pith. sign in

arxiv: 2605.16107 · v1 · pith:6C7RYJ3Wnew · submitted 2026-05-15 · 💻 cs.CL

Multi-Level Contextual Token Relation Modeling for Machine-Generated Text Detection

Chenwang Wu , Yiuming Cheung , Bo Han , Shuhai Zhang , Defu Lian This is my paper

Pith reviewed 2026-05-20 19:03 UTC · model grok-4.3

classification 💻 cs.CL
keywords machine-generated text detectionmetric-based methodstoken relation modelingmulti-level inferenceMarkov calibrationrule-support reasoningcross-domain detection
0
0 comments X

The pith

Modeling local and global relations among token detection scores corrects randomness bias in machine-generated text detection.

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

The paper unifies various metric-based approaches to detecting machine-generated text under one framework. It identifies that single-token scores are distorted by the random nature of text generation. To address this, the authors derive how scores transition over multiple hops and then model both nearby and distant relations between these scores. They introduce a calibration step using Markov models for local context and logical rules for global context, combining them in a joint inference system. This leads to better detection performance in diverse real-world conditions without much added computation.

Core claim

By deriving the multi-hop transitions of token-level detection scores and exploring their local and global relations, a multi-level contextual token relation modeling framework is proposed, consisting of a lightweight Markov-informed calibration module for local relations and a rule-support reasoning module for global relations, which are combined in a joint multi-level inference framework for improved MGT detection.

What carries the argument

The multi-level contextual token relation modeling framework that refines token-level evidence using local Markov calibration and global rule-support reasoning before aggregation.

If this is right

  • Improved detection accuracy holds in cross-LLM and cross-domain scenarios.
  • The approach maintains low computational overhead relative to model-based alternatives.
  • Joint local calibration and global rule reasoning yields more stable detection signals.
  • The unified framework reveals common limitations across existing metric-based detectors.

Where Pith is reading between the lines

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

  • The same relation-modeling idea could extend to spotting generated content in other modalities such as code or images.
  • Hybrid systems that layer this calibration on top of current detectors might achieve both efficiency and robustness gains.
  • Further tests on entirely unseen model families would strengthen the cross-LLM generalization claim.

Load-bearing premise

The token-level detection score is easily biased by the inherent randomness of the MGT generation process, and that theoretically derived multi-hop transitions plus local/global relation modeling can reliably correct this bias before aggregation.

What would settle it

A controlled experiment on generated texts where disabling the multi-level relation modules produces no gain over simple token-score aggregation under varying randomness levels.

Figures

Figures reproduced from arXiv: 2605.16107 by Bo Han, Chenwang Wu, Defu Lian, Shuhai Zhang, Yiuming Cheung.

Figure 1
Figure 1. Figure 1: The detection score distances (Mean Absolute Difference) of neighbors [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The detection score distances (Mean Absolute Difference) of 1-hop [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the global statistics for detection scores of MGT and HGT. Here, DetectLLM scores are used and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The workflow of the Multi-level contextual token relation modeling framework. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The performance improvement compared with original detector (left) and our preliminary work. Here the base detector is DetectLLM. Values greater [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Detection performance concerning AUROC under different mixed texts. All detectors are trained on pure Llama-2-70b texts. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Detection performance under Dipper and Polish paraphrasing texts. All detectors are trained on Llama-2-70b texts. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Detection performance concerning AUROC under adversarial texts concerning character perturbation. All detectors are trained on Llama-2-70b texts. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The empirical MGT probability associated with different threshold atoms induced by the global statistics. DetectLLM scores are used here. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance comparison with probabilistic rule-based method on the Essay dataset. The reported results are the average performance across all [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Ablation results concerning AUROC on the Essay dataset. The reported results are the average performance across all LLM-generated texts. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Detection performance at different numbers of rules on the Essay dataset. All detectors are trained on ChatGPT texts. [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Detection performance at different initial part lengths on the Essay dataset. All detectors are trained on ChatGPT texts. [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
read the original abstract

Machine-generated texts (MGTs) pose risks such as disinformation and phishing, underscoring the need for reliable detection. Metric-based methods, which extract statistically distinguishable features of MGTs, are often more practical than complex model-based methods that are prone to overfitting. Given their diverse designs, we first place representative metric-based methods within a unified framework, enabling a clear assessment of their advantages and limitations. Our analysis identifies a core challenge across these methods: the token-level detection score is easily biased by the inherent randomness of the MGTs generation process. Then, we theoretically derive the multi-hop transitions of the token-level detection score and explore their local and global relations. Based on these findings, we propose a multi-level contextual token relation modeling framework for MGT detection. Specifically, for local relations, we model them through a lightweight Markov-informed calibration module that refines token-level evidence before aggregation. For global relations, we introduce a rule-support reasoning module that uses explicit logical rules derived from contextual score statistics. Finally, we combine the local calibrated score and the global rule-support reasoning signal in a joint multi-level inference framework. Extensive experiments show broad and substantial improvements across various real-world scenarios, including cross-LLM and cross-domain settings, with low computational overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript unifies representative metric-based methods for machine-generated text (MGT) detection under a single framework, identifies the core issue that token-level detection scores are biased by randomness in the generation process, theoretically derives multi-hop transitions of these scores along with their local and global relations, and proposes a multi-level contextual token relation modeling framework. The framework uses a lightweight Markov-informed calibration module to refine local token-level evidence before aggregation and a rule-support reasoning module that applies explicit logical rules from contextual score statistics; these signals are combined in a joint multi-level inference step. Extensive experiments are reported to demonstrate broad and substantial gains across real-world scenarios including cross-LLM and cross-domain settings, with low computational overhead.

Significance. If the theoretical derivation of the multi-hop transitions is sound and the local/global modules reliably correct the identified bias, the work would provide a practical, lightweight advance over existing metric-based detectors. It avoids the overfitting risks of model-based approaches while delivering measurable improvements in challenging transfer settings, which could be valuable for deployment in disinformation and content-moderation pipelines.

major comments (2)
  1. [§4] §4 (Theoretical Derivation): The multi-hop transition derivation and subsequent Markov-informed calibration module implicitly rely on stationary token-score statistics. Under modern non-stationary sampling regimes (temperature, top-p/nucleus sampling) the conditional distributions evolve within a sequence, which can break the Markov assumption used to characterize how generation randomness propagates to token scores. The paper must demonstrate, via analysis or targeted ablation, that the derived transitions remain valid or can be adapted in these regimes; otherwise the local calibration cannot be guaranteed to remove bias before the global rule-support signal is applied.
  2. [§6] §6 (Experiments): The claim of broad improvements across cross-LLM and cross-domain settings is central, yet the reported results do not isolate performance under varying generation hyperparameters (different temperatures or top-p values). Without such controls it is unclear whether the observed gains stem from the multi-level modeling correcting the randomness bias or from other factors; adding these ablations would directly test the load-bearing assumption identified in the weakest point of the argument.
minor comments (2)
  1. [Abstract / §3] The abstract and §3 would benefit from a short explicit statement of the precise baselines and metrics (e.g., AUROC, F1) used to quantify the “substantial improvements,” to allow readers to gauge effect sizes immediately.
  2. [§2 / §4] Notation for token-level scores, transition matrices, and rule-support signals should be introduced once in §2 or §4 and used consistently thereafter to avoid minor ambiguity when the local and global modules are combined.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments, which help clarify the assumptions underlying our theoretical derivation and strengthen the experimental validation of our claims. We address each major comment below and outline the specific revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [§4] §4 (Theoretical Derivation): The multi-hop transition derivation and subsequent Markov-informed calibration module implicitly rely on stationary token-score statistics. Under modern non-stationary sampling regimes (temperature, top-p/nucleus sampling) the conditional distributions evolve within a sequence, which can break the Markov assumption used to characterize how generation randomness propagates to token scores. The paper must demonstrate, via analysis or targeted ablation, that the derived transitions remain valid or can be adapted in these regimes; otherwise the local calibration cannot be guaranteed to remove bias before the global rule-support signal is applied.

    Authors: We appreciate the referee's identification of the stationarity assumption in the multi-hop transition derivation. The Markov process is used to model local score transitions for tractable calibration of randomness bias, but we acknowledge that temperature and nucleus sampling can introduce non-stationarity in conditional distributions. Our framework's lightweight calibration module is empirically robust across the generation settings tested, yet to rigorously address this, we will add a new analysis subsection in §4 discussing adaptation of the transitions under non-stationary regimes (e.g., via time-varying transition matrices) and include targeted ablations measuring bias reduction as sampling parameters vary. These changes will clarify the scope of the local calibration's guarantees. revision: yes

  2. Referee: [§6] §6 (Experiments): The claim of broad improvements across cross-LLM and cross-domain settings is central, yet the reported results do not isolate performance under varying generation hyperparameters (different temperatures or top-p values). Without such controls it is unclear whether the observed gains stem from the multi-level modeling correcting the randomness bias or from other factors; adding these ablations would directly test the load-bearing assumption identified in the weakest point of the argument.

    Authors: We agree that explicit controls for generation hyperparameters are necessary to isolate the contribution of the multi-level contextual modeling and confirm that gains arise from bias correction rather than other factors. The current experiments use representative generation settings for cross-LLM and cross-domain transfer, but we will expand §6 with new ablation studies that vary temperature (e.g., 0.5, 0.8, 1.0) and top-p (e.g., 0.85, 0.95, 1.0) while reporting detection performance for our framework versus baselines. This will directly test the load-bearing assumption and demonstrate consistent improvements attributable to the local and global relation modules. revision: yes

Circularity Check

0 steps flagged

No circularity: theoretical derivation and framework remain independent of fitted outputs

full rationale

The paper first unifies existing metric-based detectors, identifies the token-level bias challenge from generation randomness, then states that it theoretically derives multi-hop transitions of the detection score before introducing local Markov calibration and global rule-support modules. No equations, self-citations, or fitted parameters are shown that would make the multi-hop derivation reduce to a re-expression of the same data or prior author results by construction. The claimed improvements rest on the subsequent empirical evaluation rather than any self-definitional or load-bearing self-citation step.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so specific free parameters, axioms, and invented entities cannot be extracted. The central approach rests on the domain assumption that token-level scores carry recoverable local and global relational structure that can be modeled separately from the generation randomness.

pith-pipeline@v0.9.0 · 5761 in / 1066 out tokens · 35698 ms · 2026-05-20T19:03:35.482481+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    theoretically derive the multi-hop transitions of the token-level detection score and explore their local and global relations... lightweight Markov-informed calibration module... rule-support reasoning module that uses explicit logical rules derived from contextual score statistics

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages · 3 internal anchors

  1. [1]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskeveret al., “Language models are unsupervised multitask learners,”OpenAI blog, vol. 1, no. 8, p. 9, 2019

    work page 2019

  3. [3]

    Disinformation capabilities of large language models,

    I. Vykopal, M. Pikuliak, I. Srba, R. Moro, D. Macko, and M. Bielikov ´a, “Disinformation capabilities of large language models,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2024, pp. 14 830–14 847

    work page 2024

  4. [4]

    Codeipprompt: intellectual property infringement assessment of code language models,

    Z. Yu, Y . Wu, N. Zhang, C. Wang, Y . V orobeychik, and C. Xiao, “Codeipprompt: intellectual property infringement assessment of code language models,” inProceedings of the International Conference on Machine Learning. PMLR, 2023, pp. 40 373–40 389

    work page 2023

  5. [5]

    The state of phishing attacks,

    J. Hong, “The state of phishing attacks,”Communications of the ACM, vol. 55, no. 1, pp. 74–81, 2012. MULTI-LEVEL CONTEXTUAL TOKEN RELATION MODELING FOR MACHINE-GENERATED TEXT DETECTION 15

    work page 2012

  6. [6]

    Semstamp: A se- mantic watermark with paraphrastic robustness for text generation,

    A. Hou, J. Zhang, T. He, Y . Wang, Y .-S. Chuang, H. Wang, L. Shen, B. Van Durme, D. Khashabi, and Y . Tsvetkov, “Semstamp: A se- mantic watermark with paraphrastic robustness for text generation,” in Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume 1: L...

    work page 2024

  7. [7]

    Release Strategies and the Social Impacts of Language Models

    I. Solaiman, M. Brundage, J. Clark, A. Askell, A. Herbert-V oss, J. Wu, A. Radford, G. Krueger, J. W. Kim, S. Krepset al., “Release strategies and the social impacts of language models,”arXiv preprint arXiv:1908.09203, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1908

  8. [8]

    How close is chatgpt to human experts? comparison corpus, evaluation, and detection,

    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

    work page arXiv 2023

  9. [9]

    Seqxgpt: Sentence-level ai-generated text detection,

    P. Wang, L. Li, K. Ren, B. Jiang, D. Zhang, and X. Qiu, “Seqxgpt: Sentence-level ai-generated text detection,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 1144–1156

    work page 2023

  10. [10]

    Coco: Coherence-enhanced machine-generated text detection under data limita- tion with contrastive learning,

    X. Liu, Z. Zhang, Y . Wang, H. Pu, Y . Lan, and C. Shen, “Coco: Coherence-enhanced machine-generated text detection under data limita- tion with contrastive learning,”arXiv preprint arXiv:2212.10341, 2022

    work page arXiv 2022

  11. [11]

    Detectgpt: Zero-shot machine-generated text detection using probability curvature,

    E. Mitchell, Y . Lee, A. Khazatsky, C. D. Manning, and C. Finn, “Detectgpt: Zero-shot machine-generated text detection using probability curvature,” inProceedings of the International Conference on Machine Learning. PMLR, 2023, pp. 24 950–24 962

    work page 2023

  12. [12]

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

    G. Bao, Y . Zhao, Z. Teng, L. Yang, and Y . Zhang, “Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature,” inProceedings of the Twelfth International Con- ference on Learning Representations, 2024, pp. 1–9

    work page 2024

  13. [13]

    Spotting llms with binoc- ulars: Zero-shot detection of machine-generated text,

    A. Hans, A. Schwarzschild, V . Cherepanova, H. Kazemi, A. Saha, M. Goldblum, J. Geiping, and T. Goldstein, “Spotting llms with binoc- ulars: Zero-shot detection of machine-generated text,” inProceedings of the International Conference on Machine Learning. PMLR, 2024, pp. 17 519–17 537

    work page 2024

  14. [14]

    GLTR: Statistical Detection and Visualization of Generated Text

    S. Gehrmann, H. Strobelt, and A. M. Rush, “Gltr: Statistical detection and visualization of generated text,”arXiv preprint arXiv:1906.04043, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  15. [15]

    Biscope: Ai-generated text detection by checking memorization of preceding tokens,

    H. Guo, S. Cheng, X. Jin, Z. Zhang, K. Zhang, G. Tao, G. Shen, and X. Zhang, “Biscope: Ai-generated text detection by checking memorization of preceding tokens,”Proceedings of the Advances in Neural Information Processing Systems, vol. 37, pp. 104 065–104 090, 2024

    work page 2024

  16. [16]

    Dna- detectllm: Unveiling ai-generated text via a dna-inspired mutation-repair paradigm,

    X. Zhu, Y . Ren, F. Fang, Q. Tan, S. Wang, and Y . Cao, “Dna- detectllm: Unveiling ai-generated text via a dna-inspired mutation-repair paradigm,” inProceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025, pp. 1–13

    work page 2025

  17. [17]

    Beyond raw detection scores: Markov-informed calibration for boosting machine- generated text detection,

    C. Wu, Y .-m. Cheung, S. Zhang, B. Han, and D. Lian, “Beyond raw detection scores: Markov-informed calibration for boosting machine- generated text detection,” inProceedings of the F ourteenth International Conference on Learning Representations, 2026, pp. 1–10

    work page 2026

  18. [18]

    A watermark for large language models,

    J. Kirchenbauer, J. Geiping, Y . Wen, J. Katz, I. Miers, and T. Gold- stein, “A watermark for large language models,” inProceedings of the International Conference on Machine Learning. PMLR, 2023, pp. 17 061–17 084

    work page 2023

  19. [19]

    A resilient and accessible distribution-preserving watermark for large language models,

    Y . Wu, Z. Hu, J. Guo, H. Zhang, and H. Huang, “A resilient and accessible distribution-preserving watermark for large language models,” inProceedings of the International Conference on Machine Learning. PMLR, 2024, pp. 53 443–53 470

    work page 2024

  20. [20]

    {REMARK- LLM}: A robust and efficient watermarking framework for generative large language models,

    R. Zhang, S. S. Hussain, P. Neekhara, and F. Koushanfar, “{REMARK- LLM}: A robust and efficient watermarking framework for generative large language models,” inProceedings of the 33rd USENIX Security Symposium (USENIX Security 24), 2024, pp. 1813–1830

    work page 2024

  21. [21]

    Adaptive text watermark for large language mod- els,

    Y . Liu and Y . Bu, “Adaptive text watermark for large language mod- els,” inProceedings of the 41st International Conference on Machine Learning, 2024, pp. 30 718–30 737

    work page 2024

  22. [22]

    Gptzero official website,

    GPTZero, “Gptzero official website,” [Online], 2023, https://gptzero.me

    work page 2023

  23. [23]

    G3detector: General gpt-generated text detector,

    H. Zhan, X. He, Q. Xu, Y . Wu, and P. Stenetorp, “G3detector: General gpt-generated text detector,”arXiv preprint arXiv:2305.12680, 2023

    work page arXiv 2023

  24. [24]

    Llm paternity test: Generated text detection with llm genetic inheritance,

    X. Yu, Y . Qi, K. Chen, G. Chen, X. Yang, P. Zhu, W. Zhang, and N. Yu, “Llm paternity test: Generated text detection with llm genetic inheritance,”arXiv preprint arXiv:2305.12519, 2023

    work page arXiv 2023

  25. [25]

    Llmdet: A third party large language models generated text detection tool,

    K. Wu, L. Pang, H. Shen, X. Cheng, and T.-S. Chua, “Llmdet: A third party large language models generated text detection tool,” in Proceedings of the Findings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 2113–2133

    work page 2023

  26. [26]

    Multiscale positive-unlabeled detection of ai-generated texts,

    Y . Tian, H. Chen, X. Wang, Z. Bai, Q. ZHANG, R. Li, C. Xu, and Y . Wang, “Multiscale positive-unlabeled detection of ai-generated texts,” inProceedings of the Twelfth International Conference on Learning Representations, 2024, pp. 1–9

    work page 2024

  27. [27]

    Radar: Robust ai-text detection via adversarial learning,

    X. Hu, P.-Y . Chen, and T.-Y . Ho, “Radar: Robust ai-text detection via adversarial learning,”Advances in neural information processing systems, vol. 36, pp. 15 077–15 095, 2023

    work page 2023

  28. [28]

    Ghostbuster: Detecting text ghostwritten by large language models,

    V . Verma, E. Fleisig, N. Tomlin, and D. 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 (V olume 1: Long Papers), 2024, pp. 1702–1717

    work page 2024

  29. [29]

    Detecting and grounding multi-modal media manipulation and beyond,

    R. Shao, T. Wu, J. Wu, L. Nie, and Z. Liu, “Detecting and grounding multi-modal media manipulation and beyond,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5556– 5574, 2024

    work page 2024

  30. [30]

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

    J. Su, T. Zhuo, D. Wang, and P. Nakov, “Detectllm: Leveraging log rank information for zero-shot detection of machine-generated text,” inProceedings of the Findings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 12 395–12 412

    work page 2023

  31. [31]

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

    H. Zhou, J. Zhu, P. Su, K. Ye, Y . Yang, S. Gavioli-Akilagun, and C. Shi, “Adadetectgpt: Adaptive detection of llm-generated text with statistical guarantees,” inProceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025, pp. 1–10

    work page 2025

  32. [32]

    Dna-gpt: Divergent n-gram analysis for training-free detection of gpt- generated text,

    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,” inProceedings of the Twelfth International Conference on Learning Representations, 2024, pp. 1–9

    work page 2024

  33. [33]

    Zero-shot detection of machine-generated codes,

    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

    work page arXiv 2023

  34. [34]

    Simllm: Detecting sentences generated by large language models using similarity between the generation and its re-generation,

    H.-Q. Nguyen-Son, M.-S. Dao, and K. Zettsu, “Simllm: Detecting sentences generated by large language models using similarity between the generation and its re-generation,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 22 340–22 352

    work page 2024

  35. [35]

    Learn-to- distance: Distance learning for detecting llm-generated text,

    H. Zhou, J. Zhu, K. Ye, Y . Yang, E. Xu, and C. Shi, “Learn-to- distance: Distance learning for detecting llm-generated text,”arXiv preprint arXiv:2601.21895, 2026

    work page arXiv 2026

  36. [36]

    In- trinsic dimension estimation for robust detection of ai-generated texts,

    E. Tulchinskii, K. Kuznetsov, L. Kushnareva, D. Cherniavskii, S. Nikolenko, E. Burnaev, S. Barannikov, and I. Piontkovskaya, “In- trinsic dimension estimation for robust detection of ai-generated texts,” Proceedings of the Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  37. [37]

    Zero-shot detection of llm-generated text using to- ken cohesiveness,

    S. Ma and Q. Wang, “Zero-shot detection of llm-generated text using to- ken cohesiveness,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 17 538–17 553

    work page 2024

  38. [38]

    Text fluoroscopy: Detecting llm-generated text through intrinsic features,

    X. Yu, K. Chen, Q. Yang, W. Zhang, and N. Yu, “Text fluoroscopy: Detecting llm-generated text through intrinsic features,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 15 838–15 846

    work page 2024

  39. [39]

    Repreguard: Detecting llm-generated text by revealing hidden representation patterns,

    X. Chen, J. Wu, S. Yang, R. Zhan, Z. Wu, Z. Luo, D. Wang, M. Yang, L. S. Chao, and D. F. Wong, “Repreguard: Detecting llm-generated text by revealing hidden representation patterns,”arXiv preprint arXiv:2508.13152, 2025

    work page arXiv 2025

  40. [40]

    Training-free llm-generated text detection by mining token probability sequences,

    Y . Xu, Y . Wang, Y . Bi, H. Cao, Z. Lin, Y . Zhao, and F. Wu, “Training-free llm-generated text detection by mining token probability sequences,” inProceedings of the Thirteenth International Conference on Learning Representations, 2025, pp. 1–10

    work page 2025

  41. [41]

    Detecting subtle differences between human and model languages using spectrum of relative likeli- hood,

    Y . Xu, Y . Wang, H. An, Z. Liu, and Y . Li, “Detecting subtle differences between human and model languages using spectrum of relative likeli- hood,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 10 108–10 121

    work page 2024

  42. [42]

    Moses: Uncertainty-aware ai-generated text detection via mixture of stylistics experts with conditional thresholds,

    J. Wu, J. Wang, Z. Liu, B. Chen, D. Hu, H. Wu, and S.-T. Xia, “Moses: Uncertainty-aware ai-generated text detection via mixture of stylistics experts with conditional thresholds,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025, pp. 5797–5816

    work page 2025

  43. [43]

    Detree: Detecting human-ai collaborative texts via tree-structured hierarchical representation learn- ing,

    Y . He, S. Zhang, Y . Cao, L. Ma, and P. Luo, “Detree: Detecting human-ai collaborative texts via tree-structured hierarchical representation learn- ing,” inProceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025, pp. 1–10

    work page 2025

  44. [44]

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

    C. Zeng, S. Tang, Y . Chen, Z. Shen, W. Yu, X. Zhao, H. Chen, W. Cheng et al., “Human texts are outliers: Detecting llm-generated texts via out- of-distribution detection,” inProceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025, pp. 1–10

    work page 2025

  45. [45]

    Ipad: Inverse prompt for ai detection–a robust and explainable llm-generated text detector,

    Z. Chen, Y . Feng, C. He, Y . Deng, H. Pu, and B. Li, “Ipad: Inverse prompt for ai detection–a robust and explainable llm-generated text detector,”arXiv e-prints, pp. arXiv–2502, 2025. MULTI-LEVEL CONTEXTUAL TOKEN RELATION MODELING FOR MACHINE-GENERATED TEXT DETECTION 16

    work page 2025

  46. [46]

    Hld: Approx- imate hierarchical linguistic distribution modeling for llm-generated text detection,

    R. Guo, W. Zeng, F. Wu, Y . Kong, Y . Wu, W. Donget al., “Hld: Approx- imate hierarchical linguistic distribution modeling for llm-generated text detection,” inProceedings of the F ourteenth International Conference on Learning Representations, 2026, pp. 1–10

    work page 2026

  47. [47]

    Scissorhands: Exploiting the persistence of importance hypothesis for llm kv cache compression at test time,

    Z. Liu, A. Desai, F. Liao, W. Wang, V . Xie, Z. Xu, A. Kyrillidis, and A. Shrivastava, “Scissorhands: Exploiting the persistence of importance hypothesis for llm kv cache compression at test time,”Advances in Neural Information Processing Systems, vol. 36, pp. 52 342–52 364, 2023

    work page 2023

  48. [48]

    Bayesian image super-resolution with deep modeling of image statistics,

    S. Gao and X. Zhuang, “Bayesian image super-resolution with deep modeling of image statistics,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 2, pp. 1405–1423, 2022

    work page 2022

  49. [49]

    Deeplogic: Joint learning of neural perception and logical reasoning,

    X. Duan, X. Wang, P. Zhao, G. Shen, and W. Zhu, “Deeplogic: Joint learning of neural perception and logical reasoning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4321– 4334, 2022

    work page 2022

  50. [50]

    Truthfulqa: Measuring how models mimic human falsehoods,

    S. Lin, J. Hilton, and O. Evans, “Truthfulqa: Measuring how models mimic human falsehoods,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2022, pp. 3214–3252

    work page 2022

  51. [51]

    Detectrl: Benchmarking llm-generated text detection in real-world sce- narios,

    J. Wu, R. Zhan, D. Wong, S. Yang, X. Yang, Y . Yuan, and L. Chao, “Detectrl: Benchmarking llm-generated text detection in real-world sce- narios,”Proceedings of the Advances in Neural Information Processing Systems, vol. 37, pp. 100 369–100 401, 2024

    work page 2024

  52. [52]

    Para- phrasing evades detectors of ai-generated text, but retrieval is an effective defense,

    K. Krishna, Y . Song, M. Karpinska, J. Wieting, and M. Iyyer, “Para- phrasing evades detectors of ai-generated text, but retrieval is an effective defense,”Proceedings of Advances in Neural Information Processing Systems, vol. 36, pp. 27 469–27 500, 2023

    work page 2023

  53. [53]

    Can AI-Generated Text be Reliably Detected?

    V . S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi, “Can ai-generated text be reliably detected?”arXiv preprint arXiv:2303.11156, 2023

    work page Pith review arXiv 2023

  54. [54]

    Slack federated ad- versarial training,

    J. Zhu, B. Han, J. Yao, Q. Yao, T. Liu, and J. Xu, “Slack federated ad- versarial training,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025