Pith. sign in

REVIEW 3 major objections 3 minor 82 references

An essay's reasoning structure, not its wording, is the durable fingerprint of its LLM author: a graph neural network over argument-relation graphs beats a text-only transformer by up to 27 F1 points under obfuscation and 19 on unseen model

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 ↗

Reasoning graphs extracted by an argument-mining pipeline let a GNN attribute LLM-generated essays to the correct model family with more robustness to paraphrasing/backtranslation and better cross-version generalization than a Longformer text baseline.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Reasoning graphs for authorship attribution is a fresh idea with a useful corpus, but the central claim is unproven: the graphs are near-complete and the paper omits the trivial-edge ablation that would isolate the contribution of reasoning structure. the 3 major comments →

arxiv 2607.14905 v2 pith:5KLEB5UX submitted 2026-07-16 cs.CL cs.AI

Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution

Zlata Kikteva , Artur Romazanov , Annette Hautli-Janisz , Ramon Ruiz-Dolz This is my paper
classification cs.CL cs.AI
keywords LLM authorship attributionreasoning graphsargument mininggraph neural networksobfuscation robustnesscross-version generalizationLLM-generated text detectionargumentative essays
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 reading

The paper attempts to establish that how an LLM structures its reasoning — which claims support, attack, or restate one another — is a more stable signature of authorship than the words it chooses. It converts essays into reasoning graphs using an argument-mining pipeline and trains a graph neural network to assign each essay to one of four model families (Gemma, Qwen, Llama, Phi). In the reported experiments, the graph approach keeps working under paraphrasing and backtranslation, where a text-only Longformer baseline loses up to 52 F1 points, and it transfers better to unseen versions of the same families, gaining up to 27 and 19 points respectively. A reader should care because authorship attribution built on surface style fails against cheap rewriting and against the steady release of new model versions, whereas a structural fingerprint promises to survive both.

Core claim

The central claim is that reasoning graphs carry an authorship signal that surface text lacks. For each essay the authors build a directed graph whose nodes are argumentative discourse units embedded as sentence vectors, and whose edges carry predicted probabilities of inference, conflict, rephrase, and no relation, produced by the open argument-mining pipeline oAMF and pruned by confidence thresholds. A graph neural network — most effectively a one-layer graph transformer — classifies these graphs into four model families. On a new 8,960-essay corpus spanning eight model versions, the approach beats a Longformer baseline by up to 27 F1 points under paraphrasing and backtranslation, and by u

What carries the argument

The reasoning graph: nodes are argumentative discourse units (text segments roughly the size of sentences) embedded with a sentence transformer, and directed edges join earlier to later segments carrying predicted probabilities for the relations inference, conflict, rephrase, and no relation, with low-certainty edges filtered by thresholds (sets T1, T2, T3). A shallow graph neural network — typically one layer of graph transformer or attention with global attention pooling — aggregates the structure into a graph-level embedding classified into the four model families. The machinery works by replacing lexico-syntactic surface features with a pragmatic skeleton, the argumentative relations amo

Load-bearing premise

The whole approach rests on the unverified premise that the argument-mining pipeline extracts argument components and relation labels accurately and stably enough for genuine cross-family differences to appear in the graphs — the paper itself concedes in its Limitations section that argument component detection and relation identification remain open problems, and it does not check the extracted graphs against gold argument structures.

What would settle it

Two concrete checks would settle the claim. First, rebuild the same graphs but replace the predicted relation probabilities with random or constant edge labels, keeping node embeddings and topology, and retrain: if the F1 advantage over the text-only baseline persists, the argument relations are not the carrier of the signal. Second, compare the pipeline's edge predictions on a sample of this corpus against manual gold argument-relation annotation: if agreement is near chance, or if the advantage disappears on another text genre, the 'reasoning fingerprint' is an artifact of pipeline bias rath

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

If this is right

  • If the results hold, authorship attribution can be made to survive paraphrase and backtranslation without adversarial retraining: up to +27 F1 over the text-only baseline in the same-version setting (Table 2).
  • Cross-version generalization (up to +19 F1, Table 3) means a system trained on today's models can keep working on tomorrow's releases of the same families, addressing a known failure mode of transformer detectors.
  • The consistent superiority of shallow (1-layer) GNNs on these dense graphs supplies a concrete design rule for graph-based text attribution.
  • The released LLM-OWL-AE corpus gives the community a benchmark that jointly tests robustness to obfuscation and generalization across model versions.

Where Pith is reading between the lines

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

  • The paper does not isolate whether the gains come from the argument relations specifically or from any dense graph structure built over sentence embeddings; a control that keeps the same nodes and topology but replaces the predicted relation labels with random or uniform edge weights would settle that, and is not reported.
  • Because the relation probabilities come from a classifier trained elsewhere and are never validated against gold argument annotations, some of the discriminative signal may be systematic behavior of that classifier; swapping the argument-mining pipeline or using manually annotated essays would reveal how much of the fingerprint is genuinely reasoning-specific.
  • The corpus contains no human-authored essays, so an untested extension is whether reasoning graphs also separate humans from machines — and which family a human imitator would be mistaken for.
  • The obfuscations used here attack the linguistic surface; a stronger stress test would target the structure itself by reordering arguments, deleting or splicing in support statements, or translating through a language that flattens argumentative connectives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper introduces LLM-OWL-AE, a corpus of 2,240 argumentative essays generated by eight open LLMs (four families, two versions each), augmented with paraphrase and backtranslation obfuscations. The proposed method extracts so-called reasoning graphs using the oAMF argument-mining pipeline (TARGER segmentation + ARIR relation prediction), encodes nodes with all-mpnet-base-v2 sentence embeddings and edges with either argmax or raw probability relation labels, and trains GNNs to attribute essays to one of four model families. The authors evaluate multiple GNN architectures, edge encodings, and edge thresholds under same-version and cross-version splits, comparing against a fine-tuned Longformer text baseline. They report improved robustness to obfuscation (up to +27 F1 points) and improved cross-version generalization (up to +19 F1 points), concluding that reasoning structure is a more stable authorship fingerprint than surface text.

Significance. If the causal claim were established, this would be a meaningful advance: moving authorship attribution beyond surface lexical features to discourse-level structure, with practical implications for robustness and generalization to new model versions. The evaluation is honest and fairly extensive: two data partitions, four GNN architectures, two edge encodings, four threshold settings, three seeds, and extended results in the appendix. The release of the corpus and code is also a valuable contribution. However, the central interpretation — that reasoning relations, rather than dense graph structure or sentence embeddings, drive the gains — is not currently supported by the experiments. The paper needs targeted ablations and a more disciplined model-selection protocol before the contribution can be fully assessed.

major comments (3)
  1. [§6, Table 6; §4; §7] The extracted graphs are nearly complete DAGs. For original Gemma3 essays, the 'None' threshold retains 304 edges and T1 retains 258 edges out of roughly 300 possible directed edges among 25 nodes; across Table 6, T1 retains 80–90% of all possible edges. Combined with the finding that 1-layer GNNs produce about 90% of the best results (§7), the model is essentially aggregating over an almost complete graph built on sentence embeddings. The paper never ablates the ARIR edge labels against a trivial complete DAG (uniform, random, or no edge features) or against a mean-pooled sentence-embedding classifier. Without such an ablation, the reported robustness and generalization cannot be attributed to reasoning structure; it may be an artifact of dense graph pooling of sentence embeddings. This concern directly affects the interpretation of every reported improvement.
  2. [§5.1, Tables 2 and 3] The headline gains are selected post hoc as the best among 4 GNN architectures × 2 edge encodings × 4 thresholds × up to 7 layers, with the 'best' threshold column chosen per condition. No multiple-comparison control, held-out threshold selection, or statistical significance testing is reported. With only 3 seeds and no variance reporting, a 19–27 point gap could reflect selection noise. The authors should either pre-specify a single configuration or select thresholds/architectures on a development split, and report confidence intervals across seeds.
  3. [Limitations; §4] The paper states in Limitations that argument component detection and argument relation identification remain open problems, but it never evaluates the extracted graphs against gold argument structures or against a null model for the edge labels. Since ARIR is an external model trained elsewhere, systematic biases in relation prediction (e.g., over-predicting inference) could become the real signal. A minimal validation — for example, comparing GNN performance with shuffled or randomly permuted edge labels, or with edges replaced by a trivial complete graph — is necessary to establish that the reasoning relations, not the external pipeline's artifacts, carry the signal.
minor comments (3)
  1. [§4] The description of edge encoding is inconsistent: the argmax vector is said to contain three elements, while the probabilities strategy contains four elements, and the relation to W = (w_inf, w_conf, w_rephr, w_no_rel) is not fully clarified.
  2. [§5.3] GNN training hyperparameters (hidden dimension, learning rate, optimizer, dropout schedule) are not reported; only epochs and early stopping are mentioned. This limits reproducibility.
  3. [Appendix C, Tables 7–8] The appendix says the best results were achieved with 1–3 layers, and the 'best' rows in Tables 2 and 3 are selected across layers and thresholds. The selection procedure should be made explicit, and the layer count for each reported 'best' value should be indicated in the main tables.

Circularity Check

0 steps flagged

No circularity: the GNN is trained on fixed external argument-mining features and tested on held-out essays and LLM versions.

full rationale

The claimed derivation chain is an empirical supervised pipeline, not a derivation that reduces to its inputs. Essays are generated by known LLMs; oAMF (TARGER + ARIR) converts them into graphs; all-mpnet-base-v2 yields node embeddings; a GNN is trained on family labels and evaluated on held-out essays and unseen model versions. No equation in §4 defines the target label in terms of the extracted graph, nor is any fitted parameter renamed as a prediction. The ARIR relation classifier (Ruiz-Dolz et al., 2021) and oAMF (Gemechu et al., 2025) are self-citations in the sense that a co-author is on those tools, but they are fixed external components not trained on LLM-OWL-AE labels; per the review rules this is independent support and does not raise the circularity score. The Limitations section explicitly concedes that argument component detection and relation identification are open problems, and Appendix B shows the graphs are dense (>80% of possible edges at T1) with >90% single source-sink paths; these facts weaken the causal interpretation that 'reasoning structure' rather than sentence embeddings or graph pooling drives the gains, and the 'best' threshold reporting in Tables 2/3 is a selection concern. However, absence of an ablation is an external-validity / confound issue, not a case where a prediction is equivalent to its input by construction. No circular step can be quoted from the paper.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new theoretical entities are invented. The paper's contribution is an application of existing tools (oAMF, sentence transformers, GNNs). The load-bearing assumptions are empirical: that ARIR's relation predictions are accurate enough to be meaningful, that the reasoning graphs retain their discriminative structure after obfuscation, and that the headline deltas (picked across many configurations) represent a stable advantage rather than selection noise.

free parameters (3)
  • Edge probability thresholds T1/T2/T3 = (0.5,0.5,0.5), (0.7,0.7,0.7), (0.9,0.7,0.7)
    Thresholds are chosen by hand; the 'best' column in Tables 2/3 picks whichever threshold set performs best per condition, and the paper states 'we could not observe a consistent optimal configuration' — i.e., the reported headline deltas are partly a function of this selection.
  • Number of GNN layers (1–3) = 1–3 (best results; ~90% of best F1 are 1-layer)
    Layer count is tuned on the data; the paper reports only selected configurations in the main tables rather than all 1–7 results.
  • Obfuscation severity settings (DIPPER L60/O60) = L60, O60
    Taken directly from the DIPPER authors' recommended setting — a fixed external choice, not fit to the target data, so it is included here only for completeness.
axioms (4)
  • domain assumption The oAMF pipeline (TARGER segmentation + ARIR relation classifier) produces argument structures accurate enough to serve as a discriminative fingerprint.
    The whole method rests on this. The paper never evaluates graph quality against gold argument annotations and explicitly calls argument mining 'an open problem' (Limitations). If ARIR's predictions are noisy, the GNN is fitting a noisy proxy, not reasoning structure.
  • domain assumption Sentence embeddings from all-mpnet-base-v2 preserve enough family-specific semantic content after paraphrase/backtranslation.
    Node features come only from these embeddings (Section 4, Node encoding); no ablation separates the contribution of the embeddings from the contribution of the graph structure.
  • domain assumption The 140 topics and two stances produce balanced, representative essays for each model family.
    The corpus is generated from one prompt template (Purdue OWL instructions) and topics from Ruiz-Dolz et al. (2024). Generalisation to other genres and domains is explicitly deferred to future work (Limitations).
  • domain assumption Valid English text after DIPPER paraphrase and OPUS-MT backtranslation preserves the reasoning structure of the original essay.
    The paper only 'conducted a manual sanity check on a sample' (Section 3.2); the graph statistics in Section 6 show that obfuscation changes graph density and connectivity substantially, so the preservation of argument relations is not established.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution." pith.science (2026). https://pith.science/paper/5KLEB5UX

@misc{pith2026260714905,
  author       = {Pith},
  title        = {Pith review of: Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5KLEB5UX}},
  note         = {Machine review of arXiv:2607.14905}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Given the current trend to employ large language models (LLMs) in almost any imaginable context, LLM-generated text detection and authorship attribution have become a pressing issue. Prior work has primarily focused on surface-level linguistic features, an approach shown to be susceptible to paraphrasing and other obfuscation techniques. In this paper, we go beyond the linguistic surface, extracting and analysing reasoning structures in LLM-generated texts with the goal of capturing more complex signals of LLM authorship. We propose a graph neural network approach that leverages reasoning graphs extracted by an argument mining pipeline, demonstrating improved robustness and generalisation over a traditional Longformer baseline. Our approach outperforms the baseline by up to 27 percentage points under the obfuscation attacks such as paraphrasing and backtranslation, and 19 percentage points when evaluated on the texts generated by the unseen model versions, simulating real-world conditions in which new LLM versions are continuously released.

Figures

Figures reproduced from arXiv: 2607.14905 by Annette Hautli-Janisz, Artur Romazanov, Ramon Ruiz-Dolz, Zlata Kikteva.

Figure 1
Figure 1. Figure 1: Reasoning graph-based approach using GNN architecture for LLM authorship attribution. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Essay linguistic features. Violin plots on the left are for the LLM-OWL-AE-I data partition; violin plots [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ratio of the average number of edges per node across different thresholds. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Proportion of isolated components in graphs. [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Proportion of the graphs forming a single source-sink path. [PITH_FULL_IMAGE:figures/full_fig_p016_5.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

82 extracted references · 4 canonical work pages

  1. [1]

    Companion Proceedings of the ACM Web Conference 2024 , pages=

    Online disinformation and generative language models: Motivations, challenges, and mitigations , author=. Companion Proceedings of the ACM Web Conference 2024 , pages=

    2024

  2. [2]

    Advances in neural information processing systems , volume=

    Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense , author=. Advances in neural information processing systems , volume=

  3. [3]

    Learning and individual differences , volume=

    ChatGPT for good? On opportunities and challenges of large language models for education , author=. Learning and individual differences , volume=. 2023 , publisher=

    2023

  4. [4]

    Computational Linguistics , volume=

    A survey on llm-generated text detection: Necessity, methods, and future directions , author=. Computational Linguistics , volume=

  5. [5]

    arXiv preprint arXiv:2304.07666 , year=

    ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models , author=. arXiv preprint arXiv:2304.07666 , year=

    Pith/arXiv arXiv

  6. [6]

    LLM - D etect AI ve: a Tool for Fine-Grained Machine-Generated Text Detection

    Abassy, Mervat and Elozeiri, Kareem and Aziz, Alexander and Ta, Minh Ngoc and Tomar, Raj Vardhan and Adhikari, Bimarsha and Ahmed, Saad El Dine and Wang, Yuxia and Mohammed Afzal, Osama and Xie, Zhuohan and Mansurov, Jonibek and Artemova, Ekaterina and Mikhailov, Vladislav and Xing, Rui and Geng, Jiahui and Iqbal, Hasan and Mujahid, Zain Muhammad and Mahm...

    2024

  7. [7]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Raid: A shared benchmark for robust evaluation of machine-generated text detectors , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  8. [8]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Evobench: Towards real-world llm-generated text detection benchmarking for evolving large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  9. [9]

    34th USENIX Security Symposium (USENIX Security 25) , pages=

    \ LLMmap \ : Fingerprinting for large language models , author=. 34th USENIX Security Symposium (USENIX Security 25) , pages=

  10. [10]

    2023 International conference on cyber-enabled distributed computing and knowledge discovery (cyberc) , pages=

    Neural authorship attribution: Stylometric analysis on large language models , author=. 2023 International conference on cyber-enabled distributed computing and knowledge discovery (cyberc) , pages=. 2023 , organization=

    2023

  11. [11]

    ACM SIGKDD Explorations Newsletter , volume=

    Authorship attribution in the era of llms: Problems, methodologies, and challenges , author=. ACM SIGKDD Explorations Newsletter , volume=. 2025 , publisher=

    2025

  12. [12]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    The Two Paradigms of LLM Detection: Authorship Attribution vs Authorship Verification , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  13. [13]

    arXiv preprint arXiv:2304.14072 , year=

    Origin tracing and detecting of llms , author=. arXiv preprint arXiv:2304.14072 , year=

    Pith/arXiv arXiv

  14. [14]

    The Twelfth International Conference on Learning Representations , year=

    Dna-gpt: Divergent n-gram analysis for training-free detection of gpt-generated text , author=. The Twelfth International Conference on Learning Representations , year=

  15. [15]

    OPUS - MT -- Building open translation services for the World

    Tiedemann, J. OPUS - MT -- Building open translation services for the World. Proceedings of the 22nd Annual Conference of the European Association for Machine Translation. 2020

    2020

  16. [16]

    The Tatoeba Translation Challenge -- Realistic Data Sets for Low Resource and Multilingual MT

    Tiedemann, J. The Tatoeba Translation Challenge -- Realistic Data Sets for Low Resource and Multilingual MT. Proceedings of the Fifth Conference on Machine Translation. 2020

    2020

  17. [17]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Authorship obfuscation in multilingual machine-generated text detection , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  18. [18]

    Altakrori, Malik and Scialom, Thomas and Fung, Benjamin C. M. and Cheung, Jackie Chi Kit. A Multifaceted Framework to Evaluate Evasion, Content Preservation, and Misattribution in Authorship Obfuscation Techniques. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.153

    work page doi:10.18653/v1/2022.emnlp-main.153 2022

  19. [19]

    Proceedings of the 36th ACM Conference on Hypertext and Social Media , pages=

    Esperanto: Evaluating synthesized phrases to enhance robustness in ai detection for text origination , author=. Proceedings of the 36th ACM Conference on Hypertext and Social Media , pages=

  20. [20]

    PRD etect: Perturbation-Robust LLM -generated Text Detection Based on Syntax Tree

    Li, Xiang and Yin, Zhiyi and Tan, Hexiang and Jing, Shaoling and Su, Du and Cheng, Yi and Shen, Huawei and Sun, Fei. PRD etect: Perturbation-Robust LLM -generated Text Detection Based on Syntax Tree. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.464

    work page doi:10.18653/v1/2025.findings-naacl.464 2025

  21. [21]

    Data in Brief , volume=

    Nlas-multi: A multilingual corpus of automatically generated natural language argumentation schemes , author=. Data in Brief , volume=. 2024 , publisher=

    2024

  22. [22]

    arXiv preprint arXiv:2004.05150 , year=

    Longformer: The long-document transformer , author=. arXiv preprint arXiv:2004.05150 , year=

    Pith/arXiv arXiv 2004

  23. [23]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    M4gt-bench: Evaluation benchmark for black-box machine-generated text detection , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  24. [24]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    MULTITuDE: Large-scale multilingual machine-generated text detection benchmark , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  25. [25]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    OpenTuringBench: An Open-Model-based Benchmark and Framework for Machine-Generated Text Detection and Attribution , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  26. [26]

    ACM Transactions on Information and System Security (TISSEC) , volume=

    Adversarial stylometry: Circumventing authorship recognition to preserve privacy and anonymity , author=. ACM Transactions on Information and System Security (TISSEC) , volume=. 2012 , publisher=

    2012

  27. [27]

    Proceedings of the COLING/ACL 2006 Main Conference Poster Sessions , pages=

    Obfuscating document stylometry to preserve author anonymity , author=. Proceedings of the COLING/ACL 2006 Main Conference Poster Sessions , pages=

    2006

  28. [28]

    , author=

    Author Obfuscation: Attacking the State of the Art in Authorship Verification. , author=. CLEF (Working Notes) , pages=

  29. [29]

    , author=

    Overview of the Author Obfuscation Task at PAN 2018: A New Approach to Measuring Safety. , author=. CLEF (Working Notes) , year=

    2018

  30. [30]

    , author=

    Overview of the Author Obfuscation Task at PAN 2017: Safety Evaluation Revisited. , author=. CLEF (Working Notes) , year=

    2017

  31. [31]

    9th USENIX Security Symposium (USENIX Security 00) , year=

    Can pseudonymity really guarantee privacy? , author=. 9th USENIX Security Symposium (USENIX Security 00) , year=

  32. [32]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    Humanizing machine-generated content: evading AI-text detection through adversarial attack , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    2024

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    , author=

    Author Masking through Translation. , author=. CLEF (Working Notes) , volume=

  35. [35]

    CLEF 2016 Evaluation Labs and Workshop--Working Notes Papers , pages=

    Author Obfuscation using WordNet and language models—notebook for PAN at CLEF 2016 , author=. CLEF 2016 Evaluation Labs and Workshop--Working Notes Papers , pages=

    2016

  36. [36]

    , author=

    SU@ PAN'2016: Author Obfuscation. , author=. CLEF (Working Notes) , pages=

    2016

  37. [37]

    ACM SIGKDD Explorations Newsletter , volume=

    Attribution and obfuscation of neural text authorship: A data mining perspective , author=. ACM SIGKDD Explorations Newsletter , volume=. 2023 , publisher=

    2023

  38. [38]

    arXiv preprint arXiv:2303.11156 , year=

    Can AI-generated text be reliably detected? , author=. arXiv preprint arXiv:2303.11156 , year=

    Pith/arXiv arXiv

  39. [39]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    RAFT: Realistic attacks to fool text detectors , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  40. [40]

    Transactions of the Association for Computational Linguistics , volume=

    Red teaming language model detectors with language models , author=. Transactions of the Association for Computational Linguistics , volume=. 2024 , publisher=

    2024

  41. [41]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Your language model can secretly write like humans: Contrastive paraphrase attacks on llm-generated text detectors , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  42. [42]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Navigating the shadows: Unveiling effective disturbances for modern ai content detectors , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  43. [43]

    Counter T uring Test ( CT 2): AI -Generated Text Detection is Not as Easy as You May Think - Introducing AI Detectability Index ( ADI )

    Chakraborty, Megha and Tonmoy, S.M Towhidul Islam and Zaman, S M Mehedi and Gautam, Shreya and Kumar, Tanay and Sharma, Krish and Barman, Niyar and Gupta, Chandan and Jain, Vinija and Chadha, Aman and Sheth, Amit and Das, Amitava. Counter T uring Test ( CT 2): AI -Generated Text Detection is Not as Easy as You May Think - Introducing AI Detectability Inde...

    work page doi:10.18653/v1/2023.emnlp-main.136 2023

  44. [44]

    The Open Argument Mining Framework

    Gemechu, Debela and Ruiz-Dolz, Ramon and G \'o rska, Kamila and Moslemnejad, Somaye and Maguire, Eimear and Zografistou, Dimitra and Jo, Yohan and Lawrence, John and Reed, Chris. The Open Argument Mining Framework. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2025. doi:10.18653/...

    doi:10.18653/v1/2025.acl-demo.31 2025

  45. [45]

    TARGER : Neural Argument Mining at Your Fingertips

    Chernodub, Artem and Oliynyk, Oleksiy and Heidenreich, Philipp and Bondarenko, Alexander and Hagen, Matthias and Biemann, Chris and Panchenko, Alexander. TARGER : Neural Argument Mining at Your Fingertips. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations. 2019. doi:10.18653/v1/P19-3031

    work page doi:10.18653/v1/p19-3031 2019

  46. [46]

    IEEE Intelligent Systems , volume=

    Transformer-based models for automatic identification of argument relations: A cross-domain evaluation , author=. IEEE Intelligent Systems , volume=. 2021 , publisher=

    2021

  47. [47]

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

    Reimers, Nils and Gurevych, Iryna. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. 2019

    2019

  48. [48]

    AI open , volume=

    Graph neural networks: A review of methods and applications , author=. AI open , volume=. 2020 , publisher=

    2020

  49. [49]

    , journal=

    Wu, Zonghan and Pan, Shirui and Chen, Fengwen and Long, Guodong and Zhang, Chengqi and Yu, Philip S. , journal=. A Comprehensive Survey on Graph Neural Networks , year=

  50. [50]

    2013 , issue_date =

    Peldszus, Andreas and Stede, Manfred , title =. 2013 , issue_date =. doi:10.4018/jcini.2013010101 , month = jan, pages =

    doi:10.4018/jcini.2013010101 2013

  51. [51]

    Argument Mining: A Survey

    Lawrence, John and Reed, Chris. Argument Mining: A Survey. Computational Linguistics. 2019. doi:10.1162/coli_a_00364

    doi:10.1162/coli_a_00364 2019

  52. [52]

    10th International Conference on Computational Models of Argument (COMMA 2024) , year=

    Graph Convolutional Networks and Graph Attention Networks for Approximating Arguments Acceptability , author=. 10th International Conference on Computational Models of Argument (COMMA 2024) , year=

    2024

  53. [53]

    International Conference on Scalable Uncertainty Management , pages=

    Using graph convolutional networks for approximate reasoning with abstract argumentation frameworks: A feasibility study , author=. International Conference on Scalable Uncertainty Management , pages=. 2019 , organization=

    2019

  54. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Heterogeneous Graph Neural Networks for Assumption-Based Argumentation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  55. [55]

    , author=

    Determining the Acceptability of Abstract Arguments with Graph Convolutional Networks. , author=. SAFA@ COMMA , pages=

  56. [56]

    , author=

    Graph Neural Networks for Algorithm Selection in Abstract Argumentation. , author=. ArgML@ COMMA , pages=

  57. [57]

    arXiv preprint arXiv:2110.00124 , year=

    Tree-constrained graph neural networks for argument mining , author=. arXiv preprint arXiv:2110.00124 , year=

    Pith/arXiv arXiv

  58. [58]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Automatic debate evaluation with argumentation semantics and natural language argument graph networks , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  59. [59]

    2023 , isbn =

    Zhang, Gechuan and Nulty, Paul and Lillis, David , title =. 2023 , isbn =. doi:10.1145/3594536.3595152 , booktitle =

    arXiv 2023

  60. [60]

    Neural Computing and Applications , volume=

    Seeing both sides: context-aware heterogeneous graph matching networks for extracting-related arguments , author=. Neural Computing and Applications , volume=. 2024 , publisher=

    2024

  61. [61]

    Kipf and Max Welling , title =

    Thomas N. Kipf and Max Welling , title =. 5th International Conference on Learning Representations,. 2017 , url =

    2017

  62. [62]

    6th International Conference on Learning Representations , keywords =

    Veličković, Petar and Cucurull, Guillem and Casanova, Arantxa and Romero, Adriana and Liò, Pietro and Bengio, Yoshua , biburl =. 6th International Conference on Learning Representations , keywords =

  63. [63]

    Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence , pages=

    Masked Label Prediction: Unified Message Passing Model for Semi-Supervised Classification , author=. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence , pages=. 2021 , organization=

    2021

  64. [64]

    Advances in Neural Information Processing Systems , volume=

    Recipe for a general, powerful, scalable graph transformer , author=. Advances in Neural Information Processing Systems , volume=

  65. [65]

    2010 , publisher=

    McCarthy, Philip M and Jarvis, Scott , journal=. 2010 , publisher=

    2010

  66. [66]

    Are We in the AI -Generated Text World Already? Quantifying and Monitoring AIGT on Social Media

    Sun, Zhen and Zhang, Zongmin and Shen, Xinyue and Zhang, Ziyi and Liu, Yule and Backes, Michael and Zhang, Yang and He, Xinlei. Are We in the AI -Generated Text World Already? Quantifying and Monitoring AIGT on Social Media. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/...

    doi:10.18653/v1/2025.acl-long.1120 2025

  67. [67]

    arXiv preprint arXiv:2604.26965 , year=

    The Impact of AI-Generated Text on the Internet , author=. arXiv preprint arXiv:2604.26965 , year=

    Pith/arXiv arXiv

  68. [68]

    All That ' s `Human' Is Not Gold: Evaluating Human Evaluation of Generated Text

    Clark, Elizabeth and August, Tal and Serrano, Sofia and Haduong, Nikita and Gururangan, Suchin and Smith, Noah A. All That ' s `Human' Is Not Gold: Evaluating Human Evaluation of Generated Text. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing...

    doi:10.18653/v1/2021.acl-long.565 2021

  69. [69]

    Do humans identify AI-generated text better than machines? Evidence based on excerpts from German theses☆ , journal =

    Alexandra Fiedler and Jörg Döpke , keywords =. Do humans identify AI-generated text better than machines? Evidence based on excerpts from German theses☆ , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.iree.2025.100321 , url =

    arXiv 2025

  70. [70]

    Disinformation Capabilities of Large Language Models

    Vykopal, Ivan and Pikuliak, Mat \'u s and Srba, Ivan and Moro, Robert and Macko, Dominik and Bielikova, Maria. Disinformation Capabilities of Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.793

    doi:10.18653/v1/2024.acl-long.793 2024

  71. [71]

    npj Artificial Intelligence , volume=

    Large language models reflect the ideology of their creators , author=. npj Artificial Intelligence , volume=. 2026 , publisher=

    2026

  72. [72]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Measuring political bias in large language models: What is said and how it is said , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  73. [73]

    International conference on machine learning , pages=

    Detectgpt: Zero-shot machine-generated text detection using probability curvature , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  74. [74]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Spotting LLMs with binoculars: zero-shot detection of machine-generated text , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  75. [75]

    S em E val-2024 Task 8: Multidomain, Multimodel and Multilingual Machine-Generated Text Detection

    Wang, Yuxia and Mansurov, Jonibek and Ivanov, Petar and Su, Jinyan and Shelmanov, Artem and Tsvigun, Akim and Mohammed Afzal, Osama and Mahmoud, Tarek and Puccetti, Giovanni and Arnold, Thomas. S em E val-2024 Task 8: Multidomain, Multimodel and Multilingual Machine-Generated Text Detection. Proceedings of the 18th International Workshop on Semantic Evalu...

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

  76. [76]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2018 , month=. doi:10.1609/aaai.v32i1.11604 , abstractNote=

    doi:10.1609/aaai.v32i1.11604 2018

  77. [77]

    Honnibal, Matthew and Montani, Ines and Van Landeghem, Sofie and Boyd, Adriane , doi =

  78. [78]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Alison: Fast and effective stylometric authorship obfuscation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  79. [79]

    Journal of Cultural Analytics , volume=

    Can GPT-3 pass a writer’s Turing test? , author=. Journal of Cultural Analytics , volume=

  80. [80]

    BioRxiv , year=

    Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers , author=. BioRxiv , year=

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.