Pith. sign in

REVIEW 2 major objections 5 minor 53 references

AgenTag: Attribution of AI Coding Agents from Behavioral Fingerprints

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Identify the AI agent behind a pull request from text alone

desk verdict Solid, careful empirical work with a genuinely new text-over-code finding, but the headline attribution claim is only demonstrated for autonomous, self-identifying agents; the human-mediated case, which motivates the paper, is barely above chance. read the letter →

arxiv 2608.00966 v1 pith:PA3WFDTU submitted 2026-08-02 cs.SE

Taher A. Ghaleb This is my paper · ORCID
classification cs.SE
keywords AIcodingagentsauthorshipattributionopen-setrecognitioncontrastivelearningpullrequestsbehavioralfingerprintsfew-shotenrollmentGitHub
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 claims that AI coding agents leave distinctive behavioral fingerprints in the pull requests they produce, and that these fingerprints are enough to identify which agent wrote a change—even when the agent is new, undisclosed, or running under a human account. AgenTag, the proposed method, learns a metric space in which pull requests from the same author cluster, and on 33,580 agent-authored and 6,618 human-authored PRs it identifies the authoring agent with a weighted F1 of 0.96 (macro F1 of 0.84), separates AI from human authorship with a balanced F1 of 0.89, and flags previously unseen agents with an AUC of 0.84. The paper's central and somewhat counterintuitive finding is that the identifying signal lives almost entirely in PR descriptions and commit messages, while code diffs contribute little across four code representations. The fingerprint also persists after explicit self-disclosure markers (trailers, tool URLs, vendor templates, and even the agent's own name) are stripped, showing that attribution reads latent style rather than a watermark.

What carries the argument

The load-bearing object is a supervised contrastive metric space: a two-layer projection head maps multimodal PR representations to a 128-dimensional embedding trained with supervised contrastive loss, pulling same-agent PRs together and pushing different agents apart. Each known agent is represented by the centroid (prototype) of its training embeddings, and classification is nearest-prototype by cosine similarity. This geometry supplies the open-world abilities: an input whose maximum similarity to any prototype falls below a threshold is flagged as an unseen agent, and a new agent is enrolled by adding a centroid computed from ten labeled PRs, with no retraining. A second mechanism is equ

What would settle it

Take a held-out set of human PRs whose authorship is verified by direct contact with contributors, remove any agent-assisted cases, and re-measure AI-versus-human balanced F1; if it falls well below 0.89, part of the reported separation was contamination. Alternatively, enroll a sixth coding agent that is instructed via its configuration files to write terse, human-style PR descriptions and commit messages, then check whether unseen-agent detection AUC drops from 0.84 toward chance.

Watch

Extended reading notes

Core claim

AgenTag frames agent attribution as an open-world problem, not a closed roster. It represents each PR as four streams—title/body text, commit messages, code diffs, and 41 behavioral features—and trains a supervised contrastive embedding so same-agent PRs cluster. Inference assigns a PR to the nearest agent prototype; if the nearest prototype is too far, the PR is rejected as unknown; a new agent can be enrolled by averaging ten labeled PRs. On AIDev, the two text streams together reach 0.890 weighted F1, nearly matching full fusion at 0.909, while code-diff embeddings alone reach only 0.589 and adding them does not help. The same modality ordering holds for AI-versus-human detection (balance

Load-bearing premise

AIDev's ground-truth labels for the five agents are accurate and, except for the explicitly handled Claude Code case, independent of the textual markers AgenTag reads; if unobserved label noise affects agent labels even at the same modest rate as the 5% contamination found in human labels, every reported F1 and AUC is an upper bound on true attribution accuracy.

Editorial extensions

If this is right

  • A maintainer can screen a PR at open time using only its title and description, and flag likely agent authorship without fetching commits or diffs.
  • Researchers should re-validate submitter-identity labels in datasets such as AIDev: at least 5% of 'human' PRs carry explicit AI markers, and a marker-blind model concentrates disclosed cases nearly eightfold at high confidence.
  • An attribution system can be built to expect new agents: a never-before-seen agent is detected (AUC 0.84) and can be enrolled from ten labeled PRs, so the system does not need retraining for each new tool.
  • Code stylometry methods tuned to human developers will not transfer to AI agents: the paper finds the ordering is inverted, with communication style carrying the signal and code content contributing little.
  • The fingerprints are not permanent identities: macro F1 drops from 0.805 to 0.749 on a chronological split, so deployed attributors need periodic revalidation as agents update their templates and models.

Reading between the lines

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

  • If the behavioral fingerprint lives mainly in descriptions and commit style, then a low-effort evasion path is prompt-level: instructing an agent through a repository file or system prompt to write plain, human-like PR text may degrade attribution more cheaply than changing the tool's code generation, a possibility the paper itself acknowledges as a limitation.
  • The modality reversal suggests a testable prediction: agents that share an underlying model and communication conventions but differ in product wrapper should be harder to separate than agents with distinct description habits; a follow-up study could measure attribution accuracy among same-family agents.
  • The 5% explicit-marker contamination in human PRs is a floor; the marker-blind classifier's concentration (39.1% explicit markers among the highest-confidence flags) suggests the true contamination could be several times higher, implying existing comparisons of 'human' versus 'AI' code may have a hidden mixing effect.
  • Because the method is text-centric, it may extend beyond code: the same contrastive approach could attribute AI-generated commit messages, issue comments, or documentation text to source tools, offering a cheap audit for other human-facing AI writing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes AgenTag, a multimodal framework for attributing GitHub pull requests to AI coding agents or humans. It uses four modalities (PR text, commit messages, code diffs, and behavioral features) and a supervised contrastive metric space to support closed-set identification, AI-versus-human detection, open-set rejection of unseen agents, and few-shot enrollment of new agents. Evaluated on the AIDev dataset (33,580 agentic and 6,618 human PRs), the paper reports a weighted F1 of 0.957 for five-agent identification, balanced F1 of 0.89 for AI-versus-human detection, and AUC of 0.837 for unseen-agent detection. It further reports that PR and commit text carry nearly all the signal, that code diffs contribute little across four code representations, and that behavioral fingerprints persist after stripping self-disclosed markers.

Significance. If the results hold, AgenTag is a useful step toward open-world attribution of AI coding agents, and the modality analysis is a valuable negative result for transferring human code-stylometry methods to AI agents. The study has notable strengths: five-times-repeated stratified cross-validation, per-class and macro metrics alongside weighted averages, repository-disjoint and star-matching confound controls, a replication package, and explicit acknowledgment of the Claude Code label circularity and of the classifier-based contamination estimate as circular and hence an upper bound. The main weakness is that the evaluation population does not match the motivating deployment scenario, which substantially limits the central claim as currently stated.

major comments (2)
  1. [Section IV-A, V-C, Tables III and V] Population mismatch between the motivating scenario and the evaluation data. The paper motivates attribution by human-mediated agents submitting PRs under developers' own accounts (Section I), yet four of the five AIDev agent labels come from dedicated bot accounts, platform integration, or committer identity (Section IV-A). Only Claude Code is identified via a trailer in PRs submitted under developer accounts. The headline weighted F1 of 0.957 (Table III) is therefore measured almost entirely on autonomous, self-identifying agents. The only directly relevant human-mediated agent, Claude Code, has a marker-stripped latent F1 of only 0.446 (Table V), far below the headline. No same-agent transfer experiment (training on bot-account PRs and testing on human-mediated PRs of the same agent) is reported for Codex, Copilot, Devin, or Cursor, and Section VII does not list submission mode as a v
  2. [Section IV-A, V-C, Table V] The Claude Code label circularity is not an isolated artifact; it is the only evidence for the human-mediated mode. Because Claude Code's AIDev label is derived from the 'Co-Authored-By: Claude' trailer present in 100% of its PRs, the latent estimate (F1 = 0.446) is the only non-circular number for this agent. Restoring trailers raises Claude Code's F1 to 0.717 (Table V), indicating that for this agent the fingerprint is largely a watermark. The paper's general conclusion that 'behavioral fingerprints persist after removing explicit self-disclosed markers' is therefore driven by the four autonomous agents. Please present the latent Claude Code result as the key evidence for human-mediated attribution and temper the latent-behavior conclusion accordingly.
minor comments (5)
  1. [All results sections] Numerical values are often run together with parentheses or units (e.g., '0.957(median', '0.9%of PRs', '+0.026weighted F1'). Please add consistent spacing for readability.
  2. [Table V] The tier labels 'Latent', '+Names', '+Boiler-', '+Trailers' are compact; consider a footnote defining 'Boilerplate' explicitly.
  3. [Figures 2 and 4] The 'single split (no spread)' marker for the chronological split may confuse readers; consider using a distinct symbol with a clear note that no cross-validation spread exists by design.
  4. [Table IV] The verbatim example for Copilot contains 'Fixes #. . .' with an ellipsis inside a code block; clarify whether this is literal or an editorial ellipsis.
  5. [Section VII] The statement that 'the third decimal ... should not be read as precise' is useful; consider moving this caveat to Section V-C where the 0.446 value is first reported.

Circularity Check

2 steps flagged · score 3.0 of 10

Two circular elements are explicitly disclosed and contained: Claude Code's dataset label is defined by the trailer the model reads, and the RQ4 contamination audit reuses the training labels; the central latent-attribution claim is evaluated marker-stripped and is not reduced to a fit.

  1. self definitional [Section IV-A (Dataset) and Section V-C (RQ3)]
    "Claude Code instead commits under the developer's account, so AIDev can identify its PRs only by the trailer it appends: a “Generated with Claude Code” line and a “Co-Authored-By: Claude” line. The “Co-Authored-By” line is present in every one of its 458 PRs, vs. 0.2% of other agentic PRs, so the label and the marker coincide exactly."

    For one of the five agents, the ground-truth label is defined by the exact self-disclosed string a text-based fingerprint can read. Any evaluation that does not strip the marker scores the model against a label that is a restatement of the input, so Claude Code's raw per-agent F1 (0.717 once trailers are restored) is circular by construction. The paper acknowledges this in Section V-C, strips the marker by default, and reports the latent F1 of 0.446 as the only non-circular figure for Claude Code, so the circularity is localized and disclosed rather than hidden in the headline claim.

  2. fitted input called prediction [Section V-D (RQ4 Results) and Section VII (Internal validity)]
    "This is circular (i.e., human labels are both the training signal and the object under audit) so we treat it as an upper bound."

    The contamination estimate is produced by training the AI-versus-human classifier on the same human/agent labels whose noise is being audited, then reading high agent-probability scores on the human-labeled PRs as undisclosed agent authorship. The scores are generated by the very labels used to fit the model, so the 10.5%/1.3% figures are not independent evidence; high confidence may reflect the learned boundary rather than true contamination. The paper explicitly labels this circular and reports it only as an upper bound, while anchoring the claim with an independent marker-based floor (5.0%) and showing marker rate rises with classifier confidence, so the step is disclosed and not load-bearing for the lower-bound claim.

full rationale

The core agent-identification claim is not circular: four of five AIDev labels come from bot accounts, committer identity, or platform integration, and the evaluation strips self-disclosed markers before feature extraction by default. The strongest numbers (weighted F1 0.96, macro 0.84) are measured under this marker-stripped latent setting, and the protocol uses stratified repeated cross-validation with repository-disjoint and chronological splits, so the main attribution result has independent content. The acknowledged circularities are (1) Claude Code's label coincides exactly with its trailer, and (2) the classifier-based contamination audit uses the training labels as the audited object; both are explicitly disclosed, mitigated by marker stripping and by an independent marker floor, and neither is used to derive the central latent-fingerprint conclusion. The larger concern about the motivating human-mediated setting—that only Claude Code represents it and its latent F1 is 0.446—is a population/validity gap, not a circularity, so it does not raise the circularity score beyond the localized, disclosed circular steps.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the AIDev labels, the completeness of marker stripping, and the choice of encoders; these are dataset and method premises rather than fitted scientific parameters. The model hyperparameters listed above are standard ML choices, none of which are tuned to produce the target F1 values.

free parameters (6)
  • Contrastive embedding dimension = 128
    Projection head output size chosen by hand; standard for metric learning, not tuned to the target result.
  • Supervised contrastive temperature = not reported
    Temperature hyperparameter of the SupCon loss; the text does not list its value, affecting embedding geometry and all downstream F1/AUC numbers.
  • Behavioral feature correlation threshold = 0.70
    Correlation-based hierarchical clustering at |rho| >= 0.70 reduces 53 features to 41; this threshold is arbitrary and affects the behavioral representation.
  • R-squared redundancy filter threshold = not reported
    Second-stage feature reduction; the R^2 cutoff is not specified in the text.
  • Code encoder truncation length = 4000 characters
    Code diffs are truncated to the first 4000 added characters before mean pooling, which may understate the code signal.
  • Few-shot enrollment size = 10 PRs
    A new agent is enrolled from ten labeled PRs; chosen as a round number, and sensitivity to this number is not explored.
assumptions (6)
  • domain assumption AIDev ground-truth labels correctly identify the authoring agent for the five studied agents
    All accuracy results are computed against these labels; the paper shows a partial exception for Claude Code (label from trailer) and contamination in human labels (Section IV-A, Section V-D).
  • domain assumption PR title/description, commit messages, diffs, and behavioral features capture agent identity
    The four modalities define the signal space; if identity is expressed elsewhere (for example, code semantics beyond added lines, branch metadata, or edit timing), the framework cannot see it.
  • ad hoc to paper Three-tier marker stripping removes all self-disclosed identity
    RQ3's latent-fingerprint claim assumes Tier 1-3 stripping is complete; the paper itself notes residual tool-name substrings in identifiers (5.9% of Claude Code PRs, Section VII).
  • domain assumption GitHub API collection for human PRs matches AIDev's agent-side procedure
    RQ4 compares agent and human PRs using human commit/diff data collected separately; collection bias is acknowledged as a threat (Section VII).
  • standard math Supervised contrastive embeddings preserve agent-relevant style
    The method assumes the learned metric space clusters PRs from the same agent; this is the standard SupCon/Prototypical-Net assumption (Section III).
  • domain assumption The five agents studied represent the current landscape
    External validity is explicitly limited to these agents, GitHub, and one dataset snapshot (Section VII).

how reviews work

0 comments
Cite this review

Pith. "Pith review of AgenTag: Attribution of AI Coding Agents from Behavioral Fingerprints." pith.science (2026). https://pith.science/paper/PA3WFDTU

@misc{pith2026260800966,
  author       = {Pith},
  title        = {Pith review of: AgenTag: Attribution of AI Coding Agents from Behavioral Fingerprints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PA3WFDTU}},
  note         = {Machine review of arXiv:2608.00966}
}
read the original abstract

AI coding agents increasingly author pull requests (PRs), often under developers' own accounts, obscuring who actually produced a change. Reliable attribution is important for repository governance, empirical studies of AI-assisted software development, and measuring the impact of AI coding agents. Existing work focuses on closed-set identification of known agents, leaving the practical limits of open-world AI coding agent attribution largely unexplored. In this paper, we present AgenTag, a multimodal framework for open-world AI coding agent attribution, evaluated on AIDev, comprising 33,580 PRs from five AI coding agents and 6,618 human-authored PRs. We represent each PR using textual, behavioral, and code-based modalities, and compare conventional classification with supervised contrastive learning for open-set recognition and few-shot enrollment of previously unseen agents. AgenTag identifies authoring agents with a weighted F1 of 0.96 (macro F1 of 0.84), distinguishes AI- from human-authored PRs with a balanced F1 of 0.89, and detects previously unseen agents with an AUC of 0.84. We further show that PR descriptions and commit messages provide nearly all of the attribution signal, whereas code diffs contribute little across multiple representations, indicating that coding agents are distinguished primarily by how they communicate changes rather than by the code they generate. Moreover, these behavioral fingerprints persist after removing explicit self-disclosed markers, demonstrating that attribution relies largely on latent stylistic characteristics. These findings show that reliable attribution of AI coding agents is feasible and clarify the practical trade-offs between attribution accuracy and the information required to achieve it.

Figures

Figures reproduced from arXiv: 2608.00966 by the authors.

Figure 1
Figure 1. Overview of our study: the four modalities are stripped of tool markers and encoded, then a supervised-contrastive metric space and classical baselines [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Robustness of agent identification (all text, contrastive, natural [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Behavioral signatures per agent (one-vs-rest, Cliff’s delta). Dark bars [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Weighted F1 by modality for AI coding agent identification (RQ1) and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Accuracy versus data-access cost. Filled markers linked by line: the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 40 canonical work pages

  1. [1]

    The rise of AI teammates in software engineering (SE) 3.0: How autonomous coding agents are reshaping software engineering,

    H. Li, H. Zhang, and A. E. Hassan, “The rise of AI teammates in software engineering (SE) 3.0: How autonomous coding agents are reshaping software engineering,”arXiv preprint arXiv:2507.15003, 2025

    arXiv 2025

  2. [2]

    Agen- tic much? adoption of coding agents on GitHub,

    R. Robbes, T. Matricon, T. Degueule, A. Hora, and S. Zacchiroli, “Agen- tic much? adoption of coding agents on GitHub,”ACM Transactions on Software Engineering and Methodology, 2026

    work page 2026

  3. [3]

    GitHub terms of service,

    GitHub, “GitHub terms of service,” https://docs.github.com/en/ site-policy/github-terms/github-terms-of-service, 2024

    work page 2024

  4. [4]

    Studying the impact of noises in build breakage data,

    T. A. Ghaleb, D. A. Da Costa, Y . Zou, and A. E. Hassan, “Studying the impact of noises in build breakage data,”IEEE Transactions on Software Engineering, vol. 47, no. 9, pp. 1998–2011, 2019

    1998

  5. [5]

    Grounded Copilot: How programmers interact with code-generating models,

    S. Barke, M. B. James, and N. Polikarpova, “Grounded Copilot: How programmers interact with code-generating models,”Proceedings of the ACM on Programming Languages, vol. 7, no. OOPSLA1, pp. 85–111, 2023

    work page 2023

  6. [6]

    Fingerprinting AI coding agents on GitHub,

    T. A. Ghaleb, “Fingerprinting AI coding agents on GitHub,” inPro- ceedings of the 23rd International Conference on Mining Software Repositories (MSR), 2026

    work page 2026

  7. [7]

    De-anonymizing programmers via code stylometry,

    A. Caliskan-Islam, R. Harang, A. Liu, A. Narayanan, C. V oss, F. Ya- maguchi, and R. Greenstadt, “De-anonymizing programmers via code stylometry,” in24th USENIX Security Symposium (USENIX Security 15), 2015, pp. 255–270

    work page 2015

  8. [8]

    Source code authorship attribution using n-grams,

    S. Burrows and S. M. Tahaghoghi, “Source code authorship attribution using n-grams,” inProceedings of the Twelfth Australasian Document Computing Symposium, Melbourne, Australia, RMIT University. Cite- seer, 2007, pp. 32–39

    work page 2007

Show all 53 references
  1. [9]

    Large-scale and language-oblivious code authorship identification,

    M. Abuhamad, T. AbuHmed, A. Mohaisen, and D. Nyang, “Large-scale and language-oblivious code authorship identification,” inProceedings of the 2018 ACM SIGSAC Conference on Computer and Communica- tions Security, 2018, pp. 101–114

    2018

  2. [10]

    Supervised contrastive learn- ing,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,”Advances in neural information processing systems, vol. 33, pp. 18 661–18 673, 2020

    2020

  3. [11]

    Sentence-BERT: Sentence embeddings using siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using siamese BERT-networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th In- ternational Joint Conference on Natural Language Processing (EMNLP- IJCNLP), 2...

    2019

  4. [12]

    CodeBERT: A pre-trained model for programming and natural languages,

    Z. Feng, D. Guo, D. Tang, N. Duan, X. Feng, M. Gong, L. Shou, B. Qin, T. Liu, D. Jianget al., “CodeBERT: A pre-trained model for programming and natural languages,” inFindings of the Association for Computational Linguistics: EMNLP 2020, 2020, pp. 1536–1547

    2020

  5. [13]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,”Advances in neural information processing systems, vol. 30, 2017

    2017

  6. [14]

    Appropriate statistics for ordinal level data: Should we really be using t-test and Cohen’s d for evaluating group differences on the NSSE and other surveys,

    J. Romano, J. D. Kromrey, J. Coraggio, and J. Skowronek, “Appropriate statistics for ordinal level data: Should we really be using t-test and Cohen’s d for evaluating group differences on the NSSE and other surveys,” inAnnual Meeting of the Florida Association of Institutional...

    2006

  7. [15]

    AgenTag: Attribution of AI coding agents from behavioral fingerprints (replication package),

    T. A. Ghaleb, “AgenTag: Attribution of AI coding agents from behavioral fingerprints (replication package),” https://figshare.com/s/ 6c4fb60ba3f5abf1ffdf, 2026

    2026

  8. [16]

    Comparing AI cod- ing agents: A task-stratified analysis of pull request acceptance,

    G. Pinna, J. Gong, D. Williams, and F. Sarro, “Comparing AI cod- ing agents: A task-stratified analysis of pull request acceptance,” in Proceedings of the 23rd International Conference on Mining Software Repositories, 2026

    2026

  9. [17]

    How AI coding agents modify code: A large-scale study of GitHub pull requests,

    D. Ogenrwot and J. Businge, “How AI coding agents modify code: A large-scale study of GitHub pull requests,” inProceedings of the 23rd International Conference on Mining Software Repositories, 2026

    2026

  10. [18]

    Code change characteristics and description alignment: A comparative study of agentic versus human pull requests,

    D. Pham and T. A. Ghaleb, “Code change characteristics and description alignment: A comparative study of agentic versus human pull requests,” inProceedings of the 23rd International Conference on Mining Software Repositories, 2026

    2026

  11. [19]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  12. [20]

    FaceNet: A unified embedding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “FaceNet: A unified embedding for face recognition and clustering,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 815–823

    2015

  13. [21]

    X- Vectors: Robust DNN embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X- Vectors: Robust DNN embeddings for speaker recognition,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5329–5333

    2018

  14. [22]

    Toward open set recognition,

    W. J. Scheirer, A. de Rezende Rocha, A. Sapkota, and T. E. Boult, “Toward open set recognition,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 7, pp. 1757–1772, 2012

    2012

  15. [23]

    Conventional commits specification,

    C. Commits, “Conventional commits specification,” https://www. conventionalcommits.org, 2024

    2024

  16. [24]

    An introduction to variable and feature selection,

    I. Guyon and A. Elisseeff, “An introduction to variable and feature selection,”Journal of machine learning research, vol. 3, no. Mar, pp. 1157–1182, 2003

    2003

  17. [25]

    Collinearity: a review of methods to deal with it and a simulation study evaluating their performance,

    C. F. Dormann, J. Elith, S. Bacher, C. Buchmann, G. Carl, G. Carr ´e, J. R. G. Marqu ´ez, B. Gruber, B. Lafourcade, P. J. Leit ˜aoet al., “Collinearity: a review of methods to deal with it and a simulation study evaluating their performance,”Ecography, vol. 36, no. 1, pp. 27– 46, 2013

    2013

  18. [26]

    Harrell, Frank E.,Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Anal- ysis, 2nd ed

    J. Harrell, Frank E.,Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Anal- ysis, 2nd ed. Springer, 2015

    2015

  19. [27]

    Kuhn and K

    M. Kuhn and K. Johnson,Applied predictive modeling. Springer, 2013

    2013

  20. [28]

    GraphCodeBERT: Pre-training code representations with data flow,

    D. Guo, S. Ren, S. Lu, Z. Feng, D. Tang, S. Liu, L. Zhou, N. Duan, A. Svyatkovskiy, S. Fuet al., “GraphCodeBERT: Pre-training code representations with data flow,” inProceedings of the 9th International Conference on Learning Representations, 2021

    2021

  21. [29]

    SMOTE: Synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,”Journal of Artificial Intelligence Research, vol. 16, pp. 321–357, 2002

    2002

  22. [30]

    Transformers: State- of-the-art natural language processing,

    T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowiczet al., “Transformers: State- of-the-art natural language processing,” inProceedings of the 2020 conference on empirical methods in natural language processing: system demo...

    2020

  23. [31]

    PyTorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “PyTorch: An imperative style, high-performance deep learning library,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  24. [32]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourget al., “Scikit-learn: Machine learning in Python,”the Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011

    2011

  25. [33]

    A simulation study of the number of events per variable in logistic regression analysis,

    P. Peduzzi, J. Concato, E. Kemper, T. R. Holford, and A. R. Feinstein, “A simulation study of the number of events per variable in logistic regression analysis,”Journal of clinical epidemiology, vol. 49, no. 12, pp. 1373–1379, 1996

    1996

  26. [34]

    Inference for the generalization error,

    C. Nadeau and Y . Bengio, “Inference for the generalization error,” Advances in neural information processing systems, vol. 12, 1999

    1999

  27. [35]

    Exposing LLM user privacy via traffic fingerprint analysis: A study of privacy risks in LLM agent interactions,

    Y . Zhang, X. Deng, Z. Gu, Y . Chen, K. Xu, Q. Li, and J. Wu, “Exposing LLM user privacy via traffic fingerprint analysis: A study of privacy risks in LLM agent interactions,”arXiv preprint arXiv:2510.07176, 2025

    2025

  28. [36]

    XGBoost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2016, pp. 785–794. 15

    2016

  29. [37]

    The power of bots: Characterizing and understanding bots in OSS projects,

    M. Wessel, B. M. De Souza, I. Steinmacher, I. S. Wiese, I. Polato, A. P. Chaves, and M. A. Gerosa, “The power of bots: Characterizing and understanding bots in OSS projects,”Proceedings of the ACM on Human-Computer Interaction, vol. 2, no. CSCW, pp. 1–19, 2018

    2018

  30. [38]

    On the accuracy of bot detection techniques,

    M. Golzadeh, A. Decan, and N. Chidambaram, “On the accuracy of bot detection techniques,” inProceedings of the Fourth International Workshop on Bots in Software Engineering, 2022, pp. 1–5

    2022

  31. [39]

    Is ChatGPT the ultimate programming assistant–how far is it?

    H. Tian, W. Lu, T. O. Li, X. Tang, S.-C. Cheung, J. Klein, and T. F. Bissyand´e, “Is ChatGPT the ultimate programming assistant–how far is it?”arXiv preprint arXiv:2304.11938, 2023

    2023 arXiv

  32. [40]

    GPTSniffer: A CodeBERT-based classifier to detect source code written by ChatGPT,

    P. T. Nguyen, J. Di Rocco, C. Di Sipio, R. Rubei, D. Di Ruscio, and M. Di Penta, “GPTSniffer: A CodeBERT-based classifier to detect source code written by ChatGPT,”Journal of Systems and Software, vol. 214, p. 112059, 2024

    2024

  33. [41]

    I know which LLM wrote your code last summer: LLM generated code stylometry for authorship attribution,

    T. Bisztray, B. Cherif, R. A. Dubniczky, N. Gruschka, B. Borsos, M. A. Ferrag, A. Kovacs, V . Mavroeidis, and N. Tihanyi, “I know which LLM wrote your code last summer: LLM generated code stylometry for authorship attribution,” inProceedings of the 18th ACM Workshop on Artific...

    2025

  34. [42]

    A practical guide for using statistical tests to assess randomized algorithms in software engineering,

    A. Arcuri and L. Briand, “A practical guide for using statistical tests to assess randomized algorithms in software engineering,” inProceedings of the 33rd international conference on software engineering, 2011, pp. 1–10

    2011

  35. [43]

    Security in the age of AI teammates: An empirical study of agentic pull requests on GitHub,

    M. L. Siddiq, X. Zhao, V . C. Lopes, B. Casey, and J. C. d. S. Santos, “Security in the age of AI teammates: An empirical study of agentic pull requests on GitHub,”arXiv preprint arXiv:2601.00477, 2026

    2026

  36. [44]

    How AI coding agents communicate: A study of pull request description characteristics and human review responses,

    K. Watanabe, R. Tsuchida, T. Monno, B. Huang, K. Yamasaki, Y . Fan, K. Shimari, and K. Matsumoto, “How AI coding agents communicate: A study of pull request description characteristics and human review responses,” inProceedings of the 23rd International Conference on Mining So...

    2026

  37. [45]

    Analyzing message-code inconsistency in AI coding agent-authored pull requests,

    J. Gong, G. Pinna, Y . Bian, and J. M. Zhang, “Analyzing message-code inconsistency in AI coding agent-authored pull requests,” inProceedings of the 23rd International Conference on Mining Software Repositories (MSR), 2026

    2026

  38. [46]

    Where do AI coding agents fail? an empirical study of failed agentic pull requests in GitHub,

    R. Ehsani, S. Pathak, S. Rawal, A. A. Mujahid, M. M. Imran, and P. Chatterjee, “Where do AI coding agents fail? an empirical study of failed agentic pull requests in GitHub,” inProceedings of the 23rd International Conference on Mining Software Repositories, 2026

    2026

  39. [47]

    When AI agents touch CI/CD configurations: Frequency and success,

    T. A. Ghaleb, “When AI agents touch CI/CD configurations: Frequency and success,” inProceedings of the 23rd International Conference on Mining Software Repositories (MSR), 2026

    2026

  40. [48]

    Effective identification of source code authors using byte-level information,

    G. Frantzeskou, E. Stamatatos, S. Gritzalis, and S. Katsikas, “Effective identification of source code authors using byte-level information,” in Proceedings of the 28th international conference on Software engineer- ing, 2006, pp. 893–896

    2006

  41. [49]

    A novel neural source code representation based on abstract syntax tree,

    J. Zhang, X. Wang, H. Zhang, H. Sun, K. Wang, and X. Liu, “A novel neural source code representation based on abstract syntax tree,” in 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE). IEEE, 2019, pp. 783–794

    2019

  42. [50]

    Authorship attribution of source code: A language-agnostic approach and applicability in software engineering,

    E. Bogomolov, V . Kovalenko, Y . Rebryk, A. Bacchelli, and T. Bryksin, “Authorship attribution of source code: A language-agnostic approach and applicability in software engineering,” inProceedings of the 29th ACM joint meeting on European software engineering conference and s...

    2021

  43. [51]

    Reassessing code authorship attribution in the era of language models,

    A. K. Dipongkor, Z. Yao, and K. Moran, “Reassessing code authorship attribution in the era of language models,”ACM Transactions on Software Engineering and Methodology (TOSEM), 2025, to appear

    2025

  44. [52]

    Code authorship attribution: Methods and challenges,

    V . Kalgutkar, R. Kaur, H. Gonzalez, N. Stakhanova, and A. Matyukhina, “Code authorship attribution: Methods and challenges,”ACM Computing Surveys (CSUR), vol. 52, no. 1, pp. 1–36, 2019

    2019

  45. [53]

    Bridging behavioral biometrics and source code stylometry: A survey of programmer at- tribution,

    M. Horvath, E. Pietrikova, and D. Spinellis, “Bridging behavioral biometrics and source code stylometry: A survey of programmer at- tribution,”arXiv preprint arXiv:2603.11150, 2026

    2026

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.