Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Attention maps from a proxy model reveal local patterns that separate human from AI-written text.

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 · grok-4.5

2026-07-12 21:38 UTC pith:TBQ5GCTT

load-bearing objection Solid empirical methods paper: proxy attention maps as CNN inputs give a real complementary signal, strongest under generator match, with honest motif analysis and clear soft spots on transfer. the 4 major comments →

arxiv 2606.00016 v1 pith:TBQ5GCTT submitted 2026-04-13 cs.CL cs.AI

AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection

Aria Nourbakhsh , Adelaide Danilov , Christoph Schommer , Salima Lamsiyah This is my paper
classification cs.CL cs.AI
keywords AI-generated text detectionattention attributionproxy TransformerCNN classifierauthorship motifscross-generator transferrobustness
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.

Modern language models write so fluently that detectors based on word frequencies, perplexity, or likelihood often fail. This paper argues that a better clue is how a fixed proxy Transformer pays attention while reading a passage: the resulting two-dimensional attention maps carry systematic differences between human and machine text. The method, AEyeDE, extracts those maps, treats them as images, and trains a small convolutional network to classify authorship. On machine-translation tasks it beats a text-only baseline; on open-ended generation it reaches near-ceiling accuracy when the proxy matches the generator family, stays competitive when generators are mixed or held out, and remains strong under character-level spelling changes. Clustering small patches of the maps further uncovers recurring local motifs whose frequencies shift consistently between human and AI text. The practical claim is that internal attention structure supplies a complementary, partially interpretable detection signal that does not require watermarking or access to the true generator.

Core claim

Attention-based attribution maps extracted from a single proxy Transformer contain discriminative structure that a lightweight CNN can use to tell human-written from AI-generated text. The maps alone already outperform a text-only baseline on translation detection, deliver near-ceiling generator-matched performance, remain competitive under cross-generator and cross-dataset transfer, and exhibit recurring local motifs whose relative frequencies differ systematically by authorship.

What carries the argument

The attention-derived attribution map: queries and keys from every layer and head of a fixed proxy Transformer are averaged into a single target-by-source matrix; diagonal blocks of this matrix are fed to a multi-scale CNN with attention pooling that produces an authorship score.

Load-bearing premise

A fixed proxy Transformer, even from a different family than the true generator, still produces attention maps that differ systematically enough between human and AI text for a CNN to exploit without seeing the original prompt.

What would settle it

Train and evaluate AEyeDE on a large held-out generator family never seen during proxy selection or training; if accuracy and TPR at low false-positive rate collapse to chance while text-only or likelihood baselines remain above chance, the claim that proxy attention maps transfer fails.

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

Share X Bluesky LinkedIn Reddit HN

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

4 major / 6 minor

Summary. The paper proposes AEyeDE, a supervised detector that extracts averaged multi-head multi-layer attention attribution maps from a fixed white-box proxy Transformer and classifies them with a lightweight multi-scale CNN (optionally fused with a text branch). The central claim is that these 2D attribution maps encode a complementary, transferable human-vs-AI signal beyond surface statistics or likelihood scores. Empirically, the method outperforms a text-only baseline on Marian-MT translation pairs (Table 1), reaches near-ceiling generator-matched performance on RAID (Table 2, F1 96.34–99.25) and strong results on HC3 (Table 4), remains competitive under mixed-generator training with a held-out generator (Table 3) and on external Beemo transfer (Table 9), and is robust to alternative-spelling attacks (Tables 7–8). Section 5 further clusters 8×8 patches of the CNN feature maps into recurring “motifs” whose mean rates differ systematically between human and AI text (Table 10).

Significance. If the proxy-attention signal is genuinely complementary and at least partially generator-agnostic, the work opens a useful middle ground between pure black-box text classifiers and watermarking or true-generator likelihood methods. Strengths include a clean architectural separation of attribution extraction from detection, broad evaluation across encoder–decoder and decoder-only regimes, explicit low-FPR operating points, public-code commitment, and an interpretable motif analysis that goes beyond accuracy tables. The individual-generator results and alternative-spelling robustness are particularly clear empirical contributions. The significance is tempered by the large performance gap between matched-proxy and cross-generator/cross-dataset settings, so the paper’s main lasting value is likely as a strong generator-specific detector and as evidence that attention geometry carries authorship-related structure, rather than as a fully transferable black-box solution.

major comments (4)
  1. Tables 2–3 and the abstract’s transfer claim: generator-matched (individual) F1 is 96–99, but the unified held-out-generator setting drops to F1 ≈74.7–76.2 with high precision and substantially lower recall. The abstract and §4.4 still describe the method as remaining competitive under mixed-generator generalization. This gap is load-bearing for the Sec. 3 hypothesis that proxy attention induces a transferable authorship signal rather than a family-match artifact; either the claim should be narrowed to generator-specific / same-family detection, or additional out-of-family proxy ablations (proxy family ≠ generator family, matched training budget) should be reported with the same metrics.
  2. Tables 5–6 (paraphrasing) vs. Tables 7–8 (alternative spelling): the paper claims robustness under adversarial perturbations, yet under paraphrasing the individual advantage shrinks and the unified setting collapses toward near-all-positive behavior (AEyeDE AUC 62.21, below RoBERTa and Curvature). Because paraphrasing is the more realistic evasion threat, the robustness narrative should be stratified: strong for character-level spelling, weak for paraphrase. Presenting both under a single “adversarial” umbrella overstates the result that supports the central claim.
  3. Sec. 5, Table 10 and the Grad-CAM/ablation analysis: motif rate gaps Δr(c) are statistically reliable, but the paper reports near-zero / unstable correlations between prevalence and both Grad-CAM and zeroing-ablation importance (Pearson/Spearman near 0 across top clusters). The interpretability claim that “recurring local structures” are the signature the detector exploits is therefore not supported by the saliency evidence the authors themselves compute. Either the motif section should be reframed as descriptive structure analysis only, or the authors should identify which motifs (or combinations) actually drive decisions under controlled ablations.
  4. Sec. 3 hypothesis and Limitations: the detector discards prompts and relies on a proxy that need not be the true generator, yet almost all strong decoder-only numbers use same-family proxies (Llama/Cohere/GPT-neo/Mistral for RAID families; GPT-neo for HC3). No systematic table quantifies performance when the proxy is deliberately out-of-family for every generator. Without that measurement, the claim that white-box access to “a” proxy (not the generator) is sufficient remains under-supported relative to the abstract’s positioning against likelihood-based and supervised text detectors.
minor comments (6)
  1. Figure 1 caption and §3: the diagonal-block sampling (wa=128) is well motivated by entropy statistics in footnote 1, but the number of blocks per sequence and the exact stride/overlap policy are not stated; a short equation or algorithm box would aid reproducibility.
  2. Table 4 / HC3: the comparison with the SuperAnnotate RoBERTa detector should note more prominently that HC3 was part of that model’s training data, as the authors briefly acknowledge; otherwise the near-parity reading is misleading.
  3. §4.1: class-balanced entropy loss is mentioned but not formalized; give the loss equation or the per-class weight formula.
  4. Throughout: “attribution” is used for averaged attention; a one-sentence clarification that this is not gradient×input or Integrated Gradients would avoid confusion with the XAI literature cited later (Grad-CAM).
  5. Typos / polish: “Inseqlibrary” spacing (p. 5); “arrangment” in Table 2/3 captions; “hidding” in Peng et al. reference; inconsistent hyphenation of encoder–decoder vs encoder-decoder.
  6. Beemo experiment (Table 9): state the training set size and which RAID/HC3 models were used to train the transferred detector so the “relatively small training set” claim can be checked.

Circularity Check

0 steps flagged

No circularity: standard supervised empirical detection on held-out labels with post-hoc descriptive motif analysis; no derivation reduces the target to its inputs by construction.

full rationale

AEyeDE is an empirical supervised classification pipeline, not a first-principles derivation. The load-bearing steps are: (1) extract averaged multi-head multi-layer attention maps A(x;θ) from a fixed proxy Transformer G_θ (Sec. 3 equations for ˜A^(ℓ,h) and A); (2) feed the maps (or diagonal blocks) as single-channel images into a CNN encoder E_attr whose parameters ϕ are trained by binary cross-entropy on gold human/AI labels; (3) evaluate accuracy/F1/AUC/TPR@FPR=0.01 on held-out splits, adversarial variants, and an external dataset (Tables 1–9); (4) post-hoc cluster 8×8 patches from the trained CNN’s final feature map and report class-conditional motif rates ¯r_g(c) and Δr(c) (Sec. 5, Table 10). The gold labels y are external to the attention extraction and to the clustering; the motif rates are purely descriptive statistics of a trained feature space and are never fed back as predictions or uniqueness claims. No parameter is fitted to a quantity and then re-presented as a prediction of a closely related quantity; no uniqueness theorem or ansatz is imported via self-citation to force the result; no equation equates the detection target to an input by definition. External baselines (Curvature, Binoculars, GLTR, RoBERTa, etc.) and cross-dataset transfer further keep the evaluation independent. The minor self-referential aspect that the same CNN supplies the embedding space for motif clustering does not constitute circularity under the stated criteria, because class membership remains an independent gold label and the analysis is not used to derive or force the primary detection claim. The paper is therefore self-contained against its own benchmarks with score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard Transformer attention math, public datasets, and a small set of design choices (block size, CNN depth, averaging over all heads/layers, white-box proxy). No new physical entities are postulated; 'motifs' are empirical clusters. Free parameters are architectural and threshold choices that affect both detection and the motif analysis.

free parameters (4)
  • attention block size wa
    Fixed to 128 for all decoder-only experiments; determines which local token interactions enter the CNN and the motif grid.
  • CNN channel progression and pooling schedule
    1→32→64→128→128→256 with 2×2 max-pools yielding a 16×16 final map; architecture is chosen by hand and defines both the detector and the motif embedding space.
  • motif informativeness thresholds τμ=0.01, τσ=0.001
    Hand-set filters that decide which 8×8 patches enter HDBSCAN clustering and thus which motifs are reported as discriminative.
  • layer/head averaging weights (uniform 1/LH)
    All layers and heads are averaged equally; authors note finer selection is future work, so this uniform mix is an unoptimized free design choice.
axioms (4)
  • domain assumption Averaged multi-head multi-layer attention of a fixed proxy Transformer encodes authorship-discriminative structure for both human and AI text.
    Stated as the core hypothesis in Sec. 3; not derived, only tested empirically.
  • domain assumption White-box access to some proxy Transformer (not necessarily the true generator) is available at detection time.
    Required by the framework definition and listed as a limitation; without it the method cannot run.
  • domain assumption In decoder-only models the most informative attribution mass lies near the main diagonal, justifying 128×128 diagonal blocks.
    Motivated by cited attention-sink / local-interaction work and a brief entropy comparison on HC3; treated as design premise.
  • standard math Standard softmax attention and binary cross-entropy training are valid for this classification task.
    Background ML assumptions used throughout Sec. 3.
invented entities (1)
  • Attention motifs (HDBSCAN clusters of 8×8 attribution patches) no independent evidence
    purpose: Provide an interpretable localized signature of human vs AI authorship in proxy attention maps.
    Defined post-hoc from the trained CNN feature space; not an a priori physical entity, but a paper-introduced construct whose independent existence outside this pipeline is not established.

pith-pipeline@v1.1.0-grok45 · 33093 in / 3026 out tokens · 33578 ms · 2026-07-12T21:38:26.391243+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection." pith.science (2026). https://pith.science/paper/TBQ5GCTT

@misc{pith2026260600016,
  author       = {Pith},
  title        = {Pith review of: AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TBQ5GCTT}},
  note         = {Machine review of arXiv:2606.00016}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Detecting AI-generated text is becoming increasingly challenging as modern language models approach human-level fluency and can evade detectors that rely on surface statistics or likelihood-based signals. We propose \textsc{AEyeDE}, an attribution-driven approach to human-AI authorship detection that leverages model attention as a discriminative signal. Specifically, we extract attention-based attribution matrices for both human- and AI-generated text using a \emph{proxy} Transformer model with white-box access and train a lightweight Convolutional Neural Network to learn representations from these attribution maps. Across encoder-decoder translation settings, our method consistently outperforms a text-only baseline. In decoder-only settings, it performs strongly in generator-specific detection, remains competitive on standard benchmarks, and shows robustness under cross-dataset transfer and alternative-spelling perturbations. We further show that attention maps exhibit recurring local structures whose relative frequencies differ consistently between human- and AI-generated text across datasets and proxy models. These findings suggest that attention-based attribution maps provide a complementary and interpretable signal for AI-generated text detection. We will make the code publicly available to support future research.

Figures

Figures reproduced from arXiv: 2606.00016 by Adelaide Danilov, Aria Nourbakhsh, Christoph Schommer, Salima Lamsiyah.

Figure 1
Figure 1. Figure 1: Overview of AEyeDE. Given a text sample and white-box access to a proxy generator model Gθ, we extract an attention-derived attribution matrix A (top left). For decoder-only models, we summarize A by sampling fixed-size square blocks (e.g., 128 × 128) along the main diagonal, where the strongest local token-token interactions are concentrated (orange boxes) (Xiao et al., 2023; Ivanitskiy et al.; Qi et al.,… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of the top motif cluster (by absolute mean prevalence gap [PITH_FULL_IMAGE:figures/full_fig_p014_2.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

46 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Ditto: A spoofing attack framework on watermarked llms via knowledge distillation.arXiv preprint arXiv:2510.10987,

    Hyeseon Ahn, Shinwoo Park, Suyeon Woo, and Yo-Sub Han. Ditto: A spoofing attack framework on watermarked llms via knowledge distillation.arXiv preprint arXiv:2510.10987,

    arXiv

  2. [2]

    URLhttps://aclanthology.org/2025.genaidetect-1.0/

    International Conference on Computational Linguistics. URLhttps://aclanthology.org/2025.genaidetect-1.0/. Muhammad Zain Ali, Yuxia Wang, Bernhard Pfahringer, and Tony C Smith. Detection of human and machine-authored fake news in Urdu. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Moham- mad Taher Pilehvar (eds.),Proceedings of the 63rd Annual Mee...

    2025

  3. [3]

    ISBN 979-8-89176-251-0

    Association for Compu- tational Linguistics. ISBN 979-8-89176-251-0. URLhttps://aclanthology.org/2025.acl-long.170/. EkaterinaArtemova, JasonSLucas, SaranyaVenkatraman, JooyoungLee, SergeiTilga, AdakuUchendu, and Vladislav Mikhailov. Beemo: Benchmark of expert-edited machine-generated outputs. In Luis Chiruzzo, Alan Ritter, and Lu Wang (eds.),Proceedings ...

    2025

  4. [4]

    Beyond Easy Wins: A Text Hardness-Aware Benchmark for LLM-generated Text Detection

    Association for Computational Linguistics. ISBN979-8-89176-189-6. doi: 10.18653/v1/2025.naacl-long.357. URLhttps://aclanthology.org/2025. naacl-long.357/. Navid Ayoobi, Sadat Shahriar, and Arjun Mukherjee. Beyond easy wins: A text hardness-aware benchmark for llm-generated text detection.arXiv preprint arXiv:2507.15286,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2025.naacl-long.357 2025

  5. [5]

    Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature.arXiv preprint arXiv:2310.05130,

    16 Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature.arXiv preprint arXiv:2310.05130,

    Pith/arXiv arXiv

  6. [6]

    URLhttps://doi.org/10.5281/zenodo. 5297715. If you use this software, please cite it using these metadata. Ondřej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, et al. Findings of the 2014 workshop on statistical machine translation. InProceedings of th...

    doi:10.5281/zenodo 2014

  7. [7]

    Joseph Cornelius, Oscar Lithgow-Serrano, Sandra Mitrović, Ljiljana Dolamic, and Fabio Rinaldi

    URL https://arxiv.org/abs/2504.00698. Joseph Cornelius, Oscar Lithgow-Serrano, Sandra Mitrović, Ljiljana Dolamic, and Fabio Rinaldi. Bust: Benchmark for the evaluation of detectors of llm-generated text. InProceedings of the 2024 Confer- ence of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volum...

    Pith/arXiv arXiv 2024

  8. [8]

    Robust data watermarking in language models by injecting fictitious knowledge

    Xinyue Cui, Johnny Wei, Swabha Swayamdipta, and Robin Jia. Robust data watermarking in language models by injecting fictitious knowledge. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 14292–14306,

    2025

  9. [9]

    BERT: Pre-training of deep bidirec- tionaltransformersforlanguageunderstanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirec- tionaltransformersforlanguageunderstanding. InJillBurstein, ChristyDoran, andThamarSolorio(eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Lon...

    2019

  10. [10]

    doi: 10.18653/v1/N19-1423

    Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URLhttps://aclanthology.org/N19-1423/. Liam Dugan, Daphne Ippolito, Arun Kirubarajan, Sherry Shi, and Chris Callison-Burch. Real or fake text?: Investigating human ability to detect boundaries between human-written and machine-generated text. In Proceedings of the AAAI Conference on Art...

    doi:10.18653/v1/n19-1423

  11. [12]

    ISBN 979-8-89176-256-5

    Association for Computational Lin- guistics. ISBN 979-8-89176-256-5. URLhttps://aclanthology.org/2025.findings-acl.98/. Sebastian Gehrmann, Hendrik Strobelt, and Alexander Rush. GLTR: Statistical detection and visualization of generated text. In Marta R. Costa-jussà and Enrique Alfonseca (eds.),Proceedings of the 57th Annual Meeting of the Association for...

    2025

  12. [13]

    doi: 10.18653/v1/P19-3019

    Association for Computational Linguistics. doi: 10.18653/v1/P19-3019. URLhttps: //aclanthology.org/P19-3019/. Wooyoung Go, Hyoungshick Kim, Alice Oh, and Yongdae Kim. XDAC: XAI-driven detection and attribution of LLM-generated news comments in Korean. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.),Proceedings of the ...

    doi:10.18653/v1/p19-3019 2025

  13. [14]

    Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu

    URLhttps://arxiv.org/abs/2407.21783. Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. How close is chatgpt to human experts? comparison corpus, evaluation, and detection.arXiv preprint arXiv:2301.07597,

    Pith/arXiv arXiv

  14. [15]

    Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang

    URLhttps://arxiv.org/abs/2401.12070. Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. Mgtbench: Benchmarking machine- generated text detection. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pp. 2251–2265,

    Pith/arXiv arXiv 2024

  15. [16]

    Low-entropy watermark detection via bayes’ rule derived detector

    Beining Huang, Du Su, Fei Sun, Qi Cao, Huawei Shen, and Xueqi Cheng. Low-entropy watermark detection via bayes’ rule derived detector. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 14330–14344, 2025a. He Huang, Nan Sun, Massimiliano Tani, Yu Zhang, Jiaojiao Jiang, and Sanjay Jha. Can llm-generated misinformation be detected: A...

    work page doi:10.1016/j.future.2025.107877 2025

  16. [17]

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

    Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.164. URLhttps://aclanthology.org/2020.acl-main.164/. Michael Ivanitskiy, Cecilia Diniz Behn, and Samy Wu Fung. Motifs in attention patterns of large language models. InMechanistic Interpretability Workshop at NeurIPS

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

  17. [18]

    Kaijie Jiao, Quan Wang, Licheng Zhang, Zikang Guo, and Zhendong Mao

    URLhttps://arxiv.org/abs/2310.06825. Kaijie Jiao, Quan Wang, Licheng Zhang, Zikang Guo, and Zhendong Mao. M-RangeDetector: Enhancing generalizationinmachine-generatedtextdetectionthroughmulti-rangeattentionmasks. InWanxiangChe, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.),Findings of the Association for Computational Linguistics: A...

    Pith/arXiv arXiv 2025

  18. [19]

    ISBN 979-8-89176-256-5

    Association for Compu- tational Linguistics. ISBN 979-8-89176-256-5. URLhttps://aclanthology.org/2025.findings-acl. 469/. John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. InInternational Conference on Machine Learning, pp. 17061–17084. PMLR,

    2025

  19. [20]

    RankGen: Improving text generation with large ranking models

    Kalpesh Krishna, Yapei Chang, John Wieting, and Mohit Iyyer. RankGen: Improving text generation with large ranking models. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (eds.),Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 199–232, Abu Dhabi, United Arab Emirates, December

    2022

  20. [21]

    doi: 10.18653/v1/2022.emnlp-main

    Association for Computational Linguistics. doi: 10.18653/v1/2022.emnlp-main

    doi:10.18653/v1/2022.emnlp-main 2022

  21. [22]

    Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer

    URLhttps://aclanthology.org/2022.emnlp-main.15/. Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.),Advances in Neural Information Processing Systems, volume 3...

    2022

  22. [23]

    Kristian Kuznetsov, Laida Kushnareva, Anton Razzhigaev, Polina Druzhinina, Anastasia Voznyuk, Irina Pi- ontkovskaya, Evgeny Burnaev, and Serguei Barannikov

    URLhttps://proceedings.neurips.cc/paper_files/ paper/2023/file/575c450013d0e99e4b0ecf82bd1afaa4-Paper-Conference.pdf. Kristian Kuznetsov, Laida Kushnareva, Anton Razzhigaev, Polina Druzhinina, Anastasia Voznyuk, Irina Pi- ontkovskaya, Evgeny Burnaev, and Serguei Barannikov. Feature-level insights into artificial text detection with sparse autoencoders. In...

    2023

  23. [24]

    ISBN 979-8-89176-256-5

    Association for Computational Linguistics. ISBN 979-8-89176-256-5. URL https://aclanthology.org/2025.findings-acl.1321/. Salima Lamsiyah, Saad Ezzini, Abdelkader El Mahdaouy, Hamza Alami, Abdessamad Benlahbib, Samir El Amrany, Salmane Chafik, and Hicham Hammouchi. M-daigt: A shared task on multi-domain detection of ai-generated text.arXiv preprint arXiv:2...

    arXiv 2025

  24. [25]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. URLhttps://aclanthology.org/2025.acl-long.1292/. Yuanfan Li, Zhaohan Zhang, Chengzhengxu Li, Chao Shen, and Xiaoming Liu. Iron sharpens iron: De- fending against attacks in machine-generated text detection with adversarial training. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mo...

    2025

  25. [26]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. URL https://aclanthology.org/2025.acl-long.155/. 21 Aiwei Liu, Leyi Pan, Yijian Lu, Jingjing Li, Xuming Hu, Xi Zhang, Lijie Wen, Irwin King, Hui Xiong, and Philip Yu. A survey of text watermarking in the era of large language models.ACM Computing Surveys, 57(2):1–36, 2024a. Yinhan Liu, Myl...

    Pith/arXiv arXiv 2025

  26. [27]

    On the detectability of chatgpt content: Benchmarking, methodology, and evaluation through the lens of academic writing

    Zeyan Liu, Zijun Yao, Fengjun Li, and Bo Luo. On the detectability of chatgpt content: Benchmarking, methodology, and evaluation through the lens of academic writing. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pp. 2236–2250, 2024b. Xinyang Lu, Jingtan Wang, Zitong Zhao, Zhongxiang Dai, Chuan-Sheng Foo, See ...

    2024

  27. [28]

    doi: 10.18653/v1/2024

    Association for Computational Linguistics. doi: 10.18653/v1/2024. acl-long.630. URLhttps://aclanthology.org/2024.acl-long.630/. Dominik Macko, Jakub Kopál, Robert Moro, and Ivan Srba. MultiSocial: Multilingual benchmark of machine-generated text detection of social-media texts. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar...

    doi:10.18653/v1/2024 2024

  28. [29]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. URLhttps://aclanthology.org/ 2025.acl-long.36/. Minjia Mao, Dongjun Wei, Zeyu Chen, Xiao Fang, and Michael Chau. Watermarking large language models: An unbiased and low-risk method. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2025

  29. [30]

    Andrea Pedrotti, Michele Papucci, Cristiano Ciaccio, Alessio Miaschi, Giovanni Puccetti, Felice Dell’Orletta, and Andrea Esuli

    Large language model. Andrea Pedrotti, Michele Papucci, Cristiano Ciaccio, Alessio Miaschi, Giovanni Puccetti, Felice Dell’Orletta, and Andrea Esuli. Stress-testing machine generated text detection: Shifting language models writing style to fool detectors. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.),Findings of th...

    2025

  30. [31]

    ISBN 979-8-89176-256-5

    Association for Computational Linguistics. ISBN 979-8-89176-256-5. URL https://aclanthology.org/2025.findings-acl.156/. Xinlin Peng, Ying Zhou, Ben He, Le Sun, and Yingfei Sun. Hidding the ghostwriters: An adversarial evaluation of ai-generated student essay detection. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processin...

    2025

  31. [32]

    Deltallm: A training-free framework exploiting temporal sparsity for efficient edge llm inference.arXiv preprint arXiv:2507.19608,

    22 Jiawen Qi, Chang Gao, Zhaochun Ren, and Qinyu Chen. Deltallm: A training-free framework exploiting temporal sparsity for efficient edge llm inference.arXiv preprint arXiv:2507.19608,

    Pith/arXiv arXiv

  32. [33]

    Chen, and Nicholas Andrews

    Rafael Alberto Rivera Soto, Barry Y. Chen, and Nicholas Andrews. Mitigating paraphrase attacks on machine-text detection via paraphrase inversion. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.),Findings of the Association for Computational Linguistics: ACL 2025, pp. 4421–4433, Vienna, Austria, July

    2025

  33. [34]

    ISBN 979- 8-89176-256-5

    Association for Computational Linguistics. ISBN 979- 8-89176-256-5. URLhttps://aclanthology.org/2025.findings-acl.227/. Gabriele Sarti, Nils Feldhus, Ludwig Sickert, and Oskar van der Wal. Inseq: An interpretability toolkit for sequence generation models. In Danushka Bollegala, Ruihong Huang, and Alan Ritter (eds.),Pro- ceedings of the 61st Annual Meeting...

    2025

  34. [35]

    doi: 10.18653/v1/2023.acl-demo.40

    Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-demo.40. URLhttps://aclanthology.org/2023.acl-demo.40/. Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision...

    doi:10.18653/v1/2023.acl-demo.40 2023

  35. [36]

    TextDefense: Adversarial Text Detection based on Word Importance Entropy

    ISSN 1573-1405. doi: 10.1007/s11263-019-01228-7. URLhttp://dx.doi.org/10.1007/s11263-019-01228-7. LujiaShen, XuhongZhang, ShoulingJi, YuwenPu, ChunpengGe, XingYang, andYangheFeng. Textdefense: Adversarial text detection based on word importance entropy.arXiv preprint arXiv:2302.05892,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s11263-019-01228-7

  36. [37]

    DetectLLM: Leveraging log rank information for zero-shot detection of machine-generated text

    Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov. DetectLLM: Leveraging log rank information for zero-shot detection of machine-generated text. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 12395–12412, Singapore, December

    2023

  37. [38]

    doi: 10.18653/v1/2023.findings-emnlp.827

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.827. URLhttps://aclanthology.org/2023.findings-emnlp.827/. Zhixiong Su, Yichen Wang, Herun Wan, Zhaohan Zhang, and Minnan Luo. HACo-det: A study towards fine-grained machine-generated text detection under human-AI coauthoring. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova,...

    doi:10.18653/v1/2023.findings-emnlp.827 2023

  38. [39]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. URL https://aclanthology.org/2025.acl-long.1069/. Vasiliki Tassopoulou, George Retsinas, and Petros Maragos. Enhancing handwritten text recognition with n-gram sequence decomposition and multitask learning. In2020 25th International Conference on Pattern Recognition (ICPR), pp. 10555–10560. IEEE,

    2025

  39. [40]

    URLhttps://aclanthology.org/2020.eamt-1.61/

    European Association for Machine Translation. URLhttps://aclanthology.org/2020.eamt-1.61/. Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson, and Maksim Kuprashevich. Llmtrace: A corpus for classification and fine-grained localization of ai-written text.arXiv preprint arXiv:2509.21269,

    arXiv 2020

  40. [41]

    Turingbench: A benchmark environment for turing test in the age of neural text generation.arXiv preprint arXiv:2109.13296,

    Adaku Uchendu, Zeyu Ma, Thai Le, Rui Zhang, and Dongwon Lee. Turingbench: A benchmark environment for turing test in the age of neural text generation.arXiv preprint arXiv:2109.13296,

    Pith/arXiv arXiv

  41. [42]

    TrustAI at SemEval-2024 task 8: A comprehensive analysis of multi- domain machine generated text detection techniques

    Ashok Urlana, Aditya Saibewar, Bala Mallikarjunarao Garlapati, Charaka Vinayak Kumar, Ajeet Singh, and Srinivasa Rao Chalamala. TrustAI at SemEval-2024 task 8: A comprehensive analysis of multi- domain machine generated text detection techniques. In Atul Kr. Ojha, A. Seza Doğruöz, Harish Tay- yar Madabushi, Giovanni Da San Martino, Sara Rosenthal, and Aia...

    2024

  42. [43]

    doi: 10.18653/v1/2024.semeval-1.134

    Association for Computational Linguistics. doi: 10.18653/v1/2024.semeval-1.134. URL https://aclanthology.org/2024.semeval-1.134/. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProceedings of the 31st International Conference on Neural Information P...

    work page doi:10.18653/v1/2024.semeval-1.134 2024

  43. [44]

    ISBN 9781510860964

    Curran Associates Inc. ISBN 9781510860964. Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mo- hammed Afzal, Tarek Mahmoud, Giovanni Puccetti, and Thomas Arnold. SemEval-2024 task 8: Multido- main, multimodel and multilingual machine-generated text detection. In Atul Kr. Ojha, A. Seza Doğruöz, Harish Tayyar Mada...

    work page doi:10.18653/v1/2024.semeval-1.279 2024

  44. [45]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long.240. URLhttps://aclanthology. org/2025.acl-long.240/. Junchao Wu, Runzhe Zhan, Derek Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia Chao. Detec- trl: Benchmarking llm-generated text detection in real-world scenarios.Advances in Neural Information Processin...

    work page doi:10.18653/v1/2025.acl-long.240 2025

  45. [46]

    URLhttps://aclanthology.org/2025.cl-1.8/

    doi: 10.1162/coli_a_00549. URLhttps://aclanthology.org/2025.cl-1.8/. Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks.arXiv preprint arXiv:2309.17453,

    Pith/arXiv arXiv doi:10.1162/coli_a_00549 2025

  46. [47]

    ISBN 979-8- 89176-251-0

    Association for Computational Linguistics. ISBN 979-8- 89176-251-0. URLhttps://aclanthology.org/2025.acl-long.601/. Michał Ziemski, Marcin Junczys-Dowmunt, and Bruno Pouliquen. The United Nations parallel corpus v1.0. In Nicoletta Calzolari, Khalid Choukri, Thierry Declerck, Sara Goggi, Marko Grobelnik, Bente Maegaard, Joseph Mariani, Helene Mazo, Asuncio...

    2025