds 16,851 assistant-response interactions, labels each student situation and state, each reply style, and the next student turn, then tests how those styles associate with productive versus unresolved continuation. Globally the association is statistically significant but small: verification feedback has the highest productive-continuation rate and direct answers the lowest. The spread across styles is tiny for low-confusion conceptual questions and largest when students show high cognitive load or are debugging. Situation-specific checks show patterns such as stepwise guidance followed by more confusion decrease in high-load code requests, and direct answers followed by more unresolved continuation in high-load debugging. The claim a reader should care about is that AI programming tutors need context-aware evaluation and design rather than a single global ranking of reply styles.

Core claim

Assistant response style is significantly associated with productive continuation and with unresolved continuation across the full set of interactions, though effect sizes are small. Verification feedback shows the highest productive-continuation rate (82.4 percent) and direct answers the lowest (62.7 percent). Effectiveness-score ranges are smallest in low-confusion conceptual contexts and largest in high-cognitive-load contexts; detailed comparisons further show that the same style is not followed by the same outcomes in every help-seeking situation.

What carries the argument

The effectiveness score ES = productive-continuation rate minus unresolved-continuation rate, computed per response style inside each help-seeking context (situation combined with student state), plus present-versus-absent risk differences tested with FDR-corrected Fisher exact tests. That score range is what reveals where style differences are small versus large.

Load-bearing premise

The central claim treats next-turn outcome labels produced mainly by an automated annotator, checked at 82 percent agreement on only one hundred samples, as reliable proxies for whether a reply actually helped the student continue productively.

What would settle it

Randomly assign response styles within matched student situations and states in a controlled tutoring study; if productive continuation, confusion decrease, and longer-term learning or debugging success show no reliable differences between verification feedback and direct answers—especially under high cognitive load and debugging—the claimed context-dependent association collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Global rankings of LLM reply styles are weak guides for programming tutors; context must be part of evaluation.
  • High cognitive load, high confusion, debugging, and code-request moments are where style choice is most consequential.
  • Verification feedback and explanation-before-answer are more often followed by productive continuation than bare direct answers.
  • Adaptive tutors that detect situation and state can select strategies rather than apply one fixed style.
  • Next-turn engagement metrics should sit alongside answer correctness when judging LLM tutoring quality.

Where Pith is reading between the lines

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

  • If next-turn productive continuation is a usable short proxy, real-time detectors of load and confusion could route high-load debugging turns away from direct answers toward stepwise or diagnostic replies.
  • Small global effect sizes imply most practical gain will come from context-sensitive routing, not from crowning a single best style for every student.
  • The same annotation-and-score pipeline could be run on other STEM tutoring dialogues to test whether high-load contexts again show the largest style ranges.
  • Assignment design that surfaces debugging or code-generation modes could trigger a temporary shift in the tutor’s default reply policy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper studies how ChatGPT tutoring response styles relate to students’ immediate next-turn engagement in authentic programming help-seeking dialogues. Using StudyChat, the authors construct 16,851 assistant-response interactions from 203 students, annotate student situation/state, assistant response style, and next-turn outcomes with Gemma 4 (human agreement 82%, κ=.74), and report global associations of response style with productive and unresolved continuation (χ² tests, small Cramer’s V) plus larger descriptive effectiveness-score ranges and FDR-corrected present/absent comparisons in high-load, high-confusion, debugging, and code-request contexts. The central claim is that response-style differences are statistically detectable but small globally, and that effectiveness is situation-dependent, supporting context-aware evaluation of AI tutors.

Significance. The question—what kind of LLM response is followed by productive student continuation in programming help—is timely and practically important for CS education and adaptive tutoring design. Strengths include use of a large naturalistic, course-embedded public dataset rather than synthetic prompts; transparent observational framing with reported effect sizes rather than overstated causal claims; FDR correction for many Fisher comparisons; and a clear move from global rankings to context-conditioned patterns (e.g., stepwise guidance vs. direct answers under high load/debugging). If the associations hold under stronger reliability and dependence-aware analysis, the work provides useful empirical priors for designing and evaluating context-sensitive LLM tutors. The contribution is primarily empirical association mining, not a new theory or causal identification strategy.

major comments (3)
  1. [Methods, Data Preparation; Results §4] All primary quantitative claims (global χ² associations, ranked productive/unresolved rates in Fig. 2, effectiveness-score ranges in Fig. 3, and Table 1 risk differences) rest on multi-class Gemma-4 labels for response style and next-turn outcomes. Human validation covers only 100 of 16,851 interactions (~0.6%), reports overall 82% agreement and κ=.74, and does not provide per-category agreement, confusion matrices, human–human reliability, or stratification by rare styles or high-load/debugging cells. Systematic category-specific noise could reorder the verification-feedback vs. direct-answer ranking and distort score ranges. Please expand validation (larger stratified sample), report category-level reliability, and ideally add a sensitivity check showing whether the main rankings and Table 1 patterns are stable under plausible label error.
  2. [Methods, Data Analysis; Results, Global Response-Style Patterns] The global χ² tests treat 16,851 assistant-response interactions as independent observations, but they come from only 203 students and 2,214 conversations. Within-student and within-conversation dependence will inflate χ² and understate uncertainty; with already small V (.078/.087), significance may be overstated. Please re-estimate associations with a dependence-aware approach (e.g., mixed-effects logistic models with student/conversation random effects, cluster-robust inference, or student-level bootstrap) and report whether the global style associations and key Table 1 contrasts remain after accounting for clustering.
  3. [Methods, Analysis Variables and Outcome Construction] Productive continuation is defined broadly (continuing the task, meaningful follow-up, applying the response, requesting verification, or indicating progress). Combined with generally high productive rates (often ~75–82%) and low unresolved rates, this breadth may compress style differences and make the outcome partly a measure of continued chat engagement rather than tutoring effectiveness. Please clarify coding decision rules with examples, report the prevalence of each productive sub-signal if available, and discuss how alternative, stricter operationalizations would change Fig. 2 and the effectiveness scores.
minor comments (6)
  1. [Methods, Data Analysis] State explicitly how many interactions lacked a following student turn and were excluded from outcome-rate calculations; selection into continued dialogue may bias productive/unresolved rates.
  2. [Figure 2] Figure 2 plots unresolved rates below zero for visual contrast; add a note in the caption that the negative axis is a display convention, not a signed rate.
  3. [Abstract; References [17]] The abstract and introduction cite StudyChat as UMass 2026 / McNichols et al. 2026 while the paper is framed for SIGCSE TS 2027; ensure citation status and dataset version are consistent and citable.
  4. [Methods, Data Preparation] Define or briefly exemplify secondary state labels (cognitive-load signal, context richness, answer-seeking level) so readers can judge construct validity without the annotation prompts.
  5. [Results, Global Response-Style Patterns / Figure 2] Report cell counts (n) for each response style in the global rate plot or an appendix table so readers can assess precision of the 82.4% vs. 62.7% comparison.
  6. [Abstract; Methods] Minor wording: abstract says “Gemma 4” while Methods says “Gemma4 with 26B parameters”—standardize the model name and cite the exact checkpoint if possible.

Circularity Check

0 steps flagged

No circularity: purely observational associations; ES = PCR − UCR is a transparent descriptive summary, not a forced prediction.

full rationale

The paper reports chi-square associations and descriptive rates between Gemma-4-annotated assistant response styles and next-turn student outcomes on the external StudyChat corpus. The only constructed quantities are the effectiveness score ES = PCR − UCR and the score range SR = max(ES) − min(ES); both are explicit arithmetic summaries of observed rates, not parameters fitted to data and then re-presented as predictions. There is no self-definitional equation, no fitted input called a prediction, no uniqueness theorem imported from the authors’ prior work, and no ansatz smuggled via self-citation. Related-work citations (including StudyChat itself) are to independent authors and serve as background, not load-bearing premises that force the reported contingency tables. The derivation chain is therefore self-contained observational statistics; any weakness lies in label reliability, not circular construction.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 2 invented entities

Empirical observational study; load-bearing elements are annotation reliability, outcome definitions, and analysis thresholds rather than free physical parameters or new ontological entities. The ledger records the modeling choices that convert raw dialogues into the reported associations.

free parameters (2)
  • minimum cell size for score-range calculation = 50
    Only response-style cells with ≥50 observations enter the max–min effectiveness-score range; the threshold is chosen by the authors and affects which contexts appear high-variation.
  • FDR correction threshold for Fisher comparisons
    Multiple-testing correction determines which present/absent risk differences are reported as significant; standard but still a free analysis choice.
axioms (3)
  • domain assumption Next-turn student message is a valid short-term proxy for whether an assistant response was productively helpful or left the issue unresolved.
    Stated in Methods (Analysis Variables and Outcome Construction) and Limitations; underpins all outcome rates.
  • domain assumption Gemma-4 structured JSON labels for situation, state, response style, and next-turn outcome are sufficiently accurate for aggregate statistical analysis after 82%/κ=.74 human validation on 100 samples.
    Core of the annotation pipeline (Data Preparation); entire 16k-row analysis rests on it.
  • ad hoc to paper The six focal help-seeking situations and seven response styles exhaust the pedagogically meaningful distinctions for the reported comparisons.
    Taxonomy defined by the authors; rarer labels collapsed to 'other'.
invented entities (2)
  • effectiveness score ES = productive-continuation rate − unresolved-continuation rate no independent evidence
    purpose: Single scalar summarizing favorable vs unfavorable next-turn balance within each context × style cell.
    Defined in Data Analysis; used to compute score ranges that drive the context-variation claim.
  • help-seeking context (situation × student-state combination) no independent evidence
    purpose: Stratification unit that reveals larger response-style differences under high load/confusion.
    Constructed by merging situation and state labels; central to RQ2.

pith-pipeline@v1.1.0-grok45 · 17285 in / 2562 out tokens · 37785 ms · 2026-07-14T14:35:47.728802+00:00 · methodology

0 comments
read the original abstract

As students increasingly use LLM tutors in computer science education, one question becomes especially important: what kind of response helps a student continue productively? Prior work has studied how students use LLMs in computer science education, but less is known about how tutoring response styles are associated with student follow-up across programming help-seeking contexts. This paper analyzes StudyChat (UMass, 2026), a public dataset of student and ChatGPT tutoring conversations from an artificial intelligence course. We transformed StudyChat into 16,851 assistant-response interactions from 203 students and 2,214 conversations. Using local LLM-assisted annotation with Gemma 4, we labeled student help-seeking situations, student state, assistant response style, and student next-turn outcome. Human validation showed 82\% agreement with the LLM-assisted labels (Cohen's $\kappa=.74$). We analyzed productive continuation and unresolved continuation across the full dataset and across help-seeking contexts. Globally, response style was significantly associated with productive continuation, $\chi^2(7)=100.39$, $p<.001$, $V=.078$, and unresolved continuation, $\chi^2(7)=125.77$, $p<.001$, $V=.087$, though effect sizes were small. Verification feedback had the highest productive-continuation rate (82.4\%), while direct answers had the lowest (62.7\%). Descriptively, response-style score ranges were smallest in low-confusion conceptual contexts (.017) and largest in high-cognitive-load contexts (.203). More detailed comparisons showed situation-dependent response patterns. For example, stepwise guidance was followed by greater confusion decrease in high-cognitive-load code requests, while direct answers were followed by more unresolved continuation in high-load debugging. These findings support context-aware evaluation and design of AI tutoring responses for programming education.

Figures

Figures reproduced from arXiv: 2607.09919 by Mohammad Fahim Abrar, Roghayeh Leila Barmaki, Shayla Sharmin.

Figure 1
Figure 1. Figure 1: Overview of the analysis pipeline. Authentic student-ChatGPT tutoring conversations from StudyChat were trans [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Productive continuation and unresolved continua [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Response-style effectiveness score ranges across [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

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

31 extracted references · 7 canonical work pages

  1. [1]

    McLaren, and Kenneth R

    Vincent Aleven, Ido Roll, Bruce M. McLaren, and Kenneth R. Koedinger. 2016. Help Helps, But Only So Much: Research on Help Seeking with Intelligent Tu- toring Systems.International Journal of Artificial Intelligence in Education26, 1 (2016), 205–223. doi:10.1007/s40593-015-0089-1

    doi:10.1007/s40593-015-0089-1 2016

  2. [2]

    Isaac Alpizar-Chacon and Hieke Keuning. 2025. Student’s Use of Generative AI as a Support Tool in an Advanced Web Development Course. InProceedings of the 30th ACM Conference on Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2025). 312–318. doi:10.1145/3724363.3729106

    doi:10.1145/3724363.3729106 2025

  3. [3]

    Patrick Bassner, Eduard Frankford, and Stephan Krusche. 2024. Iris: An AI- Driven Virtual Tutor for Computer Science Education. InProceedings of the 2024 Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2024). 394–400. doi:10.1145/3649217.3653543

    doi:10.1145/3649217.3653543 2024

  4. [4]

    Anael Kuperwajs Cohen, Alannah Oleson, and Amy J. Ko. 2024. Factors Influenc- ing the Social Help-Seeking Behavior of Introductory Programming Students in a Competitive University Environment.ACM Transactions on Computing Education 24, 1 (2024), 11:1–11:27. doi:10.1145/3639059

    work page doi:10.1145/3639059 2024

  5. [5]

    Paul Denny, Viraj Kumar, and Nasser Giacaman. 2023. Conversing with Copilot: Exploring Prompt Engineering for Solving CS1 Problems Using Natural Lan- guage. InProceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE ’23). 1136–1142. doi:10.1145/3545945.3569823

    doi:10.1145/3545945.3569823 2023

  6. [6]

    Paul Denny, Stephen MacNeil, Jaromir Savelka, Leo Porter, and Andrew Luxton- Reilly. 2024. Desirable Characteristics for AI Teaching Assistants in Programming Education. InProceedings of the 2024 Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2024). 408–414. doi:10.1145/3649217.3653574

    doi:10.1145/3649217.3653574 2024

  7. [7]

    Sidney D’Mello, Blair Lehman, Reinhard Pekrun, and Art Graesser. 2014. Confu- sion Can Be Beneficial for Learning.Learning and Instruction29 (2014), 153–170. doi:10.1016/j.learninstruc.2012.05.003

    work page doi:10.1016/j.learninstruc.2012.05.003 2014

  8. [8]

    Matthew Frazier, Kostadin Damevski, and Lori Pollock. 2024. Customizing Chat- GPT to Help Computer Science Principles Students Learn Through Conversation. InProceedings of the 2024 Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2024). 633–639. doi:10.1145/3649217.3653570

    doi:10.1145/3649217.3653570 2024

  9. [9]

    John Hattie and Helen Timperley. 2007. The Power of Feedback.Review of Educational Research77, 1 (2007), 81–112. doi:10.3102/003465430298487

    doi:10.3102/003465430298487 2007

  10. [10]

    Arto Hellas, Juho Leinonen, Sami Sarsa, Charles Koutcheme, Lilja Kujanpää, and Juha Sorva. 2023. Exploring the responses of large language models to beginner programmers’ help requests. InProceedings of the 2023 ACM Conference on International Computing Education Research-Volume 1. 93–105. doi:10.1145/ 3568813.3600139

    arXiv 2023

  11. [11]

    Henley, Paul Denny, Michelle Craig, and Tovi Grossman

    Majeed Kazemitabaar, Runlong Ye, Xiaoning Wang, Austin Z. Henley, Paul Denny, Michelle Craig, and Tovi Grossman. 2024. CodeAid: Evaluating a Classroom Deployment of an LLM-based Programming Assistant that Balances Student and Educator Needs. InProceedings of the CHI Conference on Human Factors in Computing Systems (CHI ’24). 650:1–650:20. doi:10.1145/3613...

    doi:10.1145/3613904.3642773 2024

  12. [12]

    Natalie Kiesler, Dominic Lohr, and Hieke Keuning. 2023. Exploring the potential of large language models to generate formative programming feedback. In2023 IEEE frontiers in education conference (FIE). IEEE, 1–5. doi:10.1109/FIE58773.2023. 10343457

    doi:10.1109/fie58773.2023 2023

  13. [13]

    Natalie Kiesler and Daniel Schiffner. 2023. Large Language Models in Introductory Programming Education: ChatGPT’s Performance and Implications for Assess- ments.arXiv preprint arXiv:2308.08572(2023). doi:10.48550/arXiv.2308.08572

    Pith/arXiv arXiv doi:10.48550/arxiv.2308.08572 2023

  14. [14]

    Oka Kurniawan, Cyrille Jégourel, Norman Tiong Seng Lee, Matthieu De Mari, and Christopher M. Poskitt. 2022. Steps Before Syntax: Helping Novice Programmers Solve Problems Using the PCDIT Framework. InProceedings of the Annual Hawaii International Conference on System Sciences. Hawaii International Conference on System Sciences. doi:10.24251/HICSS.2022.121

    work page doi:10.24251/hicss.2022.121 2022

  15. [15]

    Rongxin Liu, Carter Zenke, Charlie Liu, Andrew Holmes, Patrick Thornton, and David J. Malan. 2024. Teaching CS50 with AI: Leveraging Generative Artificial Intelligence in Computer Science Education. InProceedings of the 55th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2024). 750–756. doi:10.1145/3626252.3630938

    doi:10.1145/3626252.3630938 2024

  16. [16]

    Jakub Macina, Nico Daheim, Sankalan Pal Chowdhury, Tanmay Sinha, Manu Kapur, Iryna Gurevych, and Mrinmaya Sachan. 2023. MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems. InFindings of the Association for Computational Linguistics: EMNLP

    2023

  17. [17]

    doi:10.18653/v1/2023.findings-emnlp.372

    5602–5621. doi:10.18653/v1/2023.findings-emnlp.372

    doi:10.18653/v1/2023.findings-emnlp.372 2023

  18. [18]

    Hunter McNichols, Fareya Ikram, and Andrew S. Lan. 2026. The StudyChat Dataset: Analyzing Student Dialogues With ChatGPT in an Artificial Intelligence Course. InProceedings of the 16th International Learning Analytics and Knowledge Conference (LAK 2026). 53–63. doi:10.1145/3785022.3785029

    work page doi:10.1145/3785022.3785029 2026

  19. [19]

    Valeria Ramirez Osorio, Angela Zavaleta Bernuy, Bogdan Simion, and Michael Liut. 2025. Understanding the Impact of Using Generative AI Tools in a Database Course. InProceedings of the 56th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE TS 2025). 959–965. doi:10.1145/3641554.3701785

    work page doi:10.1145/3641554.3701785 2025

  20. [20]

    Fred Paas, Alexander Renkl, and John Sweller. 2003. Cognitive Load Theory and Instructional Design: Recent Developments.Educational Psychologist38, 1 (2003), 1–4. doi:10.1207/S15326985EP3801_1

    doi:10.1207/s15326985ep3801_1 2003

  21. [21]

    Jacob Penney, Pawan Acharya, Peter Hilbert, Priyanka Parekh, Anita Sarma, Igor Steinmacher, and Marco Aurélio Gerosa. 2025. Understanding Programming Students’ Help-Seeking Preferences in the Era of Generative AI. InCompEd 2025: Proceedings of the ACM Global Computing Education Conference. 15–21. doi:10.1145/3736181.3747165

    doi:10.1145/3736181.3747165 2025

  22. [22]

    Tung Phung, Heeryung Choi, Mengyan Wu, Christopher Brooks, Sumit Gulwani, and Adish Singla. 2026. Closing the Loop: An Instructor-in-the-Loop AI Assis- tance System for Supporting Student Help-Seeking in Programming Education. In Proceedings of the 57th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2026). 852–858. doi:10.1145/3770762.3772612

    work page doi:10.1145/3770762.3772612 2026

  23. [23]

    Price, Zhongxiu Liu, Veronica Cateté, and Tiffany Barnes

    Thomas W. Price, Zhongxiu Liu, Veronica Cateté, and Tiffany Barnes. 2017. Factors Influencing Students’ Help-Seeking Behavior while Programming with Human and Computer Tutors. InProceedings of the 2017 ACM Conference on International Computing Education Research (ICER ’17). 127–135. doi:10.1145/ 3105726.3106179

    arXiv 2017

  24. [24]

    Lianne Roest, Hieke Keuning, and Johan Jeuring. 2024. Next-step hint generation for introductory programming using large language models. InProceedings of the 26th Australasian Computing Education Conference. 144–153. doi:10.1145/3636243. 3636259

    work page doi:10.1145/3636243 2024

  25. [25]

    Griswold, and Adalbert Gerald Soosai Raj

    Anshul Shah, Anya Chernova, Elena Tomson, Leo Porter, William G. Griswold, and Adalbert Gerald Soosai Raj. 2025. Students’ Use of GitHub Copilot for Working with Large Code Bases. InProceedings of the 56th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE TS 2025). 1050–1056. doi:10.1145/3641554.3701800

    doi:10.1145/3641554.3701800 2025

  26. [26]

    Brad Sheese, Mark Liffiton, Jaromir Savelka, and Paul Denny. 2024. Patterns of student help-seeking when using a large language model-powered programming assistant. InProceedings of the 26th Australasian computing education conference. 49–57. doi:10.1145/3636243.3636249

    doi:10.1145/3636243.3636249 2024

  27. [27]

    Valerie J. Shute. 2008. Focus on Formative Feedback.Review of Educational Research78, 1 (2008), 153–189. doi:10.3102/0034654307313795

    doi:10.3102/0034654307313795 2008

  28. [28]

    John Sweller. 1988. Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science12, 2 (1988), 257–285. doi:10.1207/s15516709cog1202_4

    doi:10.1207/s15516709cog1202_4 1988

  29. [29]

    Bruner, and Gail Ross

    David Wood, Jerome S. Bruner, and Gail Ross. 1976. The Role of Tutoring in Problem Solving.Journal of Child Psychology and Psychiatry17, 2 (1976), 89–100. doi:10.1111/j.1469-7610.1976.tb00381.x

    doi:10.1111/j.1469-7610.1976.tb00381.x 1976

  30. [30]

    Stephanie Yang, Hanzhang Zhao, Yudian Xu, Karen Brennan, and Bertrand Schnei- der. 2024. Debugging with an AI Tutor: Investigating Novice Help-seeking Behaviors and Perceived Learning. InProceedings of the 2024 ACM Confer- ence on International Computing Education Research V. 1 (ICER 2024). 84–94. doi:10.1145/3632620.3671092

    doi:10.1145/3632620.3671092 2024

  31. [31]

    J. D. Zamfirescu-Pereira, Laryn Qi, Björn Hartmann, John DeNero, and Narges Norouzi. 2025. 61A Bot Report: AI Assistants in CS1 Save Students Homework Time and Reduce Demands on Staff. (Now What?). InProceedings of the 56th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE TS 2025). 1309–1315. doi:10.1145/3641554.3701864

    doi:10.1145/3641554.3701864 2025