Pith. sign in

REVIEW 2 major objections 7 minor 61 references

A detector built from uncertainty-filtered training data and partial-AUROC fine-tuning scores 0.974 overall, ranking second on the 2026 shared task.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:09 UTC pith:F63AFL62

load-bearing objection Solid shared-task system with externally anchored results, but the causal claim about dataset curation is not actually tested—no unfiltered baseline. the 2 major comments →

arxiv 2607.17382 v1 pith:F63AFL62 submitted 2026-07-19 cs.CL cs.LG

Team DACTYL at PAN 2026: Bayesian Data Mixing and Empirical X-risk Minimization for AI-text Detection

Shantanu Thorat This is my paper
classification cs.CL cs.LG
keywords AI-generated text detectionout-of-distribution generalizationBayesian neural networksdata curationpartial AUROCempirical X-risk minimizationmulticalibrationtext classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tackles a known failure of AI-text detectors: they do well on in-distribution data but degrade on out-of-distribution text, and naively mixing more datasets can make this worse by reinforcing shortcuts. The author proposes using tiny BERT models with Bayesian classification heads to measure prediction uncertainty across three datasets, then removing texts the weak models classify incorrectly with high confidence. The remaining texts train large DeBERTa-V3-large and ModernBERT-large classifiers optimized with a two-way partial-AUROC loss, and the ModernBERT model is further calibrated with an uncertainty-based method. On the 2026 shared task test set, the best system scores 0.974 overall, ranking second and beating the prior winner. The paper's claim is that careful, uncertainty-driven dataset curation, not just adding more data, is what drives strong out-of-distribution performance.

Core claim

The paper claims that 'incorrect and certain' texts — those a Bayesian BERT-tiny ensemble misclassifies with low predictive variance — are outliers that harm large classifiers, and that discarding them before fine-tuning prevents data mixing from inducing shortcut learning. The resulting ModernBERT-large system with MCGrad calibration scores 0.974 mean on the 2026 shared task test set across five metrics (AUROC 0.993, F1 0.975, C@1 0.962, Brier 0.970, F0.5u 0.969), placing second overall, 0.001 behind the top system and ahead of the prior winner at 0.968. The uncalibrated ModernBERT-large scores 0.96 and the DeBERTa-V3-large model 0.882; the calibration step is therefore responsible for the

What carries the argument

The load-bearing component is an ensemble of five small Bayesian BERT classifiers (BERT-tiny with a Bayesian linear classification head trained with an ELBO loss combined with a two-way partial-AUROC loss). For each candidate text, five sampling passes produce a mean score and a standard deviation; the standard deviation's percentile against a classifier's test-set distribution determines a keep, reject, or abstain vote. A text enters the final training set only if at least three of five classifiers vote keep and no more than one votes reject. The three datasets with the best cross-dataset AUROC (DACTYL-complete, LLMTrace, MAGA-Bench) are filtered this way, and the surviving texts fine-tune

Load-bearing premise

The curation benefit rests on an untested transfer assumption: that texts a 4-million-parameter Bayesian model gets wrong with high confidence are also harmful for the much larger DeBERTa-V3-large and ModernBERT-large classifiers; if that transfer fails, the gains come from bigger models and more data rather than from the filtering.

What would settle it

Train the two large models on the unfiltered union of DACTYL-complete, LLMTrace, and MAGA-Bench, and compare the 2026 test-set score with the filtered model's 0.974; if the unfiltered model matches or exceeds it, the curation step is not responsible for the result. A second check: label a sample of rejected texts to see whether they are genuinely mislabeled, as the dataset-cartography intuition predicts.

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

If this is right

  • Data mixing can be made safe by removing texts that weak models are confidently wrong about, rather than relying on simple concatenation of datasets.
  • Tiny Bayesian models can act as a cheap curation layer for expensive large classifiers, pruning training data while improving out-of-distribution scores.
  • Partial-AUROC fine-tuning and uncertainty-based calibration improve the official five-metric score even when AUROC is already high.
  • The full pipeline places second on the 2026 leaderboard and is one of two systems to beat the previous winner.
  • The same filtering and calibration recipe is proposed as a plug-in for other text classification tasks, such as prompt-safety detection.

Where Pith is reading between the lines

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

  • If the transfer assumption holds, the filter alone should yield detectable gains on small classifiers too; a practitioner could test the curation step without fine-tuning billion-scale models.
  • The keep/reject thresholds (5th/95th percentile abstain, 50th percentile reject, ensemble rule) are hand-set; ablating them would show how much of the 0.974 depends on those exact choices.
  • The paper's external-set results, where calibration helps only six of twelve datasets, suggest calibration gains are concentrated on near-distribution shifts; a per-domain trigger for calibration might extend the gains.
  • Dataset cartography predicts many 'incorrect and certain' texts are mislabeled, but the paper never inspects rejected texts; labeling them is a direct way to verify the mechanism.

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

2 major / 7 minor

Summary. The paper presents Team DACTYL's system for the PAN 2026 Voight-Kampff AI-generated text detection task. The authors construct DACTYLv2, an extension of their earlier DACTYL corpus, and train five Bayesian BERT-tiny classifiers on candidate training sets. Using the mean and standard deviation of each Bayesian model's predictions, they define a voting rule that keeps or rejects texts from DACTYL-complete, LLMTrace, and MAGA-Bench, forming a consolidated filtered training set. They then fine-tune DeBERTa-V3-large and ModernBERT-large with a two-way partial-AUROC empirical X-risk minimization objective, and additionally apply MCGrad multicalibration to ModernBERT-large using BNN uncertainties as features. The reported overall scores on the PAN 2026 test set are 0.882 for DeBERTa-V3-large, 0.960 for ModernBERT-large, and 0.974 for the MCGrad-calibrated model, which ranks second on the official leaderboard and beats the PAN 2025 winning baseline. The paper's central interpretive claim is that careful dataset curation leads to strong out-of-distribution performance.

Significance. If the results hold, the paper demonstrates a practically effective pipeline: BNN-uncertainty-based filtering, EXM-style partial-AUROC fine-tuning, and MCGrad calibration yield a detector that ranks second on an external shared task and outperforms the previous year's winner. The headline numbers are anchored to the official PAN 2026 leaderboard (Table 11), and the five-metric means reported in Tables 8-10 are arithmetically consistent. The release of the two large fine-tuned models and the construction of DACTYLv2 are useful resources for the community. However, the paper's main claimed contribution—that the Bayesian curation step improves OOD performance—is not isolated by any ablation, and the filtering machinery contains several hand-set components without sensitivity analysis. The external leaderboard result is credible, but the causal attribution in the abstract is not yet supported.

major comments (2)
  1. [Abstract, §3.2, Tables 7–10] The paper's interpretive claim that 'careful dataset curation can lead to strong OOD performance' is not directly tested. The large models are trained only on the kept subset (Table 7: 727,235 / 146,908 / 297,994 texts); there is no baseline trained on the unfiltered union of DACTYL-complete + LLMTrace + MAGA-Bench, nor on a random subset of the same size. Without that control, the 0.974 / 0.96 / 0.882 leaderboard results (Table 11) could be due to model capacity, data volume, the partial-AUROC objective, or the particular dataset union rather than to the Bayesian filter. The authors should add an ablation: train the same two large architectures on (a) the unfiltered superset and (b) a randomly filtered superset matched on keep rate, and report the PAN 2026 and external-set scores. This is required to support the causal attribution in the abstract and conclusion.
  2. [§3.2, Table 6] The filter rests on an unvalidated transfer assumption: that 'incorrect and certain' predictions from a 4M-parameter Bayesian BERT-tiny head identify texts that are mislabeled, noisy, or otherwise harmful for DeBERTa-V3-large and ModernBERT-large trained with two-way partial-AUROC loss. The dataset-cartography intuition [53] concerns the training dynamics of the same model, not cross-architecture transfer. No experiment shows that rejected texts are misclassified by the large models, that they increase the partial-AUROC loss, or that the rejection threshold is stable. Moreover, the voting thresholds (P_t < 5 or > 95, P_t <= 50, ensemble rule) and the per-dataset BNN hyperparameters (rho, f in Table 4) are hand-set with no sensitivity analysis. Since the filtering step is the paper's main contribution, a small sensitivity study (e.g., varying thresholds, replacing rejection with random dr
minor comments (7)
  1. [Abstract] Typo: 'DeBERTa-V3-large-large' should read 'DeBERTa-V3-large'.
  2. [§2.3, Table 3] The table columns are unclear: 'Count' and 'LLMs Used' appear as bare integers. The column headers should be named explicitly, and the domain information should be described in the caption or text.
  3. [§3.1] The number of Monte Carlo samples used during BNN training is not stated. The authors mention 5 runs at inference and 30 runs for constructing S_D, but the ELBO loss in Eq. (1) requires sampling; please report the sample count.
  4. [§4.2] The temperature T = 2.14555 is reported to five decimals without the search interval or the optimized log-loss value. Reporting these details would improve reproducibility.
  5. [Tables 8–10] DACTYL-complete, LLMTrace, and MAGA-Bench are used both as training sources and as test sets. The paper labels them in-distribution, but no explicit train/test split is described; please clarify whether any overlapping examples remain.
  6. [Resources and Code] Only the two large models are released; the filtering code and DACTYLv2 corpus are not. Since the filtering step is the novel contribution, releasing the filtering scripts and dataset composition would significantly aid reproducibility.
  7. [§5, Tables 8–9] The MCGrad calibration gains are selective: the overall score improves on PAN 2025 Test and PAN 2026 Test but degrades on ELOQUENT and on several external sets. The paper acknowledges this in the text, but the abstract's 'best score' framing should be qualified with this caveat.

Circularity Check

0 steps flagged

No constructional circularity; the rank-2 score is anchored to the external PAN 2026 leaderboard, and the filter-attribution gap is an evidential issue rather than a circular reduction.

full rationale

The paper's central quantitative claims (ModernBERT 0.96, MCGrad 0.974, rank 2) are evaluated on the external PAN 2026 test set, so they are not constructed from fitted parameters or from the filtering rule. The Bayesian BERT-tiny filter is an unverified transfer assumption: the paper never shows that 'incorrect and certain' small-model votes identify texts harmful to DeBERTa-V3-large/ModernBERT-large, and no unfiltered-superset ablation isolates the curation effect. That is a missing-control/validity problem, not an equivalence-by-construction: no equation defines the reported score in terms of the filter, and the filter itself uses held-out uncertainty percentiles plus true training labels. The temperature T=2.14555 and MCGrad are fit on a disjoint calibration set (Table 2) before held-out evaluation, which is legitimate calibration. Self-citations to [5] supply a dataset and a previously used loss variant; the loss is also cited to external LibAUC [10,55], so the self-citation is not load-bearing. Under the strict reduction test, no circular step is exhibited.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 1 invented entities

The ledger shows the paper's honest accounting: the headline result rests on a stack of hand-chosen thresholds (8 free parameter groups), several domain assumptions imported from the dataset-cartography and partial-AUROC literatures, and one unreleased new corpus. The most consequential input is the small-model uncertainty signal (axioms 1–2), which carries both the filtering and the calibration improvements but is never validated against the large models.

free parameters (8)
  • BNN initial posterior ρ per dataset = ρ ∈ {-3, -2, -3/80}; Table 4
    Manually tuned per dataset 'so that the classifier's uncertainty increased on OOD texts' (§3.1); sets the BNN head's weight-variance initialization. Rendered Table 4 is ambiguous for Personagen+SE.
  • Training sample fraction f per dataset = f ∈ {1/5, 1, 1/80}; Table 4
    Chosen per dataset for the Bayesian BERT-tiny training samples; affects which texts the voters saw and hence the uncertainty distributions S_D.
  • Vote thresholds (abstain/reject) = abstain if P_t<5 or P_t>95; reject if P_t≤50 and y_pred≠y
    Hand-chosen percentile boundaries in Table 6 (§3.2); no sensitivity analysis, and S_D itself is computed from only 30 runs per text.
  • Ensemble voting rule = keep if ≥3 keep votes and ≤1 reject vote
    Hand-chosen aggregation rule (§3.2, Table 7 'Rejected' column); examples show the rule is asymmetric and untested.
  • Temperature T for score transform = T=2.14555
    Fitted on the 25,938-text calibration set by minimizing log loss (Eqs. 2–3, §4.2); standard calibration, but a fitted constant that enters all reported MCGrad scores.
  • BNN inference runs = 5 runs (votes, MCGrad features); 30 runs (S_D)
    Chosen to limit inference time (§3.2); 5-run σ estimates are noisy inputs to both the filter and MCGrad.
  • BNN training hyperparameters = lr 1e-4, wd 0.01, batch 64, 1 epoch, pos-proportion 0.5, β=1
    Reported in §3.1; β=1 chosen so KL term does not dominate ELBO.
  • Large-model fine-tuning hyperparameters = lr 1e-5, batch 16, sampling rate 0.5, 1 epoch
    Reported in §4.1; a single epoch for both large models with no sweep or early stopping.
axioms (6)
  • domain assumption Dataset-cartography premise: samples a small confident model gets wrong are harmful outliers whose removal aids training
    Adopted from Swayamdipta et al. [53] and extended to the Bayesian/partial-AUROC setting in §3.2; the filter's central logic rests on this without verifying it for the large models.
  • domain assumption Uncertainty transfer: predictive σ of a 4M-parameter BNN head is a reliable signal of what larger non-Bayesian models will misclassify
    Load-bearing for both the filter (§3.2) and MCGrad σ-features (§4.2); unchecked empirically.
  • domain assumption Two-way partial-AUROC (DRO-KL) loss optimizes OOD generalization, and its outlier sensitivity motivates the filter
    Borrowed from LibAUC [10,55] and the author's DACTYL [5]; the outlier-sensitivity statement in §3.2 is asserted, not measured.
  • standard math Variational inference formalism (ELBO, Eq. 1; Blundell et al. [52])
    Standard Bayesian deep learning machinery; priors σ1=0.1, σ2=1.5, π=0.5 follow BLiTZ [51].
  • domain assumption Calibration-set representativeness: the 25,938-text mixture (Table 2) transfers to the PAN 2026 test distribution
    T and MCGrad are fit on this mixture; the ELOQUENT result (overall 0.92 vs 0.974) shows sensitivity to shift, and the paper notes MCGrad 'allows for correction' (§4.2).
  • ad hoc to paper The ad-hoc voting machinery is sound
    The criteria in Table 6 and the ≥3-keep/≤1-reject ensemble rule (§3.2) are constructed for this paper with no external grounding or robustness check.
invented entities (1)
  • DACTYLv2 corpus (DACTYL-complete = v1+v2) no independent evidence
    purpose: New one-shot LLM-rewrite training corpus across six domains and eight LLMs (Table 1); largest source in the filtered superset and a central input to every reported system
    Construction is described (mirror prompting, style-rewrite recipes) and counts are tabulated, but the dataset is not released and no external validation exists; all quality claims rest on the author's description.

pith-pipeline@v1.3.0-alltime-deepseek · 17273 in / 26704 out tokens · 249124 ms · 2026-08-01T18:09:24.998034+00:00 · methodology

0 comments
read the original abstract

Existing research shows that AI-generated text detection classifiers achieve strong in-distribution (ID) performance but do not maintain the same performance on out-of-distribution (OOD) texts, suggesting overfitting to dataset-specific features. However, combining different training datasets doesn't always improve performance and, in some cases, can even encourage shortcut learning. To address this issue, we fine-tune BERT-tiny models with Bayesian classification heads to select texts across three different datasets to use as a consolidated training set. We trained three different classifiers: fine-tuned DeBERTa-V3-large and ModernBERT-large classifiers via empirical X-risk minimization, and an MCGrad model that calibrates the predictions from the ModernBERT-large classifier. The DeBERTa-V3-large-large classifier achieves a mean score of 0.882 on the PAN 2026 test set across five metrics: AUROC, $F_1$, C@1, Brier score, and $F_{0.5u}$. ModernBERT-large achieves a score of 0.96 while MCGrad achieves the best score of the three with a mean score of 0.974, ranking second on the leaderboard. Our results highlight that careful dataset curation can lead to strong OOD performance. We release our ModernBERT-large and DeBERTa-V3-large models at https://huggingface.co/collections/ShantanuT01/panclef-2026 .

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

61 extracted references · 1 canonical work pages

  1. [1]

    Bevendorff, M

    J. Bevendorff, M. Fröbe, A. Greiner-Petter, A. Jakoby, M. Mayerl, P. Nakov, H. Plutz, M. Ta, Y. Wang, E. Zangerle, Overview of PAN 2026: Voight-Kampff Generative AI Detection, Text Watermarking, Multi-author Writing Style Analysis, Generative Plagiarism Detection, and Reasoning Trajectory Detection – Extended Abstract, in: R. Campos, A. Jatowt, Y. Lan, M....

    2026

  2. [2]

    Fröbe, M

    M. Fröbe, M. Wiegmann, N. Kolyada, B. Grahm, T. Elstner, F. Loebe, M. Hagen, B. Stein, M. Potthast, Continuous Integration for Reproducible Shared Tasks with TIRA.io, in: J. Kamps, L. Goeuriot, F. Crestani, M. Maistro, H. Joho, B. Davis, C. Gurrin, U. Kruschwitz, A. Caputo (Eds.), Advances in Information Retrieval. 45th European Conference on IR Research ...

    2023

  3. [3]

    Bevendorff, M

    J. Bevendorff, M. Fröbe, A. Greiner-Petter, A. Jakoby, M. Mayerl, P. Nakov, H. Plutz, M. Potthast, B. Stein, M. N. Ta, Y. Wang, E. Zangerle, in: M. Hagen, M. Potthast, B. Stein, P. Schaer, E. Zangerle, A. Barrón-Cedeño, A. G. S. de Herrera, E. S. Salido, S. MacAvaney, J. M. Struß (Eds.), Experimental IR Meets Multilinguality, Multimodality, and Interactio...

    2026

  4. [4]

    Voight- Kampff

    J. Bevendorff, R. R. Gunti, J. Karlgren, M. Fröbe, M. Potthast, B. Stein, Overview of the third “Voight- Kampff” Generative AI / LLM Detection Task at PAN and ELOQUENT 2026, in: A. Barrón-Cedeño, A. G. S. de Herrera, E. S. Salido, S. MacAvaney, J. M. Struß (Eds.), Working Notes of CLEF 2026 – Conference and Labs of the Evaluation Forum, CEUR Workshop Proc...

    2026

  5. [5]

    Thorat, A

    S. Thorat, A. Caines, Dactyl: Diverse adversarial corpus of texts yielded from large language models, arXiv preprint arXiv:2508.00619 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    Dugan, A

    L. Dugan, A. Hwang, F. Trhlík, A. Zhu, J. M. Ludan, H. Xu, D. Ippolito, C. Callison-Burch, Raid: A shared benchmark for robust evaluation of machine-generated text detectors, in: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 12463–12492

    2024

  7. [7]

    Compton, L

    R. Compton, L. Zhang, A. Puli, R. Ranganath, When more is less: Incorporating additional datasets can hurt performance by introducing spurious correlations, in: Machine learning for healthcare conference, PMLR, 2023, pp. 110–127

    2023

  8. [8]

    B. Emi, M. Spero, Technical report on the pangram ai-generated text classifier, arXiv preprint arXiv:2402.14873 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Korakakis, A

    M. Korakakis, A. Vlachos, A. Weller, Mitigating shortcut learning with interpolated learning, in: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 9191–9206

    2025

  10. [10]

    Z. Yuan, D. Zhu, Z.-H. Qiu, G. Li, X. Wang, T. Yang, Libauc: A deep learning library for x-risk optimization, in: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’23, ACM, 2023, p. 5487–5499. URL: http://dx.doi.org/10.1145/3580305.3599861. doi:10.1145/3580305.3599861

    arXiv 2023

  11. [11]

    Warner, A

    B. Warner, A. Chaffin, B. Clavié, O. Weller, O. Hallström, S. Taghadouini, A. Gallagher, R. Biswas, F. Ladhak, T. Aarsen, G. T. Adams, J. Howard, I. Poli, Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference, in: W. Che, J. Nabende, E. Shutova, M. T. Pilehvar (Eds.), Proceedi...

    2025

  12. [12]

    P. He, X. Liu, J. Gao, W. Chen, Deberta: Decoding-enhanced bert with disentangled attention, in: International Conference on Learning Representations, 2021. URL: https://openreview.net/forum? id=XPZIaotutsD

    2021

  13. [13]

    N. Tax, L. Perini, F. Linder, D. Haimovich, D. Karamshuk, N. Okati, M. Vojnovic, P. A. Apos- tolopoulos, MCGrad: Multicalibration at Web Scale, in: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 (KDD 2026), 2026. doi: 10.1145/ 3770854.3783954

    arXiv 2026

  14. [14]

    Schler, M

    J. Schler, M. Koppel, S. Argamon, J. W. Pennebaker, Effects of age and gender on blogging., in: AAAI spring symposium: Computational approaches to analyzing weblogs, volume 6, Stanford, CA, USA, 2006, pp. 199–205

    2006

  15. [15]

    Zhang, X

    G. Zhang, X. Du, Z. Yu, Z. Wang, Z. Wang, S. Guo, T. Zheng, K. Zhu, J. Liu, S. Yue, B. Liu, Z. Peng, Y. Yao, J. Yang, Z. Li, B. Zhang, M. Liu, T. Liu, Y. Gao, W. Chen, X. Zhou, Q. Liu, T. Wang, W. Huang, Finefineweb: A comprehensive study on fine-grained domain web corpus, 2024. URL: https://huggingface.co/datasets/m-a-p/FineFineWeb, version v0.1.0

    2024

  16. [16]

    Chiusano, 190k+ Medium Articles — kaggle.com, https://www.kaggle.com/datasets/ fabiochiusano/medium-articles/data, 2022

    F. Chiusano, 190k+ Medium Articles — kaggle.com, https://www.kaggle.com/datasets/ fabiochiusano/medium-articles/data, 2022. [Accessed 17-05-2026]

    2022

  17. [17]

    Saxena, F

    R. Saxena, F. Keller, MovieSum: An abstractive summarization dataset for movie screenplays, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Findings of the Association for Computational Linguistics: ACL 2024, Association for Computational Linguistics, Bangkok, Thailand, 2024, pp. 4043–4050. URL: https://aclanthology.org/2024.findings-acl.239/. doi:10.18653/...

    work page doi:10.18653/v1/2024.findings-acl.239 2024

  18. [18]

    Edwindra, GitHub - ledwindra/nber: Scrape and analyzie NBER working papers

    L. Edwindra, GitHub - ledwindra/nber: Scrape and analyzie NBER working papers. — github.com, https://github.com/ledwindra/nber, 2025. [Accessed 17-05-2026]

    2025

  19. [19]

    Yelp, Yelp Dataset — kaggle.com, https://www.kaggle.com/datasets/yelp-dataset/yelp-dataset,

  20. [20]

    [Accessed 28-05-2026]

    Anthropic, Introducing Claude Haiku 4.5 — anthropic.com, https://www.anthropic.com/news/ claude-haiku-4-5, 2025. [Accessed 28-05-2026]

    2025

  21. [21]

    [Accessed 28-05-2026]

    Anthropic, Introducing Sonnet 4.6 — anthropic.com, https://www.anthropic.com/news/ claude-sonnet-4-6, 2026. [Accessed 28-05-2026]

    2026

  22. [22]

    [Accessed 28-05-2026]

    OpenAI, Introducing GPT-5.4, https://openai.com/index/introducing-gpt-5-4/, 2026. [Accessed 28-05-2026]

    2026

  23. [23]

    GLM-5-Team, :, A. Zeng, X. Lv, Z. Hou, Z. Du, Q. Zheng, B. Chen, D. Yin, C. Ge, C. Huang, C. Xie, C. Zhu, C. Yin, C. Wang, G. Pan, H. Zeng, H. Zhang, H. Wang, H. Chen, J. Zhang, J. Jiao, J. Guo, J. Wang, J. Du, J. Wu, K. Wang, L. Li, L. Fan, L. Zhong, M. Liu, M. Zhao, P. Du, Q. Dong, R. Lu, Shuang-Li, S. Cao, S. Liu, T. Jiang, X. Chen, X. Zhang, X. Huang,...

    Pith/arXiv arXiv 2026

  24. [24]

    MiniMax, :, A. Chen, A. Li, B. Zhou, B. Gong, B. Jiang, B. Dan, C. Yu, C. Wang, C. Ma, C. Zhong, C. Zhu, C. Xiao, C. Yang, C. Du, C. Zhang, C. Zhang, C. Huang, C. Zhang, C. Du, C. Zhao, C. Guo, D. Chen, D. Ding, D. Sun, D. Zhang, E. Yang, F. Yu, G. Zheng, G. Zheng, G. Li, H. Zhu, H. Zhou, H. Zhang, H. Ding, H. Zhang, H. Sun, H. Lyu, H. Lu, H. Wang, H. Shi...

  25. [25]

    I. Labs, S. Khanna, S. Kharbanda, S. Li, H. Varma, E. Wang, S. Birnbaum, Z. Luo, Y. Miraoui, A. Palrecha, S. Ermon, A. Grover, V. Kuleshov, Mercury: Ultra-fast language models based on diffusion, 2025. URL: https://arxiv.org/abs/2506.17298.arXiv:2506.17298

    Pith/arXiv arXiv 2025

  26. [26]

    [Accessed 28-05-2026]

    OpenAI, Introducing GPT-5.4 mini and nano — openai.com, https://openai.com/index/ introducing-gpt-5-4-mini-and-nano/, 2026. [Accessed 28-05-2026]

    2026

  27. [27]

    K. Team, T. Bai, Y. Bai, Y. Bao, S. H. Cai, Y. Cao, Y. Charles, H. S. Che, C. Chen, G. Chen, H. Chen, J. Chen, J. Chen, J. Chen, J. Chen, K. Chen, L. Chen, R. Chen, X. Chen, Y. Chen, Y. Chen, Y. Chen, Y. Chen, Y. Chen, Y. Chen, Y. Chen, Y. Chen, Z. Chen, Z. Chen, D. Cheng, M. Chu, J. Cui, J. Deng, M. Diao, H. Ding, M. Dong, M. Dong, Y. Dong, Y. Dong, A. D...

    Pith/arXiv arXiv 2026

  28. [28]

    [Accessed 28-05-2026]

    TogetherAI, Together AI | The AI Native Cloud — together.ai, https://www.together.ai/, 2026. [Accessed 28-05-2026]

    2026

  29. [29]

    FireworksAI, Fireworks AI - Fastest Inference for Generative AI — fireworks.ai, https://fireworks.ai/,

  30. [30]

    Tolstykh, A

    I. Tolstykh, A. Tsybina, S. Yakubson, M. Kuprashevich, Llmtrace: A corpus for classification and fine-grained localization of ai-written text, arXiv preprint arXiv:2509.21269 (2025)

    arXiv 2025

  31. [31]

    Gugliotta, L

    C. Gugliotta, L. La Cava, A. Tagarelli, Personagen: A persona-driven open-ended machine- generated text dataset, in: Proceedings of the 34th ACM International Conference on Information and Knowledge Management, 2025, pp. 6397–6401

    2025

  32. [32]

    [Accessed 28-05-2026]

    2026

  33. [33]

    A. Song, Y. Cheng, Y. Xu, R. Feng, Maga-bench: Machine-augment-generated text via alignment detection benchmark, arXiv preprint arXiv:2601.04633 (2026)

    Pith/arXiv arXiv 2026

  34. [34]

    J. Wu, R. Zhan, D. F. Wong, S. Yang, X. Yang, Y. Yuan, L. S. Chao, Detectrl: Benchmarking llm- generated text detection in real-world scenarios, Advances in Neural Information Processing Systems 37 (2024) 100369–100401

    2024

  35. [35]

    [Accessed 17-05-2026]

    HuggingFaceGECLM, HuggingFaceGECLM/StackExchange_Mar2023 ·Datasets at Hug- ging Face — huggingface.co, https://huggingface.co/datasets/{H}ugging{F}ace{G}{E}{C}{L}{M}/ {S}tack{E}xchange_{M}ar2023, 2023. [Accessed 17-05-2026]

    2023

  36. [36]

    L. L. Cava, D. Costa, A. Tagarelli, Is contrasting all you need? contrastive learning for the detection and attribution of ai-generated text, in: U. Endriss, F. S. Melo, K. Bach, A. J. B. Diz, J. M. Alonso- Moral, S. Barro, F. Heintz (Eds.), ECAI 2024 - 27th European Conference on Artificial Intelligence, 19-24 October 2024, Santiago de Compostela, Spain ...

    2024

  37. [37]

    Bevendorff, D

    J. Bevendorff, D. Dementieva, M. Fröbe, B. Gipp, A. Greiner-Petter, J. Karlgren, M. Mayerl, P. Nakov, A. Panchenko, M. Potthast, et al., Overview of pan 2025: Generative ai detection, multilingual text detoxification, multi-author writing style analysis, and generative plagiarism detection, in: European Conference on Information Retrieval, Springer, 2025,...

    2025

  38. [38]

    Z. Wang, M. Xiong, J. Lian, Z. Dou, Reasoning-aware aigc detection via alignment and reinforce- ment, 2026. URL: https://arxiv.org/abs/2604.19172.arXiv:2604.19172

    Pith/arXiv arXiv 2026

  39. [39]

    R. Roy, N. Imanpour, A. Aziz, S. Bajpai, G. Singh, S. Biswas, K. Wanaskar, P. Patwa, S. Ghosh, S. Dixit, N. R. Pal, V. Rawte, R. Garimella, G. Jena, A. Sheth, V. Sharma, A. N. Reganti, V. Jain, A. Chadha, A. Das, A comprehensive dataset for human vs. ai generated text detection, 2026. URL: https://arxiv.org/abs/2510.22874.arXiv:2510.22874

    Pith/arXiv arXiv 2026

  40. [40]

    Tripathi, EndLessTime/SAGE·Datasets at Hugging Face — huggingface.co, https://huggingface

    A. Tripathi, EndLessTime/SAGE·Datasets at Hugging Face — huggingface.co, https://huggingface. co/datasets/{E}nd{L}ess{T}ime/{S}{A}{G}{E}, 2025. [Accessed 17-05-2026]

    2025

  41. [41]

    Hamborg, N

    F. Hamborg, N. Meuschke, C. Breitinger, B. Gipp, news-please: A generic news crawler and extractor, in: Proceedings of the 15th International Symposium of Information Science, 2017, pp. 218–223. doi:10.5281/zenodo.4120316

    doi:10.5281/zenodo.4120316 2017

  42. [42]

    Artemova, J

    E. Artemova, J. Lucas, S. Venkatraman, J. Lee, S. Tilga, A. Uchendu, V. Mikhailov, Beemo: Bench- mark of expert-edited machine-generated outputs, 2025. URL: https://arxiv.org/abs/2411.04032. arXiv:2411.04032

    Pith/arXiv arXiv 2025

  43. [43]

    Y. Chen, J. Chen, G. Mo, X. Chen, B. He, X. Han, L. Sun, CoCoNUTS: Concentrating on Content while Neglecting Uninformative Textual Styles for AI-Generated Peer Review Detection, 2025. URL: https://arxiv.org/abs/2509.04460.arXiv:2509.04460

    Pith/arXiv arXiv 2025

  44. [44]

    Gromadzki, A

    M. Gromadzki, A. Wróblewska, A. Kaliska, On the effectiveness of llm-specific fine-tuning for detecting ai-generated text, 2026. URL: https://arxiv.org/abs/2601.20006.arXiv:2601.20006

    arXiv 2026

  45. [45]

    Ben Allal, A

    L. Ben Allal, A. Lozhkov, G. Penedo, T. Wolf, L. von Werra, Cosmopedia, 2024. URL: https:// huggingface.co/datasets/HuggingFaceTB/cosmopedia

    2024

  46. [46]

    [Accessed 18-05-2026]

    OriginalityAI, GitHub - OriginalityAI/AI-detector-research-tool — github.com, https://github.com/ {O}riginality{A}{I}/{A}{I}-detector-research-tool, 2023. [Accessed 18-05-2026]

    2023

  47. [47]

    Conover, M

    M. Conover, M. Hayes, A. Mathur, J. Xie, J. Wan, S. Shah, A. Ghodsi, P. Wendell, M. Zaharia, R. Xin, Free dolly: Introducing the world’s first truly open instruction-tuned llm, 2023. URL: https://www. databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm

    2023

  48. [48]

    Jabarian, A

    B. Jabarian, A. Imas, Artificial writing and automated detection, Technical Report, National Bureau of Economic Research, 2025

    2025

  49. [49]

    Bhargava, A

    P. Bhargava, A. Drozd, A. Rogers, Generalization in nli: Ways (not) to go beyond simple heuristics, 2021.arXiv:2110.01518

    Pith/arXiv arXiv 2021

  50. [50]

    X. Zhu, Y. Ren, Y. Cao, X. Lin, F. Fang, Y. Li, Reliably bounding false positives: A zero-shot machine- generated text detection framework via multiscaled conformal prediction, in: W. Che, J. Nabende, E. Shutova, M. T. Pilehvar (Eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Associat...

    2025

  51. [51]

    Esposito, Blitz - bayesian layers in torch zoo (a bayesian deep learing library for torch), https: //github.com/piEsposito/blitz-bayesian-deep-learning/, 2020

    P. Esposito, Blitz - bayesian layers in torch zoo (a bayesian deep learing library for torch), https: //github.com/piEsposito/blitz-bayesian-deep-learning/, 2020

    2020

  52. [52]

    Blundell, J

    C. Blundell, J. Cornebise, K. Kavukcuoglu, D. Wierstra, Weight uncertainty in neural networks,

  53. [53]

    I. Turc, M. Chang, K. Lee, K. Toutanova, Well-read students learn better: The impact of student initialization on knowledge distillation, CoRR abs/1908.08962 (2019). URL: http://arxiv.org/abs/ 1908.08962.arXiv:1908.08962

    Pith/arXiv arXiv 1908

  54. [54]

    org/api/libauc.losses.html#libauc.losses.auc.tp{A}{U}{C}_{K}{L}_{L}oss, ???? [Accessed 28-06-2026]

    LibAUC, libauc.losses &#x2014; libauc 1.0.0 documentation — docs.libauc.org, https://docs.libauc. org/api/libauc.losses.html#libauc.losses.auc.tp{A}{U}{C}_{K}{L}_{L}oss, ???? [Accessed 28-06-2026]

    2026

  55. [55]

    D. Zhu, G. Li, B. Wang, X. Wu, T. Yang, When auc meets dro: Optimizing partial auc for deep learning with non-convex convergence guarantee, in: International Conference on Machine Learning, PMLR, 2022, pp. 27548–27573

    2022

  56. [56]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, ...

    2020

  57. [57]

    Swayamdipta, R

    S. Swayamdipta, R. Schwartz, N. Lourie, Y. Wang, H. Hajishirzi, N. A. Smith, Y. Choi, Dataset cartography: Mapping and diagnosing datasets with training dynamics, in: B. Webber, T. Cohn, Y. He, Y. Liu (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Association for Computational Linguistics, Online, 2...

    doi:10.18653/v1/2020.emnlp-main.746 2020

  58. [61]

    [Accessed 29-05-2026]

    Lambda, The Superintelligence Cloud | Lambda — lambda.ai, https://lambda.ai/, 2026. [Accessed 29-05-2026]

    2026

  59. [2015]

    URL: https://arxiv.org/abs/1505.05424.arXiv:1505.05424

    Pith/arXiv arXiv

  60. [2022]

    [Accessed 17-05-2026]

    2026

  61. [2026]

    URL: https://arxiv.org/abs/2605.26494.arXiv:2605.26494

    Pith/arXiv arXiv