Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Transforming Chatbot Text: A Sequence-to-Sequence Approach

T0 review · 4 major / 5 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 Fine-tuned T5-small and BART can rewrite GPT output so that six GPT-text classifiers lose up to about 19 percent relative accuracy, while retraining on the rewritten text restores detection to near-baseline levels.

desk verdict A competent but incremental attack study whose same-corpus evaluation limits the 'humanization' claim. read the letter →

arxiv 2506.12843 v1 pith:S6EKTPMB submitted 2025-06-15 cs.CL cs.LG

Natesh Reddy , Mark Stamp This is my paper
classification cs.CLcs.LG
keywords AI-generatedtextdetectionsequence-to-sequencetransformationT5BARTadversarialattackclassificationdataaugmentationGPTevasion
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

This paper argues that a sequence-to-sequence rewrite can serve as an adversarial transformation for AI-text detection. The authors fine-tune T5-small and BART on a paired dataset in which each GPT-generated sentence has a human-written counterpart, then use the models to rewrite GPT output into a 'humanized' form. Across six classifiers and three embedding types, accuracy of detectors trained on original GPT versus human text falls by up to about 19 percent relative to baseline when tested on the rewritten text. The same rewrite is not a true disguise, however: classifiers retrained on the rewritten text distinguish it from human text at accuracies close to the original baseline. If correct, this establishes a concrete attack/defense cycle for GPT-text detection and points to adversarial data augmentation as a workable defense.

What carries the argument

The central object is a fine-tuned encoder-decoder sequence-to-sequence transformation. A pretrained T5-small or BART model is fine-tuned in a supervised way on pairs where the input is a GPT-generated sentence prefixed with the instruction 'humanize' and the target is the corresponding human-written sentence from the same context. At inference the model rewrites GPT sentences into a style meant to resemble human prose, and the evaluation grid—six classifiers (logistic regression, random forest, XGBoost, MLP, DNN, LSTM) times three embeddings (Word2Vec, GloVe, BERT)—is what turns the rewrites into measurable accuracy changes.

What would settle it

Apply the same fine-tuning and evaluation to a second paired corpus with a different register, such as news or fiction, and compare the accuracy drop and the retrained accuracy. If the drop shrinks well below 19 percent or retrained classifiers fail to return to near-baseline accuracy, the paper's claim does not generalize. A complementary check: test the transformed text against a detector outside the six classifiers—a stronger pretrained detector or human judges—since the paper only evaluates the six classifiers trained on these embeddings.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that post-processing GPT-generated text with fine-tuned Seq2Seq models makes it measurably harder for trained classifiers to separate AI from human writing. Using T5-small and BART, each fine-tuned to map GPT sentences to paired human rewrites, the transformed GPT datasets lower detector accuracy from the 0.93-0.98 baseline range to as low as roughly 0.75 for word-vector embeddings and 0.72-0.91 for BERT embeddings, depending on the classifier. The method also shows a limit: when the same classifiers are retrained on the transformed text versus human text, accuracy returns to nearly baseline levels, with the best BERT plus deep-network configuration reaching about 0.98. The paper reads this as evidence that the transformation shifts the distribution of GPT text rather than erasing the difference, and that retraining on transformed text is a viable defense.

Load-bearing premise

The load-bearing premise is that each GPT-generated sentence in the training data has a valid human-written counterpart expressing the same content; if that pairing is loose or the WikiHow style is distinctive, the measured attack and defense numbers may describe one corpus rather than general human-likeness.

Editorial extensions

If this is right

  • A detector that has never seen rewritten text can be degraded by post-processing, so evaluation benchmarks for AI-text detection should include transformed text as a separate test condition.
  • Retraining on transformed text restores detection to near-baseline accuracy, so data augmentation with Seq2Seq-rewritten text is a usable defense against this attack.
  • The accuracy drop appears across all three embeddings and all six classifiers tested, so the effect is not specific to one classifier or feature representation.
  • The residual gap after retraining implies that the rewrite removes some GPT-like signals but leaves others that a classifier can still learn.

Reading between the lines

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

  • The paired-corpus design means the attack is only as portable as the pairing: on text types without clean human rewrites, the same fine-tuning recipe may fail, and on stylistically uniform corpora the apparent drop could partly reflect a shift away from the training distribution rather than toward human style.
  • One testable implication the author leaves implicit is that the residual signal learned by retrained classifiers is likely stylistic rather than semantic; measuring perplexity, vocabulary richness, or sentence-length variance before and after transformation would help isolate that signal.
  • An adversarial loop that feeds classifier feedback back into the Seq2Seq fine-tuning—mentioned in the paper as future work—would directly test whether the defense's recovered accuracy can itself be eroded, turning the reported static attack into a co-evolution.
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

4 major / 5 minor

Summary. The paper proposes fine-tuning T5-small and BART as sequence-to-sequence post-processors that rewrite GPT-generated text with the aim of making it more human-like, using paired human/GPT sentences from a WikiHow-derived dataset. It evaluates six classifiers (LR, RF, XGB, MLP, DNN, LSTM) with three embeddings (Word2Vec, GloVe, BERT) under three conditions: baseline human-vs-GPT classification, baseline classifiers tested on the Seq2Seq-transformed text, and classifiers retrained on the transformed text. The reported results show accuracy drops after transformation (Table 4), followed by a return to near-baseline accuracy after retraining (Table 5). The authors frame the transformation as both an adversarial attack on detectors and a data-augmentation defense, while acknowledging in Section 5.4 that the transformed text is not sufficiently human-like to evade retrained classifiers.

Significance. If the empirical claims hold, the paper demonstrates a useful and clearly reported result: fine-tuned Seq2Seq post-processing shifts GPT text out of the decision regions of static classifiers, while retraining on the shifted distribution restores detection. The systematic grid over six classifiers and three embeddings, with complete accuracy tables and confusion matrices, is a strength, as is the candid discussion in Section 5.4 that the transformed text remains distinguishable from human text. However, the closed-loop, single-corpus evaluation and the absence of control conditions mean the broader 'humanization' claim is not established; the evidence is best interpreted as a distribution-shift attack on static detectors rather than as evidence that the output is more human-like. The retraining results, especially BERT/DNN accuracy returning to 0.9845 after retraining, are informative for the attack/defense framing and give the paper value for the AI-text detection community.

major comments (4)
  1. [Section 5.2 / Table 4; Section 4] The accuracy drops in Table 4 are measured against human text drawn from the same WikiHow distribution that supplied the paired targets used to fine-tune the Seq2Seq models in Section 4. The transformed text is therefore optimized to resemble exactly the human distribution used as the 'human' class in the attack evaluation, so the observed drop could be a same-corpus distribution-shift artifact rather than evidence of general human-likeness. I ask for an out-of-corpus human test set and a control transformation (e.g., generic paraphrase or synonym substitution) to distinguish the humanization hypothesis from the hypothesis that the transformation simply moves GPT text outside the classifiers' training manifold.
  2. [Table 5 / Section 5.4] The retraining results are the clearest indicator of what the transformation actually does: after retraining on transformed text, the best models recover to essentially baseline accuracy (BERT/DNN: 0.9845 for T5 and 0.9825 for BART versus 0.9840 baseline). Section 5.4 correctly concedes that this shows the transformed text can still be distinguished from human text, which is the signature of a distribution shift, not of humanization. The abstract and Section 5.4 nevertheless describe the transformation as producing text that is 'more human-like'; this claim is not supported by the presented evidence and should be reframed or explicitly scoped to 'harder for static classifiers trained on the original GPT distribution.'
  3. [Section 4; Tables 3–5] The evaluation reports no variance, confidence intervals, or significance tests. Section 4 states that accuracies are based on 2,000 samples, and each entry in Tables 3–5 appears to come from a single run, yet the discussion treats small differences as meaningful (e.g., Table 5 BERT/MLP 0.9805 vs 0.9810). Because the headline claim is about significant degradation, the paper should report multiple runs with error bars and use a paired test such as McNemar's test for the key baseline-versus-transformed comparisons.
  4. [Table 4; Section 5.2] Even after the attack, the strongest detector (BERT/DNN) retains accuracy of 0.8940 for T5-transformed text and 0.9080 for BART-transformed text, which is far from evasion. The statement that the Seq2Seq approach 'significantly degrades' classifiers should be scoped to classifiers that have not been retrained and should explicitly acknowledge that the best detector still classifies nearly 90% of transformed samples correctly. In addition, the anomalously low BERT/LSTM baseline (0.8495 in Table 3) persists after retraining (0.8510 and 0.8390 in Table 5) and deserves an explanation, since that row drives the largest reported relative drops for BERT.
minor comments (5)
  1. [Section 3.1.1] The classifier is described as 'Linear Regression (LR)' in the opening sentence; it should be 'Logistic Regression' for consistency with the binary classification task.
  2. [Section 3.1.5] The definition of DNN as 'an MLP architecture with multiple hidden layers' is fine, but it would be clearer to state the exact number of layers used in the experiments, since the distinction between MLP and DNN is otherwise purely terminological.
  3. [Section 3.1.6] The acronym LSTM is defined in the section title but the term 'Long Short Term Memory' is used without a hyphen; consider standardizing to 'Long Short-Term Memory'.
  4. [Appendix] The appendix contains formatting issues such as 'T ransformed BAR T' and missing spaces after colons; these should be corrected in the final version.
  5. [Section 3.2.4 / 3.2.5] The subsection headers place T5-small and BART under 'Word Embeddings,' which is misleading; consider moving them to a separate 'Sequence-to-Sequence Models' subsection.

Circularity Check

1 steps flagged · score 6.0 of 10

Closed-loop evaluation: the 'human' test class is the same corpus used as the Seq2Seq fine-tuning target, so the attack drop measures in-corpus fit, not general human-likeness.

  1. fitted input called prediction [Section 3.3 (dataset [10]); Section 4 (paired fine-tuning); Section 5.2 (evaluation on same human data)]
    ""Each input is a GPT-generated sentence prefixed with a task-specific instruction (i.e., 'humanize') and the corresponding target is the human-written version which was taken from the same semantic context." ... "We then test each of the six classification models trained in the baseline case on these transformed samples, with the human-generated data again serving as the other class in each of these binary classification experiment.""

    The transform is fine-tuned to map GPT sentences onto human-written targets from the [10]/WikiHow corpus, and the attack evaluation uses the same corpus's human-generated data as the human class. Table 4's accuracy drop therefore measures how well the generated text matches the exact training-target distribution: in the limit of perfect training, output converges to the human class by construction, so classifier misclassification is forced. The retraining experiment (Table 5) confirms the transformed text is a new, separable distribution, so the drop is a same-corpus distribution-shift effect, not evidence of general human-likeness. The 'humanization' prediction is fitted to and evaluated on the same data, with no independent human-text corpus or control transformation tested.

full rationale

The paper is an empirical attack/defense study rather than a derivation, and most of its machinery is non-circular: the classifiers, embeddings, and retraining control are real experiments, and the retraining results actually provide an honest internal check that the transformed text is not genuinely human-like. However, the central claim that Seq2Seq post-processing makes text 'more human-like' and degrades detectors is evaluated in a closed loop. The Seq2Seq models are trained to map GPT sentences to human-written sentences taken from [10]/WikiHow, and the 'human' class in the attack evaluation is the same human-generated data from [10]. Thus the accuracy drop in Table 4 is, by construction, a measure of how well the transformation reproduces its own training-target distribution; it does not establish human-likeness relative to any other human corpus. The retraining control (Table 5) supports this reading: once classifiers see the transformed distribution, they separate it from human text at baseline accuracy, showing the transformed text is its own machine-like distribution rather than a genuinely humanized one. The self-citation to [10] is not itself the circularity, since the dataset is externally constructed and described; the circularity lies in using the same corpus as both the fine-tuning target and the evaluation's human class. This is a partial, evaluation-level circularity rather than equation-level derivational circularity, so a mid-range score is appropriate.

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

The central claim rests on the validity of the paired dataset and the operationalization of human-likeness as classifier accuracy. No new entities are introduced, and no free parameters are fitted to a derivation.

assumptions (3)
  • domain assumption Paired human/GPT data from [10] provides valid humanized targets.
    Section 4 assumes each GPT sentence has a human-written counterpart from the same semantic context, but the quality of this pairing is not validated beyond the dataset description.
  • domain assumption Classifier accuracy is a valid proxy for human-likeness.
    The paper defines human-likeness operationally as the inability of classifiers to distinguish, with no human evaluation, perplexity, or linguistic metrics.
  • domain assumption The WikiHow and GPT datasets from [10] are representative of human and AI text respectively.
    All conclusions rest on this single dataset, which comes from the same author group, so the representativeness and generalizability are not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transforming Chatbot Text: A Sequence-to-Sequence Approach." pith.science (2026). https://pith.science/paper/S6EKTPMB

@misc{pith2026250612843,
  author       = {Pith},
  title        = {Pith review of: Transforming Chatbot Text: A Sequence-to-Sequence Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6EKTPMB}},
  note         = {Machine review of arXiv:2506.12843}
}
read the original abstract

Due to advances in Large Language Models (LLMs) such as ChatGPT, the boundary between human-written text and AI-generated text has become blurred. Nevertheless, recent work has demonstrated that it is possible to reliably detect GPT-generated text. In this paper, we adopt a novel strategy to adversarially transform GPT-generated text using sequence-to-sequence (Seq2Seq) models, with the goal of making the text more human-like. We experiment with the Seq2Seq models T5-small and BART which serve to modify GPT-generated sentences to include linguistic, structural, and semantic components that may be more typical of human-authored text. Experiments show that classification models trained to distinguish GPT-generated text are significantly less accurate when tested on text that has been modified by these Seq2Seq models. However, after retraining classification models on data generated by our Seq2Seq technique, the models are able to distinguish the transformed GPT-generated text from human-generated text with high accuracy. This work adds to the accumulating knowledge of text transformation as a tool for both attack -- in the sense of defeating classification models -- and defense -- in the sense of improved classifiers -- thereby advancing our understanding of AI-generated text.

Figures

Figures reproduced from arXiv: 2506.12843 by the authors.

Figure 1
Figure 1. Baseline confusion matrices (GloVe embeddings) [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Confusion matrices after transformation (GloVe and T5-small) [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Accuracy comparison for transformed text [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Confusion matrices after retraining (GloVe and T5-small) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Accuracy comparison for retraining 5.4 Discussion Our experiments demonstrate that by post-processing GPT-generated text us￾ing a Seq2Seq approach, we can significantly degrade the performance of classi￾fication models that have been trained to distinguish GPT-generate…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [36]

    I wish I had not given up my evening plans to watch my sister’s children. I really needed that time to relax and I feel angry at her for asking me to give that up

    Ying Zhou, Ben He, and Le Sun. Humanizing machine-generated content: Evading AI-text detection through adversarial attack. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Lan- guage Resources and Evaluation , LREC–COLING, pages 8427–8437, 2024. Appendix This appendix provides representative examples showing the orig...

    work page 2024

  2. [16]

    Kadhim, Lei Jiao, Rishad A

    Ahmed K. Kadhim, Lei Jiao, Rishad A. Shafik, and Ole-Christoffer Granmo. Adversarial attacks on AI-generated text detection models: A token probability-based approach using embeddings. https://arxiv.org/abs/ 2501.18998, 2025

    arXiv 2025

  3. [1]

    GPT-4 technical report

    Josh Achiam et al. GPT-4 technical report. https://arxiv.org/abs/ 2303.08774, 2024

    arXiv 2024

  4. [2]

    Model card and evaluations for Claude models

    Anthropic. Model card and evaluations for Claude models. https://www- cdn.anthropic.com/bd2a28d2535bfb0494cc8e2a3bf135d2e7523226/ Model-Card-Claude-2.pdf, 2023. 17

    work page 2023

  5. [3]

    Ernest Yeboah Boateng and Daniel A. Abaye. A review of the logistic re- gression model with emphasis on medical research. Journal of Data Analysis and Information Processing , 7:190–207, 2019

    work page 2019

  6. [4]

    Yaping Chai, Haoran Xie, and Joe S. Qin. Text data augmentation for large language models: A comprehensive survey of methods, challenges, and opportunities. https://arxiv.org/abs/2501.18845, 2025

    arXiv 2025

  7. [5]

    XGBoost: A scalable tree boosting sys- tem

    Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting sys- tem. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , KDD ’16, pages 785–794, 2016

    work page 2016

  8. [6]

    PaLM: Scaling language modeling with path- ways

    Aakanksha Chowdhery et al. PaLM: Scaling language modeling with path- ways. Journal of Machine Learning Research , 24(1), 2023

    work page 2023

Show all 36 references
  1. [7]

    Scaling instruction-finetuned language models

    Hyung Won Chung et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research , 25(1):1–53, 2024

    2024

  2. [8]

    BERT: Pre-training of deep bidirectional transformers for language under- standing

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan...

    2019

  3. [9]

    HotFlip: White- box adversarial examples for text classification

    Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. HotFlip: White- box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages 31–36, 2018

    2018

  4. [10]

    Dis- tinguishing chatbot from human

    Gauri Anil Godghase, Rishit Agrawal, Tanush Obili, and Mark Stamp. Dis- tinguishing chatbot from human. In Mark Stamp and Martin Jureˇ cek, edi- tors, Machine Learning, Deep Learning, and AI for Cybersecurity . Springer, 2025

    2025

  5. [11]

    Detection and moderation of detrimental content on social media platforms: Current status and future directions

    Vaishali Gongane, Mousami Munot, and Devidas Anuse. Detection and moderation of detrimental content on social media platforms: Current status and future directions. Social Network Analysis and Mining , 12(1), 2022

    2022

  6. [12]

    Garrido-Merch´ an

    Roberto Gozalo-Brizuela and Eduardo C. Garrido-Merch´ an. A survey of generative AI applications. Journal of Computer Science , 20(8):801–818, 2024

    2024

  7. [13]

    Devising and detecting phishing emails using large language models

    Fredrik Heiding et al. Devising and detecting phishing emails using large language models. IEEE Access, 12:42131–42146, 2024

    2024

  8. [14]

    AI, write an essay for me: A large-scale comparison of human-written versus ChatGPT-generated essays

    Steffen Herbold et al. AI, write an essay for me: A large-scale comparison of human-written versus ChatGPT-generated essays. https://arxiv.org/ abs/2304.14276, 2023

    2023 arXiv

  9. [15]

    Jiang et al

    Albert Q. Jiang et al. Mistral 7B. https://arxiv.org/abs/2310.06825, 2023. 18

    2023 arXiv

  10. [17]

    WikiHow: A large scale text summarization dataset

    Mahnaz Koupaee and William Yang Wang. WikiHow: A large scale text summarization dataset. https://arxiv.org/abs/1810.09305, 2018

    2018 arXiv

  11. [18]

    Miles McCain, and Miles Brundage

    Sarah Kreps, R. Miles McCain, and Miles Brundage. All the news that’s fit to fabricate: AI-generated text as a tool of media misinformation. Journal of Experimental Political Science , 9(1):104–117, 2022

    2022

  12. [19]

    BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension

    Mike Lewis et al. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. InProceedings of the 58th Annual Meeting of the Association for Computational Linguis- tics, pages 7871–7880, 2020

    2020

  13. [20]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. https://arxiv.org/abs/1711.05101, 2019

    2019 arXiv

  14. [21]

    Intellectual property rights and AI-generated content — issues in human authorship, fair use doctrine, and output liability

    Adnan Masood. Intellectual property rights and AI-generated content — issues in human authorship, fair use doctrine, and output liability. https://medium.com/@adnanmasood/intellectual-property-rights- and-ai-generated-content-issues-in-human-authorship-fair-use- 8c7ec9d6fdc3, 2025

    2025

  15. [22]

    Efficient esti- mation of word representations in vector space

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient esti- mation of word representations in vector space. https://arxiv.org/abs/ 1301.3781, 2013

    2013 arXiv

  16. [23]

    GloVe: Global Vectors for Word Representation

    Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global Vectors for Word Representation. InProceedings of the 2014 Confer- ence on Empirical Methods in Natural Language Processing , EMNLP, pages 1532–1543, 2014

    2014

  17. [24]

    Exploring the limits of transfer learning with a uni- fied text-to-text transformer

    Colin Raffel et al. Exploring the limits of transfer learning with a uni- fied text-to-text transformer. Journal of Machine Learning Research , 21(1), 2020

    2020

  18. [25]

    Adversarial training using feedback loops

    Ali Haisam Muhammad Rafid and Adrian Sandu. Adversarial training using feedback loops. https://arxiv.org/abs/2308.11881, 2023

    2023 arXiv

  19. [26]

    The power of human-like virtual-influencer- generated content: Impact on consumers’ willingness to follow and purchase intentions

    Warinrampai Rungruangjit, Kulachet Mongkol, Intaka Piriyakul, and Kitti Charoenpornpanichkul. The power of human-like virtual-influencer- generated content: Impact on consumers’ willingness to follow and purchase intentions. Computers in Human Behavior Reports , 16:100523, 2024

    2024

  20. [27]

    Introduction to Machine Learning with Applications in Infor- mation Security

    Mark Stamp. Introduction to Machine Learning with Applications in Infor- mation Security. Chapman and Hall/CRC, 2nd edition, 2022

    2022

  21. [28]

    Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learn- ing with neural networks. In Proceedings of the 28th International Con- ference on Neural Information Processing Systems — Volume 2 , NIPS’14, page 3104–3112, 2014. 19

    2014

  22. [29]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron et al. Llama 2: Open foundation and fine-tuned chat models. https://arxiv.org/abs/2307.09288, 2023

    2023 arXiv

  23. [30]

    Word representa- tions: A simple and general method for semi-supervised learning

    Joseph Turian, Lev-Arie Ratinov, and Yoshua Bengio. Word representa- tions: A simple and general method for semi-supervised learning. In Pro- ceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 384–394, 2010

    2010

  24. [31]

    Long Short-term Memory RNN

    Christian Bakke Vennerød, Adrian Kjærran, and Erling Stray Bugge. Long Short-term Memory RNN. https://arxiv.org/abs/2105.06756, 2021

    2021 arXiv

  25. [32]

    Universal adversarial triggers for attacking and analyz- ing NLP

    Eric Wallace et al. Universal adversarial triggers for attacking and analyz- ing NLP. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP–IJCNLP, pages 2153–2...

    2019

  26. [33]

    A survey on LLM-generated text detection: Necessity, methods, and future directions

    Junchao Wu et al. A survey on LLM-generated text detection: Necessity, methods, and future directions. Computational Linguistics , 51(1):275–338, 2025

    2025

  27. [34]

    Multi-layer perceptron trainability explained via variability

    Yueyao Yu and Yin Zhang. Multi-layer perceptron trainability explained via variability. https://arxiv.org/abs/2105.08911, 2023

    2023 arXiv

  28. [35]

    Defending against neural fake news

    Rowan Zellers et al. Defending against neural fake news. https://arxiv. org/abs/1905.12616, 2020

    1905 arXiv

Pith tools

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