GENCNIPPET: Automated Generation of Code Snippets for Supporting Programming Questions

Chanchal K. Roy; Saikat Mondal

REVIEW 4 major objections 5 minor 31 references

GENCNIPPET: Automated Generation of Code Snippets for Supporting Programming Questions

T0 review · 4 major / 5 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 The paper proposes GENCNIPPET, a browser plug-in that uses a fine-tuned Llama-3 model to generate relevant example code snippets for Stack Overflow questions that lack them, aiming to help those questions get timely answers.

desk verdict A well-motivated proposal for generating missing code snippets on Stack Overflow, but the planned training and evaluation never actually test the missing-snippet scenario, and the paper contains no results. read the letter →

arxiv 2504.16292 v1 pith:EFQEOL5B submitted 2025-04-22 cs.SE

Saikat Mondal , Chanchal K. Roy This is my paper · ORCID

classification cs.SE

keywords StackOverflowquestionqualitycodesnippetsgenerationfine-tuningLlama-3browserpluginuserstudy

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 tackles a documented problem: Stack Overflow questions that omit code snippets receive acceptable answers far less often than questions that include them, and many askers omit code because of time pressure, employer policy, confidentiality, or uncertainty about what to share. It proposes GENCNIPPET, a browser plug-in that reads the question's problem description and programming language, sends them to a fine-tuned Llama-3 model, and returns a short example code snippet the asker can paste into the question. The authors plan to fine-tune on hundreds of thousands of high-scoring Java and Python questions, each paired with the single snippet it contains, and to evaluate output through similarity metrics, expert review of 400 samples, and live submissions of 50 generated snippets to Stack Overflow as suggested edits. The paper is a plan: it specifies the architecture and evaluation rather than reporting trained-model results. If the proposal works, it would lower a common barrier to getting help without replacing the human Q&A platform.

What carries the argument

The central object is the fine-tuned snippet generator: Llama-3-8B adapted with LoRA on structured pairs of (question text, language, date) → (code snippet), where the snippets come from positive-scoring questions that genuinely need code, as judged by an existing machine-learning filter. Around that core sits GENCNIPPET's two-part architecture: a browser-client script captures the problem description and language tag, a server builds a prompt asking for a concise example that demonstrates the problem 'including intentional faults,' and the model's output is returned to the user for pasting into the question. The evaluation machinery is three-layered: ROUGE, BLEU, and BERTScore for an initial similarity screen; a 400-sample expert review on a 5-point Likert scale; and a live test that submits generated snippets as Stack Overflow suggested edits and tracks acceptance and response times.

What would settle it

Execute 100 generated snippets against the problem descriptions they are meant to support and record how many reproduce the reported fault; a majority failing to reproduce it would falsify the central claim that GENCNIPPET generates relevant supporting examples.

Watch

Extended reading notes

Core claim

The paper's central claim is that an open-weights language model tailored to Stack Overflow's question style can generate a relevant, concise example snippet for a question that lacks one, and that supplying that snippet during submission improves the question's chances of a timely and appropriate answer. The authors ground this in earlier findings that only 23.8% of snippet-missing questions receive acceptable answers, versus 61.4% of questions that include code, and they design GENCNIPPET around the missing piece: a fine-tuned Llama-3-8B that takes a problem description, language tag, and date as input and produces 'Code: [code]' as output. The paper does not yet present evidence that the model can generate such snippets; it lays out how that evidence will be obtained, through automatic similarity scores, manual expert ratings, and a wild test in which 50 generated snippets are submitted as suggested edits on Stack Overflow.

Load-bearing premise

The plan rests on treating the single code snippet attached to each high-scoring question as the correct, representative example: if those snippets are arbitrary or unrepresentative, the model learns to mimic them rather than to produce something genuinely useful.

Editorial extensions

If this is right

If the fine-tuned model performs as planned, askers who cannot share real code can still post questions with a relevant example snippet, and the known gap in acceptable-answer rates between snippet-missing and snippet-including questions should narrow.
A working GENCNIPPET would strengthen Stack Overflow as a human-curated knowledge base instead of substituting automated answers for expert discussion, preserving the community-review mechanism.
The planned comparison against zero-shot and few-shot GPT-4 and CodeLlama would quantify how much domain-specific fine-tuning adds over general-purpose code generation for this task.
If community moderators accept a substantial share of the 50 wild-test suggested edits, that acceptance would be direct evidence that generated snippets meet real Stack Overflow quality expectations.

Reading between the lines

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

A stronger test than similarity metrics would be execution: run each generated snippet against the described problem and check whether it reproduces the reported error; the paper's planned evaluation does not include this check, and the training objective of matching one attached snippet is only a proxy for usefulness.
Because the prompt explicitly asks for code that contains the intentional fault, a natural extension is paired generation: produce the failing snippet together with a corrected version, which would serve diagnosis as well as the learning goals Stack Overflow is meant to support.
The same pipeline could generalize beyond Stack Overflow to issue trackers and discussion forums, where missing minimal reproductions are a common barrier; the paper limits itself to Java and Python questions on one platform.

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.

Desk Editor's Note

read the letter

Read this if you want to see a concrete LLM-for-SO proposal with a clear blind spot. The paper is a plan, not a study: no model, no experiments, no data analysis. The motivation is solid — questions missing code get far fewer answers, prior work identified them, and generating examples could help. Credit where due: the dataset construction is reasonable (filtering via a learned classifier, positive-score questions, single snippets for low noise), fine-tuning Llama-3 with LoRA is sensible, and the three-way evaluation (automatic metrics, manual review, wild testing with suggested edits) shows they understand the limits of similarity scores alone. The citation pattern is fine; the self-citations point to their own prior classifier, which is the natural basis.

The soft spot is load-bearing. The dataset consists of questions that already contain code snippets (Section III-A), and the fine-tuning pairs are "problem descriptions and their corresponding code snippets" (Section III-B). The paper never says code is stripped from the problem description before training. If it isn't, the model learns to copy the snippet, and automatic evaluation against that same snippet measures copying. If it is stripped, the descriptions are still written for questions that include code and often reference it, so they don't look like the real missing-code cases the tool targets. Either way, the proposed evaluation does not test the actual use case. Manual review and wild testing are good ideas but remain promises. This is not a minor flaw; it's the central claim.

Other soft spots are minor. Restricting to single-snippet positive-score questions may create a clean but unrepresentative training set. The comparison with GPT-4 is standard. The survey plan is detailed but 20 participants is small, and snowball sampling is biased; they acknowledge this.

Who is this for? Someone working on question-quality tools for Q&A forums. The paper gives a clear design to critique, but there is no result to verify. I would not cite it in the next year. If it were submitted as a full paper, I would desk reject it — not enough substance. The authors would be better served by a workshop or a registered report where reviewers can shape the experimental design before the work is done.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes GENCNIPPET, a browser plug-in for Stack Overflow (SO) that is intended to generate example code snippets for user questions that lack them. The authors plan to filter SO questions that require code snippets and have positive scores, fine-tune a Llama-3-8B model on problem-description/ground-truth-snippet pairs, and evaluate the generated snippets using automatic metrics (ROUGE, BLEU, BERTScore), manual expert review, and a live wild test. The paper presents dataset construction statistics (558,552 questions across Java and Python), a detailed fine-tuning and evaluation plan, the tool architecture, and a survey design for user studies. No experiments, model training, or evaluation results are reported; the entire contribution is a proposal in future tense.

Significance. If the proposed approach worked, GENCNIPPET could address a real problem documented in the software engineering literature: SO questions that miss required code snippets receive acceptable answers far less often (23.8% vs. 61.4%) and with greater delay. The paper is clearly motivated and the planned evaluation is unusually comprehensive, including user studies and wild testing. The authors also exhibit good awareness of threats to validity and limitations of automatic metrics. However, the manuscript contains no empirical evidence whatsoever, and, more importantly, the proposed training and evaluation protocol has a fundamental mismatch with the deployment scenario: the model is trained and evaluated on questions that already contain code snippets, but the tool is intended to serve questions that omit code. As submitted, the paper is a research plan rather than a completed study, so its significance cannot yet be assessed.

major comments (4)

[III-A, III-B, III-C] The training and evaluation protocol does not match the deployment use case. The dataset consists exclusively of questions that already contain code snippets (Table I: 'Questions with Code Snippets'), and the fine-tuning input/output pairs are 'problem descriptions and their corresponding code snippets' (Section III-B). The manuscript never states that code blocks are stripped from the problem description before being fed to the model. If code remains in the input, the model can learn to copy it, and the automatic evaluation against that same snippet (ROUGE/BLEU/BERTScore, Section III-C) would measure copying fidelity, not generation relevance. If code is stripped, the remaining descriptions are still written for questions that include code and often reference it (e.g., 'in the following code'), so they are not representative of descriptions of questions that omit code. The proposed protocol, as described, cannot support the central claim that GENCNIPPET will generate relevant snippets for questions that miss them.
[Entire manuscript] The paper contains no experiments, no results, and no validation. The abstract states 'We plan to introduce GENCNIPPET', and the methodology is phrased entirely in future tense ('We will fine-tune', 'We will evaluate', 'will be assessed'). The central claim that the fine-tuned model will generate relevant code snippets is asserted, not demonstrated. For a full research paper, this is a proposal rather than a completed study, and the reader cannot verify any of the claims about the tool's effectiveness.
[III-C] The automatic evaluation against a single ground-truth snippet is not a valid measure of snippet relevance for the intended task. For a given problem description, many different code snippets may be equally relevant and correct; ROUGE, BLEU, and BERTScore reward surface similarity to one reference and will penalize equally valid alternative snippets. The manual review of 400 samples is a reasonable complement, but it is still applied to questions that already have code, so it does not assess the tool's behavior on questions that lack code. The planned wild test on recent questions without code is the only evaluation component that matches the deployment scenario, but it is only a plan and is not reported.
[III-D] The comparison with foundation models (GPT-4, CodeLlama) uses 'the same set of questions from Section III-C', i.e., questions already containing code snippets. This evaluates the models' ability to reproduce existing snippets, not their ability to generate snippets for questions that miss them. Since the research questions RQ1 and RQ2 explicitly target 'questions that miss required code snippets', the proposed comparison does not directly answer those research questions.

minor comments (5)

[References] Reference [14] is cited as 'StackExchage. StackExchage API' with a URL to data.stackexchange.com; the spelling and the URL-to-citation match should be corrected.
[IV-A, Fig. 3] Figure 3 shows a mock interface with a small matplotlib example. It would help readers if the figure caption explained the source of this example and noted it is illustrative only, not an output of the proposed system.
[IV-C, Eq. (1)] The weighted Net Promoter Score formula uses response scores 2, 1, 0, -1, -2, which differs from the standard 0-10 NPS scale. The text should justify this choice or clarify the mapping to the standard NPS calculation.
[IV-C] The term 'Open Circular' is capitalized without explanation; this appears to be a label for open recruitment and should be described more precisely.
[III-A] The paper states 'we will incorporate more recent data up to the point when we begin fine-tuning', but the dataset statistics in Table I are fixed as of April 2024. It would be clearer to state whether the reported counts are final or preliminary.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the paper is a proposal whose planned held-out evaluation and wild testing are not reducible to its training inputs.

full rationale

This manuscript is a research proposal rather than a derivation: it describes a planned pipeline for fine-tuning Llama-3 on Stack Overflow questions that already contain code snippets and then evaluating the generated snippets. The potentially circular-looking elements are (1) training on questions that already contain snippets and evaluating against those same snippets, and (2) relying on the authors' own earlier SANER paper [10] for the code-snippet-necessity classifier and for motivating statistics. Neither reduces to a self-justifying loop. The evaluation plan uses a held-out test split (10% of 558,552 questions), manual expert review of 400 samples, and a wild test on 50 questions that originally lacked snippets, tracking community acceptance rate. These are external checks, not the training target re-labeled as a prediction. The citation to [10] is a peer-reviewed, published study external to this paper; the authors do not invoke a uniqueness theorem or define a quantity in terms of itself. The mismatch between training distribution (questions with snippets) and deployment distribution (questions without snippets) is a legitimate validity threat, but it is not circularity, because the paper does not claim that the held-out automatic metrics alone establish real-world relevance. Score 1 reflects the presence of self-citation in the methodology without any load-bearing circular reduction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper is a proposal, so all central quantities are planned, not measured. The main free parameters are the fine-tuning hyperparameters. The key assumptions are about the validity of the ground-truth snippets and the filtering model.

free parameters (4)

learning rate = 2e-5
Chosen for the planned fine-tuning; not derived from data, and affects whether the model converges.
batch size = 32
Selected for the planned fine-tuning; influences training dynamics and is set by hand.
epochs = 3 (with early stopping)
Selected for the planned fine-tuning; early stopping is also discretionary.
train/validation/test split ratio = 80/10/10
Chosen by hand; affects evaluation reliability and sample size.

assumptions (4)

domain assumption The ML filter from Mondal et al. [10] correctly identifies questions that require code snippets.
The dataset construction uses this filter to select training and evaluation questions; an inaccurate filter would contaminate the dataset.
domain assumption The existing code snippet in a high-scoring SO question is a valid ground-truth output for the generation task.
The model is trained to produce snippets matching the original snippet; this assumes the original snippet is the correct representative example.
ad hoc to paper Fine-tuning Llama-3 on the filtered SO data will produce a model that generalizes to unseen questions.
This is the core premise of the proposed approach; no pilot or experimental evidence is provided.
domain assumption Automatic similarity metrics (ROUGE, BLEU, BERTScore) and two-expert manual review adequately measure snippet relevance.
The evaluation plan relies on these measures; the paper acknowledges they may not fully capture relevance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GENCNIPPET: Automated Generation of Code Snippets for Supporting Programming Questions." pith.science (2026). https://pith.science/paper/EFQEOL5B

@misc{pith2026250416292,
  author       = {Pith},
  title        = {Pith review of: GENCNIPPET: Automated Generation of Code Snippets for Supporting Programming Questions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFQEOL5B}},
  note         = {Machine review of arXiv:2504.16292}
}

read the original abstract

Context: Software developers often ask questions on Technical Q&A forums like Stack Overflow (SO) to seek solutions to their programming-related problems (e.g., errors and unexpected behavior of code). Problem: Many questions miss required code snippets due to the lack of readily available code, time constraints, employer restrictions, confidentiality concerns, or uncertainty about what code to share. Unfortunately, missing but required code snippets prevent questions from getting prompt and appropriate solutions. Objective: We plan to introduce GENCNIPPET, a tool designed to integrate with SO's question submission system. GENCNIPPET will generate relevant code examples (when required) to support questions for their timely solutions. Methodology: We first downloaded the SO April 2024 data dump, which contains 1.94 million questions related to Python that have code snippets and 1.43 million questions related to Java. Then, we filter these questions to identify those that genuinely require code snippets using a state-of-the-art machine learning model. Next, we select questions with positive scores to ensure high-quality data. Our plan is to fine-tune Llama-3 models (e.g., Llama-3-8B), using 80% of the selected questions for training and 10% for validation. The primary reasons for choosing Llama models are their open-source accessibility and robust fine-tuning capabilities, which are essential for deploying a freely accessible tool. GENCNIPPET will be integrated with the SO question submission system as a browser plugin. It will communicate with the fine-tuned model to generate code snippets tailored to the target questions. The effectiveness of the generated code examples will be assessed using automatic evaluation against ground truth, user perspectives, and live (wild) testing in real-world scenarios.

Figures

Figures reproduced from arXiv: 2504.16292 by the authors.

**Figure 1.** Motivating examples A large number of SO questions discuss the programming problems (e.g., coding errors, unexpected behavior) that warrant code snippets for a resolution [8, 9]. Unfortunately, question submitters often miss the required code snippets, which prevents these questions from getting appropriate answers promptly [8, 10, 11, 12, 13]. For example, the question shown in [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗

**Figure 2.** Study methodology genuinely require code snippets by leveraging the state-of-theart machine learning model proposed by Mondal et al. [10]. To further ensure the quality and relevance of code snippets, we select questions with positive scores. To reduce noise, we initially excluded questions containing multiple code snippets. As a result, our dataset consists of 558,552 questions (242,494 in Java and 316,058 in Pyth… view at source ↗

**Figure 3.** Proposed GENCNIPPET interface. IV. GENCNIPPET: A TOOL FOR GENERATING EXAMPLE CODE SNIPPETS (RQ3) We propose GENCNIPPET, a web-based browser plugin designed to enhance SO question submission by generating example code snippets. Powered by a fine-tuned LLM, GENCNIPPET analyzes the problem descriptions in question texts and generates relevant code snippets to improve question clarity and quality. To assess its practi… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: An overview of the GENCNIPPET system architecture. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

[1]

How social q&a sites are changing knowledge sharing in open source software communities

Bogdan Vasilescu, Alexander Serebrenik, Prem Devanbu, and Vladimir Filkov. How social q&a sites are changing knowledge sharing in open source software communities. In ACM conference on Computer supported cooperative work & social computing , pages 342–354, 2014

work page 2014
[2]

What are developers talking about? an analysis of topics and trends in stack overflow

Anton Barua, Stephen W Thomas, and Ahmed E Has- san. What are developers talking about? an analysis of topics and trends in stack overflow. Empirical software engineering, 19:619–654, 2014

work page 2014
[3]

Mining stack overflow to turn the ide into a self-confident programming prompter

Luca Ponzanelli, Gabriele Bavota, Massimiliano Di Penta, Rocco Oliveto, and Michele Lanza. Mining stack overflow to turn the ide into a self-confident programming prompter. In 11th working conference on mining software repositories , pages 102–111, 2014

work page 2014
[4]

Asking and answering questions during a programming change task

Jonathan Sillito, Gail C Murphy, and Kris De V older. Asking and answering questions during a programming change task. IEEE Transactions on Software Engineer- ing, 34(4):434–451, 2008

work page 2008
[5]

Chatgpt vs llama: Impact, reliability, and challenges in stack overflow discussions

Leuson Da Silva, Jordan Samhi, and Foutse Khomh. Chatgpt vs llama: Impact, reliability, and challenges in stack overflow discussions. arXiv preprint arXiv:2402.08801, 2024

arXiv 2024
[6]

The impact of generative ai on critical thinking: Self-reported reductions in cognitive effort and confi- dence effects from a survey of knowledge workers

Hao-Ping Hank Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. The impact of generative ai on critical thinking: Self-reported reductions in cognitive effort and confi- dence effects from a survey of knowledge workers. 2025

work page 2025
[7]

Which is a better programming assistant? a comparative study between chatgpt and stack overflow

Jinrun Liu, Xinyu Tang, Linlin Li, Panpan Chen, and Yepang Liu. Which is a better programming assistant? a comparative study between chatgpt and stack overflow. arXiv preprint arXiv:2308.13851 , 2023

arXiv 2023
[8]

Treude, O

C. Treude, O. Barzilay, and M. Storey. How do program- mers ask and answer questions on the web?: Nier track. In Proc. ICSE, pages 804–807, 2011

work page 2011

Show all 31 references

[9]

Squire and C

M. Squire and C. Funkhouser. ” A bit of code”: How the stack overflow community creates quality postings. In Proc. HICSS, pages 1425–1434, 2014

2014
[10]

Can we identify stack overflow ques- tions requiring code snippets? investigating the cause & effect of missing code snippets

Saikat Mondal, Mohammad Masudur Rahman, and Chanchal K Roy. Can we identify stack overflow ques- tions requiring code snippets? investigating the cause & effect of missing code snippets. In International Confer- ence on Software Analysis, Evolution and Reengineering (SANER), 2024

2024
[11]

Calefato, F

F. Calefato, F. Lanubile, and N. Novielli. How to ask for technical help? evidence-based guidelines for writing questions on stack overflow. IST, 94:186–207, 2018

2018
[12]

Quality questions need quality code: Classifying code fragments on stack overflow

Maarten Duijn, Adam Kucera, and Alberto Bacchelli. Quality questions need quality code: Classifying code fragments on stack overflow. In 2015 IEEE/ACM 12th Working Conference on Mining Software Repositories , pages 410–413. IEEE, 2015

2015
[13]

Asaduzzaman, A

M. Asaduzzaman, A. S. Mashiyat, C. K. Roy, and K. A Schneider. Answering questions about unanswered questions of stack overflow. In Proc. MSR, pages 97–100, 2013

2013
[14]

StackExchage API, 2024

StackExchage. StackExchage API, 2024. URL http:// data.stackexchange.com/stackoverflow

2024
[15]

M. M. Rahman and C. K. Roy. An insight into the unresolved questions at stack overflow. In Proc. MSR , pages 426–429, 2015

2015
[16]

Is stack overflow obsolete? an empirical study of the characteristics of chatgpt answers to stack overflow questions

Samia Kabir, David N Udo-Imeh, Bonan Kou, and Tianyi Zhang. Is stack overflow obsolete? an empirical study of the characteristics of chatgpt answers to stack overflow questions. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1–17, 2024

2024
[17]

Can llm replace stack overflow? a study on robustness and reliability of large language model code generation

Li Zhong and Zilong Wang. Can llm replace stack overflow? a study on robustness and reliability of large language model code generation. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 21841–21849, 2024

2024
[18]

Investigating the utility of chatgpt in the issue tracking system: An exploratory study

Joy Krishan Das, Saikat Mondal, and Chanchal Roy. Investigating the utility of chatgpt in the issue tracking system: An exploratory study. In Proceedings of the 21st International Conference on Mining Software Reposito- ries, pages 217–221, 2024

2024
[19]

Why do developers engage with chatgpt in issue-tracker? investigating usage and reliance on chatgpt-generated code

Joy Krishan Das, Saikat Mondal, and Chanchal Roy. Why do developers engage with chatgpt in issue-tracker? investigating usage and reliance on chatgpt-generated code. In 32nd edition of the IEEE International Confer- ence on Software Analysis, Evolution, and Reengineering (SANE...

2025
[20]

Large language model vs

Sahrima Jannat Oishwee, Natalia Stakhanova, and Zadia Codabux. Large language model vs. stack overflow in addressing android permission related challenges. In Proceedings of the 21st International Conference on Mining Software Repositories , pages 373–383, 2024

2024
[21]

Answers or no answers: Studying question answerability in stack overflow

Alton YK Chua and Snehasish Banerjee. Answers or no answers: Studying question answerability in stack overflow. Journal of Information Science, 41(5):720–731, 2015

2015
[22]

Predict- ing the quality of questions on stackoverflow

Antoaneta Baltadzhieva and Grzegorz Chrupała. Predict- ing the quality of questions on stackoverflow. InProceed- ings of the international conference recent advances in natural language processing, pages 32–40, 2015

2015
[23]

Inferfix: End-to-end program repair with llms

Matthew Jin, Syed Shahriar, Michele Tufano, Xin Shi, Shuai Lu, Neel Sundaresan, and Alexey Svyatkovskiy. Inferfix: End-to-end program repair with llms. In Pro- ceedings of the 31st ACM Joint European Software Engi- neering Conference and Symposium on the Foundations of Softwar...

2023
[24]

Exploring parameter-efficient fine- tuning of large language model on automated program repair

Guochang Li, Chen Zhi, Jialiang Chen, Junxiao Han, and Shuiguang Deng. Exploring parameter-efficient fine- tuning of large language model on automated program repair. In Proceedings of the 39th IEEE/ACM Interna- tional Conference on Automated Software Engineering , pages 719–731, 2024

2024
[25]

Dlfix: Context-based code transformation learning for auto- mated program repair

Yi Li, Shaohua Wang, and Tien N Nguyen. Dlfix: Context-based code transformation learning for auto- mated program repair. In Proceedings of the ACM/IEEE 42nd international conference on software engineering , pages 602–614, 2020

2020
[26]

Cure: Code- aware neural machine translation for automatic program repair

Nan Jiang, Thibaud Lutellier, and Lin Tan. Cure: Code- aware neural machine translation for automatic program repair. In 2021 IEEE/ACM 43rd International Confer- ence on Software Engineering (ICSE), pages 1161–1173. IEEE, 2021

2021
[27]

Towards efficient fine-tuning of language models with organizational data for automated software review

Mona Nashaat and James Miller. Towards efficient fine-tuning of language models with organizational data for automated software review. IEEE Transactions on Software Engineering, 2024

2024
[28]

Fine-tuning large language models to improve accuracy and comprehensibility of automated code re- view

Yongda Yu, Guoping Rong, Haifeng Shen, He Zhang, Dong Shao, Min Wang, Zhao Wei, Yong Xu, and Juhong Wang. Fine-tuning large language models to improve accuracy and comprehensibility of automated code re- view. ACM transactions on software engineering and methodology, 2024

2024
[29]

Correlating automated and human evaluation of code documentation generation quality

Xing Hu, Qiuyuan Chen, Haoye Wang, Xin Xia, David Lo, and Thomas Zimmermann. Correlating automated and human evaluation of code documentation generation quality. ACM Transactions on Software Engineering and Methodology (TOSEM), 31(4):1–28, 2022

2022
[30]

Population research: convenience sampling strategies

Samuel J Stratton. Population research: convenience sampling strategies. Prehospital and disaster Medicine , 36(4):373–374, 2021

2021
[31]

Accessibility in software practice: A practitioner’s perspective

Tingting Bi, Xin Xia, David Lo, John Grundy, Thomas Zimmermann, and Denae Ford. Accessibility in software practice: A practitioner’s perspective. arXiv preprint arXiv:2103.08778, 2021

2021 arXiv

Pith tools

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

[1] [1]

How social q&a sites are changing knowledge sharing in open source software communities

Bogdan Vasilescu, Alexander Serebrenik, Prem Devanbu, and Vladimir Filkov. How social q&a sites are changing knowledge sharing in open source software communities. In ACM conference on Computer supported cooperative work & social computing , pages 342–354, 2014

work page 2014

[2] [2]

What are developers talking about? an analysis of topics and trends in stack overflow

Anton Barua, Stephen W Thomas, and Ahmed E Has- san. What are developers talking about? an analysis of topics and trends in stack overflow. Empirical software engineering, 19:619–654, 2014

work page 2014

[3] [3]

Mining stack overflow to turn the ide into a self-confident programming prompter

Luca Ponzanelli, Gabriele Bavota, Massimiliano Di Penta, Rocco Oliveto, and Michele Lanza. Mining stack overflow to turn the ide into a self-confident programming prompter. In 11th working conference on mining software repositories , pages 102–111, 2014

work page 2014

[4] [4]

Asking and answering questions during a programming change task

Jonathan Sillito, Gail C Murphy, and Kris De V older. Asking and answering questions during a programming change task. IEEE Transactions on Software Engineer- ing, 34(4):434–451, 2008

work page 2008

[5] [5]

Chatgpt vs llama: Impact, reliability, and challenges in stack overflow discussions

Leuson Da Silva, Jordan Samhi, and Foutse Khomh. Chatgpt vs llama: Impact, reliability, and challenges in stack overflow discussions. arXiv preprint arXiv:2402.08801, 2024

arXiv 2024

[6] [6]

The impact of generative ai on critical thinking: Self-reported reductions in cognitive effort and confi- dence effects from a survey of knowledge workers

Hao-Ping Hank Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. The impact of generative ai on critical thinking: Self-reported reductions in cognitive effort and confi- dence effects from a survey of knowledge workers. 2025

work page 2025

[7] [7]

Which is a better programming assistant? a comparative study between chatgpt and stack overflow

Jinrun Liu, Xinyu Tang, Linlin Li, Panpan Chen, and Yepang Liu. Which is a better programming assistant? a comparative study between chatgpt and stack overflow. arXiv preprint arXiv:2308.13851 , 2023

arXiv 2023

[8] [8]

Treude, O

C. Treude, O. Barzilay, and M. Storey. How do program- mers ask and answer questions on the web?: Nier track. In Proc. ICSE, pages 804–807, 2011

work page 2011