Pith. sign in

REVIEW 4 major objections 42 references

Students who type prompts for LLM coding tasks succeed more often on the first try than those who submit unedited voice prompts.

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

2026-07-08 19:47 UTC pith:YYECRVCI

load-bearing objection Large-N exploratory CS-ed study: typed prompts beat unedited voice on first try for 2/3 Prompt Problems; free choice confounds causal modality claims, but the edited-voice result is the useful bit. the 4 major comments →

arxiv 2607.05808 v1 pith:YYECRVCI submitted 2026-07-07 cs.CY

Say What? Examining Text and Voice Input Modalities for Prompt-Based Programming in Computing Education

Kaitlin Riegel , Yan Cathy Hua , Paul Denny , Victor-Alexandru P\u{a}durean , Juho Leinonen , James Prather , Adish Singla This is my paper
classification cs.CY
keywords prompt-based programmingvoice inputtext inputcomputing educationlarge language modelsPrompt Problemsmultimodal interfacesintroductory programming
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 how voice versus text input changes outcomes when introductory programming students solve Prompt Problems—tasks where the work is writing a natural-language prompt that makes an LLM produce correct code. In a free-choice study of 919 students, typed text prompts were more likely to succeed on the first attempt than unedited voice transcripts for two of three problems; editing the transcribed voice prompt before submission removed that gap. Most students only tried and preferred text, while those who tried voice often mixed modalities rather than switching fully. The finding matters because voice interfaces are entering LLM tools fast, yet nearly all prior computing-education research has assumed typed interaction. Designers and instructors therefore need evidence about when voice helps, when it hurts, and how editing fits into the workflow.

Core claim

For two of three Prompt Problems, students who typed prompts were more likely to succeed on the first attempt than students who submitted unedited voice prompts; there was no first-attempt success difference when students edited their transcribed voice prompts before submission. Across problems, voice users varied in how they mixed modalities—suggesting a complementary rather than exclusive use—yet most students only tried and reported preferring text. Qualitative responses framed text as better for precision and revision and voice as faster or more natural for some, with drawbacks around transcription errors and environment.

What carries the argument

Prompt Problems: programming tasks that require students to craft natural-language prompts so an LLM generates correct code. The study compares three input paths—typed text, unedited voice transcription, and edited voice transcription—under free modality choice among N=919 introductory students, measuring first-attempt success, persistence, and self-reported preferences.

Load-bearing premise

That first-attempt success differences mainly come from the input channel itself rather than from which students freely chose each channel—voice users may already differ in skill, confidence, language, environment, or strategy.

What would settle it

A randomized trial that assigns the same students to forced text, forced unedited voice, or forced voice-then-edit on identical Prompt Problems and compares first-attempt success; if the text advantage vanishes under random assignment, free-choice self-selection—not modality—explains the observed gap.

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

If this is right

  • Multimodal LLM tools for intro programming should keep text as a reliable path and make post-transcription edit of voice prompts easy and default-friendly.
  • Instruction can treat voice as a complementary option rather than a drop-in replacement for typed prompting.
  • Students who only submit unedited voice may need scaffolding—edit prompts, quality checks—to match text first-attempt rates.
  • Preference patterns imply most intro students will stay with text unless tools lower the friction of voice-to-edit workflows.
  • Tool design should surface transcription quality and invite revision before code is generated.

Where Pith is reading between the lines

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

  • Free-choice designs can understate true modality effects if less confident or less fluent students self-select into voice; randomized assignment would isolate the channel.
  • The unedited-voice penalty may partly track speech-recognition quality and language fluency rather than “thinking in voice” itself—separating ASR error from prompt content is a clean next measurement.
  • If editing equalizes success, the educational lever may be revision habits more than the audio channel; teaching students to review transcripts could transfer to better prompt writing generally.
  • As on-device speech models improve, the unedited-voice gap may shrink, which would change both tool defaults and classroom norms.

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

4 major / 0 minor

Summary. This exploratory empirical study examines how introductory programming students (N=919) use text versus voice input when solving Prompt Problems—tasks that require writing natural-language prompts so an LLM generates correct code. Students freely chose and could switch between modalities. The authors compare first-attempt prompt success, persistence, and post-activity survey perspectives across text, unedited voice (ASR submitted as-is), and edited voice. For two of three problems, text users had higher first-attempt success than unedited-voice users; edited-voice success did not differ from text. Most students preferred and only tried text. Qualitative analysis of open responses describes perceived roles, advantages, and drawbacks of each modality. The paper discusses implications for multimodal LLM tools and instructional design in computing education.

Significance. Voice input for LLM-based programming tools is understudied relative to text, and the paper addresses a timely gap with a large authentic classroom sample and a clear operational distinction between unedited and edited voice. The descriptive pattern—that unedited ASR lags text while edited transcriptions do not—is useful for tool designers if framed carefully. The free-choice design and qualitative perspectives add ecological validity about preference and complementary use. Strengths include the large N, the three-problem series, and the explicit hedge that effects appear for only two of three problems. Significance for causal or prescriptive design claims is limited by self-selection; as an observational map of modality use and student perspectives, the contribution is solid and publishable with tighter claims and fuller reporting.

major comments (4)
  1. The free-choice design is load-bearing for any modality-effect or design implication. Students self-selected and switched modalities; unedited-voice users are a non-exchangeable minority who may differ in skill, confidence, language background, environment, or care. The abstract and results frame first-attempt success differences as modality contrasts, and the discussion draws tool/instructional implications. Without randomization, matching, propensity adjustment, or a clear within-subject modality contrast that holds student fixed, those gaps cannot cleanly support causal or design claims about input channel. Either add such analyses (e.g., within-student switches; covariates for prior performance/language) or systematically rewrite results and discussion in strictly observational language and drop or heavily qualify design prescriptions that assume channel effects.
  2. The edited-voice ≈ text pattern, if robust, suggests the gap may be transcription quality or cleanup rather than voice-as-prompting-modality per se. That mechanism is central to what tool builders should fix (ASR quality / edit UX vs. voice prompting pedagogy). The manuscript should test or at least structure the discussion around this alternative: e.g., edit distance / ASR error rates vs. success; comparison of edited-voice prompts to text prompts on linguistic features; and whether “unedited voice” failures are mostly ASR artifacts. Without that, the headline modality contrast risks being misread as a voice-prompting deficit when it may be a transcription pipeline issue.
  3. Success is defined on first-attempt prompt correctness for Prompt Problems, but the paper must fully specify the automated or human check that a generated program is correct, including test suites, partial credit, and how multi-file or multi-function outputs were judged. If success coding is automated against fixed tests, report pass criteria and any flaky LLM nondeterminism controls (temperature, retries, seed). If human-judged, report reliability. This is load-bearing for the primary outcome; incomplete operationalization undermines the N=919 success comparisons.
  4. Only two of three problems show the text > unedited-voice first-attempt pattern. The manuscript needs a problem-level account (difficulty, required prompt structure, length, domain vocabulary, ASR-hostile tokens) and should avoid pooling language that implies a general modality effect. Report per-problem Ns by modality arm, effect sizes with CIs, and multiplicity handling. If the third problem is a boundary condition, elevate that in the abstract and discussion rather than treating it as a minor hedge.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful, constructive review that correctly identifies both the contribution and the limits of our free-choice observational design. We agree that causal or strong design prescriptions are not warranted and will systematically reframe the abstract, results, and discussion in observational language, elevate the two-of-three problem pattern, and add the requested operational detail on success coding. We will also expand the discussion of the edited-voice ≈ text pattern as evidence that transcription quality (rather than voice-as-modality per se) is a primary practical bottleneck. We cannot add randomization or propensity adjustment post hoc, but we will report within-student switch patterns, per-problem Ns and effect sizes with CIs, and clearer multiplicity language. With these revisions the paper will present a tighter, more carefully hedged observational map of modality use that remains useful for tool designers and instructors.

read point-by-point responses

  1. Referee: The free-choice design is load-bearing for any modality-effect or design implication. Students self-selected and switched modalities; unedited-voice users are a non-exchangeable minority who may differ in skill, confidence, language background, environment, or care. The abstract and results frame first-attempt success differences as modality contrasts, and the discussion draws tool/instructional implications. Without randomization, matching, propensity adjustment, or a clear within-subject modality contrast that holds student fixed, those gaps cannot cleanly support causal or design claims about input channel. Either add such analyses (e.g., within-student switches; covariates for prior performance/language) or systematically rewrite results and discussion in strictly observational language and drop or heavily qualify design prescriptions that assume channel effects.

    Authors: We agree. The design is free-choice and observational; unedited-voice users are a self-selected minority, so between-modality first-attempt differences cannot support causal claims about input channel. We will systematically rewrite the abstract, results framing, and discussion in strictly observational language (e.g., “students who submitted unedited voice prompts were less likely to succeed on first attempt than students who submitted text” rather than modality-effect language). Design and instructional implications will be heavily qualified as hypotheses for future controlled work and as practical observations about complementary use and preference, not as prescriptions that assume channel effects. We will add descriptive within-student switch analyses (who switched, in which direction, and associated success patterns) and report available covariates (where collected) without claiming they identify causal effects. We cannot add randomization or propensity matching post hoc; that limitation will be stated explicitly in the limitations section. revision: yes

  2. Referee: The edited-voice ≈ text pattern, if robust, suggests the gap may be transcription quality or cleanup rather than voice-as-prompting-modality per se. That mechanism is central to what tool builders should fix (ASR quality / edit UX vs. voice prompting pedagogy). The manuscript should test or at least structure the discussion around this alternative: e.g., edit distance / ASR error rates vs. success; comparison of edited-voice prompts to text prompts on linguistic features; and whether “unedited voice” failures are mostly ASR artifacts. Without that, the headline modality contrast risks being misread as a voice-prompting deficit when it may be a transcription pipeline issue.

    Authors: We agree this alternative mechanism is central and under-emphasized. The edited-voice ≈ text pattern is consistent with transcription quality / cleanup mattering more than voice-as-prompting-modality per se. We will restructure the discussion around this interpretation and state clearly that the unedited-voice gap may largely reflect ASR artifacts rather than a deficit of spoken prompting. Where data allow, we will report descriptive comparisons of prompt length and surface features across text, unedited voice, and edited voice, and we will discuss qualitative evidence that students edited for ASR errors. We do not have ground-truth audio or forced-alignment ASR error rates for the full corpus, so we cannot compute edit distance against a gold transcript for every submission; we will state that limitation and treat the edited-vs-unedited contrast as the best available proxy. Tool-design implications will be reframed toward ASR quality and edit UX rather than assuming a voice-prompting pedagogy deficit. revision: yes

  3. Referee: Success is defined on first-attempt prompt correctness for Prompt Problems, but the paper must fully specify the automated or human check that a generated program is correct, including test suites, partial credit, and how multi-file or multi-function outputs were judged. If success coding is automated against fixed tests, report pass criteria and any flaky LLM nondeterminism controls (temperature, retries, seed). If human-judged, report reliability. This is load-bearing for the primary outcome; incomplete operationalization undermines the N=919 success comparisons.

    Authors: We agree that full operationalization of the primary outcome is required. Success was determined automatically by the Prompt Problems platform: the LLM-generated program was executed against a fixed, hidden test suite for that problem; a submission counted as successful only if all tests passed (no partial credit). Outputs were single-file solutions matching the problem’s expected interface; multi-file or multi-function cases were not used in this activity. We will add a dedicated subsection specifying the test-suite pass criteria, the LLM configuration used for generation (model, temperature, and any retry policy), and how nondeterminism was handled. We will also note that first-attempt success is defined on the first submitted prompt for that problem under the chosen modality labeling. These details will appear in Methods so that the N=919 comparisons rest on a fully specified outcome. revision: yes

  4. Referee: Only two of three problems show the text > unedited-voice first-attempt pattern. The manuscript needs a problem-level account (difficulty, required prompt structure, length, domain vocabulary, ASR-hostile tokens) and should avoid pooling language that implies a general modality effect. Report per-problem Ns by modality arm, effect sizes with CIs, and multiplicity handling. If the third problem is a boundary condition, elevate that in the abstract and discussion rather than treating it as a minor hedge.

    Authors: We agree and will elevate the two-of-three pattern rather than treating it as a minor hedge. The abstract and discussion will state up front that the text vs. unedited-voice first-attempt difference appeared for two of three problems and will characterize the third as a boundary condition. We will add a problem-level account covering relative difficulty, expected prompt structure and length, domain vocabulary, and features that may be ASR-hostile. Results will report per-problem Ns by modality (text / unedited voice / edited voice), effect sizes with confidence intervals, and explicit language that we do not pool into a single general modality effect; multiplicity will be acknowledged (three related comparisons) with appropriate caution. This keeps the contribution as a descriptive, problem-sensitive map rather than a blanket modality claim. revision: yes

Circularity Check

0 steps flagged

Empirical free-choice user study; success outcomes are externally checkable and not definitionally forced by inputs.

full rationale

This paper reports an exploratory classroom study of text vs. voice input for Prompt Problems (N=919), with free modality choice, first-attempt success of LLM-generated code, persistence, and survey/qualitative perspectives. The central claims are observational comparisons (e.g., typed prompts more often succeeded on first attempt than unedited voice for two of three problems; edited voice matched text) plus thematic findings about student preferences and process. Success is defined by whether the generated program is correct for the assigned Prompt Problem—an external, checkable criterion—not by a fitted parameter, self-referential normalization, or uniqueness theorem. There is no derivation chain of equations, no parameter fitted to a subset then re-presented as a prediction of a closely related quantity, no load-bearing uniqueness claim imported from the authors’ prior work, and no ansatz smuggled in via self-citation. Self-selection and free choice are validity/confounding concerns for causal modality-effect interpretations, not circularity of the reported results. Per the analyzer rules, honest non-finding is the correct outcome: score 0 with empty steps.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

Empirical education/HCI study. No free parameters in the mathematical sense; no invented physical entities. Load-bearing background is domain assumptions about Prompt Problems, LLM code generation as the success oracle, and automatic speech-to-text as the voice path. The free-choice design is a methodological assumption that weakens causal attribution of modality effects. Ledger stays lean because the contribution is observational comparison, not a fitted theory.

axioms (4)
  • domain assumption Prompt Problems are a valid and representative task for studying student–LLM interaction in introductory programming (success = natural-language prompt yields correct generated code).
    Central outcome metric and educational framing rest on this task genre; invoked throughout the study design as described in the abstract.
  • domain assumption Automatic transcription of voice input is accurate enough that unedited voice submissions meaningfully represent spoken prompting rather than pure ASR failure noise.
    The key contrast between unedited and edited voice prompts assumes transcription quality is a controllable intermediate, not the sole driver of failure; not independently validated in the abstract.
  • ad hoc to paper Students' free selection and switching among modalities yields interpretable group comparisons for modality effects on first-attempt success.
    Exploratory free-choice design is a paper-specific methodological premise; if self-selection dominates, comparative success claims about modality weaken.
  • domain assumption Standard educational research practices for surveys and qualitative thematic analysis of open responses are sufficient to support reported perspectives on modality roles.
    Qualitative claims about perceived roles, drawbacks, and advantages of voice vs text depend on unstated coding validity from the abstract alone.

pith-pipeline@v0.9.1-grok · 6395 in / 3008 out tokens · 64028 ms · 2026-07-08T19:47:44.371218+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) are increasingly integrated into computing education, yet nearly all prior research has focused on text-based interactions. As voice-enabled interfaces become more capable and more common, there is growing interest in understanding how voice input might shape students' use of LLM-powered tools. In this exploratory study, we investigated how introductory programming students interact with Prompt Problems, which are programming tasks that require crafting natural-language prompts to generate correct code. Students (N = 919) solved a series of Prompt Problems with the freedom to select or switch between text and voice input modalities. We collected their prompt submissions as well as post-activity survey responses, then analysed differences in prompt accuracy, persistence, and perspectives by modality. For two of the three problems, we found that students who typed their prompts using text were more likely to have those prompts succeed on the first attempt than students who submitted unedited voice prompts. There was no difference in success rate if students edited their transcribed voice prompts before submission. Across the problems, we found evidence that students who tried voice prompting varied in their usage of modality - perhaps indicating a complementary, or non-preferential approach. However, most students only tried and reported preferring text. Our qualitative analysis revealed how students' perceived the roles of voice and text input in shaping their problem-solving process, as well as the reported drawbacks and advantages of each modality. We discuss implications for future multimodal tools and instructional design in computing education.

Figures

Figures reproduced from arXiv: 2607.05808 by Adish Singla, James Prather, Juho Leinonen, Kaitlin Riegel, Paul Denny, Victor-Alexandru P\u{a}durean, Yan Cathy Hua.

Figure 1
Figure 1. Figure 1: The Prompt Programming platform: the Prompt Problem (left), including the required function signature and visual [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Student input modality preferences. 4.3 Students’ Perceptions of Prompting Modalities (RQ3) [PITH_FULL_IMAGE:figures/full_fig_p004_2.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

42 extracted references · 42 canonical work pages · 4 internal anchors

  1. [1]

    Kiran Bisra, Qing Liu, John C Nesbit, Farimah Salimi, and Philip H Winne. 2018. Inducing self-explanation: A meta-analysis.Ed. Psych. Rev.30, 3 (2018), 703–725

    work page 2018

  2. [2]

    Ulagamuthalvi

    Devineni Chandu, Devarasetty Sai Sriram, and V. Ulagamuthalvi. 2025. Voice- Controlled Smart Programming Assistant. In2025 7th Int. Conf. on Intelligent Sustainable Systems (ICISS). 79–84. doi:10.1109/ICISS63372.2025.11076259

    work page doi:10.1109/iciss63372.2025.11076259 2025

  3. [3]

    Fred D. Davis. 1989. Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology.MIS Quarterly13, 3 (1989), 319–340. http://www.jstor.org/stable/249008

    work page 1989

  4. [4]

    Promptly: Using Prompt Problems to Teach Learners How to Effectively Utilize AI Code Generators

    Paul Denny, Juho Leinonen, James Prather, Andrew Luxton-Reilly, Thezyrie Amarouche, Brett A. Becker, and Brent N. Reeves. 2023. Promptly: Using Prompt Problems to Teach Learners How to Effectively Utilize AI Code Generators. arXiv:2307.16364 [cs.HC] https://arxiv.org/abs/2307.16364

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    Becker, and Brent N

    Paul Denny, Juho Leinonen, James Prather, Andrew Luxton-Reilly, Thezyrie Amarouche, Brett A. Becker, and Brent N. Reeves. 2024. Prompt Problems: A New Programming Exercise for the Generative AI Era. InProceedings of the 55th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2024). ACM, 296–302. doi:10.1145/3626252.3630909

    work page doi:10.1145/3626252.3630909 2024

  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 on Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2024). ACM, 408–414. doi:10.1145/3649217.3653574

    work page doi:10.1145/3649217.3653574 2024

  7. [7]

    Xin Luna Dong, Seungwhan Moon, Yifan Ethan Xu, Kshitiz Malik, and Zhou Yu

    work page

  8. [8]

    InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23)

    Towards Next-Generation Intelligent Assistants Leveraging LLM Tech- niques. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). ACM, 5792–5793. doi:10.1145/3580305.3599572

    work page doi:10.1145/3580305.3599572

  9. [9]

    Yan Cathy Hua, Paul Denny, Jörg Wicker, and Katerina Taškova. 2024. A system- atic review of aspect-based sentiment analysis: domains, methods, and trends.Ar- tificial Intelligence Review57, 11 (Sept. 2024), 296. doi:10.1007/s10462-024-10906-z

    work page doi:10.1007/s10462-024-10906-z 2024

  10. [10]

    Yan Cathy Hua, Paul Denny, Jörg Wicker, and Katerina Taškova. 2025. Data- Efficient Adaptation and a Novel Evaluation Method for Aspect-based Sentiment Analysis. arXiv:2511.03034 [cs.CL] https://arxiv.org/abs/2511.03034

    work page arXiv 2025

  11. [11]

    Sven Jacobs and Natalie Kiesler. 2025. GenAI Voice Mode in Programming Education. InProceedings of the 25th Koli Calling International Conference on Computing Education Research (Koli Calling ’25). ACM, New York, NY, USA, Article 43, 8 pages. doi:10.1145/3769994.3770001

    work page doi:10.1145/3769994.3770001 2025

  12. [12]

    Rahul Jampala, Devisri Santosh Kola, Adithya Nagendra Gummadi, Meghana Bhavanam, and Ithaya Rani Pannerselvam. 2024. The Evolution of Voice As- sistants: From Text-to-Speech to Conversational AI. In2024 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT). 1332–1338. doi:10.1109/IDCIoT59759.2024.10467739

    work page doi:10.1109/idciot59759.2024.10467739 2024

  13. [13]

    Ericson, David Weintrop, and Tovi Grossman

    Majeed Kazemitabaar, Justin Chow, Carl Ka To Ma, Barbara J. Ericson, David Weintrop, and Tovi Grossman. 2023. Studying the effect of AI Code Generators on Supporting Novice Learners in Introductory Programming. InProceedings of the 2023 CHI Conf. on Human Factors in Computing Systems (CHI ’23). ACM, Article 455, 23 pages. doi:10.1145/3544548.3580919

    work page doi:10.1145/3544548.3580919 2023

  14. [14]

    Majeed Kazemitabaar, Runlong Ye, Xiaoning Wang, Austin Zachary 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 2024 CHI Conference on Human Factors in Computing Systems (CHI ’24). ACM. doi:10.1145/361...

    work page doi:10.1145/3613904.3642773 2024

  15. [15]

    Chris Kerslake, Paul Denny, IV Smith, David H., James Prather, Juho Leinonen, Andrew Luxton-Reilly, and Stephen MacNeil. 2024. Integrating Natural Language Prompting Tasks in Introductory Programming Courses. InProceedings of the 2024 on ACM Virtual Global Computing Education Conference V. 1 (SIGCSE Virtual 2024). ACM, 88–94. doi:10.1145/3649165.3690125

    work page doi:10.1145/3649165.3690125 2024

  16. [16]

    Leon Korkmaz, Susanne Schmidt, Lucie Kruse, and Frank Steinicke. 2024. I Talk-You Write. Exploring Asymmetric Text-and Voice-Based Communication Between Humans and Virtual Agents. InProceedings of the 12th International Conference on Human-Agent Interaction. 25–33

    work page 2024

  17. [17]

    Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, and Pattie Maes. 2025. Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task.arXiv preprint arXiv:2506.088724 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Ying Li, Runze Yang, Xiantao Zhang, Peng Shi, Da Yang, and Xuefei Huang. 2024. ProgMate: An Intelligent Programming Assistant Based on LLM. In2024 IEEE Frontiers in Education Conference (FIE). 1–5. doi:10.1109/FIE61694.2024.10893312

    work page doi:10.1109/fie61694.2024.10893312 2024

  19. [19]

    2012.Sentiment Analysis and Opinion Mining

    Bing Liu. 2012.Sentiment Analysis and Opinion Mining. Springer Cham. XIV + 167 pages. doi:10.1007/978-3-031-02145-9

    work page doi:10.1007/978-3-031-02145-9 2012

  20. [20]

    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). ACM, 750–756. doi:10.1145/3626252.3630938

    work page doi:10.1145/3626252.3630938 2024

  21. [21]

    Wenhan Lyu, Yimeng Wang, Tingting (Rachel) Chung, Yifan Sun, and Yixuan Zhang. 2024. Evaluating the Effectiveness of LLMs in Introductory Computer Science Education: A Semester-Long Field Study. InProceedings of the Eleventh ACM Conference on Learning @ Scale (L@S ’24). ACM, 63–74. doi:10.1145/3657604. 3662036

    work page doi:10.1145/3657604 2024

  22. [22]

    Amama Mahmood, Junxiang Wang, Bingsheng Yao, Dakuo Wang, and Chien- Ming Huang. 2025. User Interaction Patterns and Breakdowns in Conversing with LLM-Powered Voice Assistants.International Journal of Human-Computer Studies195 (Jan. 2025), 103406. doi:10.1016/j.ijhcs.2024.103406

    work page doi:10.1016/j.ijhcs.2024.103406 2025

  23. [23]

    Margulieux, James Prather, Brent N

    Lauren E. Margulieux, James Prather, Brent N. Reeves, Brett A. Becker, Gozde Cetin Uzun, Dastyni Loksa, Juho Leinonen, and Paul Denny. 2024. Self-Regulation, Self-Efficacy, and Fear of Failure Interactions with How Novices Use LLMs to Solve Programming Problems. InProceedings of the Innovation and Technology in Computer Science Education V. 1 (ITiCSE 2024...

    work page arXiv 2024

  24. [24]

    Hollebeek

    Cristina Mele, Tiziana Russo-Spena, Angelo Ranieri, Irene Di Bernardo, and Linda D. Hollebeek. 2025. Talking vs typing: how voice- vs text-based educational robots shape student engagement.Journal of Service Management(09 2025), 1–24. doi:10.1108/JOSM-02-2024-0082

    work page doi:10.1108/josm-02-2024-0082 2025

  25. [25]

    OpenAI. [n. d.]. OpenAI Developers: Create transcirption. https: //developers.openai.com/api/reference/resources/audio/subresources/ transcriptions/methods/create#audio-createtranscription-model

    work page

  26. [26]

    OpenAI. 2022. Introducing Whisper. https://openai.com/index/whisper/

    work page 2022

  27. [27]

    Victor-Alexandru Padurean, Paul Denny, Andrew Luxton-Reilly, Alkis Go- tovos, and Adish Singla. 2025. Interleaving Natural Language Prompting with Code Editing for Solving Programming Tasks with Generative AI Mod- els. arXiv:2509.14088 [cs.CY] https://arxiv.org/abs/2509.14088

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Reinhard Pekrun. 2006. The control-value theory of achievement emotions: Assumptions, corollaries, and implications for educational research and practice. Educational psychology review18, 4 (2006), 315–341

    work page 2006

  29. [29]

    Tung Phung, Heeryung Choi, Mengyan Wu, Adish Singla, and Christopher Brooks. 2025. Plan More, Debug Less: Applying Metacognitive Theory to AI- Assisted Programming Education. InArtificial Intelligence in Education - 26th International Conference, AIED 2025, Palermo, Italy (Lecture Notes in Computer Science, Vol. 15877). Springer, 3–17. doi:10.1007/978-3-0...

    work page doi:10.1007/978-3-031-98414-3_1 2025

  30. [30]

    Farman Ali Pirzado, Awais Ahmed, Román Alejandro Mendoza-Urdiales, and Hugo Terashima-Marin. 2024. Navigating the Pitfalls: Analyzing the Behavior of LLMs as a Coding Assistant for Computer Science Students—A Systematic Review of the Literature.IEEE Access12 (2024), 112605–112625. doi:10.1109/ ACCESS.2024.3443621

    work page arXiv 2024

  31. [31]

    James Prather, Brent N Reeves, Paul Denny, Juho Leinonen, Stephen MacNeil, Andrew Luxton-Reilly, João Orvalho, Amin Alipour, Ali Alfageeh, Thezyrie Amarouche, Bailey Kimmel, Jared Wright, Musa Blake, and Gweneth Barbre

    work page

  32. [32]

    InProc of the 27th Australasian Comp

    Breaking the Programming Language Barrier: Multilingual Prompting to Empower Non-Native English Learners. InProc of the 27th Australasian Comp. Ed. Conference. ACM, 74–84. doi:10.1145/3716640.3716649

    work page doi:10.1145/3716640.3716649

  33. [33]

    and Kimmel, Bailey and Wright, Jared and Briggs, Ben , year = 2024, month = aug, pages =

    James Prather, Brent N Reeves, Juho Leinonen, Stephen MacNeil, Arisoa S Ran- drianasolo, Brett A. Becker, Bailey Kimmel, Jared Wright, and Ben Briggs. 2024. The Widening Gap: The Benefits and Harms of Generative AI for Novice Pro- grammers. InProc of the 2024 ACM Conf. on Int. Comp. Ed. Research (ICER ’24). ACM, 469–486. doi:10.1145/3632620.3671116

    work page doi:10.1145/3632620.3671116 2024

  34. [34]

    Victor-Alexandru Pădurean, Paul Denny, Alkis Gotovos, and Adish Singla. 2025. Prompt Programming: A Platform for Dialogue-based Computational Problem Solving with Generative AI Models. InProc. of the 30th ACM Conference on Innovation and Technology in CS Education V. 1 (ITiCSE 2025). ACM, 458–464. doi:10.1145/3724363.3729094

    work page doi:10.1145/3724363.3729094 2025

  35. [35]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust Speech Recognition via Large-Scale Weak Supervision. InInternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Proceedings of Machine Learning Research), Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara En...

    work page 2023

  36. [36]

    Sherry Ruan, Jacob O Wobbrock, Kenny Liou, Andrew Ng, and James Landay

    work page

  37. [37]

    Speech is 3x faster than typing for english and mandarin text entry on mobile devices.arXiv preprint arXiv:1608.07323(2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  38. [38]

    Christine Rzepka, Benedikt Berger, and Thomas Hess. 2022. Voice assistant vs. Chatbot–examining the fit between conversational agents’ interaction modalities and information search tasks.Information Systems Frontiers24, 3 (2022), 839–856

    work page 2022

  39. [39]

    Brad Sheese, Mark Liffiton, Jaromir Savelka, and Paul Denny. 2024. Patterns of Student Help-Seeking When Using a Large Language Model-Powered Program- ming Assistant. InProc. of the 26th Australasian Comp. Ed. Conf. (ACE ’24). ACM, 49–57. doi:10.1145/3636243.3636249

    work page doi:10.1145/3636243.3636249 2024

  40. [40]

    2011.Cognitive Load Theory

    John Sweller, Paul Ayres, and Slava Kalyuga. 2011.Cognitive Load Theory. Springer, New York. doi:10.1007/978-1-4419-8126-4

    work page doi:10.1007/978-1-4419-8126-4 2011

  41. [41]

    Qwen Team. 2026. Qwen3.5: Accelerating Productivity with Native Multimodal Agents. https://qwen.ai/blog?id=qwen3.5

    work page 2026

  42. [42]

    Angela Zavaleta Bernuy, Naaz Sibia, Pan Chen, Jessica Jia-Ni Xu, Elexandra Tran, Runlong Ye, Viktoria Pammer-Schindler, Andrew Petersen, Joseph Jay Williams, and Michael Liut. 2024. Does the Medium Matter? An Exploration of Voice- Interaction for Self-Explanations. InProc. of the 2024 ACM Designing Interactive Systems Conf. (DIS ’24). ACM, 86–101. doi:10....

    work page doi:10.1145/3643834.3661596 2024