Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Fast segmentation of watermarked texts from large language models through an epidemic change-point framework

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Treating watermarked spans as epidemic change-points, WISER provably localizes them all in linear time.

desk verdict Strong algorithm and a new multi-segment consistency bound, but the i.i.d. pivot assumption doesn't cover the hashing schemes used in the experiments; fix the scope and this is publishable. read the letter →

arxiv 2509.21160 v2 pith:4AW2NW2M submitted 2025-09-25 stat.ML cs.LGstat.ME

Soham Bonnerjee , Subhrajyoty Roy , Sayar Karmakar This is my paper · ORCID
classification stat.MLcs.LGstat.ME MSC 62G1062G2062M10
keywords watermarksegmentationepidemicchange-pointslargelanguagemodelspivotstatisticsmultiplechange-pointdetectionlinear-timealgorithmmixed-sourcetextconsistencyguarantees
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that locating watermarked segments in mixed human-machine text is the same statistical problem as detecting an epidemic change-point: inside a watermarked stretch the pivot statistics of tokens shift upward, and outside it they return to baseline. On that footing it proposes WISER, a block-screening, discarding, enlargement, and local re-estimation pipeline that runs in $O(n)$ time and, under minimum-length and separation conditions, simultaneously finds the number of watermarked segments and estimates their boundaries. The main theorem says that with block length proportional to $\sqrt{n}$ and a threshold set to the $\alpha$-quantile of the null maximum block sum, the probability that WISER returns the true segment count and boundaries within error $M_\epsilon/d$ approaches at least $1-\epsilon$ as text length grows. Because existing segmentation methods are slow and lack guarantees, a provable linear-time localizer would make watermark localization practical for long documents and across watermarking schemes.

What carries the argument

The machinery is the epidemic change-point estimator for irregular signals, applied to pivot statistics $X_t = h(Y_t)$, where a pivot statistic is a score function of the token and the watermark key whose law is the same for every token when no watermark is present. Watermarked tokens are assumed to raise the mean of $X_t$ by at least $d$ over the unwatermarked level $\mu_0$, while the mean inside the segment may vary. The paper adapts an existing change-point estimator for non-constant means to the epidemic setting as the interval minimizing the sum of $(X_k - \mu_0 - \rho\tilde d)$ outside the interval, and proves its boundary error is controlled by the signal strength. WISER wraps this estimator in a blocking stage: partition into about $\sqrt{n}$ blocks, keep blocks whose sums exceed the $\alpha$-quantile of the null maximum, discard connected components shorter than $c\sqrt{n}\log n$, enlarge survivors, and restrict each local estimator to a search window of width about $\sqrt{n}$. The block-level scan bypasses arbitrary dependence among watermarked pivot statistics, which is why the consistency proof uses cumulant-generating-function bounds and an envelope argument rather than standard change-point inequalities.

What would settle it

Run WISER on synthetic mixed texts with $n$ growing, two watermarked segments of length $\lfloor C_0 n^{1/2+\gamma'}/\log n \rfloor$ separated by the same order, block length $b = \lceil\sqrt{n}\rceil$, watermark strength $d$ fixed, and $Q$ the $0.95$-quantile of the null maximum block sum. If the empirical frequency of $\{\hat K=K$ and boundary error below $M_\epsilon/d\}$ does not approach at least $0.95$ as $n$ grows, then one of Assumption 3.1, the cumulant condition (3.2), or the elevated-alternatives assumption fails and the theorem's prediction is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that watermark segmentation can be solved by an epidemic change-point estimator adapted to irregular signals, giving the first $O(n)$ algorithm with consistency guarantees for multiple watermarked intervals. For a single segment, the estimator that minimizes the biased out-of-interval sum has boundary error $O(1/\tilde d)$, where $\tilde d$ is a chosen signal-strength parameter, and $O(1/\tilde d^2)$ under local sub-Gaussianity. For multiple segments, WISER partitions the text into about $\sqrt{n}$ blocks and keeps only those whose pivot-sum exceeds the $\alpha$-quantile of the null maximum block sum; it then discards short spurious clusters, enlarges the survivors, estimates the signal strength, restricts each search to a local window, and re-runs the single-segment estimator. Theorem 3.2 says that under an elevated-alternatives assumption, a cumulant-gap condition, and a minimum-length and separation assumption, the estimated number of segments equals the true $K$ and each estimated boundary differs from the truth by at most $M_\epsilon/d$ with probability at least $1-\epsilon$ in the limit. Thus stronger watermarks yield sharper boundaries, and the number of segments need not be known in advance.

Load-bearing premise

The guarantee depends on each true watermarked segment being at least about $n^{1/2+\gamma'}/\log n$ tokens long and separated from neighbouring segments by at least the same length; shorter or adjacent watermarked spans fall outside the theorem, and the paper offers no fallback guarantee for them.

Editorial extensions

If this is right

  • A long mixed-source document can be scanned for all machine-generated stretches in one pass, since WISER runs in $O(n)$ time.
  • The user does not need to specify the number of watermarked segments; the estimated count matches the true count with probability approaching 1 under the theorem's conditions.
  • Boundary error shrinks as watermark strength $d$ grows, so stronger watermarks give both easier detection and more precise localization.
  • The same pivot statistics that power detection tests are enough for localization, so WISER applies across Gumbel, inverse-transform, red-green, and permute-and-flip watermarking schemes.
  • The single-segment rate $O(1/\tilde d)$, sharpened to $O(1/\tilde d^2)$ under local sub-Gaussianity, quantifies how choosing a larger signal-strength parameter $\tilde d$ trades off against boundary accuracy.

Reading between the lines

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

  • If the epidemic reading is right, any signal whose mean is elevated only on intervals, not just watermarks, could be segmented by the same blocking-plus-local-refinement pipeline; machine-generated spans detected by perplexity or log-rank scores without a secret key are a natural candidate.
  • Because the theorem controls boundary error, the total length of the union of estimated segments should consistently estimate the watermark proportion, connecting WISER to proportion-estimation problems in mixed AI-human text.
  • The paper's asymmetry critique implies that comparisons relying only on the Rand Index may overstate methods that return overly large watermarked regions; future evaluations should report a modified index such as MRI.
  • A stress test on segments shorter than the $C_0 n^{1/2+\gamma'}/\log n$ threshold is not covered by the guarantee; a multi-scale block choice might extend the pipeline, but that is an open possibility rather than a claim of the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper frames watermark localization in LLM-generated text as an epidemic change-point problem. The authors introduce pivot statistics for unwatermarked tokens and model watermarked segments as intervals where the mean of h(Y_t) is elevated by at least d (Assumption 2.2). They propose WISER, which performs block-level thresholding with a quantile Q of the null maximum block sum, discards short selected runs, enlarges the remaining intervals, estimates the signal strength dtilde from the enlarged intervals, and then applies a localized version of the single-segment estimator of Eq. (3.1) inside restricted search windows. Theorem 3.1 gives an O_P(dtilde^{-1}) boundary error for a single segment, and Theorem 3.2 gives consistency for multiple segments under minimum length/separation and a cumulant condition (3.2). Experiments on Gumbel, Inverse Transform, Red-Green, and Permute-and-Flip watermarks with several LLMs report higher IOU/F1/MRI and lower runtime than Aligator, SeedBS-NOT, and Waterseeker.

Significance. If the theoretical claims were fully established for the schemes tested, this would be a useful contribution: WISER is simple, runs in O(n), accommodates multiple segments, and the paper provides detailed proofs, extensive benchmarks, and a sensible modified Rand index. The finite-sample probability bounds and the explicit discussion of tuning parameters are strengths. However, the central i.i.d. pivot-statistics lemma (Lemma 2.2) is proved only under a stronger assumption than Assumption 2.1, and the hash-based watermarking used in the experiments produces dependent key variables. Until that gap is closed or the claims are restricted, the 'complete theoretical guarantees' advertised in Section 1.1 are not established for the empirical setting. The significance is therefore conditional on a fixable but load-bearing re-scoping or a new dependence analysis.

major comments (3)
  1. [Section 2.1, Lemma 2.2 and Appendix D.3] The assertion that {Y_t}_{t in S} are i.i.d. is not a consequence of Assumption 2.1. Assumption 2.1 only states that for an unwatermarked token, omega_t and zeta_t are independent conditional on omega_{1:t-1}; it does not state that the zeta_t are i.i.d. or independent of the history. The proof in Appendix D.3 explicitly uses 'the key zeta_t that are i.i.d.' and the pivot law L(Y) as the conditional law given omega_{1:t}. For the hash-based watermarking used in Section 4 and Appendix A, zeta_t = A(omega_{(t-m):(t-1)}, Key) is a deterministic function of previous tokens, so the conditional laws vary and the zeta_t are dependent. Consequently the i.i.d. null block sums used to select Q, the binomial count in Proposition 5, and the martingale-difference property of epsilon_t in the proof of Theorem D.1 (around Eqs. (D.4)-(D.7)) are not justified for those schemes. The theory currently covers only the case of genuinely i.i.d. secret keys whose distribution does not depend on the text history; please either prove the needed statements under a suitable dependence condition (e.g., a conditional pivot-law bound) or explicitly restrict the theoretical claims and present the Section 4 results as empirical evidence outside the proven regime.
  2. [Theorem 3.2 and Appendix C] Theorem 3.2 assumes that the null distribution of the pivot statistics is absolutely continuous with respect to Lebesgue measure. The paper does not verify this for the Red-Green and Inverse Transform watermarking setups used in the experiments, where the natural score functions are bounded and potentially lattice-valued (e.g., green-token counts). If the null law is discrete, the calibration condition P0(max_k S_k > Q)=alpha and the exceedance-count argument in Proposition 5 require a different treatment. Please either state which pivot statistics h(Y) are used for each scheme and verify the absolute continuity assumption, or extend the result to discrete laws (or to a suitable smoothing argument).
  3. [Condition (3.2) and Assumption 3.1] Condition (3.2) is the main quantitative control on the watermarked distribution used in Theorem 3.2, but it is not verified for any of the schemes in Appendix B; the Donsker-Varadhan remark is heuristic and does not establish the inequality for Gumbel or Inverse Transform watermarks. Since Section 1.1 claims 'complete theoretical guarantees in the most general case,' the paper should either verify (3.2) for the examples or state it as an unverified condition limiting the theorem's scope. Relatedly, Assumption 3.1 excludes adjacent or shorter-than-C0 n^{1/2+gamma'}/log n segments, so the 'most general case' phrasing overstates the coverage of Theorem 3.2.
minor comments (5)
  1. [Assumption 3.1] Assumption 3.1 uses d(I_k, I_{k-1}) for k=1, but the distance d is defined only for two disjoint intervals in Section 3.2; please define d(I_1, I_0) explicitly (e.g., as infinity or as the distance from the segment to the text boundary).
  2. [Step 3 of the proof of Theorem D.2] After choosing rho=(8 nu)^{-1}, the argument yields d >= 2 rho dtilde only with possible equality; to match the strict inequality d > 2 rho dtilde used in the proof, choose a slightly smaller rho, for instance rho=(16 nu)^{-1}.
  3. [Title and Appendix C.1.3] The title contains the apparent fragment 'REGION-A'; also Appendix C.1.3 spells 'meta-Ilama' instead of 'Meta-Llama'.
  4. [Equation (3.2)] Equation (3.2) writes E_{1,p}[exp(-theta X)] with a lowercase p; please use P to match the class notation used elsewhere.
  5. [Sections 1 and 5] The apparent circularity of estimating dtilde from screened intervals is addressed in Step 3 of the proof of Theorem D.2, where dtilde is shown to concentrate in [tau d, 4 nu d]; the main unresolved issue is not circularity but the i.i.d. assumption underlying the screening stage. The introduction's 'complete theoretical guarantees' phrasing should be tempered to acknowledge the limitations stated in Appendix A and Section 5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; WISER's derivation is self-contained given its assumptions, and the main caveat is a missing assumption rather than a circular reduction.

full rationale

The WISER derivation is self-contained and does not reduce to its inputs. The final estimator in (3.1) is adapted from the external irregular-signal change-point estimator of Kley et al. (2024), not from a self-citation. The only data-dependent nuisance parameter is dtilde, estimated in Step 28 by averaging (X_t - mu0) over the screening-stage intervals D_j. In Appendix D.2 (Step 3), the paper proves that, conditional on high-probability events A_n, B_n, E_n, F_n, dtilde concentrates in [tau d, 4 nu d]; Step 4 then applies Theorem 3.1 uniformly over that interval and obtains the consistency bound with the true d. The final error bound is therefore not obtained by tuning dtilde to force the estimates, and the theorem's probability statement is not identical by construction to any fitted quantity. The citations of Li et al. (2025a; 2024a) are for the pivot-statistic framework and human-edit reconstruction, and the sole self-citation (Bonnerjee et al., 2025) is cited alongside Bai (1994) only as a reference for Hajek-Renyi-type techniques; it is not load-bearing. The one substantive caveat is a missing assumption, not a circular one: Lemma 2.2's proof asserts 'the key zeta_t that are i.i.d.', although Assumption 2.1 states only conditional independence of omega_t and zeta_t given omega_{1:(t-1)}, and hashing-based watermarks (zeta_t = A(omega_{(t-m):(t-1)}, Key)) are not i.i.d. Appendix A admits that the general-m setting 'requires specific attention' and Section 5 lists human-edit robustness as future work. This could invalidate the claimed coverage of the experimental setting, but it is an assumption gap rather than a self-referential reduction, so the circularity score remains 0.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central guarantee is conditional on a stack of assumptions: verifier key access (2.1), elevated mean shift (2.2), minimum segment length and separation (3.1), a Donsker-Varadhan condition (3.2), and several moment and regularity conditions. Assumptions 2.2 and 3.1 are the most domain-specific; 3.2 is introduced for this proof. The data-driven dtilde is a fitted quantity but is controlled by concentration arguments rather than used as a free fit to the final error.

free parameters (6)
  • Block size b = b=65 in accuracy experiments; ceil(sqrt(n)) in runtime experiments
    First-stage screening resolution; theory requires b approximately sqrt(n), exact value is hand-chosen.
  • rho (rho) = 0.5 in accuracy experiments, 0.1 in runtime experiments; paper recommends 0.1 to 0.5
    Controls the bias term in estimator (3.1); theory requires d > 2 rho dtilde.
  • gamma (gamma) = 0.1
    Controls enlargement of detected intervals; theory requires 0 < gamma << 1/2.
  • alpha (alpha) = 0.05
    Sets the block threshold Q as the (1-alpha) quantile of the null max block sum.
  • discarding constant c = not specified numerically
    Connected intervals shorter than c sqrt(n) log n are discarded; the paper does not give a concrete rule for choosing c.
  • dtilde (estimated signal strength) = sample mean of (Xt - mu0) over enlarged intervals; data-dependent
    Estimated from detected enlarged intervals; the theorem proves it lies in [tau d, 4 nu d] with high probability, so it is not an independent free constant.
assumptions (7)
  • domain assumption Assumption 2.1: for any text, verifier has pseudo-random variables zeta_t such that unwatermarked tokens satisfy independence of omega_t and zeta_t conditional on omega_{1:(t-1)}.
    Load-bearing: it gives i.i.d. pivot statistics under the null (Lemma 2.2). The paper admits this is an idealization of hash-based reconstruction and uses the m=1 simplification.
  • domain assumption Assumption 2.2: inf_{P in P} E_{1,P}[h(Y)] >= mu0 + d for some d > 0.
    Central signal model: without a uniform mean elevation, the epidemic mean-shift interpretation fails. The paper illustrates it for Gumbel and inverse transform watermarks under restricted P, not for all schemes.
  • domain assumption Assumption 3.1: min_k {|I_k| and d(I_k, I_{k-1})} >= C0 n^{1/2+gamma'}/log n.
    Needed for the screening stage to separate segments and discard spurious blocks; short or adjacent watermarked segments invalidate the consistency guarantee.
  • ad hoc to paper Condition (3.2): kappa = inf_{theta>=0} [theta(mu0 + tau d) + log sup_P E_{1,P}[exp(-theta X)]] < 0.
    A Donsker-Varadhan-style strengthening of Assumption 2.2 used to concentrate the data-driven estimator of d and to control block-threshold behavior; it is a new condition invented for this proof.
  • ad hoc to paper Absolute continuity of the null distribution of the pivot statistics (Theorem 3.2).
    The theorem requires this, but the experiments include Red-Green watermarking whose typical pivot statistics are discrete; the paper does not reconcile this gap.
  • standard math Exponential moment, compactness, and nonzero variance conditions: sup_P E_{1,P}[exp(eta|epsilon|)] < infinity, P closed and compact, Var0(epsilon) > 0, sup_P Var1,P(epsilon) > 0.
    Technical regularity conditions for Theorem 3.1's proof via cumulant generating functions and Danskin's theorem; not domain-specific.
  • standard math E0[|X - mu0|^{3+delta}] < infinity, sup_P E1,P[X] < infinity, and bounded number of segments K.
    Technical assumptions for Theorem 3.2's block-threshold and concentration arguments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast segmentation of watermarked texts from large language models through an epidemic change-point framework." pith.science (2026). https://pith.science/paper/4AW2NW2M

@misc{pith2026250921160,
  author       = {Pith},
  title        = {Pith review of: Fast segmentation of watermarked texts from large language models through an epidemic change-point framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AW2NW2M}},
  note         = {Machine review of arXiv:2509.21160}
}
read the original abstract

With the growing use of large language models, concerns over content authenticity have spurred a variety of watermarking schemes. These schemes use secret keys to detect machine-generated text while remaining imperceptible to readers. Detection typically reduces to statistical hypothesis testing for the presence of watermarks, a topic that is now well studied. In contrast, the finer-grained task of localizing which segments of a text are watermarked is much less explored; existing approaches often lack scalability or guarantees robust to paraphrasing and post-editing. We bring a new perspective to this segmentation problem through the lens of epidemic change-points and, by exploiting this connection, propose WISER, a novel and computationally efficient watermark segmentation algorithm. We establish finite-sample error bounds and consistency for detecting multiple watermarked segments in a single text. Complementing these theoretical results, our extensive numerical experiments show that WISER outperforms state-of-the-art baseline methods, both in terms of computational speed as well as accuracy, on various benchmark datasets embedded with diverse watermarking schemes. Together, these theoretical and empirical results position WISER as an effective tool for watermark localization and illustrate how classical statistical ideas can yield theoretically valid and computationally efficient solutions to a modern problem of immediate importance.

Figures

Figures reproduced from arXiv: 2509.21160 by the authors.

Figure 1
Figure 1. (Left) A mixed source text with watermarked tokens [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (Left): The Algorithm WISER; (Right) WISER in action with key steps. this, κ lifts the minimum separation between the un-watermarked and watermarked distributions into a gap between the cumulant functions, and can therefore be understood to be mild. Equation (3.2) establishes a weak uniform control over the behavior of pivot statistics under watermarked segments. This allows us to rigorously bypass the possibly arbi… view at source ↗
Figure 3
Figure 3. Time complexity (seconds) for various algorithms as a function of completion lengths (n). Y-axis is in log-scale, with 95% confidence interval shown in shades. an epidemic change-point problem, we bridged a novel connection between classical statistical theory and a modern challenge in generative AI, and also designed a linear time algorithm with provable consistency guarantees, which were further confirmed by our e… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Effect on performance metrics (IOU and Rand Index) due to modification of the hyper [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 42 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Watermarking of large language models

    Scott Aaronson. Watermarking of large language models. https://simons.berkeley.edu/talks/scottaaronson-ut-austin-openai-2023-08-17, August 2023. Talk at the Simons Institute for the Theory of Computing

    work page 2023

  3. [3]

    Least squares estimation of a shift in linear processes

    Jushan Bai. Least squares estimation of a shift in linear processes. J. Time Ser. Anal., 15 0 (5): 0 453--472, 1994. ISSN 0143-9782,1467-9892. doi:10.1111/j.1467-9892.1994.tb00204.x. URL https://doi.org/10.1111/j.1467-9892.1994.tb00204.x

    arXiv 1994

  4. [4]

    Openai, google, others pledge to watermark ai content for safety, white house says

    Diane Bartz and Krystal Hu. Openai, google, others pledge to watermark ai content for safety, white house says. https://www.reuters.com/technology/openai-google-others-pledge-watermark-ai-content-safety-white-house-2023-07-21/, 2023. Accessed: 2023-10-03

    work page 2023

  5. [5]

    On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pp.\ 610--623, 2021

    Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pp.\ 610--623, 2021

    2021

  6. [6]

    Joseph R. Biden. Fact sheet: President biden issues executive order on safe, secure, and trustworthy artificial intelligence. https://bidenwhitehouse.archives.gov/briefing-room/statements-releases/2023/10/30/fact-sheet-president-biden-issues-executive-order-on-safe-secure-and-trustworthy-artificial-intelligence/, October 2023. The White House, October 30, 2023

    work page 2023

  7. [7]

    Testing synchronization of change-points for multiple time series

    Soham Bonnerjee, Sayar Karmakar, Maggie Cheng, and Wei Biao Wu. Testing synchronization of change-points for multiple time series. Preprint, 2025. URL https://sohamb01.github.io/drafts/test-of-synchronization.pdf

    work page 2025

  8. [8]

    A statistical hypothesis testing framework for data misappropriation detection in large language models

    Yinpeng Cai, Lexin Li, and Linjun Zhang. A statistical hypothesis testing framework for data misappropriation detection in large language models. arXiv preprint arXiv:2501.02441, 2025

    work page arXiv 2025

Show all 63 references
  1. [9]

    Towards better statistical understanding of watermarking llms

    Zhongze Cai, Shang Liu, Hanzhao Wang, Huaiyang Zhong, and Xiaocheng Li. Towards better statistical understanding of watermarking llms. arXiv preprint arXiv:2403.13027, 2024

    2024 arXiv

  2. [10]

    Can llm-generated misinformation be detected? arXiv preprint arXiv:2309.13788, 2023

    Canyu Chen and Kai Shu. Can llm-generated misinformation be detected? arXiv preprint arXiv:2309.13788, 2023

    2023 arXiv

  3. [11]

    Detecting change-points in epidemic models

    Zhenmin Chen, Zihao Li, and Min Zhou. Detecting change-points in epidemic models. Journal of advanced statistics, 1 0 (4): 0 181, 2016

    2016

  4. [12]

    Undetectable watermarks for language models

    Miranda Christ, Sam Gunn, and Or Zamir. Undetectable watermarks for language models. In The Thirty Seventh Annual Conference on Learning Theory, pp.\ 1125--1139. PMLR, 2024

    2024

  5. [13]

    Machine-generated text: A comprehensive survey of threat models and detection methods

    Evan N Crothers, Nathalie Japkowicz, and Herna L Viktor. Machine-generated text: A comprehensive survey of threat models and detection methods. IEEE Access, 11: 0 70977--71002, 2023

    2023

  6. [14]

    Limit theorems in change-point analysis

    Mikl \'o s Cs \"o rg \"o and Lajos Horv \'a th. Limit theorems in change-point analysis. 1997

    1997

  7. [15]

    John M. Danskin. The theory of max-min and its application to weapons allocation problems, volume V of Econometrics and Operations Research. Springer-Verlag New York, Inc., New York, 1967

    1967

  8. [16]

    Measuring association on topological spaces using kernels and geometric graphs

    Nabarun Deb, Promit Ghosal, and Bodhisattva Sen. Measuring association on topological spaces using kernels and geometric graphs. arXiv preprint arXiv:2010.01768, 2020

    2010 arXiv

  9. [17]

    Donsker and S

    Monroe D. Donsker and S. R. S. Varadhan. Asymptotic evaluation of certain markov process expectations for large time. IV . Comm. Pure Appl. Math., 36 0 (2): 0 183--212, 1983. ISSN 0010-3640. doi:10.1002/cpa.3160360204. URL https://doi.org/10.1002/cpa.3160360204

    1983 doi

  10. [18]

    Three bricks to consolidate watermarks for large language models

    Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. Three bricks to consolidate watermarks for large language models. In 2023 IEEE international workshop on information forensics and security (WIFS), pp.\ 1--6. IEEE, 2023

    2023

  11. [19]

    Probability inequalities for sums of independent random variables

    D Kh Fuk and Sergey V Nagaev. Probability inequalities for sums of independent random variables. Theory of Probability & Its Applications, 16 0 (4): 0 643--660, 1971

    1971

  12. [20]

    Model equality testing: Which model is this API serving? In The Thirteenth International Conference on Learning Representations, 2025

    Irena Gao, Percy Liang, and Carlos Guestrin. Model equality testing: Which model is this API serving? In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=QCDdI7X3f9

    2025

  13. [21]

    Gltr: Statistical detection and visualization of generated text

    Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019

    1906 arXiv

  14. [22]

    Edit distance robust watermarks via indexing pseudorandom codes

    Noah Golowich and Ankur Moitra. Edit distance robust watermarks via indexing pseudorandom codes. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=FZ45kf5pIA

    2024

  15. [23]

    H\' a jek and A

    J. H\' a jek and A. R\' e nyi. Generalization of an inequality of K olmogorov. Acta Math. Acad. Sci. Hungar., 6: 0 281--283, 1955. ISSN 0001-5954,1588-2632. doi:10.1007/BF02024392. URL https://doi.org/10.1007/BF02024392

    1955 doi

  16. [24]

    Hall and C

    P. Hall and C. C. Heyde. Martingale limit theory and its application. Probability and Mathematical Statistics. Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York-London, 1980. ISBN 0-12-319350-8

    1980

  17. [25]

    Spotting llms with binoculars: Zero-shot detection of machine-generated text

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Spotting llms with binoculars: Zero-shot detection of machine-generated text. arXiv preprint arXiv:2401.12070, 2024

    2024 arXiv

  18. [26]

    Unbiased watermark for large language models

    Zhengmian Hu, Lichang Chen, Xidong Wu, Yihan Wu, Hongyang Zhang, and Heng Huang. Unbiased watermark for large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=uWVC5FVidc

    2024

  19. [27]

    Towards optimal statistical watermarking

    Baihe Huang, Banghua Zhu, Hanlin Zhu, Jason Lee, Jiantao Jiao, and Michael Jordan. Towards optimal statistical watermarking. In Socially Responsible Language Modelling Research, 2023. URL https://openreview.net/forum?id=Fc2FaS9mYJ

    2023

  20. [28]

    Estimators for epidemic alternatives

    Marie Hu s kov \'a . Estimators for epidemic alternatives. Commentationes Mathematicae Universitatis Carolinae, 36 0 (2): 0 279--291, 1995

    1995

  21. [29]

    Use of cumulative sums of squares for retrospective detection of changes of variance

    Carmen Incl \'a n and George C Tiao. Use of cumulative sums of squares for retrospective detection of changes of variance. Journal of the American Statistical Association, 89 0 (427): 0 913--923, 1994. doi:10.1080/01621459.1994.10476824

    1994

  22. [30]

    A watermark for large language models

    John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. In International Conference on Machine Learning, pp.\ 17061--17084. PMLR, 2023

    2023

  23. [31]

    On the reliability of watermarks for large language models

    John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. On the reliability of watermarks for large language models. In The Twelfth International Conference on Learning Representati...

    2024

  24. [32]

    Change-point analysis with irregular signals

    Tobias Kley, Yuhan Philip Liu, Hongyuan Cao, and Wei Biao Wu. Change-point analysis with irregular signals. The Annals of Statistics, 52 0 (6): 0 2913--2930, 2024

    2024

  25. [33]

    Robust distortion-free watermarks for language models

    Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. Robust distortion-free watermarks for language models. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=FpaCL1MO2C

    2024

  26. [34]

    Detecting fake content with relative entropy scoring

    Thomas Lavergne, Tanguy Urvoy, and Fran c ois Yvon. Detecting fake content with relative entropy scoring. Pan, 8 0 (27-31): 0 4, 2008

    2008

  27. [35]

    The cusum test of homogeneity with an application in spontaneous abortion epidemiology

    Bruce Levin and Jennie Kline. The cusum test of homogeneity with an application in spontaneous abortion epidemiology. Statistics in Medicine, 4 0 (4): 0 469--488, 1985

    1985

  28. [36]

    Robust detection of watermarks for large language models under human edits

    Xiang Li, Feng Ruan, Huiyuan Wang, Qi Long, and Weijie J Su. Robust detection of watermarks for large language models under human edits. arXiv preprint arXiv:2411.13868, 2024 a

    2024 arXiv

  29. [37]

    A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules

    Xiang Li, Feng Ruan, Huiyuan Wang, Qi Long, and Weijie J Su. A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules. The Annals of Statistics, 53 0 (1): 0 322--351, 2025 a

    2025

  30. [38]

    Optimal estimation of watermark proportions in hybrid ai-human texts

    Xiang Li, Garrett Wen, Weiqing He, Jiayuan Wu, Qi Long, and Weijie J Su. Optimal estimation of watermark proportions in hybrid ai-human texts. arXiv preprint arXiv:2506.22343, 2025 b

    2025 arXiv

  31. [39]

    Segmenting watermarked texts from language models

    Xingchi Li, Guanxun Li, and Xianyang Zhang. Segmenting watermarked texts from language models. Advances in Neural Information Processing Systems, 37: 0 14634--14665, 2024 b

    2024

  32. [40]

    Monitoring ai-modified content at scale: A case study on the impact of chatgpt on ai conference peer reviews

    Weixin Liang, Zachary Izzo, Yaohui Zhang, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, et al. Monitoring ai-modified content at scale: A case study on the impact of chatgpt on ai conference peer reviews. arXiv preprint arXiv:2403.07...

    2024 arXiv

  33. [41]

    Adaptive text watermark for large language models

    Yepeng Liu and Yuheng Bu. Adaptive text watermark for large language models. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024

    2024

  34. [42]

    Architecture of a fake news detection system combining digital watermarking, signal processing, and machine learning

    David Meg \' as, Minoru Kuribayashi, Andrea Rosales, Krzysztof Cabaj, and Wojciech Mazurczyk. Architecture of a fake news detection system combining digital watermarking, signal processing, and machine learning. Journal of Wireless Mobile Networks, Ubiquitous Computing, and De...

    2022

  35. [43]

    Large language models challenge the future of higher education

    Silvia Milano, Joshua A McGrane, and Sabina Leonelli. Large language models challenge the future of higher education. Nature Machine Intelligence, 5 0 (4): 0 333--334, 2023

    2023

  36. [44]

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

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International conference on machine learning, pp.\ 24950--24962. PMLR, 2023

    2023

  37. [45]

    Empirical likelihood ratio test for the epidemic change model

    Wei Ning, Junvie Pailden, and Arjun Gupta. Empirical likelihood ratio test for the epidemic change model. Journal of Data science, 10 0 (1): 0 107--127, 2012

    2012

  38. [46]

    Leyi Pan, Aiwei Liu, Yijian LU, Zitian Gao, Yichen Di, Shiyu Huang, Lijie Wen, Irwin King, and Philip S. Yu. Waterseeker: Pioneering efficient detection of watermarked segments in large documents. In AAAI 2025 Workshop on Preventing and Detecting LLM Misinformation (PDLM), 202...

    2025

  39. [47]

    A more efficient algorithm to compute the rand index for change-point problems

    Lucas de Oliveira Prates. A more efficient algorithm to compute the rand index for change-point problems. arXiv preprint arXiv:2112.03738, 2021

    2021 arXiv

  40. [48]

    Evaluating durability: Benchmark insights into multimodal watermarking

    Jielin Qiu, William Han, Xuandong Zhao, Shangbang Long, Christos Faloutsos, and Lei Li. Evaluating durability: Benchmark insights into multimodal watermarking. CoRR, abs/2406.03728, 2024. URL https://doi.org/10.48550/arXiv.2406.03728

    2024 doi arXiv

  41. [49]

    H \"o lder norm test statistics for epidemic change

    Alfredas Ra c kauskas and Charles Suquet. H \"o lder norm test statistics for epidemic change. Journal of statistical planning and inference, 126 0 (2): 0 495--520, 2004

    2004

  42. [50]

    Testing epidemic changes of infinite dimensional parameters

    Alfredas Ra c kauskas and Charles Suquet. Testing epidemic changes of infinite dimensional parameters. Statistical Inference for Stochastic Processes, 9 0 (2): 0 111--134, 2006

    2006

  43. [51]

    Robust speech recognition via large-scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pp.\ 28492--28518. PMLR, 2023

    2023

  44. [52]

    Zero-shot statistical tests for llm-generated text detection using finite sample concentration inequalities

    Tara Radvand, Mojtaba Abdolmaleki, Mohamed Mostagir, and Ambuj Tewari. Zero-shot statistical tests for llm-generated text detection using finite sample concentration inequalities. arXiv preprint arXiv:2501.02406, 2025

    2025

  45. [53]

    Release strategies and the social impacts of language models

    Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019

    1908 arXiv

  46. [54]

    Deep kernel relative test for machine-generated text detection

    Yiliao Song, Zhenqiao Yuan, Shuhai Zhang, Zhen Fang, Jun Yu, and Feng Liu. Deep kernel relative test for machine-generated text detection. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  47. [55]

    Detect LLM : Leveraging log rank information for zero-shot detection of machine-generated text

    Jinyan Su, Terry Yue Zhuo, Di Wang, and Preslav Nakov. Detect LLM : Leveraging log rank information for zero-shot detection of machine-generated text. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. URL https://openreview.net/forum?id=Dy2mbQIdMz

    2023

  48. [56]

    Howkgpt: Investigating the detection of chatgpt-generated university student homework through context-aware perplexity analysis

    Christoforos Vasilatos, Manaar Alam, Talal Rahwan, Yasir Zaki, and Michail Maniatakos. Howkgpt: Investigating the detection of chatgpt-generated university student homework through context-aware perplexity analysis. arXiv preprint arXiv:2305.18226, 2023

    2023 arXiv

  49. [57]

    Ai is tearing wikipedia apart, May 2023

    Claire Woodcock. Ai is tearing wikipedia apart, May 2023. URL https://www.vice.com/en/article/ai-is-tearing-wikipedia-apart/. Accessed: 2025-09-14

    2023

  50. [58]

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

    Yihan Wu, Zhengmian Hu, Junfeng Guo, Hongyang Zhang, and Heng Huang. A resilient and accessible distribution-preserving watermark for large language models. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024

    2024

  51. [59]

    Tests for change-points with epidemic alternatives

    Qiwei Yao. Tests for change-points with epidemic alternatives. Biometrika, 80 0 (1): 0 179--191, 1993

    1993

  52. [60]

    Provable robust watermarking for AI -generated text

    Xuandong Zhao, Prabhanjan Vijendra Ananth, Lei Li, and Yu-Xiang Wang. Provable robust watermarking for AI -generated text. In The Twelfth International Conference on Learning Representations, 2024 a . URL https://openreview.net/forum?id=SsmT8aO45L

    2024

  53. [61]

    Efficiently identifying watermarked segments in mixed-source texts

    Xuandong Zhao, Chenwen Liao, Yu-Xiang Wang, and Lei Li. Efficiently identifying watermarked segments in mixed-source texts. In Neurips Safe Generative AI Workshop 2024, 2024 b

    2024

  54. [62]

    Permute-and-flip: An optimally stable and watermarkable decoder for LLM s

    Xuandong Zhao, Lei Li, and Yu-Xiang Wang. Permute-and-flip: An optimally stable and watermarkable decoder for LLM s. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=YyVVicZ32M

    2025

  55. [63]

    Duwak: Dual watermarks in large language models

    Chaoyi Zhu, Jeroen Galjaard, Pin-Yu Chen, and Lydia Chen. Duwak: Dual watermarks in large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Findings of the Association for Computational Linguistics: ACL 2024, pp.\ 11416--11436, Bangkok, Thailand, August...

    2024 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.