Pith. sign in

REVIEW 2 major objections 4 minor 56 references

LLMs learn a lasting aversion to token self-repetition early in training; measuring that aversion detects their text.

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 · grok-4.5

2026-07-11 21:56 UTC pith:4LWPHD5P

load-bearing objection Simple single-pass detector that is competitive with Binoculars on modern targets, with clean early-training and locality evidence that is actually new. the 2 major comments →

arxiv 2607.04061 v1 pith:4LWPHD5P submitted 2026-07-05 cs.CL cs.AIcs.LGstat.ML

Telescope: Improving Zero Shot Detection of LLM Generated Content By Measuring Token Repetition Probability

Christopher Nassif , Josh F. Cooper This is my paper
classification cs.CL cs.AIcs.LGstat.ML
keywords LLM-generated text detectionzero-shot detectiontoken repetitionvestigial heuristicTelescope Perplexitypre-training dynamicsblack-box detectors
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.

Language models are trained to sound human, yet the authors argue that early training leaves a lasting statistical scar. While they are still learning simple next-token bigrams, models develop an unusually strong aversion to a token repeating itself. That bias never fully disappears; it remains as a “vestigial heuristic” that is more strongly activated by text the models themselves produce than by human writing. Telescope Perplexity is a simple score that probes exactly this residual: for each token it asks how surprised a reference model is to see that same token again given the prefix up to and including it. Across many datasets (including new ones generated by GPT-4o Mini and DeepSeek-V3), reference models, and common perturbations, the score ranks LLM text above human text at state-of-the-art or competitive rates while needing only a single forward pass. The practical claim is that a developmental artifact of pre-training can be turned into a lightweight, zero-shot detector that stays useful even as new models appear.

Core claim

The paper claims that the extreme local aversion to token self-repetition that language models acquire in the earliest, bigram-like phase of pre-training persists as a vestigial heuristic and is differentially activated by LLM-generated versus human text. Measuring that residual with Telescope Perplexity therefore separates the two classes of text with high accuracy in a zero-shot setting.

What carries the argument

Telescope Perplexity (Eq. 1): the average negative log probability that a reference model assigns to each token given the sequence up to and including that token, P(s_i | s_1:i). It isolates the model’s learned repetition probability rather than ordinary next-token perplexity.

Load-bearing premise

That the strong early aversion to a token repeating itself is never fully unlearned and remains a detectable signature across essentially all current model families and training mixes.

What would settle it

If a modern model family, or a late-stage training checkpoint of an existing family, produced text whose Telescope Perplexity distribution fully overlapped that of human writing of the same genre, the claimed vestigial signature would be absent.

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

If this is right

  • A single-forward-pass detector can match or beat multi-model and multi-pass zero-shot methods on many contemporary target models.
  • Thresholds still need domain-specific calibration; a score tuned on one genre can lose accuracy when transferred to another.
  • Genres that force deliberate repetition (poetry, some news styles) can suppress the signature and produce confident false negatives.
  • The same local-repetition probe can also separate synthetic versus human portions of a model’s own training corpus.

Where Pith is reading between the lines

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

  • If the vestigial signal is truly fixed early, detectors built on it may remain useful longer than supervised classifiers that must be retrained for every new model family.
  • Training interventions that deliberately re-introduce or penalize self-repetition later in pre-training could be tested as a way to erase the signature and thereby evade this class of detectors.
  • The same early-local-bias logic may apply to other simple statistical regularities (e.g., local n-gram entropy) that form before long-range dependencies are mastered.

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 / 4 minor

Summary. The paper introduces Telescope Perplexity (Eq. 1), the average negative log probability that a reference model assigns to immediate self-repetition of each token given the prefix that includes that token. The authors hypothesize that early pre-training instills a strong local aversion to token repetition that persists as a “Vestigial Heuristic” and is differentially activated by LLM-generated versus human text. They support the developmental claim with training-checkpoint curves (Pythia, Amber), bigram/trigram locality ablations (Table 1), and training-data separability experiments, then show that the metric yields competitive or superior zero-shot AUROC against Perplexity, DetectLLM-LRR, Binoculars and Fast-DetectGPT across twelve reference models, public and newly generated modern-target datasets (GPT-4o Mini, DeepSeek-V3), length/perturbation/humanizer ablations, and transferability tests, all with a single forward pass.

Significance. If the empirical ranking holds, Telescope Perplexity supplies a simple, single-pass, reference-model-agnostic detector that matches or exceeds the current zero-shot SOTA (Binoculars) while roughly halving compute relative to two-model methods and avoiding the multi-pass cost of perturbation detectors. The accompanying training-dynamics and locality evidence is a novel mechanistic contribution that links early bigram statistics to a practical detection signal. Strengths include extensive multi-model/multi-dataset evaluation with confidence intervals, public code, newly generated modern-target sets, and explicit discussion of domain sensitivity and threshold calibration. These elements make the work immediately useful for practitioners and a useful probe for future interpretability studies of early training biases.

major comments (2)
  1. Table 2 averages AUROC over twelve reference models of which six are SmolLM/SmolLM2 variants; the authors themselves note the resulting architectural bias. Because the headline claim of “state-of-the-art or competitive” performance rests on these averages, the main text should also report (or prominently highlight) the non-SmolLM subset averages and the per-model spreads already present in Appendix 8.20, so that readers can judge robustness without the over-representation of one family.
  2. Section 5.4 and Table 3 show that transferability F1 can drop substantially when the threshold is tuned on other domains. The practical recommendations correctly call for domain-specific calibration, yet the abstract and introduction still present Telescope Perplexity as a ready-to-deploy zero-shot detector. A short, explicit statement of the calibration requirement (and the expected degradation under pure transfer) should appear in the abstract or early results so that the efficiency claim is not overstated relative to methods that may be more threshold-stable.
minor comments (4)
  1. Eq. (1) is called “Telescope Perplexity” yet is not exponentiated; a one-sentence clarification that the monotonic transform does not affect ranking-based metrics would remove a possible source of confusion.
  2. Figure 2 and Appendix 8.13 show early plateauing of the signature; adding a brief quantitative summary (e.g., fraction of final value reached by 10 % of training) would make the “early emergence” claim easier to cite.
  3. The poetry/stylized-text failure mode (Appendix 8.7) is interesting; a short pointer in the main-text limitations section would help practitioners anticipate the genre-specific risk already quantified for news writing.
  4. Minor typographical inconsistencies appear in the arXiv identifier formatting and a few table captions; a light copy-edit pass would polish the presentation.

Circularity Check

0 steps flagged

No circularity: Telescope Perplexity is an independently defined single-pass statistic; detection power and the vestigial-heuristic interpretation are both evaluated against external labels and checkpoints.

full rationale

The paper defines Telescope Perplexity (Eq. 1) solely from a reference model’s conditional probabilities M(s_i | s_1:i) and never fits any free parameter of that definition to detection labels. Detection is ordinary thresholding of the resulting scalar; AUROC/F1 are measured on external human/LLM corpora (including newly generated GPT-4o-mini and DeepSeek-V3 sets). The supporting “vestigial-heuristic” story is tested with independent evidence—early-training checkpoints of Pythia and Amber (Fig. 2, App. 8.13), bigram/trigram ablations (Table 1), and separability of SmolLM’s own FineWeb vs. Cosmopedia subsets (App. 8.3)—none of which reuse the detection labels. Thresholds are either optimized on held-out data or transferred across datasets (Table 3). No equation, uniqueness claim, or self-citation reduces the reported AUROC rankings to a quantity fitted on the same labels used for evaluation. The work is therefore free of the six circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The central empirical claim rests on standard language-model assumptions plus one interpretive hypothesis (vestigial heuristics) that is tested rather than assumed. No free parameters are fitted inside the metric itself; classification thresholds are ordinary operating-point choices. The only invented conceptual entity is the named “Vestigial Heuristic.”

free parameters (2)
  • classification threshold τ
    Chosen either by maximizing F1 on the test set or by logistic regression on all other datasets (transferability protocol). Ordinary detector operating point; does not enter the definition of Telescope Perplexity.
  • minimum text length filter (100 words)
    Arbitrary but fixed preprocessing choice applied uniformly; justified by commercial practice and statistical reliability of short texts.
axioms (3)
  • domain assumption Early in pre-training, language models behave approximately as bigram (or low-order n-gram) predictors and therefore learn strong local statistics such as token non-repetition.
    Invoked in Section 3 and supported by citations to Belrose et al. (2024) and Choshen et al. (2022); used to motivate why a repetition-aversion signal should appear early.
  • ad hoc to paper Statistical biases acquired in the earliest training phase are not fully unlearned later and remain detectable in the final model’s token probabilities.
    Core of the “Vestigial Heuristic” hypothesis (Section 3); tested via checkpoint curves but not derived from first principles.
  • domain assumption Reference models drawn from different families and sizes share sufficiently similar early-training biases that the same scalar probe works across them.
    Implicit in the multi-reference-model experimental design (Section 4.2) and the generality claims of Section 3.3.
invented entities (1)
  • Vestigial Heuristic independent evidence
    purpose: Name and conceptual frame for the hypothesized persistent early-training aversion to token repetition that Telescope Perplexity is designed to probe.
    Introduced in the abstract and Section 3; the paper supplies empirical correlates (early rise and plateau of the score, locality, training-data separability) but the entity itself is a new interpretive construct.

pith-pipeline@v1.1.0-grok45 · 89767 in / 2632 out tokens · 41579 ms · 2026-07-11T21:56:59.398220+00:00 · methodology

0 comments
read the original abstract

Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge. While LLMs are trained to write like humans, we hypothesize that this training leaves an indelible mark. LLMs develop a particularly strong aversion to token repetition very early in training. This bias persists as a ''Vestigial Heuristic'' (a developmental artifact) that is activated in LLM-generated text, separating LLM from human writing. To probe this phenomenon, we introduce Telescope Perplexity, a metric that evaluates the token repetition of the model, $P(s_i | s_{1:i})$ . Our empirical investigation reveals that the Telescope Perplexity signature emerges early in pre-training, and Telescope Perplexity empirically enables highly effective zero-shot LLM detection. We show state-of-the-art or competitive performance across diverse datasets (including modern evaluation sets we introduce), reference models, and perturbation schemes with greater efficiency than other methods.

Figures

Figures reproduced from arXiv: 2607.04061 by Christopher Nassif, Josh F. Cooper.

Figure 1
Figure 1. Figure 1: An overview of the “Vestigial Heuristic” hypothesis and measuring the average token repetition probability for the Telescope Perplexity. Telescope Perplexity measures the reference language model’s learned likelihood of outputting the last token it processed given its context up to that point. For a language model M and a token sequence ⃗s = (s1, ..., sL) of length L, the Telescope Perplexity is defined as… view at source ↗
Figure 2
Figure 2. Figure 2: Telescope Perplexity evaluated on text generated by Pythia-2.8B model checkpoints throughout training. Note the early stabilization of the Telescope Perplexity. emergence and subsequent persistence strongly support the hypothesis that Telescope Perplexity captures a “Vestigial” characteristic established during the foundational learning phase, rather than a property that evolves continuously with model cap… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of minimum text length on the AUROC perfor￾mance of several detectors on the Detect LLM Text dataset (top) and the HC3 Plus dataset (bottom). 5.2. Robustness [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of paraphrasing sentences (top) and swapping words with synonyms (bottom) on detector AUROC performance on the Ghostbusters dataset. generated text passed off as human). Practitioners who care about this asymmetry should set thresholds against a tar￾get false-negative rate on stylized genres specifically, rather than against the data set-averaged decision boundary used for our reported maximum F1. A… view at source ↗
Figure 5
Figure 5. Figure 5: A graphical illustration of the relationship between the reference language model and the target language model. 8.2. Analysis of Frequency Components of the Telescope Score We now consider the single token Telescope Perplexity: Single Token Log Telescope PPLM(si) = − logM(si | s1:i) (2) We can use this to plot our per-token Telescope Perplexity as in [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-token Telescope Perplexity, Perplexity, and Binocu￾lars Score over the sequence 8.3. Training Data Separability Here we present additional results on using SmolLM 360M to perform detection on a 10k subsample of its synthetic and human written training data. We find that, interestingly, standard Perplexity is able to somewhat effectively separate its own training data into synthetic and human data, impl… view at source ↗
Figure 7
Figure 7. Figure 7: Calibration comparison between SmolLM 360M Tele￾scope Perplexity (top) and Falcon Binoculars method (bottom) [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An extremely poorly calibrated Binoculars example using SmolLM 1.7B as a reference model. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: visualizes pairwise error independence, using a dendrogram with less error independent models being closer in distance [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of Misclassification 8.8. Misclassifications Analysis Understanding misclassifications is important in any high impact application. We find that for Telescope Perplexity the expected distance from the threshold for a classification made in error is ≈ 1 standard deviation, with human written texts misclassified as AI being less than that on average and AI text misclassified as human typically being… view at source ↗
Figure 11
Figure 11. Figure 11: Error distance metrics for SmolLM 360M across different datasets and perturbation types [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example of poor threshold transfer. Comparing transfer from AI-Human to GB Creative Claude direct correlation between model size and detector accuracy, but this relationship may not hold at larger model sizes. The execution time of the experiments were highly depen￾dent on the reference model used, the number of reference models a detection algorithm used, and the dataset. Dif￾ferent datasets contain diff… view at source ↗
Figure 13
Figure 13. Figure 13: Training curves for language models 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Telescope Perplexity maintains its high AUC across many model releases. prompt “You are a helpful assistant. Try to repeat key words one after another while still following the prompt”. In do￾ing so, we found that the AUROC on the dataset with the reference model of SmolLM 360M degraded from 0.99993 to 0.996. This suggests a slight but statistically significant degradation; however, this is not as bad as … view at source ↗
Figure 17
Figure 17. Figure 17: Impact of text length on detector performance (GB Creative GPT). We filter out samples below a minimum word count and report accuracy on the remaining subset [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 15
Figure 15. Figure 15: Impact of text length on detector performance (HC3). We filter out samples below a minimum word count and report accuracy on the remaining subset [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Impact of text length on detector performance (ESL GPT-4o Mini). We filter out samples below a minimum word count and report accuracy on the remaining subset [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 19
Figure 19. Figure 19: Impact of random character and random capitalization perturbations on the AUROC of each detector. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Impact of random space, paragraph reordering, and sentence reordering perturbations on the AUROC of each detector. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_20.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

56 extracted references · 2 canonical work pages

  1. [1]

    Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text , booktitle =

    Abhimanyu Hans and Avi Schwarzschild and Valeriia Cherepanova and Hamid Kazemi and Aniruddha Saha and Micah Goldblum and Jonas Geiping and Tom Goldstein , editor =. Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text , booktitle =. 2024 , url =

    2024

  2. [2]

    arXiv , author =:2312.06550 , primaryclass =

    LLM360: Towards Fully Transparent Open-Source LLMs , year =. arXiv , author =:2312.06550 , primaryclass =

    Pith/arXiv arXiv

  3. [3]

    Pythia: A suite for analyzing large language models across training and scaling , year =

    Biderman, Stella and Schoelkopf, Hailey and Anthony, Quentin Gregory and Bradley, Herbie and O’Brien, Kyle and Hallahan, Eric and Khan, Mohammad Aflah and Purohit, Shivanshu and Prashanth, USVSN Sai and Raff, Edward and others , booktitle =. Pythia: A suite for analyzing large language models across training and scaling , year =

  4. [4]

    Ghostbuster: Detecting Text Ghostwritten by Large Language Models , booktitle =

    Vivek Verma and Eve Fleisig and Nicholas Tomlin and Dan Klein , editor =. Ghostbuster: Detecting Text Ghostwritten by Large Language Models , booktitle =. 2024 , url =. doi:10.18653/V1/2024.NAACL-LONG.95 , timestamp =

    doi:10.18653/v1/2024.naacl-long.95 2024

  5. [5]

    arXiv , author =:2412.16525 , primaryclass =

    Investigating Efficacy of Perplexity in Detecting LLM-Generated Code , url =. arXiv , author =:2412.16525 , primaryclass =

    Pith/arXiv arXiv

  6. [6]

    Manning and Chelsea Finn , editor =

    Eric Mitchell and Yoonho Lee and Alexander Khazatsky and Christopher D. Manning and Chelsea Finn , editor =. DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature , booktitle =. 2023 , url =

    2023

  7. [7]

    arXiv , author =:2311.16867 , primaryclass =

    The Falcon Series of Open Language Models , url =. arXiv , author =:2311.16867 , primaryclass =

    Pith/arXiv arXiv

  8. [8]

    Loubna Ben Allal and Anton Lozhkov and Elie Bakouch , title =

  9. [9]

    arXiv , author =:2302.13971 , primaryclass =

    LLaMA: Open and Efficient Foundation Language Models , url =. arXiv , author =:2302.13971 , primaryclass =

    Pith/arXiv arXiv

  10. [10]

    arXiv , author =:2307.09288 , primaryclass =

    Llama 2: Open Foundation and Fine-Tuned Chat Models , url =. arXiv , author =:2307.09288 , primaryclass =

    Pith/arXiv arXiv

  11. [11]

    arXiv , author =:2407.21783 , primaryclass =

    The Llama 3 Herd of Models , url =. arXiv , author =:2407.21783 , primaryclass =

    Pith/arXiv arXiv

  12. [12]

    arXiv , author =:2408.00118 , primaryclass =

    Gemma 2: Improving Open Language Models at a Practical Size , url =. arXiv , author =:2408.00118 , primaryclass =

    Pith/arXiv arXiv

  13. [13]

    arXiv , author =:2310.15264 , primaryclass =

    Towards Possibilities and Impossibilities of AI-generated Text Detection: A Survey , url =. arXiv , author =:2310.15264 , primaryclass =

    Pith/arXiv arXiv

  14. [14]

    2026 , eprint=

    A Training-free Method for LLM Text Attribution , author=. 2026 , eprint=

    2026

  15. [15]

    Approximation by superpositions of a sigmoidal function , volume =

    Cybenko, G , journal =. Approximation by superpositions of a sigmoidal function , volume =

  16. [16]

    Can linguists distinguish between ChatGPT/AI and human writing?: A study of research ethics and academic publishing , url =

    J Elliott Casal and Matt Kessler , doi =. Can linguists distinguish between ChatGPT/AI and human writing?: A study of research ethics and academic publishing , url =. Research Methods in Applied Linguistics , month =

  17. [17]

    Brenner , editor =

    Kristina Radivojevic and Nicholas Clark and Paul R. Brenner , editor =. LLMs Among Us: Generative. Proceedings of the. 2024 , url =. doi:10.1609/AAAISS.V3I1.31202 , timestamp =

    work page doi:10.1609/aaaiss.v3i1.31202 2024

  18. [18]

    DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text , booktitle =

    Jinyan Su and Terry Yue Zhuo and Di Wang and Preslav Nakov , editor =. DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text , booktitle =. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-EMNLP.827 , timestamp =

    doi:10.18653/v1/2023.findings-emnlp.827 2023

  19. [19]

    arXiv , author =:2412.19437 , primaryclass =

    DeepSeek-V3 Technical Report , url =. arXiv , author =:2412.19437 , primaryclass =

    Pith/arXiv arXiv

  20. [20]

    arXiv , author =:2410.21276 , primaryclass =

    GPT-4o System Card , url =. arXiv , author =:2410.21276 , primaryclass =

    Pith/arXiv arXiv

  21. [21]

    How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection , year =

    Guo, Biyang and Zhang, Xin and Wang, Ziyuan and Jiang, Minqi and Nie, Jinran and Ding, Yuxuan and Yue, Jianwei and Wu, Yupeng , journal =. How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection , year =

  22. [22]

    arXiv , author =:2309.02731 , primaryclass =

    HC3 Plus: A Semantic-Invariant Human ChatGPT Comparison Corpus , url =. arXiv , author =:2309.02731 , primaryclass =

    Pith/arXiv arXiv

  23. [23]

    LLM - Detect AI Generated Text , year =

    Jules King and Perpetual Baffour and Scott Crossley and Ryan Holbrook and Maggie Demkin , howpublished =. LLM - Detect AI Generated Text , year =

  24. [24]

    AI Vs Human Text , year =

    Shayan Gerami , howpublished =. AI Vs Human Text , year =

  25. [25]

    The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , title =

    Guilherme Penedo and Hynek Kydl. The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , title =

  26. [26]

    Science Advances , volume =

    Dmitry Kobak and Rita González-Márquez and Emőke-Ágnes Horvát and Jan Lause , title =. Science Advances , volume =. 2025 , doi =. https://www.science.org/doi/pdf/10.1126/sciadv.adt3813 , abstract =

    doi:10.1126/sciadv.adt3813 2025

  27. [27]

    arXiv , author =:2405.17247 , primaryclass =

    An Introduction to Vision-Language Modeling , url =. arXiv , author =:2405.17247 , primaryclass =

    Pith/arXiv arXiv

  28. [28]

    Fu and Stefano Ermon and Atri Rudra and Christopher R

    Tri Dao and Daniel Y. Fu and Stefano Ermon and Atri Rudra and Christopher R. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness , booktitle =. 2022 , url =

    2022

  29. [29]

    2015 , eprint=

    Distilling the Knowledge in a Neural Network , author=. 2015 , eprint=

    2015

  30. [30]

    2024 , eprint=

    A Survey on Mixture of Experts , author=. 2024 , eprint=

    2024

  31. [31]

    arXiv , author =:2203.15556 , primaryclass =

    Training Compute-Optimal Large Language Models , url =. arXiv , author =:2203.15556 , primaryclass =

    Pith/arXiv arXiv

  32. [32]

    arXiv , author =:2409.17416 , primaryclass =

    From Deception to Detection: The Dual Roles of Large Language Models in Fake News , url =. arXiv , author =:2409.17416 , primaryclass =

    Pith/arXiv arXiv

  33. [33]

    Detecting

    Liyanage, Vijini and Buscaldi, Davide and Forcioli, Penelope , booktitle =. Detecting

  34. [34]

    An Empirical Study to Understand How Students Use ChatGPT for Writing Essays , booktitle =

    Andrew Jelson and Daniel Manesh and Alice Jang and Daniel Dunlap and Young. An Empirical Study to Understand How Students Use ChatGPT for Writing Essays , booktitle =. 2026 , url =. doi:10.1145/3772318.3791056 , timestamp =

    work page doi:10.1145/3772318.3791056 2026

  35. [35]

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

    Yafu Li and Qintong Li and Leyang Cui and Wei Bi and Zhilin Wang and Longyue Wang and Linyi Yang and Shuming Shi and Yue Zhang , editor =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2024 , url =. doi:10.18653/V1/2024.ACL-LONG.3 , timestamp =

    doi:10.18653/v1/2024.acl-long.3 2024

  36. [36]

    Patterns , volume =

    Weixin Liang and Mert Y. Patterns , volume =. 2023 , url =. doi:10.1016/J.PATTER.2023.100779 , timestamp =

    doi:10.1016/j.patter.2023.100779 2023

  37. [37]

    Feedback Prize - English Language Learning , year =

    Alex Franklin and Maggie and Meg Benner and Natalie Rambis and Perpetual Baffour and Ryan Holbrook and Scott Crossley and ulrichboser , howpublished =. Feedback Prize - English Language Learning , year =

  38. [38]

    The Grammar-Learning Trajectories of Neural Language Models , booktitle =

    Leshem Choshen and Guy Hacohen and Daphna Weinshall and Omri Abend , editor =. The Grammar-Learning Trajectories of Neural Language Models , booktitle =. 2022 , url =. doi:10.18653/V1/2022.ACL-LONG.568 , timestamp =

    doi:10.18653/v1/2022.acl-long.568 2022

  39. [39]

    Fern , editor =

    Nora Belrose and Quintin Pope and Lucia Quirke and Alex Mallen and Xiaoli Z. Fern , editor =. Neural Networks Learn Statistics of Increasing Complexity , booktitle =. 2024 , url =

    2024

  40. [40]

    Large Scale Structure of Neural Network Loss Landscapes , booktitle =

    Stanislav Fort and Stanislaw Jastrzebski , editor =. Large Scale Structure of Neural Network Loss Landscapes , booktitle =. 2019 , url =

    2019

  41. [41]

    A python library for confidence intervals , year =

    Jacob Gildenblat , howpublished =. A python library for confidence intervals , year =

  42. [42]

    Applied Intelligence , year =

    Takahashi, Kanae and Yamamoto, Kouji and Kuchiba, Aya and Koyama, Tatsuki , title =. Applied Intelligence , year =

  43. [43]

    Fast Implementation of DeLong’s Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves , volume =

    Sun, Xu and Xu, Weichao , doi =. Fast Implementation of DeLong’s Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves , volume =. IEEE Signal Processing Letters , keywords =

  44. [44]

    DeLong and David M

    Elizabeth R. DeLong and David M. DeLong and Daniel L. Clarke-Pearson , issn =. Comparing the Areas under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach , url =. Biometrics , number =

  45. [45]

    Zoom In: An Introduction to Circuits , volume =

    Olah, Chris and Cammarata, Nick and Schubert, Ludwig and Goh, Gabriel and Petrov, Michael and Carter, Shan , doi =. Zoom In: An Introduction to Circuits , volume =. Distill , month =

  46. [46]

    A Practical Examination of AI-Generated Text Detectors for Large Language Models , booktitle =

    Brian Tufts and Xuandong Zhao and Lei Li , editor =. A Practical Examination of AI-Generated Text Detectors for Large Language Models , booktitle =. 2025 , url =. doi:10.18653/V1/2025.FINDINGS-NAACL.271 , timestamp =

    doi:10.18653/v1/2025.findings-naacl.271 2025

  47. [47]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics,

    Yuxia Wang and Jonibek Mansurov and Petar Ivanov and Jinyan Su and Artem Shelmanov and Akim Tsvigun and Chenxi Whitehouse and Osama Mohammed Afzal and Tarek Mahmoud and Toru Sasaki and Thomas Arnold and Alham Fikri Aji and Nizar Habash and Iryna Gurevych and Preslav Nakov , editor =. Proceedings of the 18th Conference of the European Chapter of the Associ...

    2024

  48. [48]

    Tavan, Ehsan and Najafi, Maryam , title =

  49. [49]

    Peters and Mark Neumann and Mohit Iyyer and Matt Gardner and Christopher Clark and Kenton Lee and Luke Zettlemoyer , editor =

    Matthew E. Peters and Mark Neumann and Mohit Iyyer and Matt Gardner and Christopher Clark and Kenton Lee and Luke Zettlemoyer , editor =. Deep Contextualized Word Representations , booktitle =. 2018 , url =. doi:10.18653/V1/N18-1202 , timestamp =

    doi:10.18653/v1/n18-1202 2018

  50. [50]

    Automatic Detection of Generated Text is Easiest when Humans are Fooled , booktitle =

    Daphne Ippolito and Daniel Duckworth and Chris Callison. Automatic Detection of Generated Text is Easiest when Humans are Fooled , booktitle =. 2020 , url =. doi:10.18653/V1/2020.ACL-MAIN.164 , timestamp =

    doi:10.18653/v1/2020.acl-main.164 2020

  51. [51]

    The Twelfth International Conference on Learning Representations,

    Guangsheng Bao and Yanbin Zhao and Zhiyang Teng and Linyi Yang and Yue Zhang , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  52. [52]

    The Twelfth International Conference on Learning Representations,

    Xianjun Yang and Wei Cheng and Yue Wu and Linda Ruth Petzold and William Yang Wang and Haifeng Chen , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  53. [53]

    The Thirteenth International Conference on Learning Representations,

    Christopher Ackerman and Nina Panickssery , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  54. [54]

    A Survey on

    Wu, Junchao and Yang, Shu and Zhan, Runzhe and Yuan, Yulin and Chao, Lidia Sam and Wong, Derek Fai , doi =. A Survey on. Computational Linguistics , month =

  55. [55]

    Long Ouyang and Jeffrey Wu and Xu Jiang and Diogo Almeida and Carroll L. Wainwright and Pamela Mishkin and Chong Zhang and Sandhini Agarwal and Katarina Slama and Alex Ray and John Schulman and Jacob Hilton and Fraser Kelton and Luke Miller and Maddie Simens and Amanda Askell and Peter Welinder and Paul F. Christiano and Jan Leike and Ryan Lowe , editor =...

    2022

  56. [56]

    Ben Allal, Loubna and Lozhkov, Anton and Penedo, Guilherme and Wolf, Thomas and von Werra, Leandro , title =