Pith. sign in

REVIEW 3 major objections 6 minor 42 references

Latent Trajectory Discrimination for AI-Generated Text Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read AI text detection improves by tracking how meaning evolves, not just what the text says.

desk verdict A clearly specified trajectory-based detector with real potential, but the empirical case is not yet made: possible topic leakage in Reviews and NYT-AI, test-set hyperparameter selection, and a missing static-representation control. read the letter →

arxiv 2607.14967 v1 pith:QISNUH5U submitted 2026-07-16 cs.CL cs.AI

Gianluca Bonifazi , Christopher Buratti , Michele Marchetti , Federica Parlapiano , Giulia Quaglieri , Davide Traini , Domenico Ursino , Luca Virgili This is my paper
classification cs.CLcs.AI
keywords AI-generatedtextdetectionlatenttrajectorycontrastivelearningtemporallyalignedsimilarityautoregressivegenerationgeometryslidingwindowembeddings
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 the signal distinguishing AI-written from human-written text lies in the dynamics of how meaning evolves through a document, not in static word or sentence statistics. To test this, it proposes GTCL, which cuts a document into overlapping windows, embeds each window, and compares the sequence of embedding differences between consecutive windows. A temporally aligned similarity measure plus supervised contrastive learning trains the model to recognize generation-specific geometric patterns in these trajectories. On three benchmarks it reports accuracy and F1 scores of 0.98/0.98 on two datasets and 0.83/0.82 on an adversarial one, beating all baselines tested. The wider claim, if correct, is that modeling the 'how' of text generation—the evolving path through latent space—is a robust, model-agnostic detection strategy.

What carries the argument

The mechanism is the latent trajectory of a document, defined as the ordered sequence of embeddings of overlapping word windows, reduced to the sequence of differences between consecutive window embeddings (first-order dynamics). The similarity measure is temporally aligned: when comparing two trajectories, a local transition at position i is matched only with transitions at positions j ≤ i, weighted by an exponential decay e^{-γ(i-j)}; this enforces a forward direction consistent with autoregressive generation. This similarity feeds a supervised contrastive loss (SupCon) with hard positives and hard negatives selected from the training set, so the projection space is organized according to

What would settle it

A direct test would be to run GTCL after deduplicating the Reviews and NYT-AI datasets at the source-article level (ensuring no paper or NYT article appears in both partitions) and check whether accuracy and F1 remain near 0.98; if performance collapses, the reported gains come from content matching, not trajectory dynamics.

Watch

Extended reading notes

Core claim

The central claim is that AI-generated text can be reliably detected by modeling a document as a discrete latent trajectory and comparing first-order differences of overlapping window embeddings, rather than by compressing the whole document into one static vector. The paper introduces Geometric Trajectory and Contrastive Learning (GTCL): it segments text into ordered, fixed-length windows, encodes each window with a frozen backbone encoder followed by a trainable projection module, forms a trajectory matrix of consecutive embedding differences, and applies supervised contrastive learning using a temporally aligned similarity that only matches a transition at position i in one text with tran

Load-bearing premise

The load-bearing premise is that random 80/20 document splits prevent the same underlying article or review from appearing in both the training and test sets, so the model cannot cheat by matching content or topic rather than trajectory geometry.

Editorial extensions

If this is right

  • If GTCL generalizes as reported, detectors can work on any text without access to logits, decoding traces, or watermarking, making post-hoc detection of arbitrary LLM output feasible.
  • The approach could be extended to detect text from diffusion-based language models, as the paper itself suggests, by adapting the similarity to handle generation without a left-to-right direction.
  • The reported geometry statistics (step irregularity, curvature, directional dispersion) provide quantitative diagnostics that could be adopted by other detection systems to check whether their features actually capture dynamics.
  • The backbone-encoder ablation suggests that the trajectory signal is not tied to a specific embedding model, so the method can be instantiated with lighter or heavier encoders depending on deployment constraints.
  • The method's balanced performance on both classes (vs. baselines that collapse one class) indicates it may reduce false-positive/false-negative asymmetries that matter in real deployment.

Reading between the lines

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

  • The contrastive groups are built using the same similarity function that is being trained, so the method's success could be partly driven by the hard-example selection rather than by trajectory geometry per se; a useful test would be to compare against randomly sampled contrastive groups of equal size.
  • The RAID gains are smaller than on the other datasets, which suggests trajectory geometry may be a weaker signal under adversarial paraphrasing; one testable extension is to see whether combining trajectory differences with a static global representation recovers the gap.
  • The paper's premise that human writing is less constrained than autoregressive generation could be tested directly on other human domains (e.g., legal, medical) where writing is deliberately formulaic, to see whether the geometric signature of 'human' changes or disappears.
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

3 major / 6 minor

Summary. The paper proposes GTCL, a framework for AI-generated text detection that represents a document as an ordered sequence of overlapping word windows, encodes each window with a frozen sentence encoder followed by a trainable transformer projection, computes first-order differences between consecutive window embeddings, and applies supervised contrastive learning with a temporally aligned similarity measure. Classification is done via k-NN in the learned space. Experiments on three benchmarks (an OpenReview-based Reviews set, NYT-AI, and a RAID subset) report high accuracy/F1 (0.98/0.98 on Reviews, 0.98/0.98 on NYT-AI, 0.83/0.82 on RAID) and the paper concludes that explicitly modeling sequential dynamics provides robust discriminative signals.

Significance. If validated, the trajectory-dynamics view is a useful conceptual reframing of AIGTD and complements static detectors. The paper ships public code and data and provides geometric-statistics evidence that human and AI trajectories differ statistically. However, the significance is conditional on resolving the evaluation concerns outlined below, which include potential same-source leakage in two benchmarks, hyperparameter selection on test partitions, and the absence of a static-representation control.

major comments (3)
  1. [Section 4.1, Table 1] The Reviews and NYT-AI datasets are constructed by pairing human and AI documents that describe the same underlying content: Reviews uses OpenReview papers with AI reviews of the same papers; NYT-AI uses synthetic counterparts of the same NYT articles. The random 80/20 document-level split can place the human and AI versions of the same source in different partitions, allowing the model to match content/topic instead of trajectory dynamics. No paper/article-level deduplication or grouped split is described. This threatens the central claim on two of the three benchmarks. Please re-split at the source level and re-run.
  2. [Section 4.4, Tables 4-9] The hyperparameter analysis reports Acc/F1 on the test partitions for each value, and the default values listed in Section 4.1 are the best-performing values from these sweeps. Because no validation split is described, the test set is effectively used for model selection. This inflates the reported numbers in Table 3. Use a held-out validation split or nested cross-validation for hyperparameter selection.
  3. [Section 3.3.3-3.3.4, Eq. 3.5-3.6; Section 4.5, Table 10] The central claim is that trajectory dynamics (first-order differences and temporally aligned similarity) drive detection. Table 3 compares GTCL to baselines that differ in encoder, objective, and input representation; the only ablation (Section 4.5) varies the backbone encoder. There is no control using the same window embeddings and contrastive objective with a static pooling (e.g., mean-pooled or concatenated windows, or using absolute z_i instead of differences δ_i). The geometric statistics in Section 4.2 show group differences but do not prove the learned representation exploits them. Please add such controls to isolate the contribution of trajectory dynamics.
minor comments (6)
  1. [Section 4.1, 2.2] ModernBERT is cited as [5] and [5,37]; reference [5] is the Drayson et al. paper, not the ModernBERT paper. The correct citation for ModernBERT is [37].
  2. [Eq. (3.7)] The normalization factor Z is defined as a double sum over all i,j, but w_ij=0 for j>i. Z should be restricted to j≤i; as written the weights do not sum to 1. Since the error multiplies all similarities by a constant, it does not change rankings but should be corrected.
  3. [Section 4.2] The Mann–Whitney U test results are only reported as p<0.05; reporting exact p-values or effect sizes would be more informative, especially given the large sample sizes.
  4. [Tables 3-9] All results appear to be single runs; no standard deviations or significance tests are provided. The RAID margins (e.g., GTCL vs DeTeCtive: 0.83 vs 0.81) are small and may not be significant. Report multiple seeds and confidence intervals.
  5. [Section 3.3.5] Hard positives/negatives are mined using the same similarity s(·,·) that defines the contrastive loss, coupling the mining criterion and the optimization target. The paper should ablate this choice (e.g., random negatives or a frozen mining representation) to ensure results are not an artifact of this coupling.
  6. [Section 5] The claim of 'well-calibrated predictions' is not supported by calibration experiments; please add evidence or soften the statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the trajectory-contrastive pipeline is a standard supervised training objective, and the benchmark comparisons are external to any fitted quantity.

full rationale

GTCL's derivation chain is not circular. The representation (Eq. 3.4-3.5) is a sequence of window-embedding differences; the similarity (Eq. 3.6-3.7) is a weighted inner product of those differences; the loss (Eq. 3.9) is supervised contrastive learning with external labels; inference is k-NN on frozen projections. Hard positive/negative mining in Sec. 3.3.5 uses the same similarity function as the loss, but this is a standard training-time mining loop, not a definition of the test prediction in terms of its own output: labels are external, the projection is frozen at inference, and the reported numbers are measured on held-out test partitions. The geometric statistics in Sec. 4.2 are in-sample descriptive evidence and are presented as motivation/support rather than as the detection prediction itself; using them as supportive evidence is not circular. There are no load-bearing self-citations, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The legitimate concerns about the paper -- absence of a static-representation control with the same backbone and contrastive objective, possible topic overlap in Reviews train/test splits, and hyperparameters apparently selected on test sets -- are external-validity or overfitting risks, not reductions of the central claim to its own inputs. Under the rule that only explicit construct-level circularity counts, the paper receives 0.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The trajectory is a representation, not a new entity; no new forces, particles, dimensions, or conserved quantities are introduced. The main burden falls on the free hyperparameters, most of which are tuned on the test partitions, and on the assumption that local window embeddings capture the purported autoregressive smoothness.

free parameters (8)
  • Sliding step Q = 8
    Chosen by hyperparameter sweep in Section 4.4 on the same test partitions used for headline results.
  • Window length M = 64
    Chosen by hyperparameter sweep in Section 4.4; intermediate values preferred.
  • Number of windows N = 32
    Chosen by hyperparameter sweep in Section 4.4.
  • Decay factor gamma = 0.2
    Chosen by hyperparameter sweep in Section 4.4; too low or too high degrades performance.
  • Contrastive group size Psi = 128
    Chosen by hyperparameter sweep in Section 4.4.
  • k-NN neighbors k = 9
    Chosen by hyperparameter sweep in Section 4.4; performance is robust to k.
  • Temperature tau = not reported
    Temperature in Eq. 3.9 is never given a value or analysis.
  • Projection module dimensionality and architecture = not specified
    The transformer-based projection module is described only as 'transformer-based'; layer count, hidden size, and output dimension D are not stated.
assumptions (4)
  • domain assumption Autoregressive generation induces local continuity that makes AI latent trajectories geometrically smoother or less variable than human trajectories.
    Stated in Section 3.1 and Section 5; supported only by weak Mann-Whitney differences in Table 2 and references [13, 23]. RAID inverts trajectory length and curvature, weakening the global regularity claim.
  • domain assumption Nomic Embed Text v1.5 embeddings of 64-word windows preserve enough local semantic information for trajectory deltas to be discriminative.
    Relies on the pretrained encoder; the only test is a one-backbone swap in Section 4.5.
  • domain assumption Supervised contrastive learning with hard positive/negative mining transfers to unseen test documents via kNN.
    This is the central learning assumption, supported only by the benchmark numbers themselves.
  • standard math Standard linear algebra and probability results used in Eqs. 3.6-3.9 are correct.
    Inner products, normalization, and softmax contrastive loss are standard; no unproved mathematical claims appear.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Trajectory Discrimination for AI-Generated Text Detection." pith.science (2026). https://pith.science/paper/QISNUH5U

@misc{pith2026260714967,
  author       = {Pith},
  title        = {Pith review of: Latent Trajectory Discrimination for AI-Generated Text Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QISNUH5U}},
  note         = {Machine review of arXiv:2607.14967}
}
read the original abstract

Most existing approaches to AI-Generated Text Detection (AIGTD) treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. However, this perspective overlooks the inherently dynamic nature of autoregressive generation, where content evolves progressively through the latent space. In this paper, we reformulate AIGTD as the problem of distinguishing between latent generation trajectories. Instead of relying on static representations, we model how textual representations evolve across the sequence. To this end, we propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation. GTCL then applies contrastive learning to these trajectories to learn geometric regularities associated with the autoregressive generation. Evaluations performed on three different benchmarks and several approaches show that GTCL outperforms detection baselines consistently, which implies that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains. These results suggest that modeling trajectory differences could improve detection and open up a dynamic direction that has been underexplored in previous AIGTD literature.

Figures

Figures reproduced from arXiv: 2607.14967 by the authors.

Figure 1
Figure 1. Overview of the GTCL workflow. A document is segmented into overlapping token windows [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

  1. [1]

    Bellini, F

    V. Bellini, F. Semeraro, J. Montomoli, M. Cascella, and E. Bignami. Between human and AI: assessing the reliability of AI text detection tools.Current Medical Research and Opinion, 40(3):353–358, 2024. Taylor & Francis

    2024

  2. [2]

    Bhattacharjee, T

    A. Bhattacharjee, T. Kumarage, R. Moraffah, and H. Liu. ConDA: Contrastive Domain Adaptation for AI-generated Text Detection. InProc. of the International Joint Conference on Natural Language Processing and the Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (IJCNLP’23), pages 598–610, Nusa Dua, Bali, 2023. ACL

    2023

  3. [3]

    Bhattacharjee and H

    A. Bhattacharjee and H. Liu. Fighting fire with fire: can ChatGPT detect AI-generated text?ACM SIGKDD Explorations Newsletter, 25(2):14–21, 2024. ACM

    2024

  4. [4]

    Cheng, V.S

    Y. Cheng, V.S. Sadasivan, M. Saberi, S. Saha, and S. Feizi. Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text. InProc. of the Annual Conference on Neural Information Processing Systems (NeurIPS’25), San Diego, CA, USA, 2025

    2025

  5. [5]

    Drayson, E

    G. Drayson, E. Yilmaz, and V. Lampos. Machine-generated text detection prevents language model collapse. In Proc. of the International Conference on Empirical Methods in Natural Language Processing (EMNLP’25), pages 29645–29661, Suzhou, China, 2025

    2025

  6. [6]

    Dugan, A

    L. Dugan, A. Hwang, F. Trhl ´ ık, A. Zhu, J. Magnus Ludan, H. Xu, D. Ippolito, and C. Callison-Burch. RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors. InProc. of the Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL’24, pages 12463–12492, Bangkok, Thailand, 2024. ACL

    2024

  7. [7]

    Gehrmann, H

    S. Gehrmann, H. Strobelt, and A.M. Rush. GLTR: Statistical Detection and Visualization of Generated Text. In Proc. of the International Conference of the Association for Computational Linguistics (ACL’19), pages 111–116, Florence, Italy, 2019. ACL

    2019

  8. [8]

    B. Guo, X. Zhang, Z. Wang, M. Jiang, J. Nie, Y. Ding, J. Yue, and Y. Wu. How close is chatgpt to human experts? comparison corpus, evaluation, and detection.arXiv preprint arXiv:2301.07597, 2023

    arXiv 2023

Show all 42 references
  1. [9]

    X. Guo, Y. He, S. Zhang, T. Zhang, W. Feng, H. Huang, and C. Ma. DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, (NeurIPS’24), ...

    2024

  2. [10]

    A. Hans, A. Schwarzschild, V. Cherepanova, H. Kazemi, A. Saha, M. Goldblum, J. Geiping, and T. Goldstein. Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text. InProc. of the International Conference on Machine Learning (ICML’24), Vienna, Austria, 2024....

    2024

  3. [11]

    J. Hao, J. Qiang, Y. Zhu, Y. Li, Y.H. Yuan, X. Hu, and X. Ouyang. Robust and semantic-faithful post-hoc watermarking of text generated by black-box language models.Frontiers of Computer Science, 19(9):199357, 2025

    2025

  4. [12]

    P. He, J. Gao, and W. Chen. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing. InProc. of the International Conference on Learning Representations (ICLR’23), Kigali, Rwanda, 2023

    2023

  5. [13]

    Holtzman, J

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi. The Curious Case of Neural Text Degeneration. InProc. of the International Conference on Learning Representations (ICLR’20), Addis Abeba, Ethiopia, 2020. OpenReview.net

    2020

  6. [14]

    H. Hu, X. Wang, Y. Zhang, Q. Chen, and Q. Guan. A comprehensive survey on contrastive learning.Neurocomputing, 610:128645, 2024. Elsevier

    2024

  7. [15]

    X. Hu, P.Y. Chen, and T.Y. Ho. RADAR: Robust AI-Text Detection via Adversarial Learning. InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems, (NeurIPS’23), New Orleans, LA, USA, 2023

    2023

  8. [16]

    Khosla, P

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan. Supervised Contrastive Learning. InAdvances in Neural Information Processing Systems (NeurIPS’20), Virtual Event, 2020

    2020

  9. [17]

    Kirchenbauer, J

    J. Kirchenbauer, J. Geiping, Y. Wen, J. Katz, I. Miers, and T. Goldstein. A Watermark for Large Language Models. InProc. of the International Conference on Machine Learning, (ICML’23), volume 202 ofProceedings of Machine Learning Research, pages 17061–17084, Honolulu, HI, USA,...

    2023

  10. [18]

    Kirchenbauer, J

    J. Kirchenbauer, J. Geiping, Y. Wen, M. Shu, K. Saifullah, K. Kong, K. Fernando, A. Saha, M. Goldblum, and T. Goldstein. On the Reliability of Watermarks for Large Language Models. InProc. of the International Conference on Learning Representations (ICLR’24), Vienna, Austria, ...

    2024

  11. [19]

    La Cava, D

    L. La Cava, D. Costa, and A. Tagarelli. Is Contrasting All You Need? Contrastive Learning for the Detection and Attribution of AI-generated Text. InProc. of the European Conference on Artificial Intelligence (ECAI’24) - Including Conference on Prestigious Applications of Intel...

    2024

  12. [20]

    Z. Li, Q. Yi, Z. Ji, Y. Lu, S. Zou, Y. Li, K. Xiao, and H. Liang. Bimarker: Enhancing text watermark detection for large language models with bipolar watermarks.Neurocomputing, page 133843, 2026. Elsevier

    2026

  13. [21]

    Lv and J

    Z. Lv and J. Sun. Zero-Shot Detection of LLM-Generated Text via Multilevel Information Compression and Multidimensional Feature Fusion.Neurocomputing, page 132731, 2026. Elsevier

    2026

  14. [22]

    C. Mao, C. Vondrick, H. Wang, and J. Yang. Raidar: geneRative AI Detection viA Rewriting. InProc. of the International Conference on Learning Representations (ICLR’24), Vienna, Austria, 2024. OpenReview.net

    2024

  15. [23]

    Mitchell, Y

    E. Mitchell, Y. Lee, A. Khazatsky, C.D. Manning, and C. Finn. DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature. InProc. of the International Conference on Machine Learning (ICML’23), volume 202, pages 24950–24962, Honolulu, HI, USA, 2023. PMLR

    2023

  16. [24]

    Y. Mo, H. Qin, Y. Dong, Z. Zhu, and Zhenglin Li. Large language model (llm) ai text generation detection based on transformer deep learning algorithm.arXiv preprint arXiv:2405.06652, 2024

    2024 arXiv

  17. [25]

    Nussbaum, J

    Z. Nussbaum, J. Xavier Morris, A. Mulyar, and B. Duderstadt. Nomic Embed: Training a Reproducible Long Context Text Embedder.Transactions on Machine Learning Research, 1, 2025

    2025

  18. [26]

    Pudasaini, L

    S. Pudasaini, L. Miralles-Pechu´ an, D. Lillis, and M. Llorens Salvador. Survey on AI-generated plagiarism detection: The impact of large language models on academic integrity.Journal of Academic Ethics, pages 1–34, 2024. Springer

    2024

  19. [27]

    W. Qu, W. Zheng, T. Tao, D. Yin, Y. Jiang, Z. Tian, W. Zou, J. Jia, and J. Zhang. Provably robust multi-bit watermarking for AI-generated text. InProc. of the USENIX Security Symposium (USENIX Security 25), pages 201–220, Seattle, W A, USA, 2025. 22

    2025

  20. [28]

    R. Roy, N. Imanpour, A. Aziz, S. Bajpai, G. Singh, S. Biswas, K. Wanaskar, P. Patwa, S. Ghosh, S. Dixit, N. Ranjan Pal, V. Rawte, R. Garimella, G. Jena, A.P. Sheth, V. Sharma, A. Naresh Reganti, V. Jain, A. Chadha, and A. Das. A Comprehensive Dataset for Human vs. AI Generated...

    2025 arXiv

  21. [29]

    Sturua and I

    S. Sturua and I. Mohr and M.K. Akram and M. G¨ unther and B. Wang and M. Krimmel and F. Wang and G. Mastrapas and A. Koukounas and N. Wang and H. Xiao. jina-embeddings-v3: Multilingual embeddings with task lora.arXiv preprint arXiv:2409.10173, 2024

    2024 arXiv

  22. [30]

    Sadasivan, A

    V.S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi. Can AI-Generated Text be Reliably Detected? Stress Testing AI Text Detectors Under Various Attacks.Transactions on Machine Learning Research, 1, 2025

    2025

  23. [31]

    Z. Shi, Y. Wang, F. Yin, X. Chen, K. W. Chang, and C.J. Hsieh. Red teaming language model detectors with language models.Transactions of the Association for Computational Linguistics, 12:174–189, 2024. MIT Pressy

    2024

  24. [32]

    Sun and Z

    J. Sun and Z. Lv. Zero-shot detection of LLM-generated text via text reorder.Neurocomputing, 631:129829, 2025. Elsevier

    2025

  25. [33]

    K. Sun, G. Bao, H. Cui, and Y. Zhang. When AI Settles Down: Late-Stage Stability as a Signature of AI-Generated Text Detection.arXiv preprint arXiv:2601.04833, 2026

    2026

  26. [34]

    MacFarland and J.M

    T.W. MacFarland and J.M. Yates. Mann–Whitney U test. InIntroduction to nonparametric statistics for the biological sciences using R, pages 103–132. 2016. Springer

    2016

  27. [35]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is All you Need. InProc. of the International Conference on Advances in Neural Information Processing Systems (NIPS’17), page 30, Long Beach, CA, USA, 2017. Curran A...

    2017

  28. [36]

    P. Wang, L. Li, K. Ren, B. Jiang, D. Zhang, and X. Qiu. SeqXGPT: Sentence-Level AI-Generated Text Detection. InProc. of the International Conference on Empirical Methods in Natural Language Processing (EMNLP’23), pages 1144–1156, Singapore, 2023. ACL

    2023

  29. [37]

    Warner, A

    B. Warner, A. Chaffin, B. Clavi´ e, O. Weller, O. Hallstr¨ om, S. Taghadouini, A. Gallagher, R. Biswas, F. Ladhak, T. Aarsen, G. T. Adams, J. Howard, and I. Poli. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finet...

    2025

  30. [38]

    K. Wu, L. Pang, H. Shen, X. Cheng, and T.S. Chua. LLMDet: A Third Party Large Language Models Generated Text Detection Tool. InFindings of the Association for Computational Linguistics (EMNLP’23), pages 2113–2133, Singapore, 2023. ACL

    2023

  31. [39]

    X. Yang, W. Cheng, Y. Wu, L.R. Petzold, W.Y. Wang, and H. Chen. DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text. InProc. of the International Conference on Learning Represen- tations, (ICLR’24), Vienna, Austria, 2024. OpenReview.net

    2024

  32. [40]

    X. Yang, L. Pan, X. Zhao, H. Chen, L.R. Petzold, W.Y. Wang, and W. Cheng. A Survey on Detection of LLMs- Generated Content. InFindings of the Association for Computational Linguistics (EMNLP’24), pages 9786–9805, Miami, FL, USA, 2024. ACL

    2024

  33. [41]

    X. Yang, K. Zhang, H. Chen, L. Petzold, W. Y. Wang, and W. Cheng. Zero-shot detection of machine-generated codes.arXiv preprint arXiv:2310.05103, 2023

    2023 arXiv

  34. [42]

    Z. Yang, Z. Feng, R. Huo, H. Lin, H. Zheng, R. Nie, and H. Chen. The Imitation Game revisited: A comprehensive survey on recent advances in AI-generated text detection.Expert Systems with Applications, 272:126694, 2025. Elsevier. 23

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.