Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Detector performance measured on the standard human-vs-direct-LLM benchmark fails to transfer to human text that an LLM rewrites, with recall at a 1% false-positive rate dropping from 91-94% to 15-31%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:00 UTC pith:VQZC5RHL

load-bearing objection Solid matched benchmark showing detectors collapse on human text rewritten by an LLM, even though the origin-vs-intensity confound keeps the LLM2L comparison from being fully causal. the 3 major comments →

arxiv 2607.29539 v1 pith:VQZC5RHL submitted 2026-07-31 cs.CL cs.AI

ARB: A Matched Authorship-Rewriting Benchmark Dataset for AI-Text Detector Evaluation

Gaetano Perrone , Simon Pietro Romano This is my paper
classification cs.CL cs.AI
keywords AI-text detectionbenchmark transferLLM rewritingauthorship regimeslow false-positive evaluationparaphrase robustnesshuman-in-the-loop writingmatched benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that AI-text detection scores measured on the conventional human-versus-direct-LLM benchmark do not predict how well a detector catches human-authored text that an LLM has rewritten. In its matched benchmark, the two strongest zero-shot detectors catch over 91% of directly generated LLM text but only 15-31% of human-origin LLM-rewritten text, while LLM-only rewriting of LLM text loses only 10-13 points. The authors argue that standard evaluation therefore overestimates real-world robustness for the common workflow of polishing human drafts with an LLM. The reason to care is that such overestimation could lead to overconfident deployment of detectors in academic-integrity, plagiarism-review, and content-provenance settings.

Core claim

The paper's central discovery is that performance estimated on the conventional HUMAN-vs-FREE-LLM condition transfers to same-generator second-pass rewriting of LLM text but not to LLM rewriting of human text. FastDetectGPT and Binoculars-falcon-7b retained 78.3% and 83.0% TPR@1%FPR on LLM2L, but fell to 30.8% and 15.1% on H2L, with RADAR showing the same pattern (66.8% to 12.2%). The authors interpret this as a benchmark-transfer failure: an LLM-mediated surface alone does not explain detector degradation, because the two LLM-mediated regimes remain far apart. They carefully label the difference an 'operational source-origin-associated gap' rather than a pure causal effect of origin, since

What carries the argument

The matched four-regime quartet is the central object: each source text anchors HUMAN (original human text), FREE-LLM (direct LLM generation from a topic), H2L (LLM rewrite of the human text), and LLM2L (same-generator rewrite of the FREE-LLM output). The same-generator LLM2L control is what carries the argument: it holds the rewriting instruction and generator fixed while varying only the input's origin, turning rewriting from a generic attack into a diagnostic comparison. Paired block-level deltas, computed within dataset-by-generator blocks and macro-averaged, together with recall at a fixed 1% false-positive rate, are the measurement machinery that makes the transfer failure visible.

Load-bearing premise

The load-bearing premise is that holding the rewrite instruction and generator fixed while varying only whether the input is human or LLM text isolates content origin as the cause of the performance gap, but the two rewriting paths are not matched on transformation strength (token edit distance 0.602 versus 0.348), so the gap could partly reflect how much the text was changed rather than who wrote the original.

What would settle it

Produce H2L and LLM2L rewrites with overlapping distributions of normalized edit distance, word ratio, and lexical overlap, then recompute TPR@1%FPR. If the operational gap collapses to near zero, transformation strength, not content origin, is the driver; if a large gap persists under matched transformation, the origin-specific transfer-failure claim holds. The released benchmark's per-pair diagnostics permit a subset analysis on pairs with similar edit distance to run this check directly.

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

If this is right

  • A detector evaluated only on human-versus-direct-LLM text can overestimate its recall on AI-assisted editing workflows by 60 or more percentage points.
  • Human-origin LLM-rewritten text and LLM-origin LLM-rewritten text are different detection targets; pooling them into one 'rewritten' class hides the failure.
  • Deployers should validate detectors on H2L-style data at a fixed low false-positive rate before using them in high-stakes settings.
  • Detector-family rankings differ by regime, so a single aggregate leaderboard score is not a reliable basis for tool selection.
  • Supervised encoder baselines remain below 3% recall at the 1% FPR operating point in every regime, so they are not a safe fallback even on the standard benchmark.

Where Pith is reading between the lines

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

  • The origin-associated asymmetry suggests a complementary detection channel: if LLMs edit human text more than they edit LLM text, then rewrite-distance signals could be exploited as a detector; the paper mentions this mechanism in related work but does not build on it.
  • A testable extension is to match H2L and LLM2L rewrites on transformation strength (e.g., overlapping normalized edit distance) and re-measure the gap; the authors' own diagnostics indicate this confounding remains open.
  • If the transfer failure is generic, detectors fine-tuned on H2L examples may not generalize to unseen rewriting styles, domains, or generators; the paper calls for such tests but does not run them.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces ARB, a benchmark with 1,800 human source texts from XSum, WritingPrompts, and OpenWebText, and four open-weight LLMs, producing four matched regimes: HUMAN, FREE-LLM, H2L (LLM rewrite of human text), and LLM2L (same-generator rewrite of LLM text). Five detectors are evaluated under a fixed 1% FPR operating point using block-level paired deltas and bootstrap confidence intervals. The central empirical finding is that detectors with high recall on direct LLM text (FastDetectGPT 91.2%, Binoculars 93.5%, RADAR 66.8%) lose most low-FPR recall on H2L (30.8%, 15.1%, 12.2%), while LLM2L remains much closer to the direct-generation baseline (78.3%, 83.0%, 64.6%). The paper interprets this as a benchmark-transfer failure: conventional HUMAN-vs-FREE-LLM evaluation overestimates detector robustness for human-origin text revised by an LLM. The authors explicitly note that H2L and LLM2L differ in transformation strength and frame the H2L-LLM2L gap as an operational rather than causal estimate.

Significance. If the result holds, ARB is a valuable, reusable benchmark: it ships a large released dataset, reproducible generation/evaluation pipeline, fixed detector configurations, block-level matching, and low-FPR reporting. The direct transfer-failure finding for H2L is well supported by tight bootstrap intervals and is practically important for academic-integrity and content-provenance deployments. The paper is also careful to avoid detector-threshold tuning and to report stratified heterogeneity. However, the more novel explanatory contrast - that detectors are 'largely robust to LLM-only rewriting' and that the H2L-LLM2L gap reflects content origin - is currently confounded by transformation strength. As a benchmark resource and a cautionary evaluation, the contribution is significant; as a causal separation of authorship origin from rewriting intensity, it is not yet established.

major comments (3)
  1. [§6.4, §6.6 (Table 10), Abstract] The transformation-strength confound is load-bearing for the headline clause 'even though the same detectors remain largely robust to LLM-only rewriting.' Table 10 shows Human→H2L has token NED 0.602 and Jaccard 0.463, while Free-LLM→LLM2L has NED 0.348 and Jaccard 0.587. Section 6.4 itself states that 'the two factors are partially confounded,' and Section 7.5 lists transformation-strength confounding as a central limitation. Since LLM2L applies a much gentler surface edit, the large H2L-LLM2L gaps in Table 9 (e.g., 0.680 for Binoculars, 0.475 for FastDetectGPT) could be produced by edit intensity alone rather than by content origin. The abstract and Section 7.1 nevertheless assert robustness to LLM-only rewriting without this qualification. The authors should either add a matched-intensity rewriting control or substantially weaken the origin-specific wording in the abstract, Section 1,
  2. [§4.6, §5.4, Table 9] The 'matched' design matches source items and generator blocks, but not the input text properties of the two rewriting paths. FREE-LLM is generated from a topic rather than from the HUMAN source, and Table 5 shows H2L outputs are on average longer than LLM2L outputs (227 vs 210 words; HUMAN is 297 words). Since likelihood- and curvature-based detectors are sensitive to text length and genre, part of the H2L-LLM2L gap could be driven by the different input/output distributions of the two paths rather than by authorship origin. The paper reports length-stratified sampling and block-level pairing, but the final texts differ substantially across regimes. At minimum, the authors should explicitly discuss length and text-type differences as an additional confound, and consider a sensitivity analysis stratified by output length or a matched-length control.
  3. [§5.1, §6.3, §7.5] The 'largely robust to LLM-only rewriting' conclusion is derived from a single operationalization: a same-generator, single-pass, non-adversarial rewrite with one fixed prompt. The paper itself cites adversarial and iterative rewriting work showing much larger degradation (e.g., TempParaphraser, GradEscape, HUMPA), so the statement in the abstract is broader than the evidence. This does not invalidate the benchmark-transfer finding for the tested rewrite protocol, but the claim should be conditioned on 'non-adversarial same-generator second-pass rewriting' in the abstract and conclusion to avoid implying robustness to arbitrary LLM-only rewriting.
minor comments (5)
  1. [Abstract] Typographical issue: '93.5\%' contains a stray backslash in the abstract text.
  2. [Figures 2 and 3] The x-axis label 'OpenWebT ext' appears to be a split rendering of 'OpenWebText'. Please fix for clarity.
  3. [§4.6] The sentence 'all 7,200 quartets (23,400 individual texts)' is confusing because each source item's HUMAN text is shared across the four generator blocks; the text could clarify that the 7,200 quartets share 1,800 HUMAN items, giving 1,800 + 3×7,200 = 23,400 unique texts.
  4. [Table 6] Detector names are inconsistent: 'Binoculars-falcon-7b' and 'binoculars-falcon-7b' are used in different places (Table 6 vs. text). Standardize capitalization.
  5. [§7.5] The 'Transformation-strength confounding' paragraph is the most important limitation and should be reflected in the abstract and Section 1, not only in Section 7.5. Consider moving a version of that caveat to the front matter.

Circularity Check

0 steps flagged

No significant circularity: ARB is an empirical benchmark evaluation with no fitted-parameter-as-prediction step, no self-citation chain carrying the argument, and no equation that reduces to its own inputs.

full rationale

The paper reports a matched empirical benchmark measurement and does not claim a derivation whose output is contained in its input. Detector thresholds are fixed at TPR@1%FPR computed against the shared HUMAN reference, which is a standard operating-point definition rather than a fitted parameter; all detector scores come from released implementations with configurations fixed across regimes, and none of these configurations is tuned on ARB labels or on the headline H2L–LLM2L comparison. The central claim—that conventional HUMAN-vs-FREE-LLM performance does not transfer to human-origin LLM rewriting—is a contingent empirical finding, not a tautology. The paper's own limitations (Sections 6.4 and 7.5) concede that H2L and LLM2L are partially confounded with transformation strength (token NED 0.602 vs 0.348; Jaccard 0.463 vs 0.587). This weakens the causal/origin-specific interpretation of the operational gap, but it is a validity/confounding concern, not circularity: the transformation diagnostics are not detector inputs, and the paper explicitly reframes the result as an 'operational source-origin-associated gap' rather than a causal estimate. There are no load-bearing self-citations: the references include no prior work by the present authors, and external corroborating works (e.g., MAGRET, RAIDAR) are used only as independent supporting evidence, not as the basis for the measured results. No fitted value is renamed as a prediction, and no equation reduces to its own inputs. Honest non-finding is therefore appropriate.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No fitted parameters or invented entities are introduced. The benchmark's main epistemic burden is the domain assumption that the matched design isolates content origin despite the acknowledged transformation-strength confound, plus the usual representativeness assumptions for corpora, generators, and the 1% FPR operating point.

axioms (4)
  • domain assumption TPR@1%FPR on a fixed HUMAN reference is a valid operational measure of detector performance.
    Used throughout Section 5.3; assumes a 1% false-positive constraint is the right operating point for deployment contexts where false accusations are costly.
  • domain assumption The matched four-regime design with fixed rewrite instruction and same generator makes H2L vs LLM2L a comparison of content origin.
    Stated in Sections 4.1 and 4.6, but Section 7.5 concedes transformation strength differs (NED 0.602 vs 0.348), so the assumption is only partially met.
  • domain assumption XSum, WritingPrompts, and OpenWebText human texts represent the human-written English distribution for the domains tested.
    Section 4.2; external validity is restricted to these three corpora and to English.
  • domain assumption Four open-weight instruction-tuned generators are representative of LLM rewriting behavior.
    Section 4.4; closed-source generators, cross-model rewriting, and multilingual rewriting are explicitly out of scope.

pith-pipeline@v1.3.0-daily-deepseek · 33154 in / 11067 out tokens · 110551 ms · 2026-08-03T05:00:51.395717+00:00 · methodology

0 comments
read the original abstract

Standard AI-text detection benchmarks compare human-written text against text generated directly by large language models (LLMs). While prior work has shown that rewriting and paraphrasing can degrade detector performance, it remains unclear whether performance measured on this conventional benchmark predicts detector behavior when human-authored content is rewritten by an LLM. To address this gap, we introduce Authorship-Rewriting Benchmark (ARB), built from 1,800 human source texts (600 each from XSum, WritingPrompts, and OpenWebText) and four open-weight generators (Llama-3.2-3B, Qwen2.5-7B, Mistral-7B, Gemma-2-9B). Each source item yields four matched variants: human-written (HUMAN), direct LLM generation (Free-LLM), LLM-rewritten human text (H2L), and same-generator LLM-rewritten LLM text (LLM2L). We evaluated five detectors (FastDetectGPT, Binoculars-falcon-7b, RADAR, BERT-Defense, RoBERTa-Defense) at a strict 1%-false-positive operating point (TPR@1%FPR). FastDetectGPT and Binoculars-falcon-7b detected 91.2% and 93.5\% of direct LLM text, but only 30.8% and 15.1% of human text an LLM had rewritten, a drop of 60-78 percentage points. The same detectors retained 78.3% and 83.0% recall when LLM text was rewritten by the same model, a much smaller decline of 10-13 points. RADAR followed the same pattern (66.8% to 12.2%), while BERT-Defense and RoBERTa-Defense stayed below 3% recall across all regimes. These results show that detector performance measured on the conventional human-vs-LLM benchmark does not transfer to human-authored text revised by an LLM, even though the same detectors remain largely robust to LLM-only rewriting.

Figures

Figures reproduced from arXiv: 2607.29539 by Gaetano Perrone, Simon Pietro Romano.

Figure 1
Figure 1. Figure 1: End-to-end ARB construction and evaluation workflow. A filtered, length-stratified human source item anchors four matched regimes. Human retains the source text; Free-LLM is generated from a dataset-specific topic; H2L rewrites the human source; and LLM2L rewrites the corresponding Free-LLM output with the same generator. Only complete validated quartets proceed to detector scoring and paired block-level a… view at source ↗
Figure 2
Figure 2. Figure 2: Mean ΔTPR@1%FPR under H2L by detector and dataset. Each cell is the mean across generators of the paired block-level difference between H2L and Free-LLM (H2L minus Free-LLM). Negative values indicate degradation relative to direct LLM generation. FastDetectGPT decreases from TPR@1%FPR 0.912 in FREE-LLM to 0.783 in LLM2L, with ΔTPR@1%FPR = −0.130. Binoculars-falcon-7b decreases from 0.935 to 0.830, with ΔTP… view at source ↗
Figure 3
Figure 3. Figure 3: Mean ΔTPR@1%FPR under LLM2L by detector and dataset. Each cell is the mean across generators of the paired block-level difference between LLM2L and Free-LLM (LLM2L minus Free-LLM). Negative values indicate degradation relative to direct LLM generation. BERT-Defense RoBERTa-Defense FastDetectGPT binoculars-falcon-7b radar Detector 0.0 0.2 0.4 0.6 0.8 1.0 Mean TPR@1% FPR FreeLLM LLM2L H2L [PITH_FULL_IMAGE:f… view at source ↗
Figure 4
Figure 4. Figure 4: Detection performance across generation regimes. Bars show the detector-level mean TPR@1%FPR for Free￾LLM, LLM2L, and H2L, macro-averaged across dataset × generator blocks. Error bars denote block-structured bootstrap 95% confidence intervals. The contrast between H2L and LLM2L visualizes the source-origin gap under LLM-mediated surface. has a gap of 0.475 ([0.444, 0.511]). All three intervals exclude zero… view at source ↗
Figure 5
Figure 5. Figure 5: Textual change induced by the two rewriting paths. H2L corresponds to Human → H2L, whereas LLM2L corresponds to Free-LLM → LLM2L. The figure shows that H2L introduces stronger surface changes, while LLM2L is a more conservative second-pass rewrite. Error bars denote block-structured bootstrap 95% confidence intervals. 7. Discussion The results have implications for detector robustness, benchmark validity, … view at source ↗
Figure 6
Figure 6. Figure 6: AUROC heatmaps for the H2L regime. Each panel corresponds to one detector; rows are generator models, and columns are datasets. Values are averaged within each generator × dataset cell. All panels share the same 0–1 color scale. First Author et al.: Preprint submitted to Elsevier Page 27 of 27 [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: TPR@1%FPR heatmaps for the H2L regime. Each panel corresponds to one detector; rows are generator models and columns are datasets. Values are averaged within each generator × dataset cell. All panels share the same 0–1 color scale. First Author et al.: Preprint submitted to Elsevier Page 28 of 27 [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 7 canonical work pages · 4 internal anchors

  1. [1]

    , author Shahriar, S

    author Ayoobi, N. , author Shahriar, S. , author Mukherjee, A. ( year 2025 ). title Beyond Easy Wins: A Text Hardness-Aware Benchmark for LLM-generated Text Detection . howpublished [Preprint]. arXiv . :10.48550/arXiv.2507.15286

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.15286 2025

  2. [2]

    , author Baidya, S.S

    author Baidya, M.S. , author Baidya, S.S. , author Chawla, C. ( year 2026 ). title Detecting the Machine: A Comprehensive Benchmark of AI -Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions . howpublished [Preprint]. arXiv . :10.48550/arXiv.2603.17522

    work page doi:10.48550/arxiv.2603.17522 2026

  3. [3]

    , author Zhao, Y

    author Bao, G. , author Zhao, Y. , author Teng, Z. , author Yang, L. , author Zhang, Y. ( year 2024 ). title Fast- DetectGPT : Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature , in: booktitle The Twelfth International Conference on Learning Representations

    2024

  4. [4]

    , author Feng, Y

    author Chen, Z. , author Feng, Y. , author He, C. , author Deng, Y. , author Pu, H. , author Li, B. ( year 2025 a). title IPAD : Inverse Prompt for AI Detection -- A Robust and Explainable LLM -Generated Text Detector . howpublished [Preprint]. arXiv . :10.48550/arXiv.2502.15902

    doi:10.48550/arxiv.2502.15902 2025

  5. [5]

    , author He, K

    author Chen, Z. , author He, K. , author Huang, Y. , author Zhu, Y. , author Feng, M. ( year 2025 b). title D iv S core: Zero-Shot Detection of LLM -Generated Text in Specialized Domains , in: editor Christodoulopoulos, C. , editor Chakraborty, T. , editor Rose, C. , editor Peng, V. (Eds.), booktitle Proceedings of the 2025 Conference on Empirical Methods...

    2025

  6. [6]

    , author Kamar, E

    author Chung, J. , author Kamar, E. , author Amershi, S. ( year 2023 ). title Increasing Diversity While Maintaining Accuracy: Text Data Generation with Large Language Models and Human Interventions , in: booktitle Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher Association for Co...

    doi:10.18653/v1/2023.acl-long.34 2023

  7. [7]

    , author Hwang, A

    author Dugan, L. , author Hwang, A. , author Trhlik, F. , author Ludan, J.M. , author Zhu, A. , author Xu, H. , author Ippolito, D. , author Callison-Burch, C. ( year 2024 ). title RAID : A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors , in: booktitle Proceedings of the 62nd Annual Meeting of the Association for Computational ...

    2024

  8. [8]

    'Your AI Text is not Mine': Redefining and Evaluating AI-generated Text Detection under Realistic Assumptions

    author Dycke, N. , author Sakharova, M. , author Daheim, N. , author Gurevych, I. ( year 2026 ). title 'Your AI Text is not Mine': Redefining and Evaluating AI -generated Text Detection under Realistic Assumptions . howpublished [Preprint]. arXiv . :10.48550/arXiv.2606.04906

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.04906 2026

  9. [9]

    , author Lewis, M

    author Fan, A. , author Lewis, M. , author Dauphin, Y. ( year 2018 ). title Hierarchical Neural Story Generation , in: editor Gurevych, I. , editor Miyao, Y. (Eds.), booktitle Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher Association for Computational Linguistics , address Melbo...

    doi:10.18653/v1/p18-1082 2018

  10. [10]

    , author Strobelt, H

    author Gehrmann, S. , author Strobelt, H. , author Rush, A.M. ( year 2019 ). title GLTR : Statistical Detection and Visualization of Generated Text , in: booktitle Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , publisher Association for Computational Linguistics , address Florence, Italy . ...

    doi:10.18653/v1/p19-3019 2019

  11. [11]

    , author Cohen, V

    author Gokaslan, A. , author Cohen, V. , author Pavlick, E. , author Tellex, S. ( year 2019 ). title OpenWebText Corpus . howpublished http://Skylion007.github.io/OpenWebTextCorpus

    2019

  12. [12]

    , author Dubey, A

    author Grattafiori, A. , author Dubey, A. , author Jauhri, A. , author Pandey, A. , author Kadian, A. , author Al-Dahle, A. , author Letman, A. , author Mathur, A. , author Schelten, A. , author Vaughan, A. , author Yang, A. , author Fan, A. , author Goyal, A. , author Hartshorn, A. , author Yang, A. , author Mitra, A. , author Sravankumar, A. , author Ko...

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  13. [13]

    , author Schwarzschild, A

    author Hans, A. , author Schwarzschild, A. , author Cherepanova, V. , author Kazemi, H. , author Saha, A. , author Goldblum, M. , author Geiping, J. , author Goldstein, T. ( year 2024 ). title Spotting LLM s with Binoculars: Zero-Shot Detection of Machine-Generated Text , in: booktitle Proceedings of the 41st International Conference on Machine Learning ,...

    2024

  14. [14]

    , author Li, R

    author Hao, W. , author Li, R. , author Zhao, W. , author Yang, J. , author Mao, C. ( year 2025 ). title Learning to Rewrite: Generalized LLM -Generated Text Detection , in: editor Che, W. , editor Nabende, J. , editor Shutova, E. , editor Pilehvar, M.T. (Eds.), booktitle Proceedings of the 63rd Annual Meeting of the Association for Computational Linguist...

    2025

  15. [15]

    , author Chen, P.Y

    author Hu, X. , author Chen, P.Y. , author Ho, T.Y. ( year 2023 ). title RADAR : Robust AI -text Detection via Adversarial Learning , in: booktitle Proceedings of the 37th International Conference on Neural Information Processing Systems , publisher Curran Associates Inc. , address Red Hook, NY, USA . pp. pages 15077--15095

    2023

  16. [16]

    , author Zhang, R

    author Huang, J. , author Zhang, R. , author Su, J. , author Chen, Y. ( year 2025 a). title T emp P araphraser: ``Heating Up'' Text to Evade AI -Text Detection through Paraphrasing , in: editor Christodoulopoulos, C. , editor Chakraborty, T. , editor Rose, C. , editor Peng, V. (Eds.), booktitle Proceedings of the 2025 Conference on Empirical Methods in Na...

    doi:10.18653/v1 2025

  17. [17]

    , author Cao, J

    author Huang, Y. , author Cao, J. , author Luo, H. , author Guan, X. , author Liu, B. ( year 2025 b). title MAGRET : Machine-generated Text Detection with Rewritten Texts , in: editor Rambow, O. , editor Wanner, L. , editor Apidianaki, M. , editor Al-Khalifa, H. , editor Eugenio, B.D. , editor Schockaert, S. (Eds.), booktitle Proceedings of the 31st Inter...

    2025

  18. [18]

    , author Duckworth, D

    author Ippolito, D. , author Duckworth, D. , author Callison-Burch, C. , author Eck, D. ( year 2020 ). title Automatic Detection of Generated Text is Easiest when Humans are Fooled , in: booktitle Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , publisher Association for Computational Linguistics , address Online ....

    doi:10.18653/v1/2020.acl-main.164 2020

  19. [19]

    , author Sablayrolles, A

    author Jiang, A.Q. , author Sablayrolles, A. , author Mensch, A. , author Bamford, C. , author Chaplot, D.S. , author de las Casas, D. , author Bressand, F. , author Lengyel, G. , author Lample, G. , author Saulnier, L. , author Lavaud, L.R. , author Lachaux, M.A. , author Stock, P. , author Le Scao, T. , author Lavril, T. , author Wang, T. , author Lacro...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.06825 2023

  20. [20]

    , author Ahmad, L

    author Kirchner, J.H. , author Ahmad, L. , author Aaronson, S. , author Leike, J. ( year 2023 ). title New AI classifier for indicating AI -written text . howpublished https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/ . note Accessed: 16-Apr-2025

    2023

  21. [21]

    , author Song, Y

    author Krishna, K. , author Song, Y. , author Karpinska, M. , author Wieting, J. , author Iyyer, M. ( year 2023 ). title Paraphrasing evades detectors of AI -generated text, but retrieval is an effective defense , in: booktitle Proceedings of the 37th International Conference on Neural Information Processing Systems , publisher Curran Associates Inc. , ad...

    2023

  22. [22]

    , author Zubiaga, A

    author Lau, H.T. , author Zubiaga, A. ( year 2025 ). title Understanding the effects of human-written paraphrases in LLM -generated text detection . journal Natural Language Processing Journal volume 11 , pages 100151 . :https://doi.org/10.1016/j.nlp.2025.100151

    arXiv 2025

  23. [23]

    , author Li, Q

    author Li, Y. , author Li, Q. , author Cui, L. , author Bi, W. , author Wang, Z. , author Wang, L. , author Yang, L. , author Shi, S. , author Zhang, Y. ( year 2024 ). title MAGE : Machine-generated Text Detection in the Wild , in: booktitle Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , p...

    doi:10.18653/v1/2024.acl-long.3 2024

  24. [24]

    , author Yuksekgonul, M

    author Liang, W. , author Yuksekgonul, M. , author Mao, Y. , author Wu, E. , author Zou, J. ( year 2023 ). title GPT detectors are biased against non-native E nglish writers . journal Patterns volume 4 , pages 100779 . :https://doi.org/10.1016/j.patter.2023.100779

    arXiv 2023

  25. [25]

    Feature-Augmented Transformers for Robust AI-Text Detection Across Domains and Generators

    author Mady, M. , author Reschke, J. , author Schuller, B. ( year 2026 ). title Feature-Augmented Transformers for Robust AI -Text Detection Across Domains and Generators . howpublished [Preprint]. arXiv . :10.48550/arXiv.2605.03969

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.03969 2026

  26. [26]

    , author Vondrick, C

    author Mao, C. , author Vondrick, C. , author Wang, H. , author Yang, J. ( year 2024 ). title RAIDAR : generative AI detection via rewriting . howpublished [Preprint]. arXiv . :10.48550/arXiv.2401.12970

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.12970 2024

  27. [27]

    , author Emi, B.N

    author Masrour, E. , author Emi, B.N. , author Spero, M. ( year 2025 ). title DAMAGE : Detecting Adversarially Modified AI Generated Text , in: editor Alam, F. , editor Nakov, P. , editor Habash, N. , editor Gurevych, I. , editor Chowdhury, S. , editor Shelmanov, A. , editor Wang, Y. , editor Artemova, E. , editor Kutlu, M. , editor Mikros, G. (Eds.), boo...

    2025

  28. [28]

    , author Fan, S

    author Meng, W. , author Fan, S. , author Wei, C. , author Chen, M. , author Li, Y. , author Zhang, Y. , author Zhang, Z. , author Chen, W. ( year 2025 ). title GradEscape : A Gradient-Based Evader Against AI-Generated Text Detectors , in: booktitle 34th USENIX Security Symposium (USENIX Security 25) , publisher USENIX Association , address Seattle, WA . ...

    2025

  29. [29]

    , author Hardin, C

    author Mesnard, T. , author Hardin, C. , author Dadashi, R. , author Bhupatiraju, S. , author Pathak, S. , author Sifre, L. , author Rivière, M. , author Kale, M.S. , author Love, J. , author Tafti, P. , author Hussenot, L. , author Sessa, P.G. , author Chowdhery, A. , author Roberts, A. , author Barua, A. , author Botev, A. , author Castro-Ros, A. , auth...

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.08295 2024

  30. [30]

    , author Lee, Y

    author Mitchell, E. , author Lee, Y. , author Khazatsky, A. , author Manning, C.D. , author Finn, C. ( year 2023 ). title DetectGPT : Zero-Shot Machine-Generated Text Detection Using Probability Curvature , in: booktitle Proceedings of the 40th International Conference on Machine Learning , publisher PMLR . pp. pages 24950--24962 . https://proceedings.mlr...

    2023

  31. [31]

    , author Cohen, S.B

    author Narayan, S. , author Cohen, S.B. , author Lapata, M. ( year 2018 ). title Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization , in: editor Riloff, E. , editor Chiang, D. , editor Hockenmaier, J. , editor Tsujii, J. (Eds.), booktitle Proceedings of the 2018 Conference on Empirical Methods ...

    2018

  32. [32]

    Benchmarking of LLM Detection: Comparing Two Competing Approaches

    author Pr \"o hl, T. , author Putzier, E. , author Zarnekow, R. ( year 2024 ). title Benchmarking of LLM Detection: Comparing Two Competing Approaches . howpublished [Preprint]. arXiv . :10.48550/arXiv.2406.11670

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.11670 2024

  33. [33]

    , author Sarwar, Z

    author Pu, J. , author Sarwar, Z. , author Abdullah, S.M. , author Rehman, A. , author Kim, Y. , author Bhattacharya, P. , author Javed, M. , author Viswanath, B. ( year 2023 ). title Deepfake Text Detection: Limitations and Opportunities , in: booktitle 2023 IEEE Symposium on Security and Privacy (SP) , publisher IEEE Computer Society , address Los Alami...

    arXiv 2023

  34. [34]

    , author Yang, B

    author Qwen Team , author Yang, A. , author Yang, B. , author Zhang, B. , author Hui, B. , author Zheng, B. , author Yu, B. , author Liu, D. , author Huang, F. , author Wei, H. , author Lin, H. , author Zhou, J. ( year 2024 ). title Qwen2.5 Technical Report . howpublished [Preprint]. arXiv . :10.48550/arXiv.2412.15115

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2024

  35. [35]

    , author Ramesh, A

    author Ranganath, S. , author Ramesh, A. ( year 2026 ). title StealthRL : Reinforcement Learning Paraphrase Attacks for Multi-Detector Evasion of AI -Text Detectors . howpublished [Preprint]. arXiv . :10.48550/arXiv.2602.08934

    doi:10.48550/arxiv.2602.08934 2026

  36. [36]

    , author Kumar, A

    author Sadasivan, V.S. , author Kumar, A. , author Balasubramanian, S. , author Wang, W. , author Feizi, S. ( year 2025 ). title Can AI -Generated Text be Reliably Detected? Stress Testing AI Text Detectors Under Various Attacks . howpublished [Preprint]. arXiv . :10.48550/arXiv.2303.11156

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.11156 2025

  37. [37]

    , author Singh, A.K

    author Sahoo, P. , author Singh, A.K. , author Saha, S. , author Jain, V. , author Mondal, S. , author Chadha, A. ( year 2024 ). title A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications . howpublished [Preprint]. arXiv . :10.48550/arXiv.2402.07927

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.07927 2024

  38. [38]

    , author Schlippe, T

    author Schaaff, K. , author Schlippe, T. , author Mindner, L. ( year 2024 ). title Classification of human- and AI-generated texts for different languages and domains . journal International Journal of Speech Technology volume 27 , pages 935--956 . :10.1007/s10772-024-10143-3

    work page doi:10.1007/s10772-024-10143-3 2024

  39. [39]

    , author Ilie, M

    author Schulhoff, S. , author Ilie, M. , author Balepur, N. , author Kahadze, K. , author Liu, A. , author Si, C. , author Li, Y. , author Gupta, A. , author Han, H. , author Schulhoff, S. , author Dulepet, P.S. , author Vidyadhara, S. , author Ki, D. , author Agrawal, S. , author Pham, C. , author Kroiz, G. , author Li, F. , author Tao, H. , author Sriva...

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.06608 2024

  40. [40]

    , author Wang, Y

    author Shi, Z. , author Wang, Y. , author Yin, F. , author Chen, X. , author Chang, K.W. , author Hsieh, C.J. ( year 2024 ). title Red Teaming Language Model Detectors with Language Models . journal Transactions of the Association for Computational Linguistics volume 12 , pages 174--189 . :10.1162/tacl_a_00639

    doi:10.1162/tacl_a_00639 2024

  41. [41]

    , author Patil, K

    author Stowe, K. , author Patil, K. ( year 2026 ). title Spotlights and Blindspots: Evaluating Machine-Generated Text Detection . howpublished [Preprint]. arXiv . :10.48550/arXiv.2604.16607

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.16607 2026

  42. [42]

    , author Chen, Y

    author Tao, Z. , author Chen, Y. , author Xi, D. , author Li, Z. , author Xu, W. ( year 2026 ). title Toward Reliable Detection of LLM -Generated Texts: A Comprehensive Evaluation Framework with CUDRT . journal ACM Trans. Intell. Syst. Technol. volume 17 . :10.1145/3779427

    work page doi:10.1145/3779427 2026

  43. [43]

    , author Chen, Y

    author Wang, T. , author Chen, Y. , author Liu, Z. , author Chen, Z. , author Chen, H. , author Zhang, X. , author Cheng, W. ( year 2025 ). title Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors . howpublished [Preprint]. arXiv . :10.48550/arXiv.2410.19230

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.19230 2025

  44. [44]

    , author Mansurov, J

    author Wang, Y. , author Mansurov, J. , author Ivanov, P. , author Su, J. , author Shelmanov, A. , author Tsvigun, A. , author Mohammed Afzal, O. , author Mahmoud, T. , author Puccetti, G. , author Arnold, T. , author Aji, A. , author Habash, N. , author Gurevych, I. , author Nakov, P. ( year 2024 a). title M 4 GT -Bench: Evaluation Benchmark for Black-Bo...

    2024

  45. [45]

    , author Mansurov, J

    author Wang, Y. , author Mansurov, J. , author Ivanov, P. , author Su, J. , author Shelmanov, A. , author Tsvigun, A. , author Whitehouse, C. , author Afzal, O.M. , author Mahmoud, T. , author Sasaki, T. , author Arnold, T. , author Aji, A.F. , author Habash, N. , author Gurevych, I. , author Nakov, P. ( year 2024 b). title M4 : Multi-generator, Multi-dom...

    2024

  46. [46]

    , author Anohina-Naumeca, A

    author Weber-Wulff, D. , author Anohina-Naumeca, A. , author Bjelobaba, S. , author Folt \'y nek, T. , author Guerrero-Dib, J. , author Popoola, O. , author S igut, P. , author Waddington, L. ( year 2023 ). title Testing of Detection Tools for AI -Generated Text . journal International Journal for Educational Integrity volume 19 . :10.1007/s40979-023-00146-z

    doi:10.1007/s40979-023-00146-z 2023

  47. [47]

    DUPE: Detection Undermining via Prompt Engineering for Deepfake Text

    author Weichert, J. , author Dimobi, C. ( year 2024 ). title DUPE : Detection Undermining via Prompt Engineering for Deepfake Text . howpublished [Preprint]. arXiv . :10.48550/arXiv.2404.11408

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.11408 2024

  48. [48]

    , author Zhan, R

    author Wu, J. , author Zhan, R. , author Wong, D. , author Yang, S. , author Yang, X. , author Yuan, Y. , author Chao, L.S. ( year 2024 ). title DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios , in: editor Globerson, A. , editor Mackey, L. , editor Belgrave, D. , editor Fan, A. , editor Paquet, U. , editor Tomczak, J. , editor ...

    2024

  49. [49]

    , author Zhan, R

    author Wu, J. , author Zhan, R. , author Wong, D.F. , author Yang, S. , author Liu, X. , author Chao, L.S. , author Zhang, M. ( year 2025 ). title Who Wrote This? The Key to Zero-Shot LLM -Generated Text Detection Is GECS core , in: editor Rambow, O. , editor Wanner, L. , editor Apidianaki, M. , editor Al-Khalifa, H. , editor Eugenio, B.D. , editor Schock...

    2025

  50. [50]

    , author Li, N

    author Xiang, L. , author Li, N. , author Liu, Y. , author Hu, J. ( year 2026 ). title AI -Generated Text Detection: A Comprehensive Review of Active and Passive Approaches . journal Computers, Materials and Continua volume 86 . :https://doi.org/10.32604/cmc.2025.073347

    arXiv 2026

  51. [51]

    ( year 2019 )

    author Yadan, O. ( year 2019 ). title Hydra - A framework for elegantly configuring complex applications . howpublished Github . https://github.com/facebookresearch/hydra

    2019

  52. [52]

    , author Jin, H

    author Yang, J. , author Jin, H. , author Tang, R. , author Han, X. , author Feng, Q. , author Jiang, H. , author Zhong, S. , author Yin, B. , author Hu, X. ( year 2024 ). title Harnessing the Power of LLM s in Practice: A Survey on ChatGPT and Beyond . journal ACM Trans. Knowl. Discov. Data volume 18 . :10.1145/3649506

    doi:10.1145/3649506 2024

  53. [53]

    , author Min, R

    author Zha, Y. , author Min, R. , author Sushmita, S. ( year 2025 ). title PADBen : A Comprehensive Benchmark for Evaluating AI Text Detectors Against Paraphrase Attacks . howpublished [Preprint]. arXiv . :10.48550/arXiv.2511.00416

    doi:10.48550/arxiv.2511.00416 2025

  54. [54]

    , author Keuper, M

    author Zhou, Y. , author Keuper, M. , author Fritz, M. ( year 2025 ). title Balancing Diversity and Risk in LLM Sampling: How to Select Your Method and Parameter for Open-Ended Text Generation , in: editor Che, W. , editor Nabende, J. , editor Shutova, E. , editor Pilehvar, M.T. (Eds.), booktitle Proceedings of the 63rd Annual Meeting of the Association f...

    2025

  55. [55]

    , author Muresanu, A.I

    author Zhou, Y. , author Muresanu, A.I. , author Han, Z. , author Paster, K. , author Pitis, S. , author Chan, H. , author Ba, J. ( year 2022 ). title Large Language Models Are Human-Level Prompt Engineers , in: booktitle NeurIPS 2022 Workshop on Foundation Models for Decision Making . note [Oral presentation]

    2022

  56. [56]

    , author Wayne, L

    author Zhuang, L. , author Wayne, L. , author Ya, S. , author Jun, Z. ( year 2021 ). title A Robustly Optimized BERT Pre-training Approach with Post-training , in: editor Li, S. , editor Sun, M. , editor Liu, Y. , editor Wu, H. , editor Liu, K. , editor Che, W. , editor He, S. , editor Rao, G. (Eds.), booktitle Proceedings of the 20th Chinese National Con...

    2021