Pith. sign in

REVIEW 3 major objections 5 minor 49 references

PhantomHunter: Detecting Unseen Privately-Tuned LLM-Generated Text via Family-Aware Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PhantomHunter detects text from unseen private fine-tuned LLMs by learning a family-level token-probability signature, reaching F1 above 96%.

desk verdict PhantomHunter is the first to tackle detection of text from unseen privately-tuned LLMs of known families; the method is sound and the results strong, but the persistence claim is tested at only one fine-tuning scale. read the letter →

arxiv 2506.15683 v1 pith:W46SXXQP submitted 2025-06-18 cs.CL cs.CY

Yuhui Shi , Yehan Yang , Qiang Sheng , Hao Mi , Beizhe Hu , Chaoxi Xu , Juan Cao This is my paper
classification cs.CLcs.CY
keywords LLM-generatedtextdetectionprivatelyfine-tunedLLMsfamily-awarelearningtokenprobabilityfeaturescontrastivemixture-of-expertsunseenmodelgeneralizationLLaMA/GemmaMistral
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

PhantomHunter addresses a gap in AI-text detection: text produced by a privately fine-tuned open-source LLM, where a user has adapted a public model on private data and then generated content, causes existing detectors to degrade sharply. The paper first shows this failure empirically, with a probability-based detector losing 41% accuracy after fine-tuning on 11.3M tokens, and then identifies the cause: token probability lists of a fine-tuned model stay closer to their own base family than to other families. The proposed detector learns these family-level traits from observable derivatives and uses them to judge text from held-out fine-tuned models of the same families. If correct, it gives a practical forensic tool for catching text from private fine-tuned LLaMA, Gemma, and Mistral models using only the public base models at test time, with F1 above 96%.

What carries the argument

The load-bearing object is the family-level token probability signature. For a text x, the method runs M base LLMs and records each token's probability, forming an M×N matrix p; a CNN and a two-layer Transformer encode this into a shared representation RF. A SimCLR-style contrastive loss treats samples generated by models of the same family as positive pairs, and a family classifier predicts the source family, producing gating weights that route RF through a mixture-of-experts binary detector, one expert per family. The idea is that fine-tuning adapts behavior to a domain while preserving much of the base model's probabilistic structure, so the shared trait survives and can be learned from observable fine-tuned models.

What would settle it

Fine-tune a single base LLM on a much larger or more style-altering corpus (say, over 100M tokens of one distinctive domain) and check whether its token probability lists still cluster with the parent base rather than another family, and train PhantomHunter with one family held out entirely; if either the clustering or the never-seen-family detection collapses, the family trait is an artifact of the light fine-tuning tested here.

Watch

Extended reading notes

Core claim

The central claim is that a fine-tuned LLM retains a detectable 'family trait' in token-probability space: after full-parameter or LoRA fine-tuning, text generated by an unseen derivative model is more probable under its own base model than under other families' base models, and PhantomHunter learns to exploit exactly this. The architecture extracts per-token probabilities from several base models, encodes them with CNN and Transformer layers, pulls same-family samples together with a contrastive loss, predicts the family, and uses a mixture-of-experts detector whose expert weights are gated by that prediction. On held-out fine-tuned derivatives of LLaMA-2 7B-Chat, Gemma 7B-it, and Mistral 7B-Instruct-v0.1, trained on only one unseen domain per family, the detector reaches Macro-F1 of 96.24-97.14 on arXiv-style writing and 96.40-96.98 on Q&A, outperforming seven baselines and four commercial services.

Load-bearing premise

PhantomHunter works only if a privately tuned model stays closer in token-probability space to its own base family than to other families after fine-tuning, and only when that base family is known and its base models are deployed at test time.

Editorial extensions

If this is right

  • A detector can catch text from a privately fine-tuned model it has never seen, as long as the base family is known and its base models are available at test time.
  • Because PhantomHunter scores source family as well as human/AI, a suspicious passage can be attributed to LLaMA, Gemma, or Mistral derivatives, though family accuracy is moderate (Macro F1 61 to 81%).
  • Adding an 'others' category for seen closed models such as GPT-4o mini and Claude 3.7 Sonnet preserves performance on unseen fine-tuned models, so the design can broaden beyond open families.
  • Removing any of the three components (base probability features, contrastive family loss, mixture-of-experts gating) reportedly lowers performance, meaning the family signal is carried jointly by all of them.

Reading between the lines

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

  • Untested but plausible: heavier fine-tuning (large corpora, RLHF-style alignment, or domain shift) could erode the family trait; the paper only held out one domain per family, so the persistence claim is demonstrated under light to moderate adaptation.
  • A natural test is to measure family-trait similarity directly: compute cosine similarity between a derivative's probability lists and all candidate base models after increasingly large fine-tuning to find where the trait decays.
  • The requirement to deploy all base models locally means PhantomHunter is a forensic tool for institutions that can host large open models, not a lightweight end-user API; the paper openly trades compute and memory for detection power.
  • The contrastive family-learning idea transfers to other forensic tasks, such as attributing a text to a specific open model family when the private derivative is unavailable, which the paper's moderate family-prediction results suggest is still unsolved.
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 / 5 minor

Summary. The paper addresses the detection of text generated by privately fine-tuned open-source LLMs. It first reports a preliminary experiment showing that standard detectors degrade as the amount of fine-tuning data grows, and a similarity analysis indicating that token-probability lists of a base model and its own fine-tuned derivatives are more similar to each other than to other families. The authors then propose PhantomHunter, which extracts token-probability features using several base LLMs, applies a contrastive family-aware learning objective, and combines a family classifier with a mixture-of-experts detection head. Experiments on arXiv abstracts and HC3-derived Q&A data across the LLaMA, Gemma, and Mistral families, with held-out fine-tuning domains (cs and fin), report macro F1 scores around 96-97%, outperforming the compared baselines and commercial services.

Significance. If the results hold, the paper makes a useful practical contribution: it demonstrates a real failure mode of existing detectors on privately fine-tuned models, proposes a non-watermarking detection approach that leverages family-level probabilistic structure, and provides ablations showing that each component contributes. The preliminary risk analysis and the simple extension to seen models are also valuable. The main limitation is that the central persistence claim is tested only at a single fine-tuning scale and with one held-out domain per scenario, and the results are single runs without error bars or released code/data. The stated scope of 'unseen privately-tuned LLMs' is therefore broader than the evidence directly supports.

major comments (3)
  1. [Section 3.3 and Section 2 (Figure 1)] The load-bearing claim that fine-tuning 'preserves much of the underlying probabilistic structure' is asserted unconditionally, but the evaluation never varies the fine-tuning scale. All derivative models in Section 4.1 use corpora in the range of roughly 0.2-6.4M tokens, while Figure 1 shows that SeqXGPT's detection accuracy drops monotonically as the fine-tuning corpus grows to 11.3M tokens. This creates a direct tension: the paper's own preliminary experiment indicates that the probabilistic fingerprint erodes with heavier fine-tuning, yet the central generalization claim is only tested on comparable, relatively light fine-tuning. Please add experiments that vary fine-tuning data size, training epochs, or checkpoints, or explicitly restrict the persistence claim to the tested fine-tuning regime.
  2. [Section 4.1 and Table 3] The 'unseen' dimension in the evaluation is only the fine-tuning domain: cs is held out for arXiv and fin for Q&A. The training and test derivatives use the same fine-tuning recipe and comparable corpus sizes, so the test models differ from the training models along the domain axis but not along the adaptation-strength axis. Since the paper's opening experiment identifies adaptation strength as a critical factor, the evaluation should include held-out domains with varied fine-tuning scale and possibly different fine-tuning recipes (e.g., different LoRA ranks, different epochs) to substantiate the claim of generalizing to unseen privately-tuned models.
  3. [Section 4.2 and Table 3] All reported results are single F1 numbers without variance or statistical significance tests. The claimed improvements over the best baseline are 2.0-6.1 percentage points depending on setting; without multiple seeds, confidence intervals, or paired tests, it is difficult to assess whether these differences are stable. This is particularly important because the central claim is superiority over existing detectors on the unseen-model scenario.
minor comments (5)
  1. [Abstract and Conclusion] The abstract and conclusion say '3 industrial services', but Section 4.5 and Table 5 report comparisons with four commercial detectors (WinstonAI, Sapling, BlueEyes, HasteWire). Please correct the count.
  2. [Table 3 caption] The caption contains the typo 'Constrative learning'; it should read 'Contrastive learning'.
  3. [Section 4.1] The text contains the typo 'HC3 dateset'; it should be 'HC3 dataset'.
  4. [Equation (2)] The notation for the augmented sample is inconsistent: the numerator uses \tilde{R}^m_{F_i} but the denominator uses R_{F_j} without a tilde, making it unclear whether the augmented positive is included in the denominator. Please define all symbols explicitly.
  5. [Section 4.5, Analysis 1] The family classification results in Table 4 are reported as aggregate F1 scores without per-family breakdown or analysis of confusion patterns; a brief discussion of which families are confused would help interpret the 'moderate' family prediction and its effect on the MoE gating signal.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the detector is a supervised, held-out evaluation; the only caveat is motivational leakage in the Figure 2 heatmap, which is not definitional.

full rationale

PhantomHunter's derivation chain is an empirical supervised pipeline rather than a derivation from fitted definitions. The family-trait hypothesis (Section 2) is measured by cosine similarity of token-probability lists between fine-tuned models and their bases; the detector then extracts the same kind of probability features from base models (Eq. 1), trains a family classifier and an MoE detector on the base/phy/oth and base/med/ELI5 derivatives, and evaluates on held-out cs/fin derivatives. The held-out domains are excluded from training, and the binary detection loss (Eqs. 4-5) is not defined in terms of the fitted family predictions except as gating weights. The central persistence claim is an empirical, falsifiable assumption, not a tautology: it could have failed, and the reported ablations (w/o BFE, w/o CL, w/o MoE) show the components matter. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation. The self-citations (Shi et al. 2024; Hu et al. 2024) are background or ethical-discussion references and do not carry the argument. One methodological caveat: the motivating heatmap in Figure 2 includes the cs models later held out for testing, so the 'unseen' evaluation is not fully blind to the motivating evidence; this is a data-leakage/experimental-design concern, not a definitional circularity, and it does not force the reported F1 scores because the trained detector never sees cs labels or samples. The paper's own limitations (Section 6 and Ethical Considerations) further bound the claim to known families and deployed base models, which reduces overclaiming. Overall: no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method's actual contribution is an inductive bias (family-aware features), not a new physical or symbolic entity. The free parameters above are the hand-chosen knobs; the neural network weights themselves are fitted but unreleased. The main axioms are empirical assumptions about how fine-tuning interacts with the base model's probability distribution, and about test-time access to the parent family.

free parameters (6)
  • Loss balancing weights lambda1, lambda2, lambda3 = 1.0, 1.0, 0.5
    Chosen by hand to balance family classification, detection, and contrastive losses; no sensitivity analysis reported (Section 3.3, Eq. 5).
  • Contrastive temperature t = 0.07
    Standard SimCLR value chosen without tuning; affects family representation separation (Section 3.2).
  • Feature dimension d = 128
    Hidden size for the CNN/Transformer encoder; selected by hand with no ablation (Section 4.2).
  • Decision threshold = 0.5
    Fixed threshold used for F1 computation for all methods; but for DNA-GPT and DetectGPT the authors grid-search the threshold, so the comparison is not strictly equal (Section 4.2).
  • Fine-tuning and training hyperparameters = 10 epochs, detector lr=2e-5, fine-tuning lr=1e-5, LoRA rank 8
    Reported in Sections 4.2 and A.3; these are settings rather than fitted constants, and no sensitivity analysis is given.
  • Trained network weights (encoders, family classifier, MoE experts) = not reported (checkpoints not released)
    The central detection function is a learned classifier; no checkpoints are provided, so the exact fitted model cannot be inspected or reproduced.
assumptions (4)
  • domain assumption Fine-tuning preserves family-level probabilistic structure: token probability lists of derivatives remain more similar to their own base model than to other families (Section 2, Figure 2).
    Empirically observed on 12 models from 3 families and 3 domains; this is the core inductive premise for family-aware learning and is only tested at one fine-tuning scale per domain.
  • domain assumption The parent family of the private model is known and its base models are available at test time (Sections 3.1 and 6).
    Feature extraction requires running the suspicious text through the exact base models theta_1...theta_M; if the private model derives from an unknown family or the base model is unavailable, the method cannot be applied.
  • domain assumption Fine-tuning on domain-specific corpora (arXiv subsets, HC3 domains) with full-parameter and LoRA is representative of real private tuning (Section 4.1).
    Only one held-out domain per scenario and two tuning methods are tested; other data scales, alignment methods, or architectures may behave differently.
  • standard math Standard supervised learning assumptions: i.i.d. train/test samples and a fixed label space (Sections 3.2, 4.2).
    The training protocol uses empirical risk minimization and a fixed set of family labels; if test distributions change, the model's guarantees do not transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PhantomHunter: Detecting Unseen Privately-Tuned LLM-Generated Text via Family-Aware Learning." pith.science (2026). https://pith.science/paper/W46SXXQP

@misc{pith2026250615683,
  author       = {Pith},
  title        = {Pith review of: PhantomHunter: Detecting Unseen Privately-Tuned LLM-Generated Text via Family-Aware Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W46SXXQP}},
  note         = {Machine review of arXiv:2506.15683}
}
read the original abstract

With the popularity of large language models (LLMs), undesirable societal problems like misinformation production and academic misconduct have been more severe, making LLM-generated text detection now of unprecedented importance. Although existing methods have made remarkable progress, a new challenge posed by text from privately tuned LLMs remains underexplored. Users could easily possess private LLMs by fine-tuning an open-source one with private corpora, resulting in a significant performance drop of existing detectors in practice. To address this issue, we propose PhantomHunter, an LLM-generated text detector specialized for detecting text from unseen, privately-tuned LLMs. Its family-aware learning framework captures family-level traits shared across the base models and their derivatives, instead of memorizing individual characteristics. Experiments on data from LLaMA, Gemma, and Mistral families show its superiority over 7 baselines and 3 industrial services, with F1 scores of over 96%.

Figures

Figures reproduced from arXiv: 2506.15683 by the authors.

Figure 1
Figure 1. Detection accuracy for LLM-generated text with increasing amounts of fine-tuning data. family-aware learning approach to detect texts from unseen privately-tuned LLMs. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Cosine similarity between the token probability lists of fine-tuned models and those of the base ones. Darker colors indicate higher similarity. where R m Fi /R˜ m Fi are the embeddings of the original/augmented sample, respectively, δ is a dot-product function, and t controls the temperature. We use a multi-layer perceptron (MLP) to classify the LLM family: yˆF = softmax(MLPF(RF)), (3) where yˆF ∈ {θ1, θ2, . . . , … view at source ↗
Figure 3
Figure 3. Overall architecture of PhantomHunter. Given a text sample x, it 1) extracts the probability feature from M base models and encode them with CNN and transformer blocks; 2) predicts the family of x to determine the family gating weights; and 3) feeds the representation RF to a mixture-of-experts network controlled by the gating weights from Step 2 for final prediction of x being LLM-generated. During training, a cont… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 23 canonical work pages

  1. [1]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  2. [2]

    Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2024. https://openreview.net/forum?id=Bpcgcr8E8Z Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature . In The Twelfth International Conference on Learning Representations

    work page 2024

  3. [3]

    Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. https://doi.org/10.5281/zenodo.5297715 GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow

    doi:10.5281/zenodo.5297715 2021

  4. [4]

    Canyu Chen and Kai Shu. 2024. https://doi.org/10.1002/aaai.12188 Combating Misinformation in the Age of LLMs: Opportunities and Challenges . AI Magazine, 45(3):354--368

    doi:10.1002/aaai.12188 2024

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. https://proceedings.mlr.press/v119/chen20j.html A Simple Framework for Contrastive Learning of Visual Representations . In Proceedings of the 37th International Conference on Machine Learning, pages 1597--1607. PMLR

    work page 2020

  6. [6]

    Yutian Chen, Hao Kang, Vivian Zhai, Liangze Li, Rita Singh, and Bhiksha Raj. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.810 Token Prediction as Implicit Classification to Identify LLM-Generated Text . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13112--13120. Association for Computational Linguistics

    doi:10.18653/v1/2023.emnlp-main.810 2023

  7. [7]

    DeepSeek-AI . 2025 a . https://arxiv.org/abs/2501.12948 DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning . Preprint, arXiv:2501.12948

    arXiv 2025

  8. [8]

    DeepSeek-AI . 2025 b . https://arxiv.org/abs/2412.19437 DeepSeek-V3 Technical Report . Preprint, arXiv:2412.19437

    arXiv 2025

Show all 49 references
  1. [9]

    Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. 2019. https://doi.org/10.18653/v1/P19-1346 ELI5: Long Form Question Answering . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3558--3567...

    2019 doi

  2. [10]

    Myles Foley, Ambrish Rawat, Taesung Lee, Yufang Hou, Gabriele Picco, and Giulio Zizzo. 2023. https://doi.org/10.18653/v1/2023.acl-long.410 Matching Pairs: Attributing Fine-Tuned Models to their Pre-Trained Large Language Models . In Proceedings of the 61st Annual Meeting of th...

    2023 doi

  3. [11]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. https://aclanthology.org/2021.emnlp-main.552.pdf SimCSE: Simple Contrastive Learning of Sentence Embeddings . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, page 6894. Association for ...

    2021

  4. [12]

    Gemma Team . 2024. https://arxiv.org/abs/2403.08295 Gemma: Open Models Based on Gemini Research and Technology . Preprint, arXiv:2403.08295

    2024 arXiv

  5. [13]

    Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. https://arxiv.org/abs/2301.07597 How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection . Preprint, arXiv:2301.07597

    2023 arXiv

  6. [14]

    Xun Guo, Shan Zhang, Yongxin He, Ting Zhang, Wanquan Feng, Haibin Huang, and Chongyang Ma. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/a117a3cd54b7affad04618c77c2fb18b-Paper-Conference.pdf DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive ...

    2024

  7. [15]

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. 2024. https://arxiv.org/abs/2401.12070 Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text . Preprint, arXiv:2401.12070

    2024 arXiv

  8. [16]

    Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. 2024. https://doi.org/10.1145/3658644.3670344 MGTBench: Benchmarking Machine-Generated Text Detection . In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 2251-...

    2024

  9. [17]

    Beizhe Hu, Qiang Sheng, Juan Cao, Yuhui Shi, Yang Li, Danding Wang, and Peng Qi. 2024. https://doi.org/10.1609/aaai.v38i20.30214 Bad actor, good advisor: Exploring the role of large language models in fake news detection . In Proceedings of the AAAI Conference on Artificial In...

    2024 doi

  10. [18]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 LoRA: Low-Rank Adaptation of Large Language Models . In The Tenth International Conference on Learning Representations

    2022

  11. [19]

    I know myself better, but not really greatly

    Jiazhou Ji, Jie Guo, Weidong Qiu, Zheng Huang, Yang Xu, Xinru Lu, Xiaoyu Jiang, Ruizhe Li, and Shujun Li. 2025. https://arxiv.org/abs/2502.12743 "I know myself better, but not really greatly": Using LLMs to Detect and Explain LLM-Generated Texts . Preprint, arXiv:2502.12743

    2025 arXiv

  12. [20]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  13. [21]

    Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. https://arxiv.org/abs/2307.10169 Challenges and Applications of Large Language Models . Preprint, arXiv:2307.10169

    2023 arXiv

  14. [22]

    Ryuto Koike, Masahiro Kaneko, and Naoaki Okazaki. 2024. https://doi.org/10.1609/aaai.v38i19.30120 OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with Adversarially Generated Examples . In Proceedings of the AAAI Conference on Artificial Intelligence, volume ...

    2024 doi

  15. [23]

    Tharindu Kumarage, Garima Agrawal, Paras Sheth, Raha Moraffah, Aman Chadha, Joshua Garland, and Huan Liu. 2024. https://arxiv.org/abs/2403.01152 A Survey of AI-generated Text Forensic Systems: Detection, Attribution, and Characterization . Preprint, arXiv:2403.01152

    2024 arXiv

  16. [24]

    Tharindu Kumarage and Huan Liu. 2023. https://doi.org/10.1109/CyberC58899.2023.00019 Neural Authorship Attribution: Stylometric Analysis on Large Language Models . In 2023 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery, pages 51--54. IE...

    2023

  17. [25]

    Aiwei Liu, Leyi Pan, Yijian Lu, Jingjing Li, Xuming Hu, Xi Zhang, Lijie Wen, Irwin King, Hui Xiong, and Philip Yu. 2024. https://doi.org/10.1145/3691626 A Survey of Text Watermarking in the Era of Large Language Models . ACM Computing Surveys, 57(2):1--36

    2024 doi

  18. [26]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 RoBERTa: A Robustly Optimized BERT Pretraining Approach . Preprint, arXiv:1907.11692

    2019 arXiv

  19. [27]

    Macedo Maia, Siegfried Handschuh, Andr\' e Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. 2018. https://doi.org/10.1145/3184558.3192301 WWW'18 Open Challenge: Financial Opinion Mining and Question Answering . In Companion Proceedings of the The Web...

    2018

  20. [28]

    Meta AI . 2024. Llama Usage Doubled May Through July 2024 . https://ai.meta.com/blog/llama-usage-doubled-may-through-july-2024/. Accessed: 2025-03-21

    2024

  21. [29]

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. https://proceedings.mlr.press/v202/mitchell23a.html DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature . In Proceedings of the 40th International Confe...

    2023

  22. [30]

    Yancey, Ruidong Liu, Mirza Basim Baig, Andr \'e Kenji Horie, and James Sharpnack

    Chenhao Niu, Kevin P. Yancey, Ruidong Liu, Mirza Basim Baig, Andr \'e Kenji Horie, and James Sharpnack. 2024. https://doi.org/10.18653/v1/2024.emnlp-industry.70 Detecting LLM -Assisted Cheating on Open-Ended Writing Tasks on Language Proficiency Tests . In Proceedings of the 2...

    2024 doi

  23. [31]

    Qi Pang, Shengyuan Hu, Wenting Zheng, and Virginia Smith. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/fa86a9c7b9f341716ccb679d1aeb9afa-Paper-Conference.pdf No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices . In Advances in Neural...

    2024

  24. [32]

    Giovanni Puccetti, Anna Rogers, Chiara Alzetta, Felice Dell’Orletta, and Andrea Esuli. 2024. https://doi.org/10.18653/v1/2024.acl-long.817 AI ‘News’ Content Farms Are Easy to Make and Hard to Detect: A Case Study in Italian . In Proceedings of the 62nd Annual Meeting of the As...

    2024 doi

  25. [33]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. http://jmlr.org/papers/v21/20-074.html Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer . Journal of Machine Lear...

    2020

  26. [34]

    Areg Mikael Sarvazyan, Jos \'e \'A ngel Gonz \'a lez, Paolo Rosso, and Marc Franco-Salvador. 2023. https://doi.org/10.1007/978-3-031-42448-9_11 Supervised Machine-Generated Text Detectors: Family and Scale Matters . In International Conference of the Cross-Language Evaluation ...

    2023 doi

  27. [35]

    Mrinank Sharma, Meg Tong, Jesse Mu, Jerry Wei, Jorrit Kruthoff, Scott Goodfriend, Euan Ong, Alwin Peng, Raj Agarwal, Cem Anil, Amanda Askell, Nathan Bailey, Joe Benton, Emma Bluemke, Samuel R. Bowman, Eric Christiansen, Hoagy Cunningham, Andy Dau, Anjali Gopal, Rob Gilson, Log...

    2025 arXiv

  28. [36]

    Yuhui Shi, Qiang Sheng, Juan Cao, Hao Mi, Beizhe Hu, and Danding Wang. 2024. https://doi.org/10.24963/ijcai.2024/55 Ten Words Only Still Help: Improving Black-Box AI-Generated Text Detection via Proxy-Guided Efficient Re-Sampling . In Proceedings of the Thirty-Third Internatio...

    2024 doi

  29. [37]

    Shantanu Thorat and Tianbao Yang. 2024. https://arxiv.org/abs/2410.14875 Which LLMs are Difficult to Detect? A Detailed Analysis of Potential Factors Contributing to Difficulties in LLM Text Detection . Preprint, arXiv:2410.14875

    2024 arXiv

  30. [38]

    Yuchuan Tian, Hanting Chen, Xutao Wang, Zheyuan Bai, Qinghua Zhang, Ruifeng Li, Chao Xu, and Yunhe Wang. 2024. https://openreview.net/forum?id=5Lp6qU9hzV Multiscale Positive-Unlabeled Detection of AI-Generated Texts . In The Twelfth International Conference on Learning Representations

    2024

  31. [39]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, W...

    2023 arXiv

  32. [40]

    Brian Tufts, Xuandong Zhao, and Lei Li. 2025. https://arxiv.org/abs/2412.05139 A Practical Examination of AI-Generated Text Detectors for Large Language Models . Preprint, arXiv:2412.05139

    2025 arXiv

  33. [41]

    Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. 2023. https://aclanthology.org/2023.emnlp-main.73/ SeqXGPT: Sentence-Level AI-Generated Text Detection . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages ...

    2023

  34. [42]

    Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mohanned Afzal, Tarek Mahmoud, Giovanni Puccetti, Thomas Arnold, et al. 2024. https://doi.org/10.18653/v1/2024.acl-long.218 M4GT-Bench: Evaluation Benchmark for Black-Box Machine-Genera...

    2024 doi

  35. [43]

    Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S

    Junchao Wu, Runzhe Zhan, Derek F. Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S. Chao. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/b61bdf7e9f64c04ec75a26e781e2ad51-Paper-Datasets_and_Benchmarks_Track.pdf DetectRL: Benchmarking LLM-Generated Text Dete...

    2024

  36. [44]

    Qilong Wu and Varun Chandrasekaran. 2024. https://doi.org/10.18653/v1/2024.acl-long.464 Bypassing LLM Watermarks with Color-Aware Substitutions . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8549--858...

    2024 doi

  37. [45]

    Xianjun Yang, Wei Cheng, Yue Wu, Linda Ruth Petzold, William Yang Wang, and Haifeng Chen. 2024. https://openreview.net/forum?id=Xlayxj2fWp DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text . In The Twelfth International Conference on Learning...

    2024

  38. [46]

    Xiao Yu, Kejiang Chen, Qi Yang, Weiming Zhang, and Nenghai Yu. 2024 a . https://doi.org/10.18653/v1/2024.emnlp-main.885 Text Fluoroscopy: Detecting LLM-Generated Text through Intrinsic Features . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pr...

    2024 doi

  39. [47]

    Xiao Yu, Yuang Qi, Kejiang Chen, Guoqiang Chen, Xi Yang, Pengyuan Zhu, Xiuwei Shang, Weiming Zhang, and Nenghai Yu. 2024 b . https://openreview.net/forum?id=BZh05P2EoN DPIC: Decoupling Prompt and Intrinsic Characteristics for LLM Generated Text Detection . In The Thirty-eighth...

    2024

  40. [48]

    Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, et al. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.743 MedDialog: Large-scale Medical Dialogue Datasets . In Proceedings of the 2020 Conference ...

    2020 doi

  41. [49]

    Biru Zhu, Lifan Yuan, Ganqu Cui, Yangyi Chen, Chong Fu, Bingxiang He, Yangdong Deng, Zhiyuan Liu, Maosong Sun, and Ming Gu. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.463 Beat LLMs at Their Own Game: Zero-Shot LLM-Generated Text Detection via Querying ChatGPT . In Proce...

    2023 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.