Pith. sign in

REVIEW 3 major objections 3 minor 42 references

The paper proves that a single pivot sequence, Z_t = -log(1-U_{t,W_t}), turns Gumbel-max watermark detection into a one-pass, anytime-valid sequential test.

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 06:13 UTC pith:HEF3U222

load-bearing objection A sound but incremental one-pass e-process for Gumbel-max watermark detection; the math checks out under a pivot assumption that needs a rigorous proof for keyed reconstruction. the 3 major comments →

arxiv 2607.21958 v1 pith:HEF3U222 submitted 2026-07-24 stat.ML cs.LG

Efficient Online LLM Watermark Detection via Rao-Blackwellized E-Processes

Lu Luo , Dandan Mo , Chengdong Xu , Ting Li , Jinhan Xie , Huiqiong Li , Niansheng Tang This is my paper
classification stat.ML cs.LG MSC 62L1062F03
keywords LLM watermarkinge-processanytime-valid inferencesequential testingGumbel-max watermarkRao-Blackwellizationonline detectionoptional stopping
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's goal is an online detector for LLM watermarking: one that reads tokens as they are generated and can declare 'watermarked' at any moment with a statistically guaranteed false-positive rate. It proves that for the Gumbel-max watermark, the sequence M_t(λ) = (1-λ)^t exp(λ S_t), built from the pivot Z_t = -log(1-U_{t,W_t}), is an e-process, meaning its running maximum has time-uniform Type I error control. That turns watermark detection into a one-pass sequential test with constant memory, valid under arbitrary optional stopping. The paper also proves that under watermarking the same process grows exponentially in expectation, so the test eventually rejects with probability one, and it provides plug-in and mixture versions that choose the tilt λ adaptively. A sympathetic reader would care because existing watermark tests are mostly fixed-horizon and become invalid if used with early stopping; this gives a rigorous streaming alternative.

Core claim

The central claim is that watermark detection reduces to testing whether a single pivot sequence behaves like standard exponential noise. For the Gumbel-max watermark, at each step the detector reconstructs the key-derived uniform U_{t,W_t} associated with the emitted token and forms Z_t = -log(1-U_{t,W_t}). The paper argues that under the null of no watermarking, Z_t is conditionally Exp(1) given the past pivots, so each one-step e-value (1-λ)exp(λ Z_t) has conditional expectation 1. Their product M_t(λ) is therefore an e-process, giving an anytime-valid level-α test via the stopping rule M_t ≥ 1/α. Under the Gumbel-max alternative, the paper proves positive asymptotic log-growth for a suit

What carries the argument

The load-bearing object is the normalized exponential score E_t = (1-λ)e^{λ Z_t}, whose product forms the e-process M_t(λ) = (1-λ)^t e^{λ S_t}. The pivot Z_t = -log(1-U_{t,W_t}) is the mechanism that strips away the unknown next-token distribution: under the null it is Exp(1), and under watermarking larger values signal dependence. Rao-Blackwellization appears through the surrogate exponential-tilt family q_κ(z) = (1-κ)exp(-(1-κ)z), under which the cumulative sum S_t is a sufficient statistic, so inference depends only on the scalar S_t rather than the full token history.

Load-bearing premise

The entire false-positive guarantee hangs on the premise that, whenever text is not watermarked, the pivot Z_t = -log(1-U_{t,W_t}) has conditional Exp(1) distribution given the past; if tokens are selected with even a small dependence on the key-derived uniforms, the guarantee fails.

What would settle it

Generate unwatermarked text by a sampling rule that makes the emitted token depend on the key-derived uniforms (for example, always pick the token with maximum U, or use a biased RNG), feed the pivots through the proposed stopping rule at α=0.05, and repeat many times. If the conditional distribution of Z_t genuinely departs from Exp(1), the empirical false-positive rate under optional stopping will exceed 5%, demonstrating that the anytime-valid guarantee depends on the independence premise.

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

If this is right

  • A detector can monitor a live token stream and stop the moment evidence crosses 1/α, with no correction for multiple looks.
  • Only a single accumulator M_t and the cumulative sum S_t need to be kept, so the method runs in constant memory per token.
  • The plug-in and mixture versions automatically adapt the tilt λ to the unknown signal strength while preserving validity.
  • Fixed-horizon watermark tests that reuse the same data for repeated testing lose Type I error control; this construction gives an online alternative.
  • Under watermarking, the test is consistent: it rejects almost surely as text length grows.

Where Pith is reading between the lines

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

  • The same pivot reduction could likely be built for green-red list watermarks if a pivot with an explicit null distribution can be defined, opening a general recipe for online detection.
  • A practical deployment could use the pivot's empirical distribution as a monitoring diagnostic: if the null assumption fails on human text, the detector would need recalibration before use.
  • Anytime-valid detection enables use cases where a decision must be made mid-generation, such as halting a chat model's output or flagging content before the full response is delivered.
  • A natural extension is to turn the e-process into a confidence sequence for watermark strength, giving an anytime estimate of how strong the watermark signal is, not just a binary rejection.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes an online, anytime-valid watermark detector for the Gumbel-max watermark, based on the reduced pivot Z_t = -log(1 - U_{t,W_t}). Under the null that Z_t | G_{t-1} ~ Exp(1), it constructs the e-process M_t(λ) = (1-λ)^t e^{λ S_t}, and adds plug-in and mixture versions of λ. The main theoretical claims are anytime-valid Type I error control via Theorem 3.1 and positive asymptotic log-growth under watermarking via Theorem 4.1. The method is O(1) memory and updates token by token (Algorithm 1). Experiments on synthetic spike-model data and on OPT-1.3B/GPT-Neo-1.3B text compare power, sequential Type I error, and runtime against online and offline baselines.

Significance. If the pivot null premise is valid, the construction is elegant and practically attractive: the one-step factors are genuine conditional e-values, the proof of Theorem 3.1 is a one-line MGF calculation, and the O(1) recursive update is a concrete improvement over history-dependent e-process baselines. The mixture and plug-in extensions are standard and valid under predictability. However, the paper's central reduction is asserted rather than proved for the actual keyed reconstruction, and the proof of the consistency theorem is deferred to an appendix that is absent from the manuscript. The significance for the sequential-testing/LLM-watermark community is therefore real but conditional on filling these gaps.

major comments (3)
  1. [§3.2 and Algorithm 1] The load-bearing null premise, Z_t | G_{t-1} ~ Exp(1), is asserted but not proved for the keyed reconstruction. The argument says W_t is independent of U_t under H0; however, G_{t-1} = σ(Z_1,...,Z_{t-1}) contains functions of past keyed pseudorandom vectors, so it may carry information about the key/PRF state and future U_t. Under the concrete key schedule used in Algorithm 1, the conditional law of U_{t,W_t} given G_{t-1} is not shown to be Unif(0,1), nor is Z_t shown to be Exp(1). Since Theorem 3.1 and the anytime-valid Type I error claim depend on this conditional law (not just marginal independence), the paper needs either a proof under an explicit key/PRF model or a clearly stated additional assumption defining the null under which the guarantee holds.
  2. [§4, Theorem 4.1] The proof of Theorem 4.1 is not present; the text defers to an Appendix that is not included. The theorem is central to the consistency claim (positive asymptotic log-growth and a.s. rejection). As written, it is an unverified assertion. Please provide the proof, or state the theorem as a conjecture and adjust the claims. Also clarify whether the λ whose existence is asserted is a universal constant or can depend on the unknown NTP sequence/P_t; if the latter, the practical relevance of the existence result is limited.
  3. [§3.4, Eq. (5)-(6)] The validity of the plug-in and mixture e-processes is stated to be proved in an absent Appendix. For the mixture this follows by convexity from Theorem 3.1, and for the plug-in it follows if λ_s is predictable and in (0,1); however, the clipping constants λ_min, λ_max are introduced without discussion of their effect on asymptotic log-growth, and Eq. (4)'s oracle λ* = 1 - μ^{-1} requires μ>1. Please include the deferred arguments and discuss the clipping/μ≤1 boundary case.
minor comments (3)
  1. [§3.3 / Proposition 3.2] The label 'Rao-Blackwellized' is not justified as stated: Proposition 3.2 only shows that S_t is sufficient under the surrogate family and that M_t(λ) is a function of S_t. Please explain how Eq. (2) is obtained by Rao-Blackwellizing a full-data e-process, or soften the terminology.
  2. [§3.4, Eq. (4)] The unclipped oracle λ* = 1 - μ^{-1} is not in (0,1) unless μ>1. Please add this condition or explicitly define the maximizer over (0,1) for μ≤1.
  3. [Throughout] The term 'unwatermarking language model' is used repeatedly; it should be 'unwatermarked'. Also, the references to an Appendix should be resolved: the manuscript contains no appendix, so all 'Appendix states...' passages are currently unverifiable.

Circularity Check

0 steps flagged

No significant circularity: the e-process is derived from an explicit pivot-null assumption, and no fitted parameter is renamed as a prediction.

full rationale

The derivation chain is self-contained apart from external pivot results. Theorem 3.1 explicitly assumes Z_t | G_{t-1} ~ Exp(1) under H0 as its premise; it does not claim to prove this from a self-citation. The construction M_t(λ) = (1-λ)^t e^{λ S_t} is the standard likelihood-ratio e-process for Exp(1) versus Exp(1−λ), normalized by the null MGF, and its anytime-validity follows from the stated conditional law. No parameter is fitted to data and then called a prediction: the plug-in rule estimates μ online, but validity holds for any predictable λ_s ∈ (0,1), so the e-process property does not depend on the fit. The oracle λ* = κ0 in Proposition 4.2 is derived in closed form from the surrogate family and is used only to motivate practical choices, not to force the consistency result. Citations to Li et al. (2025), Fernandez et al. (2023), and de Roos et al. (2025) are external and not self-citations by the present authors, so the load-bearing pivot null is not an imported uniqueness claim. The 'Rao-Blackwellized' framing is presentation rather than the argument for validity. The reviewer's concern that the Exp(1) pivot may fail under a concrete keyed reconstruction is a correctness or assumption gap, not a circularity: the paper's theorem states the pivot null as a supposition, and Algorithm 1's reconstruction details are left unspecified, but the claimed e-process property does not reduce to its own input.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central derivation uses the standard e-process machinery plus the known pivot reduction for Gumbel-max watermarks; no new physical/statistical entities are postulated. The main free choices are λ and its tuning knobs.

free parameters (3)
  • λ (fixed sensitivity parameter) = any value in (0,1) is valid; oracle under surrogate q_κ is λ*=κ0; in practice selected via plug-in μ̂ or mixture prior
    Trades power against robustness; the paper gives no principled default for λmin/λmax or the mixture grid. The plug-in λ_t = clip(1 - 1/μ̂_{t-1}, λmin, λmax) is fitted from the observed pivot mean.
  • λmin, λmax (truncation constants) = unspecified constants in (0,1)
    Required for numerical stability of the plug-in; values are chosen by user, not derived.
  • Mixture prior π (grid weights) = discrete grid with weights π_k, unspecified
    The mixture e-process requires choosing a grid of λ values and prior weights; paper does not give guidance.
axioms (5)
  • standard math Ville's inequality and e-process supermartingale properties
    Used to convert e-processes into anytime-valid tests with P(sup M_t ≥ 1/α) ≤ α (Sec. 2.1).
  • domain assumption Under H0, W_t is sampled from Pt and independent of the key-generated U_t, so Y_t = U_{t,W_t} ~ Unif(0,1) and Z_t | G_{t-1} ~ Exp(1)
    This is the pivot reduction stated in Sec. 3.2 and the premise of Theorem 3.1; it is true for the generative null model but not verified for arbitrary human text.
  • domain assumption Assumption 4.1: max_w Pt(w) bounded away from 1
    Used in Theorem 4.1 to ensure non-degenerate pivot signal.
  • domain assumption Assumption 4.2: (Pt) is m-dependent
    Needed for the asymptotic log-growth result; real LLM NTPs depend on full context and do not satisfy this.
  • ad hoc to paper Surrogate reduced family q_κ(z) = (1-κ)e^{-(1-κ)z}
    Introduced in Eq. (3) to give a closed-form oracle λ* and a 'Rao-Blackwellized' interpretation; the actual Gumbel-max alternative is not in this family.

pith-pipeline@v1.3.0-alltime-deepseek · 11286 in / 19950 out tokens · 196105 ms · 2026-08-01T06:13:58.641983+00:00 · methodology

0 comments
read the original abstract

As large language models (LLMs) are increasingly deployed, reliable and efficient mechanisms for distinguishing AI-generated text from human-written content have become essential. Statistical watermarking has emerged as a promising solution, yet most existing methods are typically fixed-horizon procedures, precluding valid early stopping in streaming generation. In this paper, we develop an efficient online watermark detection framework with anytime-valid inference based on Rao-Blackwellized e-processes, enabling recursive token-level evidence updates without storing the full history. In particular, we instantiate the framework for the Gumbel-max watermark and reduce the original token-level dependence testing problem to a pivot-induced sequential testing problem with an explicit null distribution. Theoretically, we prove anytime-valid Type I error control under arbitrary optional stopping and establish positive asymptotic log-growth under watermarking, implying consistency of the proposed stopping rules. Simulations and experiments on real LLM-generated text demonstrate efficient online detection with rigorous anytime-valid guarantees.

Figures

Figures reproduced from arXiv: 2607.21958 by Chengdong Xu, Dandan Mo, Huiqiong Li, Jinhan Xie, Lu Luo, Niansheng Tang, Ting Li.

Figure 1
Figure 1. Figure 1: Overview of the proposed online watermark detection framework. The generated [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average Type II error (on a log scale), sequential Type I error, and online [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Average Type II error, sequential Type I error, and online runtime versus [PITH_FULL_IMAGE:figures/full_fig_p020_3.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

42 extracted references · 13 linked inside Pith

  1. [1]

    (2023), ‘Watermarking of large language models’

    Aaronson, S. (2023), ‘Watermarking of large language models’. https://simons. berkeley.edu/talks/scott-aaronson-ut-austin-openai-2023-08-17

    2023

  2. [2]

    Altenschmidt, J., Altman, S., Anadkat, S. et al. (2023), ‘Gpt-4 technical report’, arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  3. [3]

    and Zhang, Y

    Bao, G., Zhao, Y., Teng, Z., Yang, L. and Zhang, Y. (2023), ‘Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature’, arXiv preprint arXiv:2310.05130

    Pith/arXiv arXiv 2023

  4. [4]

    and Biderman, S

    Black, S., Leo, G., Wang, P., Leahy, C. and Biderman, S. (2021), ‘Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow’, Zenodo

    2021

  5. [5]

    Shyam, P., Sastry, G., Askell, A. et al. (2020), ‘Language models are few-shot learners’, Advances in Neural Information Processing Systems33, 1877–1901

    2020

  6. [6]

    T., Li, Y., Lundberg, S

    Lee, Y. T., Li, Y., Lundberg, S. et al. (2023), ‘Sparks of artificial general intelligence: Early experiments with gpt-4’, arXiv preprint arXiv:2303.1271210

    Pith/arXiv arXiv 2023

  7. [7]

    and Wang, J.-K

    Chen, C. and Wang, J.-K. (2024), ‘Online detection of llm-generated texts via sequential hypothesis testing by betting’, arXiv preprint arXiv:2410.22318

    Pith/arXiv arXiv 2024

  8. [8]

    and Zamir, O

    Christ, M., Gunn, S. and Zamir, O. (2024), Undetectable watermarks for language models, in ‘The Thirty Seventh Annual Conference on Learning Theory’, PMLR, pp. 1125–1139

    2024

  9. [9]

    Kaskasoli, A., Stanforth, R., Matejovicova, T. et al. (2024), ‘Scalable watermarking for identifying large language model outputs’, Nature634(8035), 818–823. de Roos, D., Chugg, B., Gr¨ unwald, P. and Ramdas, A. (2025), ‘Rao-blackwellized e-variables’, arXiv preprint arXiv:2512.16759

    arXiv 2024

  10. [10]

    and Furon, T

    Fernandez, P., Chaffin, A., Tit, K., Chappelier, V. and Furon, T. (2023), Three bricks to consolidate watermarks for large language models, in ‘2023 IEEE International Workshop on Information Forensics and Security (WIFS)’, IEEE, pp. 1–6

    2023

  11. [11]

    Mathur, A., Schelten, A., Vaughan, A. et al. (2024), ‘The llama 3 herd of models’, arXiv preprint arXiv:2407.21783 . Gr¨ unwald, P., de Heide, R. and Koolen, W. M. (2020), Safe testing,in ‘2020 Information Theory and Applications Workshop (ITA)’, IEEE, pp. 1–54

    Pith/arXiv arXiv 2024

  12. [12]

    and Zhang, X

    Guo, H., Cheng, S., Jin, X., Zhang, Z., Zhang, K., Tao, G., Shen, G. and Zhang, X. (2024), ‘Biscope: Ai-generated text detection by checking memorization of preceding tokens’, Advances in Neural Information Processing Systems37, 104065–104090

    2024

  13. [13]

    R., Ramdas, A., McAuliffe, J

    Howard, S. R., Ramdas, A., McAuliffe, J. and Sekhon, J. (2021), ‘Time-uniform, nonpara- metric, nonasymptotic confidence sequences’, The Annals of Statistics 49(2), 1055–1080

    2021

  14. [14]

    and Huang, H

    Hu, Z., Chen, L., Wu, X., Wu, Y., Zhang, H. and Huang, H. (2023), ‘Unbiased watermark for large language models’, arXiv preprint arXiv:2310.10669

    Pith/arXiv arXiv 2023

  15. [15]

    and Jordan, M

    Huang, B., Xu, E., Ramchandran, K., Jiao, J. and Jordan, M. I. (2026), ‘Towards anytime- valid statistical watermarking’, arXiv preprint arXiv:2602.17608

    arXiv 2026

  16. [16]

    and Laks Lakshmanan, V

    Jawahar, G., Abdul-Mageed, M. and Laks Lakshmanan, V. (2020), Automatic detection of machine generated text: A critical survey, in ‘Proceedings of the 28th International Conference on Computational Linguistics’, pp. 2296–2309

    2020

  17. [17]

    and Walsh, D

    Johari, R., Koomen, P., Pekelis, L. and Walsh, D. (2022), ‘Always valid inference: Continuous monitoring of a/b tests’, Operations Research70(3), 1806–1821

    2022

  18. [18]

    Kasneci, E., Seßler, K., K¨ uchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., G¨ unnemann, S., H¨ ullermeier, E. et al. (2023), ‘Chatgpt for good? on opportunities and challenges of large language models for education’, Learning and 22 Individual Differences103, 102274

    2023

  19. [19]

    and Goldstein, T

    Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I. and Goldstein, T. (2023), A watermark for large language models, in ‘International Conference on Machine Learning’, PMLR, pp. 17061–17084

    2023

  20. [20]

    and Su, W

    Li, X., Ruan, F., Wang, H., Long, Q. and Su, W. J. (2025), ‘A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules’, The Annals of Statistics53(1), 322–351

    2025

  21. [21]

    and Su, W

    Li, X., Ruan, F., Wang, H., Long, Q. and Su, W. J. (2026), ‘Robust detection of watermarks for large language models under human edits’, Journal of the Royal Statistical Society Series B: Statistical Methodology88(2), 491–515

    2026

  22. [22]

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C. et al. (2024), ‘Deepseek-v3 technical report’, arXiv preprint arXiv:2412.19437

    Pith/arXiv arXiv 2024

  23. [23]

    and Yang, J

    Mao, C., Vondrick, C., Wang, H. and Yang, J. (2024), ‘Raidar: generative ai detection via rewriting’, arXiv preprint arXiv:2401.12970

    Pith/arXiv arXiv 2024

  24. [24]

    Milano, S., McGrane, J. A. and Leonelli, S. (2023), ‘Large language models challenge the future of higher education’, Nature Machine Intelligence5(4), 333–334

    2023

  25. [25]

    Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D. and Finn, C. (2023), Detectgpt: Zero-shot machine-generated text detection using probability curvature, in ‘International Conference on Machine Learning’, PMLR, pp. 24950–24962

    2023

  26. [26]

    Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W. et al. (2021), ‘Webgpt: Browser-assisted question-answering with human feedback’, arXiv preprint arXiv:2112.09332

    Pith/arXiv arXiv 2021

  27. [27]

    and Wagner, D

    Piet, J., Sitawarin, C., Fang, V., Mu, N. and Wagner, D. (2025), Markmywords: Analyzing and evaluating language model watermarks, in ‘2025 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML)’, IEEE, pp. 68–91

    2025

  28. [28]

    and Shafer, G

    Ramdas, A., Gr¨ unwald, P., Vovk, V. and Shafer, G. (2023), ‘Game-theoretic statistics and safe anytime-valid inference’, Statistical Science38(4), 576–601

    2023

  29. [29]

    and Koolen, W

    Ramdas, A., Ruf, J., Larsson, M. and Koolen, W. M. (2022), ‘Testing exchangeability: Fork-convexity, supermartingales and e-processes’, International Journal of Approximate 23 Reasoning141, 83–109

    2022

  30. [30]

    and Wang, R

    Ramdas, A. and Wang, R. (2025), ‘Hypothesis testing with e-values’, Foundations and Trends® in Statistics1(1-2), 1–390

    2025

  31. [31]

    (2022), ‘Ai bot chatgpt writes smart essays-should professors worry?’, Nature614(7947), 214–216

    Stokel-Walker, C. (2022), ‘Ai bot chatgpt writes smart essays-should professors worry?’, Nature614(7947), 214–216

    2022

  32. [32]

    and Nakov, P

    Su, J., Zhuo, T., Wang, D. and Nakov, P. (2023), Detectllm: Leveraging log rank information for zero-shot detection of machine-generated text, in ‘Findings of the Association for Computational Linguistics: EMNLP 2023’, pp. 12395–12412

    2023

  33. [33]

    and Zhao, Z

    Su, W., Wang, R. and Zhao, Z. (2026), ‘Online llm watermark detection via e-processes’, arXiv preprint arXiv:2602.14286

    Pith/arXiv arXiv 2026

  34. [34]

    and Hu, X

    Tang, R., Chuang, Y.-N. and Hu, X. (2024), ‘The science of detecting llm-generated text’, Communications of the ACM67(4), 50–59

    2024

  35. [35]

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S. et al. (2023), ‘Llama 2: Open foundation and fine-tuned chat models’, arXiv preprint arXiv:2307.09288

    Pith/arXiv arXiv 2023

  36. [36]

    and Wang, R

    Vovk, V. and Wang, R. (2021), ‘E-values: Calibration, combination and applications’, The Annals of Statistics49(3), 1736–1754

    2021

  37. [37]

    Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E. et al. (2025), ‘The rise and potential of large language model based agents: A survey’, Science China Information Sciences68(2), 121101

    2025

  38. [38]

    and Cao, Y

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K. and Cao, Y. (2022), ‘React: Synergizing reasoning and acting in language models’, arXiv preprint arXiv:2210.03629

    Pith/arXiv arXiv 2022

  39. [39]

    Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X. V. et al. (2022), ‘Opt: Open pre-trained transformer language models’, arXiv preprint arXiv:2205.01068

    Pith/arXiv arXiv 2022

  40. [40]

    and LeCun, Y

    Zhang, X., Zhao, J. and LeCun, Y. (2015), ‘Character-level convolutional networks for text classification’, Advances in Neural Information Processing Systems28

    2015

  41. [41]

    Zhou, H., Zhu, J., Su, P., Ye, K., Yang, Y., Gavioli-Akilagun, S. A. and Shi, C. (2025), ‘Adadetectgpt: Adaptive detection of llm-generated text with statistical guarantees’, 24 arXiv preprint arXiv:2510.01268

    arXiv 2025

  42. [42]

    and Shi, C

    Zhou, H., Zhu, J., Yang, Y. and Shi, C. (2026), ‘Detecting llm-generated text with perfor- mance guarantees’, arXiv preprint arXiv:2601.06586 . 25

    arXiv 2026