Pith. sign in

REVIEW 3 major objections 5 minor 39 references

DWT-Fusion claims that localized, multiscale fluctuations in a proxy language model's token log-probabilities—not just average likelihood—carry the signal that separates machine-written from human-written text, and it shows a training-free

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

2026-08-01 06:00 UTC pith:W4ANGE26

load-bearing objection Novel training-free wavelet detector, but headline AUROCs are inflated by test-set configuration selection; the underlying idea still deserves a fair shot with a proper nested evaluation. the 3 major comments →

arxiv 2607.22026 v1 pith:W4ANGE26 submitted 2026-07-24 cs.CL

DWT-Fusion: A Signal-Based Framework for Training-Free LLM-Generated Text Detection

Mehmet Batuhan \"Ozda\c{s} , Murat Osmano\u{g}lu This is my paper
classification cs.CL
keywords LLM-generated text detectiontraining-free detectiondiscrete wavelet transformtoken log-probabilityvoting ensemblezero-shot detectionmultiresolution analysiscalibration
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 the sequential pattern of how confidently a language model predicts each next token contains detection evidence that global averages throw away. It proposes DWT-Fusion, which mean-centers the token log-probability signal and decomposes it with a discrete wavelet transform, then scores localized energy in the detail coefficients. On HC3, M4, and MAGE, the best single wavelet score reaches held-out AUROC 0.9872, 0.8185, and 0.7138, beating the evaluated likelihood, rank, entropy, LRR, and DFT baselines on the harder two datasets. A sympathetic reader would care because the method requires no detector training, no fine-tuning, and no reference database—only inference from an off-the-shelf proxy model.

Core claim

On the paper's own terms, the central discovery is that DWT-based multiresolution scoring of token log-probability signals is a competitive and interpretable training-free detector. The authors define three scalar wavelet-domain scores—first-level detail energy, multilevel detail energy, and window-energy variability—and show that multilevel detail energy is the strongest individual score on HC3 and M4, while first-level energy leads on MAGE. They further claim that combining wavelet configurations through calibration-weighted hard voting improves threshold-independent performance on all three datasets over both statistical baselines and a DFT spectral-energy baseline, with the largest gains

What carries the argument

The load-bearing object is the discrete wavelet transform (DWT) applied to the mean-centered sequence of conditional log-probabilities assigned by a proxy causal language model. The DWT decomposes the signal into approximation and detail coefficients at up to three levels; detail coefficients at fine scales capture abrupt token-level fluctuations and at coarser scales capture broader swings. Three scalar scores condense those coefficients—first-level detail energy, multilevel detail energy, and window-energy variability—and four training-free voting rules (equal-weight/calibration-weighted, hard/soft) merge configurations without learning a meta-classifier.

Load-bearing premise

The evaluation assumes that selecting the best configuration by AUROC on the held-out test split yields a valid estimate of performance, rather than an optimistic selection effect.

What would settle it

Fix the wavelet configuration on the calibration split alone (or use nested cross-validation) and check whether the reported AUROC gains over mean log-rank and LRR survive on held-out data; alternatively, shuffle the token order within each text and show that DWT scores still discriminate, which would indicate the signal is not actually sequence-localized.

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

If this is right

  • If the central claim is right, zero-shot detection can be improved without any training by re-representing token probability sequences in the wavelet domain.
  • The voting results imply that different proxy models, wavelet families, and score definitions carry complementary information; calibration-weighted voting is a cheap way to exploit that complementarity.
  • The comparison with DFT total energy suggests that localizing fluctuations matters more than global spectral content, so future signal-based detectors should keep locality.
  • Proxy model choice matters most on heterogeneous datasets, so the method's practical performance depends on matching the proxy to the expected text distribution.
  • Because the method is training-free, it can be pointed at new generators without retraining, though the paper's own results show low-FPR detection on MAGE remains weak.

Where Pith is reading between the lines

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

  • Beyond the paper: mean-centering discards the global likelihood level; a two-channel score that adds mean log-probability to wavelet energy is a natural testable extension that may close the gap on MAGE.
  • Beyond the paper: the held-out AUROC selection of the best configuration creates an optimism risk; nested cross-validation or pre-registered configuration choice would give a sharper estimate of real-world performance.
  • Beyond the paper: the 512-token truncation and three-level decomposition bound the method; sliding-window scoring over longer texts, which the paper lists as future work, could be tested immediately.
  • Beyond the paper: the window-energy variability score, though weak in AUROC, leads on TPR@5%FPR on MAGE, suggesting that dispersion features may combine with energy features in a hybrid detector.

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. This paper proposes DWT-Fusion, a training-free framework that treats the token-level conditional log-probability sequence from a proxy causal LM as a one-dimensional signal, mean-centers it, and applies a discrete wavelet transform. Three scalar scores are defined from the detail coefficients (first-level detail energy, multilevel detail energy, window-energy variability; Eqs. 8-13), optionally combined by four voting schemes (equal-weight or calibration-weighted, hard or soft; Eqs. 14-21), with directions, thresholds, and weights set on a 30% calibration split and evaluated on a 70% held-out split. Experiments on HC3, M4, and MAGE with four proxy LMs report best-configuration AUROCs of 0.9872, 0.8185, and 0.7138 and best voting AUROCs of 0.9919, 0.8477, and 0.7471, which the authors interpret as evidence that localized multiresolution variations in token log-probability signals provide complementary, training-free detection evidence over global statistics and global DFT spectral energy.

Significance. If the reported numbers were valid out-of-sample estimates, DWT-Fusion would be a useful addition to the training-free detector toolbox: the method is well specified (Eqs. 1-21), has no trainable parameters, and the paper gives a careful comparison against log-likelihood, rank, log-rank, entropy, LRR, and a DFT spectral baseline, across three large benchmarks and four proxy models. The paper also deserves credit for its explicit and honest limitations section (Section 6), which acknowledges proxy-model dependence, the calibration-guided rather than calibration-free nature of the method, truncation at 512 tokens, and the poor low-FPR results on MAGE. However, the headline empirical claims are currently vitiated by the evaluation protocol: configurations are selected on the held-out test split, so the reported AUROC values are selection maxima rather than honest estimates of deployment performance. The central value of the paper therefore depends on whether the evaluation can be corrected.

major comments (3)
  1. [§5.1, §5.3 (vs. §4.6)] The headline AUROC numbers are not out-of-sample estimates. §5.1 states that the proxy LM and wavelet family are selected according to AUROC on the held-out test split for individual wavelet scores, and likewise for baselines. The proposed method is thus the maximum of a 60-configuration grid (4 proxies × 5 wavelets × 3 scores) evaluated on the reporting data, while each baseline is selected over only 4 proxies. This contradicts §4.6's claim that the test split is not used for any configuration-specific decision-boundary adjustment, and it also affects Table 7 (DWT best-of-60 vs. DFT best-of-4). Because selection bias grows with the number of candidates, the reported advantages on M4 and MAGE (0.8185 vs. 0.7751; 0.7138 vs. 0.6907) and the claimed DFT gaps may be artifacts. Fix: use a nested protocol, selecting configurations on the calibration split only, and re-derive the headline numbe
  2. [§4.5, §5.1 (voting ensembles)] The ensemble results are also selected per dataset on the test split: §5.1 says 'we report the best-performing voting configuration for each dataset'. The voting grid is large, comprising 4 voting schemes × 4 weighting rules × 4 scopes (score-level, wavelet-family, proxy-model, full 60-configuration). The claimed fusion gains (e.g., MAGE 0.7138 → 0.7471) compare a test-selected best ensemble against a test-selected best single configuration; the ensemble has strictly more choices, so the comparison is biased in its favor. The fusion benefit is not established. The authors should pre-specify the voting scheme and scope or select them on the calibration split, and report the resulting performance rather than the best observed.
  3. [§4.6, §5.1 (repeated splits)] All conclusions rest on a single 30/70 split with one fixed seed and no error bars. With configuration selection among 60+ candidates on the evaluation split, point-estimate AUROCs cannot distinguish genuine improvement from selection noise; the M4 and MAGE gaps over the best baselines are small (roughly 0.04 and 0.02), and a different split or seed could easily reverse them. Report means and standard deviations over multiple seeds with the complete selection protocol nested inside each split, and present paired comparisons of the proposed scores against the baselines under identical selection rules.
minor comments (5)
  1. [§4.6] The sentence stating that the held-out test split is not used for any configuration-specific decision-boundary adjustment is directly contradicted by the §5.1 selection protocol; the wording must be corrected regardless of how the protocol is fixed.
  2. [Abstract, §1] The abstract and the contribution list present test-selected maxima as the method's performance without stating the selection protocol; after fixing the protocol, these passages should report the performance of a fixed or calibration-selected configuration.
  3. [§4.1] No code repository or artifact link is provided. Given the 60-configuration grid and the many voting/weighting options, releasing code and configuration metadata would considerably aid reproducibility and verification.
  4. [§2, Table 1] Minor naming inconsistency: 'WAVEDETECT' in §2 versus 'WaveDetect' in Table 1 and reference [35]; please unify.
  5. [§3.4] For the window-energy variability score (Eqs. 11-13), the paper should state the minimum number of windows a detail level must contain to count as valid, since level-3 detail sequences for short inputs can be very short; the current validity rule is underspecified.

Circularity Check

0 steps flagged

No circularity: wavelet scores are explicit, parameter-free functions of DWT coefficients; the test-set-configuration selection flagged in Section 5.1 is an evaluation-protocol bias, not a derivation loop.

full rationale

DWT-Fusion's detection scores are not defined in terms of the labels or of the reported AUROC. Equations (8), (10), and (13) define first-level detail energy, multilevel detail energy, and window-energy variability directly as functions of the DWT detail coefficients of the mean-centered log-probability sequence, and the DWT itself is a standard transform, not the authors' own result (Sections 3.3-3.4). Score direction, normalization statistics, thresholds, and voting weights are computed only on the 30% calibration split, not on the held-out test split (Sections 3.5 and 4.6). There is no self-citation chain: the cited signal-based work (SpecDetect, WaveDetect) enters only as comparison, and no uniqueness theorem or prior result by the authors is used to force the design. The one load-bearing concern is the protocol stated in Section 5.1: 'For individual wavelet-domain scores, the proxy language model and wavelet family are selected according to AUROC on the held-out test split. For statistical baselines, the best proxy language model is also selected according to AUROC. For voting methods, we report the best-performing voting configuration for each dataset.' This means the headline AUROC values are maxima over a configuration grid evaluated on the same data used for reporting, which can inflate performance estimates and is a correctness risk for the empirical claim. However, it is not circularity under the definitions in the rubric: the scores themselves are well-defined functions of the inputs, no fitted parameter or self-citation is renamed as a prediction, and no equation reduces a reported detection result to a fitted value by construction. The honest verdict is therefore no significant circularity, with the test-set-selection issue flagged as a separate methodological caveat.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The framework is a scoring pipeline, not a derivation. It relies on hand-fixed hyperparameters and a labeled calibration split; no new physical entities or theoretical objects are introduced. The main evidential weakness is test-set configuration selection, which is captured in soundness and red_flags.

free parameters (5)
  • Maximum DWT decomposition level L_max = 3
    Set by hand in Section 3.3; affects which detail coefficients enter the scores. No sensitivity analysis is provided.
  • Window size w for window-energy variability = 8
    Fixed in Section 4.1; the window_std score depends heavily on this arbitrary choice.
  • Maximum token length = 512
    Truncation limit in Section 3.1; discards any signal beyond 512 tokens and may bias longer texts.
  • Calibration split ratio = 0.30
    Fixed in Section 4.6; the labeled calibration split is used for score direction, normalization, threshold selection, and voting weights.
  • Calibration weights and score-direction signs = computed from calibration AUROC/AUPRC/TPR metrics
    Section 3.5: deterministic functions of calibration data, but they make ensemble performance partly a function of labeled calibration labels.
axioms (5)
  • standard math Standard wavelet multiresolution decomposition theory
    Invoked in Section 3.3 via Mallat and Daubechies; the DWT coefficients are assumed to be accurate for the chosen filters and boundary mode.
  • domain assumption Symmetric boundary extension and the fixed maximum level produce stable detail coefficients on variable-length signals
    Section 3.3: short sequences with boundary effects may violate this, and the paper excludes only sequences shorter than four values.
  • domain assumption Token log-probability sequences of LLM-generated and human text differ in localized multiscale fluctuation structure
    Central motivating assumption of Sections 1 and 3.1; empirically testable but not proven.
  • domain assumption Mean-centering removes only the global offset and preserves discriminative fluctuation structure
    Section 3.2: if the main human/machine difference is in global likelihood, this preprocessing would remove the signal.
  • domain assumption The calibration split is representative of the test distribution and can be used for threshold/weight selection without overfitting
    Section 4.6: single stratified split with seed 42; no repeated trials or cross-validation to check stability.

pith-pipeline@v1.3.0-alltime-deepseek · 23115 in / 12280 out tokens · 122885 ms · 2026-08-01T06:00:39.317463+00:00 · methodology

0 comments
read the original abstract

Detecting LLM-generated text remains challenging under zero-shot and training-free conditions, especially when detectors must generalize across datasets, domains, and unseen generators. While existing training-free approaches exploit language-model statistics as detection signals, they typically characterize a text through global measures that summarize overall model behavior. Consequently, potentially informative local and multiscale variations in token-level predictability may remain underutilized. Motivated by this observation, we introduce DWT-Fusion, a training-free signal-based framework for detecting LLM-generated text using discrete wavelet analysis of token-level log-probability sequences produced by a proxy causal language model. The proposed framework analyzes these sequences through wavelet-based multiresolution signal representations and derives detection signals from localized probability dynamics. We further evaluate four training-free voting variants, including equal-weight hard voting, equal-weight soft voting, calibration-weighted hard voting, and calibration-weighted soft voting, to combine multiple wavelet configurations without training a supervised meta-classifier. We evaluate the framework on HC3, M4, and MAGE using GPT-Neo-2.7B, GPT-J-6B, Falcon-7B, and LLaMA-3-8B as proxy models. The best single wavelet configurations achieve AUROC values of 0.9872, 0.8185, and 0.7138 on HC3, M4, and MAGE, respectively. With calibration-weighted voting, the best ensemble variants further improve AUROC to 0.9919, 0.8477, and 0.7471. These findings show that DWT-based multiresolution scoring and calibration-guided voting fusion provide effective and interpretable signals for training-free LLM-generated text detection.

Figures

Figures reproduced from arXiv: 2607.22026 by Mehmet Batuhan \"Ozda\c{s}, Murat Osmano\u{g}lu.

Figure 1
Figure 1. Figure 1: Overview of DWT-Fusion, the proposed DWT-based signal scoring and [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Wavelet-domain representation of the mean-centered token log-probability [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Proxy model sensitivity of the proposed wavelet-based detector on the [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Wavelet family sensitivity of the proposed method on the held-out test [PITH_FULL_IMAGE:figures/full_fig_p030_4.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

39 extracted references · 3 linked inside Pith

  1. [1]

    Fraser, Hillary Dawkins, and Svetlana Kiritchenko

    Kathleen C. Fraser, Hillary Dawkins, and Svetlana Kiritchenko. Detecting AI-generated text: Factors influencing detectability with current methods. Journal of Artificial Intelligence Research, 82:2233–2278, April 2025

    2025

  2. [2]

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

    Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia Sam Chao, and Derek Fai Wong. A survey on LLM-generated text detection: Necessity, methods, and future directions.Computational Linguistics, 51(1):275–338, March 2025

    2025

  3. [3]

    A survey on detection of LLMs- generated content

    Xianjun Yang, Liangming Pan, Xuandong Zhao, Haifeng Chen, Linda Ruth Petzold, William Yang Wang, and Wei Cheng. A survey on detection of LLMs- generated content. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9786–9805, Miami, Florida, USA, November 2024. Associati...

    2024

  4. [4]

    Manning, and Chelsea Finn

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. Detectgpt: zero-shot machine-generated text detection using 35 probability curvature. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023

    2023

  5. [5]

    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, editors,Findings of the Asso- ciation for Computational Linguistics: EMNLP 2023, pages 12395–12412, Singapore, December 2023. Association for Computational Linguistics

    2023

  6. [6]

    Fast-detectGPT: Efficient zero-shot detection of machine-generated text via conditional probability curvature

    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. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  7. [7]

    Specdetect: simple, fast, and training-free detec- tion of llm-generated text via spectral analysis

    Haitong Luo, Weiyao Zhang, Suhang Wang, Wenji Zou, Chungang Lin, Xuy- ing Meng, and Yujun Zhang. Specdetect: simple, fast, and training-free detec- tion of llm-generated text via spectral analysis. InProceedings of the Fortieth AAAI Conference on Artificial Intelligence and Thirty-Eighth Conference on Innovative Applications of Artificial Intelligence and...

    2026

  8. [8]

    How close is ChatGPT to human experts? com- parison corpus, evaluation, and detection.arXiv preprint arXiv:2301.07597, 2023

    Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. How close is ChatGPT to human experts? com- parison corpus, evaluation, and detection.arXiv preprint arXiv:2301.07597, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    M4: Multi-generator, multi-domain, and multi- lingual black-box machine-generated text detection

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shel- manov, Akim Tsvigun, Chenxi Whitehouse, Osama Mohammed Afzal, Tarek Mahmoud, Toru Sasaki, Thomas Arnold, Alham Aji, Nizar Habash, Iryna Gurevych, and Preslav Nakov. M4: Multi-generator, multi-domain, and multi- lingual black-box machine-generated text detection. InProceedings of the 18th C...

    2024

  10. [10]

    MAGE: Machine-generated text detection in the wild

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. MAGE: Machine-generated text detection in the wild. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 36–53, Bangkok, Thailand, August 2024. Association for Computational Linguistics. 36

    2024

  11. [11]

    DNA-GPT: Divergent N-gram analysis for training-free de- tection of GPT-generated text

    Xianjun Yang, Wei Cheng, Yue Wu, Linda Petzold, William Wang, and Haifeng Chen. DNA-GPT: Divergent N-gram analysis for training-free de- tection of GPT-generated text. InInternational Conference on Learning Representations, volume 2024, pages 48572–48597, 2024

    2024

  12. [12]

    Spot- ting llms with binoculars: zero-shot detection of machine-generated text

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Spot- ting llms with binoculars: zero-shot detection of machine-generated text. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  13. [13]

    Training-free LLM-generated text detection by mining token prob- ability sequences

    Yihuai Xu, Yongwei Wang, Yifei Bi, Huangsen Cao, Zhouhan Lin, Yu Zhao, and Fei Wu. Training-free LLM-generated text detection by mining token prob- ability sequences. InInternational Conference on Learning Representations, volume 2025, pages 19072–19098, 2025

    2025

  14. [14]

    Zero-shot detection of LLM-generated text using temperature sensitivity

    Shixuan Ma, Jiahao Li, Zhendong Mao, and Quan Wang. Zero-shot detection of LLM-generated text using temperature sensitivity. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 37664–37679, San Diego, California, United States, July 2026. Association for Computational Linguistics

    2026

  15. [15]

    Wong, Shu Yang, Xuebo Liu, Lidia S

    Junchao Wu, Runzhe Zhan, Derek F. Wong, Shu Yang, Xuebo Liu, Lidia S. Chao, and Min Zhang. Who wrote this? the key to zero-shot LLM-generated text detection is GECScore. In Owen Rambow, Leo Wanner, Marianna Apidi- anaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven Schockaert, editors, Proceedings of the 31st International Conference on Computational L...

    2025

  16. [16]

    SILTD: Structural information for LLM-generated text detection.International Journal of Machine Learning and Cybernetics, 16(9):6095–6110, 2025

    Jing Yang, Shi Wang, Kangli Zi, Yanshun Sun, Yuwei Huang, and Tianyu Luo. SILTD: Structural information for LLM-generated text detection.International Journal of Machine Learning and Cybernetics, 16(9):6095–6110, 2025

    2025

  17. [17]

    Zero-shot detection of LLM-generated text via text reorder.Neurocomputing, 631:129829, 2025

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

    2025

  18. [18]

    Radar: robust ai-text de- tection via adversarial learning

    Xiaomengc Hu, Pin-Yu Chen, and Tsung-Yi Ho. Radar: robust ai-text de- tection via adversarial learning. InProceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc. 37

    2023

  19. [19]

    Fdllm: A dedicated detector for black-box llms fingerprinting

    Zhiyuan Fu, Junfan Chen, Lan Zhang, Ting Yang, Jun Niu, Hongyu Sun, Ruidong Li, Peng Liu, Jice Wang, Fannv He, and Yuqing Zhang. Fdllm: A dedicated detector for black-box llms fingerprinting. In2025 IEEE 24th International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom), pages 1374–1379, 2025

    2025

  20. [20]

    Ghostbuster: Detecting text ghostwritten by large language models

    Vivek Verma, Eve Fleisig, Nicholas Tomlin, and Dan Klein. Ghostbuster: Detecting text ghostwritten by large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1702–1717, Mexico City, Mexico, June 2024. Association f...

    2024

  21. [21]

    Learning to rewrite: Generalized LLM-generated text detection

    Wei Hao, Ran Li, Weiliang Zhao, Junfeng Yang, and Chengzhi Mao. Learning to rewrite: Generalized LLM-generated text detection. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6421–6434, Vienna, Aus...

  22. [22]

    Human texts are outliers: Detecting LLM-generated texts via out-of-distribution detection

    Cong Zeng, Shengkun Tang, Yuanzhou Chen, Zhiqiang Shen, Wenchao Yu, Xujiang Zhao, Haifeng Chen, Wei Cheng, and Zhiqiang Xu. Human texts are outliers: Detecting LLM-generated texts via out-of-distribution detection. arXiv preprint arXiv:2510.08602, 2025

    arXiv 2025

  23. [23]

    LM2otifs: An explainable framework for machine-generated texts detection.arXiv preprint arXiv:2505.12507, 2025

    Xu Zheng, Zhuomin Chen, Esteban Schafir, Sipeng Chen, Hojat Allah Salehi, Haifeng Chen, Farhad Shirani, Wei Cheng, and Dongsheng Luo. LM2otifs: An explainable framework for machine-generated texts detection.arXiv preprint arXiv:2505.12507, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    LOG-AID: Logit-based statistical features for AI text detection

    Sophie Titze and Oren Halvani. LOG-AID: Logit-based statistical features for AI text detection. InNotebook for PAN at CLEF 2025, 2025

    2025

  25. [25]

    BiScope: AI-generated text detection by checking memorization of preceding tokens

    Siyuan Cheng, Hanxi Guo, Xiaolong Jin, Guangyu Shen, Guanhong Tao, Kaiyuan Zhang, Xiangyu Zhang, and Zhuo Zhang. BiScope: AI-generated text detection by checking memorization of preceding tokens. InAdvances in Neural Information Processing Systems, volume 37, pages 104065–104090. Neural Information Processing Systems Foundation, Inc., 2024

    2024

  26. [26]

    Can attention detect AI-generated text? a novel Benford’s law-based approach.Information Processing & Management, 62(4):104139, 2025

    Zhenhua Wang, Guang Xu, and Ming Ren. Can attention detect AI-generated text? a novel Benford’s law-based approach.Information Processing & Management, 62(4):104139, 2025. 38

    2025

  27. [27]

    language is the dress of thought

    Zhenhua Wang, Guang Xu, and Ming Ren. “language is the dress of thought”: A new method for automatic detection of AI-generated text.Decision Support Systems, 201:114578, 2026

    2026

  28. [28]

    A metric-based detection system for large language model texts.ACM Transactions on Management Information Systems, 16(1), February 2025

    Linh Le and Dung Tran. A metric-based detection system for large language model texts.ACM Transactions on Management Information Systems, 16(1), February 2025

    2025

  29. [29]

    De- coupling content and expression: Two-dimensional detection of AI-generated text.arXiv preprint arXiv:2503.00258, 2025

    Guangsheng Bao, Lihua Rong, Yanbin Zhao, Qiji Zhou, and Yue Zhang. De- coupling content and expression: Two-dimensional detection of AI-generated text.arXiv preprint arXiv:2503.00258, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Di- vScore: Zero-shot detection of LLM-generated text in specialized domains

    Zhihui Chen, Kai He, Yucheng Huang, Yunxiao Zhu, and Mengling Feng. Di- vScore: Zero-shot detection of LLM-generated text in specialized domains. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19231–19253, Suzhou, China...

    2025

  31. [31]

    Detective: detecting ai-generated text via multi- level contrastive learning

    Xun Guo, Shan Zhang, Yongxin He, Ting Zhang, Wanquan Feng, Haibin Huang, and Chongyang Ma. Detective: detecting ai-generated text via multi- level contrastive learning. InProceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY , USA,

  32. [32]

    Association for Computational Linguistics

  33. [33]

    Adadetectgpt: Adaptive detection of llm- generated text with statistical guarantees

    Hongyi Zhou, Jin Zhu, Pingfan Su, Kai Ye, Ying Yang, Shakeel Gavioli- Akilagun, and Chengchun Shi. Adadetectgpt: Adaptive detection of llm- generated text with statistical guarantees. InAdvances in Neural Information Processing Systems 38 (NeurIPS 2025), Advances in Neural Information Processing Systems. NeurIPS, September 2025. The Thirty-Ninth Annual Co...

    2025

  34. [34]

    Continual origin tracing of LLM-generated text

    Haoran Li and Quan Wang. Continual origin tracing of LLM-generated text. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’25, pages 479–489, New York, NY , USA, 2025. Association for Computing Machinery. 39

    2025

  35. [35]

    DETree: DEtecting human-AI collaborative texts via tree-structured hierarchical repre- sentation learning.arXiv preprint arXiv:2510.17489, 2025

    Yongxin He, Shan Zhang, Yixuan Cao, Lei Ma, and Ping Luo. DETree: DEtecting human-AI collaborative texts via tree-structured hierarchical repre- sentation learning.arXiv preprint arXiv:2510.17489, 2025

    arXiv 2025

  36. [36]

    S. G. Mallat. A theory for multiresolution signal decomposition: The wavelet representation.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 11(7):674–693, July 1989

    1989

  37. [37]

    Society for Industrial and Applied Mathematics, 1992

    Ingrid Daubechies.Ten Lectures on Wavelets. Society for Industrial and Applied Mathematics, 1992. 40

    1992

  38. [38]

    WaveDetect: Robust framework for machine-generated text detection via wavelet transform

    Zhichen Liu, Kaitong Qin, Linhan He, and Yang Xu. WaveDetect: Robust framework for machine-generated text detection via wavelet transform. In Findings of the Association for Computational Linguistics: ACL 2026, pages 8712–8727, San Diego, California, United States, July 2026. Association for Computational Linguistics

    2026

  39. [2024]

    Curran Associates Inc