Pith. sign in

REVIEW 5 major objections 4 minor 39 references

Stress-testing Machine Generated Text Detection: Shifting Language Models Writing Style to Fool Detectors

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a DPO-based fine-tuning pass on a few thousand human/machine text pairs shifts LLM writing style toward human text and sharply reduces the accuracy of state-of-the-art machine-generated-text detectors.

desk verdict Useful attack benchmark with a real confound: the F1 drops are consistent, but the human-like style explanation is undercut by unmeasured text degradation and an overclaimed human evaluation. read the letter →

arxiv 2505.24523 v1 pith:LK7GEZKT submitted 2025-05-30 cs.CL cs.AI

Andrea Pedrotti , Michele Papucci , Cristiano Ciaccio , Alessio Miaschi , Giovanni Puccetti , Felice Dell'Orletta , Andrea Esuli This is my paper
classification cs.CLcs.AI
keywords machine-generatedtextdetectionDirectPreferenceOptimizationadversarialgenerationlinguisticprofilingstylometricfeaturesLLMalignmentdetectorrobustnessin-domainbenchmark
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 tries to establish that a short Direct Preference Optimization (DPO) fine-tuning run can shift an LLM's writing style toward human-written text, and that this shift makes current machine-generated-text detectors substantially easier to fool. On the XSUM news domain, fine-tuned Llama cuts MAGE F1 from 0.76 to 0.47 and Binoculars from 0.99 to 0.33. The authors read this as evidence that detectors lean on shallow stylistic cues rather than on deeper content or coherence. The result matters because it offers a cheap, concrete way to build harder in-domain benchmarks for detection and exposes a blind spot relevant to misinformation defense.

What carries the argument

The load-bearing object is the preference pair (human text, machine text) used in Direct Preference Optimization, a technique that adjusts model weights directly from chosen and rejected text pairs without training a separate reward model. For the linguistically informed variant, the paper first trains a linear SVM on over 130 ProfilingUD linguistic features, takes the ten features with the largest absolute coefficients, and selects pairs with the largest human/machine distance on those features. Iterating the loop re-aligns the model on its own new generations and a fresh set of features. This machinery converts the measured distributional difference between human and machine text into a concrete training signal that pulls generations toward human style.

What would settle it

Run a head-to-head comparison of detector scores on DPO-aligned outputs versus deliberately degraded outputs matched for perplexity or coherence; if detectors score the degraded text as low as the aligned text, the drop is not about style. Alternatively, have independent raters score fluency on the second-iteration aligned outputs; if those outputs are judged incoherent, the paper's central interpretation does not survive.

Watch

Extended reading notes

Core claim

The central claim is that stylistically aligning LLMs to human text, using DPO with human-written text as the preferred output, is enough to defeat state-of-the-art detectors on in-domain text. With roughly seven thousand preference pairs, both Llama 3.1 8B and Gemma 2 2B produce news and abstract text on which MAGE, LLM-DetectAIve, and Binoculars lose substantial F1, while the linguistically guided variant (dpo-ling) moves selected linguistic feature distributions measurably closer to human writing. The paper also reports that human raters stay near random-guessing levels, with very low annotator agreement, and that Radar, trained adversarially against paraphrasing, is the most resistant detector in the suite. The authors conclude that the performance drop reflects a genuine shift in writing style rather than a shift out of the detector's familiar domain.

Load-bearing premise

The central load-bearing premise is that after DPO the model still produces grammatical, coherent text, so the detector's failure is evidence of human-like style instead of evidence of broken text; the paper never directly measures fluency, and its own Limitations section and Appendix F leave that open.

Editorial extensions

If this is right

  • A single DPO iteration on a few thousand examples lowers detector F1 substantially, and a second iteration helps further for some generators.
  • Detectors trained adversarially against one attack family, such as Radar against paraphrasing, transfer at least part of that robustness to this style-shift attack.
  • Benchmarks built only from original LLM outputs overestimate real-world detector performance because they omit stylistically aligned, in-domain machine text.
  • The top shifted features (type-token ratio, clause length, and part-of-speech distributions) form a concrete checklist of what current detectors key on and what future training should diversify.

Reading between the lines

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

  • Editorial inference: the paper's own Appendix F shows visibly garbled second-iteration outputs, so part of the measured detector drop may reflect text degradation rather than successful human-style mimicry.
  • A direct fluency and coherence measurement on the aligned outputs would settle whether the drop is style-driven; none is reported, and the Limitations section concedes the human study was not aimed at grammaticality.
  • If degradation is the main driver, detectors could be hardened with simple coherence checks, which is a testable alternative explanation for the observed F1 collapse.
  • The same DPO loop should transfer to other domains and languages, but detectors may prove more resistant there, so benchmark updates would need to be continuous.
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

5 major / 4 minor

Summary. The paper proposes an adversarial pipeline in which an LLM (Llama-3.1-8B-Instruct or Gemma-2-2B-it) is fine-tuned with DPO on paired human/machine texts, labeling human-written text as the preferred completion, so as to shift the writing style of generated news articles (XSUM) and scientific abstracts (arXiv) toward human-written text. Two preference-construction strategies are tested: random pairing (dpo) and linguistically informed pairing based on the top discriminating ProfilingUD features (dpo-ling), with up to two iterations of the pipeline. Detector performance is measured before and after alignment with MAGE, RADAR, LLM-DetectAIve, Binoculars, and in-domain SVM and RoBERTa detectors, and a Prolific human study is run on XSUM. The main results are large macro-F1 drops after one DPO iteration (for example, Binoculars drops from 0.99 to 0.33 and MAGE from 0.76 to 0.47 for Llama on XSUM), accompanied by linguistic-feature analyses (JS divergences, MANOVA) and example generations in Appendix F.

Significance. The empirical attack is valuable and timely. It uses off-the-shelf, externally trained detectors, so the main evasion result is not an artifact of training the generator on detector predictions. The F1 drops are consistent across several detectors and both domains, and the authors release code, models, and data, which supports reproducibility. If the interpretation is confirmed, namely that the drops reflect successful stylistic alignment while the text remains fluent and coherent, the work would provide a challenging benchmark and demonstrate that current detectors rely on superficial stylistic cues. The linguistic analysis is a useful diagnostic. However, the manuscript does not currently establish the fluency-preservation condition, and one of its own human-evaluation results points in the opposite direction, so the significance of the findings is conditional on additional controls.

major comments (5)
  1. [§4 and Appendix F; Limitations] The central interpretation of the F1 drops as evidence that detectors rely on stylistic cues presupposes that DPO-aligned outputs remain grammatical, coherent, and on-topic. The manuscript never directly verifies this. The Limitations section states that the human study is 'not primarily focused on assessing the coherence and the grammaticality of generated documents,' and Appendix F contains visibly garbled generations for llama-dpo-iter-2 and llama-dpo-iter-2-ling (for example, the Ashcroft Centre example: 'though-dayhood Marington of Ashfield Highfields may consider laying into hospital-bed once broken wall supposed by Bio-Man unconscious Atmoshalin'). Without an automatic fluency/coherence control (perplexity, grammaticality, factuality, or targeted human ratings) or an analysis showing that the F1 drop survives in a fluency-matched subset, the observed drops may be caused by out-of-distribution or degraded text rather than by successful human-like style alignment. This is load-bearing for the paper's main claim and should be addressed before publication.
  2. [§4.2 and Table 11] The Discussion in §5 claims that after the alignment process human annotators' performance 'is not significantly affected, remaining mostly random guessing,' but Table 11 reports that Llama majority-vote accuracy rises from 0.52 to 0.70 after dpo-1-ling, while Gemma falls from 0.79 to 0.65. These are substantial and opposite movements that directly contradict the claim in the main text. At minimum, the paper must report the exact human results in the main body and provide significance testing (for example, a paired test across annotators) rather than stating that performance is unchanged. The Llama result is particularly important because it undercuts the narrative that aligned text is more human-like for the model that produced the largest detector-F1 drops.
  3. [§4.1, Table 3] The MANOVA results do not support the strength of the alignment claims as presented. Pillai's trace for HWT vs dpo-1-ling is 0.7137 with p < 1e-5, so the aligned machine texts remain highly separable from human texts on the measured linguistic features. The comparison of Pillai's traces (0.7628 vs 0.7137) is reported without confidence intervals or a test of significance, so the statement that dpo-1-ling 'shows a stronger alignment' than dpo-1 is not established; moreover, the trace for dpo-1 (0.7635) is essentially the same as the baseline (0.7628). Please add uncertainty quantification and, if possible, a formal comparison across conditions.
  4. [§2.1 and §4.1] The dpo-ling evaluation has a selection-induced circularity for the feature-alignment claim: the preference pairs are selected to maximize distance on the top ten features from an SVM trained on the same ProfilingUD features, and Table 4 then reports that dpo-1-ling moves exactly those feature families closer to HWT. For the selected features, improvement over dpo-1 is partly enforced by construction. The detector-evaluation results are not affected because the detectors are external, but the linguistic-alignment interpretation in Section 4.1 should be presented as a check on the selection mechanism rather than as independent evidence of human-like style. Ideally, alignment should be measured on held-out features or with the SVM retrained after alignment.
  5. [§4, Tables 1–3] All central performance claims are based on single point estimates. No confidence intervals, standard errors, or significance tests are provided for the F1 drops in Table 1, the TPR values in Table 2, or the Pillai's traces in Table 3. Since the paper's headline is that detectors 'can be easily fooled' and that the alignment is effective, bootstrap confidence intervals or repeated evaluations should be added; otherwise the magnitude of the drop, especially for Gemma (for example, 0.71 to 0.65 on XSUM), cannot be distinguished from noise.
minor comments (4)
  1. [Table 4 and App. B] There are formatting and typographical issues: the value '.447.118' in the Llama dpo-1-ling row appears to be two numbers run together, and the caption of Table 6 spells the tool name as 'ProflingUD' instead of 'ProfilingUD'.
  2. [§4.2] The text says human evaluation covers generations 'before and after the first iteration of DPO alignment,' but Table 11 reports only the baseline and dpo-1-ling conditions, not dpo-1; please clarify whether dpo-1 was omitted or whether the table labels are incomplete.
  3. [Abstract and §4] The abstract and results sections use the phrase 'significant drop' without any statistical test; if the manuscript keeps this wording, it should be supported by the confidence intervals requested in the major comments, or the wording should be changed to 'large drop' to avoid implying inferential significance.
  4. [Appendix F] The degraded examples in Appendix F should be explicitly discussed in the main text, with a caveat that the later iterations do not always preserve factuality or coherence; currently a reader who does not open the appendix could assume all iterations are equally fluent.

Circularity Check

1 steps flagged · score 2.0 of 10

Central detector-fooling result is not circular; one minor internal evaluation loop in the dpo-ling feature-alignment analysis.

  1. other [Sec. 2.1 (dpo-ling selection), Sec. 4.1 / Table 4, App. C.1]
    "In the second one, which we call dpo-ling, we select couples of (HWT, MGT) in a linguistically informed way: using an SVM classifier trained on explicit linguistic features extracted from the texts to identify MGT, we select the top ten most discriminating features. Then, for each of those features, we take the top-k pairs where the absolute distance on that feature between the HWT and MGT is the largest. ... Specifically, we present the results for the top 10 discriminative features selected in the dpo-1-ling setting for both models."

    The dpo-ling preference pairs are deliberately chosen to maximize HWT-MGT distance on the ten SVM-selected features, and the alignment analysis in Table 4 then measures Jensen-Shannon divergence on exactly those same ten features, presenting them as evidence of stylistic alignment. This is an internal evaluation loop: the feature set used to select the training pairs is the same feature set used to demonstrate the alignment effect, so part of the reported movement toward HWT on those features is an artifact of the selection criterion rather than independent evidence of overall human-likeness.

full rationale

The paper's central claim is empirically grounded rather than circular: DPO is trained to prefer human-written texts over machine-generated texts, new text is generated with the aligned model, and the resulting detector F1 drops are measured on pre-trained external detectors that were not consulted during DPO. No fitted parameter is renamed as a prediction, and the main evaluation is not defined in terms of the training objective. The only self-referential element is the dpo-ling analysis: training pairs are selected using an SVM's top ten linguistic features and the alignment is then reported on those same features, making the feature-level JS-divergence improvements partly a by-construction consequence of the selection strategy. This is a mild internal loop in a secondary analysis, not in the detector-fooling result. Self-citations (Esuli et al. 2024) are not load-bearing because the paper independently computes its own MANOVA comparison between HWT and MGT before and after alignment. The Limitations section explicitly concedes that coherence and grammaticality were 'not primarily focused' on in the human study, and Appendix F shows visibly degraded generations, but these are validity threats to the 'human-like style' interpretation rather than circularity. Similarly, the Section 5 claim that human performance remains 'mostly random guessing' is in tension with Table 11 (Llama majority accuracy rises from 0.52 to 0.70 after dpo-1-ling), a correctness concern, not a circularity concern. Overall, the derivation is self-contained against external detectors, so the score is low: one minor internal loop gives a 2, not a higher score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper contributes an empirical attack, not a derivation, so there are no fitted physical constants or invented theoretical entities. The unpaid assumptions are about which features detectors use and whether DPO-caused style shifts keep text fluent; both are acknowledged in the paper's limitations.

free parameters (4)
  • DPO beta and learning rate = not reported
    Grid searched over beta in {0.1, 0.5, 1.0} and lr in {5e-7, 5e-6}; the selected values per model and domain are not stated, so the tuning choices behind the main results are underspecified.
  • dpo-ling top-k per feature = 1000
    The number of top pairs selected per discriminative feature was chosen by hand and is not ablated.
  • epsilon for iterative pair selection = 0.1 (XSUM), 0.2 (arXiv)
    Hand-set tolerance used in later DPO iterations to avoid re-selecting previous features; the paper says the higher value was needed because arXiv Abstracts has fewer documents.
  • LoRA adapter rank = 32
    Adapter rank set to 32 across all experiments and not varied, matching the original LoRA paper.
assumptions (3)
  • domain assumption Linguistic features from ProfilingUD are the stylistic shortcuts that MGT detectors rely on.
    The dpo-ling selection and the interpretation of detector drops assume these 130+ features are the relevant cue space; no causal test shows detectors actually use these specific features.
  • domain assumption DPO with HWT as preferred output shifts style without degrading fluency or topical coherence.
    This is needed to attribute detector performance drops to human-like style rather than to broken text; fluency is not measured and sample outputs show degradation.
  • domain assumption Paired same-title human and machine documents are comparable except for authorship.
    The pipeline pairs an original article or abstract with a generation prompted by its title, treating remaining differences in length, formatting, and factual content as irrelevant to the style comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stress-testing Machine Generated Text Detection: Shifting Language Models Writing Style to Fool Detectors." pith.science (2026). https://pith.science/paper/LK7GEZKT

@misc{pith2026250524523,
  author       = {Pith},
  title        = {Pith review of: Stress-testing Machine Generated Text Detection: Shifting Language Models Writing Style to Fool Detectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LK7GEZKT}},
  note         = {Machine review of arXiv:2505.24523}
}
read the original abstract

Recent advancements in Generative AI and Large Language Models (LLMs) have enabled the creation of highly realistic synthetic content, raising concerns about the potential for malicious use, such as misinformation and manipulation. Moreover, detecting Machine-Generated Text (MGT) remains challenging due to the lack of robust benchmarks that assess generalization to real-world scenarios. In this work, we present a pipeline to test the resilience of state-of-the-art MGT detectors (e.g., Mage, Radar, LLM-DetectAIve) to linguistically informed adversarial attacks. To challenge the detectors, we fine-tune language models using Direct Preference Optimization (DPO) to shift the MGT style toward human-written text (HWT). This exploits the detectors' reliance on stylistic clues, making new generations more challenging to detect. Additionally, we analyze the linguistic shifts induced by the alignment and which features are used by detectors to detect MGT texts. Our results show that detectors can be easily fooled with relatively few examples, resulting in a significant drop in detection performance. This highlights the importance of improving detection methods and making them robust to unseen in-domain texts.

Figures

Figures reproduced from arXiv: 2505.24523 by the authors.

Figure 1
Figure 1. Overview of the proposed methodology. We [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. The distribution of selected linguistic features [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Distribution of accuracy scores among all [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: The distribution of selected linguistic features [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 5
Figure 5. Figure 5: The distribution of selected linguistic features [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 10 canonical work pages

  1. [1]

    Mervat Abassy, Kareem Elozeiri, Alexander Aziz, Minh Ngoc Ta, Raj Vardhan Tomar, Bimarsha Adhikari, Saad El Dine Ahmed, Yuxia Wang, Osama Mohammed Afzal, Zhuohan Xie, Jonibek Mansurov, Ekaterina Artemova, Vladislav Mikhailov, Rui Xing, Jiahui Geng, Hasan Iqbal, Zain Muhammad Mujahid, Tarek Mahmoud, Akim Tsvigun, Alham Fikri Aji, Artem Shelmanov, Nizar Hab...

    doi:10.18653/v1/2024.emnlp-demo.35 2024

  2. [2]

    Dominique Brunato, Andrea Cimino, Felice Dell ' Orletta, Giulia Venturi, and Simonetta Montemagni. 2020. https://aclanthology.org/2020.lrec-1.883/ Profiling- UD : a tool for linguistic profiling of texts . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 7145--7151, Marseille, France. European Language Resources Association

    work page 2020

  3. [3]

    Cristiano Ciaccio, Felice Dell'Orletta, Alessio Miaschi, and Giulia Venturi. 2024. Controllable text generation to evaluate linguistic abilities of italian llms. In Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it)

    work page 2024

  4. [4]

    Kevyn Collins-Thompson. 2014. https://doi.org/10.1075/itl.165.2.01col Computational assessment of text readability: A survey of current and future research . ITL - International Journal of Applied Linguistics, 165:97--135

    doi:10.1075/itl.165.2.01col 2014

  5. [5]

    Manning, Joakim Nivre, and Daniel Zeman

    Marie-Catherine de Marneffe, Christopher D. Manning, Joakim Nivre, and Daniel Zeman. 2021. https://doi.org/10.1162/coli_a_00402 Universal Dependencies . Computational Linguistics, 47(2):255--308

    doi:10.1162/coli_a_00402 2021

  6. [6]

    DeepSeek-AI. 2024. https://arxiv.org/abs/2412.19437 Deepseek-v3 technical report . Preprint, arXiv:2412.19437

    arXiv 2024

  7. [7]

    Jad Doughman, Osama Mohammed Afzal, Hawau Olamide Toyin, Shady Shehata, Preslav Nakov, and Zeerak Talat. 2025. https://aclanthology.org/2025.coling-main.288/ Exploring the limitations of detecting machine-generated text . In Proceedings of the 31st International Conference on Computational Linguistics, pages 4274--4281, Abu Dhabi, UAE. Association for Com...

    work page 2025

  8. [8]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv 2024

Show all 39 references
  1. [9]

    Liam Dugan, Alyssa Hwang, Filip Trhlik, Josh Magnus Ludan, Andrew Zhu, Hainiu Xu, Daphne Ippolito, and Chris Callison-Burch. 2024. https://arxiv.org/abs/2405.07940 Raid: A shared benchmark for robust evaluation of machine-generated text detectors . Preprint, arXiv:2405.07940

    2024 arXiv

  2. [10]

    Liam Dugan, Daphne Ippolito, Arun Kirubarajan, Sherry Shi, and Chris Callison-Burch. 2023. https://doi.org/10.1609/aaai.v37i11.26501 Real or fake text?: Investigating human ability to detect boundaries between human-written and machine-generated text . Proceedings of the AAAI ...

    2023 doi

  3. [11]

    Liam Dugan, Andrew Zhu, Firoj Alam, Preslav Nakov, Marianna Apidianaki, and Chris Callison-Burch. 2025. https://arxiv.org/abs/2501.08913 Genai content detection task 3: Cross-domain machine-generated text detection challenge . Preprint, arXiv:2501.08913

    2025 arXiv

  4. [12]

    Maciej Eder and Jan Rybicki. 2011. Stylometry with r. In DH, pages 308--310

    2011

  5. [13]

    Andrea Esuli, Fabrizio Falchi, Marco Malvaldi, and Giovanni Puccetti. 2024. You write like a gpt. In Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it)

    2024

  6. [14]

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. 2024. https://arxiv.org/abs/2401.12070 Spotting llms with binoculars: Zero-shot detection of machine-generated text . Preprint, arXiv:2401.12070

    2024 arXiv

  7. [15]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representat...

    2022

  8. [16]

    Xiaomeng Hu, Pin-Yu Chen, and Tsung-Yi Ho. 2023. Radar: Robust ai-text detection via adversarial learning. Advances in Neural Information Processing Systems, 36:15077--15095

    2023

  9. [17]

    Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.55 Reformulating unsupervised style transfer as paraphrase generation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 73...

    2020 doi

  10. [18]

    Lee, Yoo Sung Jang, and Jason Lee

    Bruce W. Lee, Yoo Sung Jang, and Jason Lee. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.834 Pushing on text readability assessment: A transformer meets handcrafted linguistic features . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Proces...

    2021 doi

  11. [19]

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. 2024. https://doi.org/10.18653/v1/2024.acl-long.3 MAGE : Machine-generated text detection in the wild . In Proceedings of the 62nd Annual Meeting of the Association for ...

    2024 doi

  12. [20]

    Dominik Macko, Jakub Kopal, Robert Moro, and Ivan Srba. 2024. https://arxiv.org/abs/2406.12549 Multisocial: Multilingual benchmark of machine-generated text detection of social-media texts . Preprint, arXiv:2406.12549

    2024 arXiv

  13. [21]

    Alessio Miaschi, Dominique Brunato, Felice Dell ' Orletta, and Giulia Venturi. 2020. https://doi.org/10.18653/v1/2020.coling-main.65 Linguistic profiling of a neural language model . In Proceedings of the 28th International Conference on Computational Linguistics, pages 745--7...

    2020 doi

  14. [22]

    Alessio Miaschi, Felice Dell ' Orletta, and Giulia Venturi. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.166 Evaluating large language models via linguistic profiling . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2835--...

    2024 doi

  15. [23]

    Manning, and Chelsea Finn

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. Detectgpt: zero-shot machine-generated text detection using probability curvature. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org

    2023

  16. [24]

    Cohen, and Mirella Lapata

    Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. https://doi.org/10.18653/v1/D18-1206 Don ' t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization . In Proceedings of the 2018 Conference on Empirical Methods in Natu...

    2018 doi

  17. [25]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 GPT-4 Technical Report

    2023 arXiv

  18. [26]

    Giovanni Puccetti, Anna Rogers, Chiara Alzetta, Felice Dell'Orletta, and Andrea Esuli. 2024. https://doi.org/10.18653/v1/2024.acl-long.817 Ai `news' content farms are easy to make and hard to detect: A case study in italian . In Proceedings of the 62nd Annual Meeting of the As...

    2024 doi

  19. [27]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/a85b405ed65c6477a4fe8302b5e06ce7-Paper-Conference.pdf Direct preference optimization: Your language model is ...

    2023

  20. [28]

    Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.827 D etect LLM : Leveraging log rank information for zero-shot detection of machine-generated text . In Findings of the Association for Computational Linguistics: EMNLP 20...

    2023 doi

  21. [29]

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. https://arxiv.org/abs/2211.09085 Galactica: A large language model for science . Preprint, arXiv:2211.09085

    2022 arXiv

  22. [30]

    Gemma Team. 2024. https://arxiv.org/abs/2408.00118 Gemma 2: Improving open language models at a practical size . Preprint, arXiv:2408.00118

    2024 arXiv

  23. [31]

    Adaku Uchendu, Thai Le, and Dongwon Lee. 2023. Attribution and obfuscation of neural text authorship: A data mining perspective. ACM SIGKDD Explorations Newsletter, 25(1):1--18

    2023

  24. [32]

    Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.673 Authorship attribution for neural text generation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384--8395, Onli...

    2020 doi

  25. [33]

    Hans van Halteren. 2004. https://doi.org/10.3115/1218955.1218981 Linguistic profiling for authorship recognition and verification . In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics ( ACL -04) , pages 199--206, Barcelona, Spain

    2004

  26. [34]

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mohanned Afzal, Tarek Mahmoud, Giovanni Puccetti, Thomas Arnold, Alham Fikri Aji, Nizar Habash, Iryna Gurevych, and Preslav Nakov. 2024 a . https://arxiv.org/abs/2402.11175 M4gt-bench: ...

    2024 arXiv

  27. [35]

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mohammed Afzal, Tarek Mahmoud, Giovanni Puccetti, and Thomas Arnold. 2024 b . https://doi.org/10.18653/v1/2024.semeval-1.279 S em E val-2024 task 8: Multidomain, multimodel and multilin...

    2024 doi

  28. [36]

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Chenxi Whitehouse, Osama Mohammed Afzal, Tarek Mahmoud, Toru Sasaki, Thomas Arnold, Alham Fikri Aji, Nizar Habash, Iryna Gurevych, and Preslav Nakov. 2024 c . https://aclanthology.org/2024.ea...

    2024

  29. [37]

    Yuxia Wang, Artem Shelmanov, Jonibek Mansurov, Akim Tsvigun, Vladislav Mikhailov, Rui Xing, Zhuohan Xie, Jiahui Geng, Giovanni Puccetti, Ekaterina Artemova, Jinyan Su, Minh Ngoc Ta, Mervat Abassy, Kareem Ashraf Elozeiri, Saad El Dine Ahmed El Etter, Maiya Goloburda, Tarek Mahm...

    2025

  30. [38]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  31. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.