Pith. sign in

REVIEW 5 major objections 4 minor 28 references

Quantum Adiabatic Generation of Human-Like Passwords

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A 256-qubit neutral atom quantum computer can be trained on real-world password data and sampled to produce novel, human-like password candidates.

desk verdict A promising feasibility study with a sign error in the training update that needs correction before the central claim can be trusted. read the letter →

arxiv 2506.08917 v1 pith:Y2XADQVI submitted 2025-06-10 quant-ph cs.AI

classification quant-phcs.AI
keywords quantumadiabaticcomputingpasswordgenerationQUBOBoltzmanndistributionneutralatomcomputerUD-MIStokenizationforce-directedplacement
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 asks whether an adiabatic quantum computer can serve as a generative model for a small but real language task: producing passwords that look as if they were chosen by a human. The authors train a Quadratic Unconstrained Binary Optimization (QUBO) model on tokenized passwords from the RockYou dataset so that the QUBO's Boltzmann distribution approximates the empirical token distribution. They then convert the trained QUBO into a Unit-Disk Maximum Independent Set problem by placing atoms with a force-directed algorithm, and sample the QuEra Aquila 256-qubit neutral atom device. In 128 hardware samples, the model produced strings such as 'Tunas200992' and 'teedem28iglove', and the learned model's mean minimum edit distance to real passwords was more than one standard deviation closer than uniform token sampling. The paper reads this as evidence that the learned QUBO captures structural properties of token sequences and can be considered a basic learned language model.

What carries the argument

The load-bearing object is the Boltzmann distribution $P(z;Q) = Z^{-1}\exp(-z^\top Q z)$ of a QUBO matrix $Q$, trained by minimizing the KL divergence between $P_Q$ and the empirical distribution of tokenized passwords. Token strings are mapped to binary vectors through one of four encodings — binary, one-hot, or two stacked one-hot variants — and the gradient of the KL divergence is estimated with Gibbs/MCMC samples and applied with ADAM. To reach the hardware, the learned $Q$ is reinterpreted as a softmax over edge probabilities, and a Fruchterman-Reingold-inspired force-directed placement (Algorithm 1) positions atoms on the device; the resulting graph is a UD-MIS instance whose Rydberg blockade radius defines the independent-set constraint. Sampling bitstrings from the 256-qubit Aquila device and de-tokenizing them yields the password candidates.

What would settle it

Collect a large number of samples from the Aquila device for the same trained QUBO and atom placement, and compare the empirical sample distribution with the QUBO's Boltzmann distribution estimated by classical MCMC, using a statistical distance such as total variation; if the distance is large, the hardware samples are not coming from the learned distribution and the human-like examples would be coincidental.

Watch

Extended reading notes

Core claim

The central claim is that a QUBO matrix trained by minimizing KL divergence against a tokenized password distribution encodes enough statistical structure that sampling from its Boltzmann distribution produces novel, human-like passwords, and that this sampling can be done on real neutral atom hardware by converting the QUBO into a UD-MIS instance. The conversion re-reads the QUBO entries as logarithms of edge-appearance probabilities and uses a Fruchterman-Reingold-inspired force-directed placement to arrange atoms; the Rydberg blockade radius then imposes the MIS constraint. On the QuEra Aquila 256-qubit device, 128 samples yielded strings like 'Tunas200992' and 'teedem28iglove', and the mean minimum edit distance to human passwords beat a uniform-token baseline by more than one standard deviation. The paper frames the resulting model as a basic learned language model, and presents the atom placement as a step toward making neutral atom hardware easier to program.

Load-bearing premise

The whole pipeline rests on the assumption that the force-directed atom placement yields a UD-MIS instance whose sampling behavior on the neutral atom device is close enough to the trained QUBO's Boltzmann distribution that generated samples reflect the learned password distribution; this is asserted and left for future analysis, not demonstrated.

Editorial extensions

If this is right

  • Password security testing can use the generated strings as realistic decoys or attack candidates without exposing real user passwords, because the model reproduces the statistical habits of the training population.
  • The QUBO-to-UD-MIS conversion gives neutral atom hardware users an automatic way to turn a learned pairwise model into an atom layout, a step the paper argues is valuable beyond password generation.
  • The reported mean minimum edit distance, more than one standard deviation closer to human passwords than uniform token sampling, supplies a quantitative baseline for measuring human-likeness of generated passwords.
  • Because the model is trained on a real breach dataset, the generated passwords inherit the same common patterns and weaknesses, which is precisely what threat-model testing requires.

Reading between the lines

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

  • Inference: The human-like strings themselves are weak evidence for the UD-MIS-to-Boltzmann mapping, because such structured outputs would be unlikely if the device were sampling from a distribution very different from the learned one; a direct distributional comparison would settle this.
  • Inference: The same tokenize-train-place-sample pipeline could be applied to other short structured sequences, such as usernames or URLs; failure there would suggest the result depends on password-specific structure rather than on the general method.
  • Inference: The placement algorithm's geometric constraints (minimum distance, area bounds) act as hyperparameters; a systematic sweep could reveal whether the human-like output is robust or requires the specific default settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes a complete generative pipeline in which a QUBO-induced Boltzmann distribution over binary token encodings of passwords is learned from RockYou data, the QUBO is converted to a UD-MIS instance via a force-directed atom-placement heuristic, and samples are drawn both by MCMC and on the QuEra Aquila 256-qubit neutral-atom device. The authors evaluate the generated passwords by mean minimum edit distance to a holdout set, compare against uniform random token sequences, and report qualitative examples such as "Tunas200992" and "teedem28iglove". The central claim is that the learned QUBOs capture structural properties of password token sequences and can serve as a basic learned language model implemented on adiabatic quantum hardware.

Significance. If the pipeline were sound, the paper would be a useful early demonstration of small-scale generative modeling on analog quantum hardware, with a practical heuristic for mapping QUBO weights to atom placements. The manuscript has genuine strengths: it uses actual hardware (Aquila), makes the code available, is transparent that the model is fitted to data rather than predicting unseen structure, and the MCMC training framework is standard. However, the empirical evidence is currently too weak to support the central claim, and I identify a load-bearing sign error in the training update, a pseudocode error in the placement algorithm, and an unvalidated QUBO-to-UD-MIS mapping. These issues require substantive revision and re-validation rather than minor editing.

major comments (5)
  1. [§III, Eq. (5)-(6)] The stated gradient has the wrong sign. Since P_Q(z) ∝ exp(-z^T Q z), differentiating D_KL(P_D || P_Q) gives ∂D_KL/∂Q = E_{P_D}[zz^T] - E_{P_Q}[zz^T]. The paper defines U = E_{P_Q}[zz^T] - E_{P_D}[zz^T] and updates Q ← Q - ηU, i.e., Q ← Q + η(E_{P_D} - E_{P_Q}). This is gradient ascent, not descent, and it actively makes co-occurring token pairs less probable. This contradicts the statement in §III-A that negative Q entries favor x_v = x_w = 1. Because the training loop is the basis for every experimental result, the sign must be corrected in Eq. (5) or Eq. (6), the update rule must be fixed to Q ← Q + ηU (or the definition of U flipped), and the experiments must be re-run or verified against the corrected rule.
  2. [Algorithm 1] As printed, Algorithm 1 does not implement iterative force-directed refinement. The loop "for i ← 1 to I" only reassigns the initial circular positions; the force computation and position update for each v ∈ V appear outside this loop. Thus the returned placement is a single one-step displacement from the initial circle, not the result of I Fruchterman-Reingold iterations. This is load-bearing for the UD-MIS construction. Please fix the pseudocode nesting (or clarify if a formatting error removed an "end"), and if the implementation matches the printed version, all UD-MIS results need to be reassessed.
  3. [§III-A and §IV-C] The mapping from the learned QUBO to the UD-MIS instance is not validated. The paper reinterprets Q entries as logarithms of edge appearance probabilities and relies on force-directed placement, but gives neither a theoretical guarantee nor an empirical check that Aquila samples approximately follow P_Q. The authors explicitly postpone theoretical analysis of the placement algorithm. Without such validation, the human-like examples in Table II cannot be attributed to the learned model; they could be coincidental outputs of the hardware. Please add a quantitative comparison between MCMC samples from P_Q and Aquila samples on the same instances, for example energy histograms, marginal token frequencies, or a distributional distance on decoded passwords.
  4. [§IV-B/C, Fig. 4, Table II] The empirical evaluation is under-specified and the baseline is too weak. The abstract's central claim rests on 128 quantum samples, but the paper never states the total number of Aquila samples, the number of shots per instance, or the success rate of the placement constraints; Table II shows only ten examples per encoding. Comparing mean minimum edit distance against uniform random token sequences is not a meaningful test, since any model that exploits token frequencies would likely beat uniform. Please report exact sample sizes, per-fold statistics with confidence intervals, and compare against standard password generators such as n-gram or Markov baselines, or against the deep-learning baselines cited in the introduction.
  5. [§IV-A] The Byte Pair Encoding tokenizer is constructed on the entire RockYou set before the 5-fold split, so the evaluation fold contributes to the token vocabulary and the fixed binary code assignment. This is a leakage path: the holdout strings are not fully independent of the model. Please move tokenizer construction inside the training folds, or justify quantitatively why the tokenizer cannot inflate the reported MED results or the "human-like" examples.
minor comments (4)
  1. [§I] Please fix the typo "Large Language Modelss" in the first sentence and "training und operating" later in the same section.
  2. [Table II] Please state explicitly whether the displayed passwords are absent from the training set, since the introduction promises generation of "novel" candidates and the reader cannot determine this from the table.
  3. [Figure 3 caption] The caption says "binarysplit(8)"; this should read "binary (8)" for consistency with Table I.
  4. [Definition 1] The definition states that Q is upper triangular, but Eq. (5) and the training procedure treat Q as if it were symmetric. Please clarify whether Q is symmetrized or whether the upper-triangular convention is meant to encode both directions of each pairwise interaction.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the QUBO training-to-sampling chain is self-contained, though BPE tokenization before the fold split is a minor evaluation leakage and the UD-MIS atom placement is an unvalidated mapping.

full rationale

The derivation is not self-referential: a QUBO Boltzmann distribution P_Q (Def. 1) is explicitly fitted to the empirical token distribution by minimizing KL divergence (Eqs. 4-5), and the reported passwords are samples from that fitted model. Sampling from a fitted model resembling the training data is the intended generative use, not a disguised prediction. The MED comparison to uniform token strings provides an external baseline, giving the evaluation independent content. The self-citations ([2], [20]) are for standard techniques (modified BPE, Gibbs sampling) and are not load-bearing; no uniqueness theorem or author-specific ansatz is imported. The paper itself notes 'we postpone a theoretical analysis of our Fruchterman-Reingold-inspired atom placement algorithm to future work' (Section III-A), which is an omitted proof of the UD-MIS-to-Boltzmann mapping and a real correctness risk, but it is not circular because the placement is constructed from Q after training. The BPE tokenizer is built on the full RockYou set before the 5-fold split (Section IV-A vs IV-C), a mild leakage that makes the holdout evaluation less clean, but the QUBO parameters are still fit only on training folds, so the central claim does not reduce to its inputs. Finally, the sign convention in Eq. (5) (U = E_Q - E_D with update Q <- Q - eta U) appears inconsistent with minimizing KL(P_D || P_Q), a potentially serious training bug if implemented literally, but this is a correctness issue rather than a circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central contribution is a fitted generative model, so its parameters are the free parameters. The most fragile assumption is the unvalidated QUBO-to-UD-MIS mapping. No new physical entities are introduced.

free parameters (4)
  • QUBO matrix entries Q_ij = learned by ADAM, not reported
    All entries of the QUBO matrix are trained by gradient descent on the KL divergence in Eq. (4); the entire generative model is these parameters.
  • Initial one-hot penalty 0.1 = 0.1
    Hand-set in the initial QUBO to favor valid one-hot vectors (Section IV-C), affects training dynamics.
  • Inverse temperature beta = 1 (absorbed into Q)
    Beta is set to 1 and folded into Q (Section III), which is a scaling choice rather than a fitted constant.
  • Force-directed placement parameters (eta, I, c, distance constraints) = eta=1e-10, I=1e5, c chosen from Fig. 3, constraints: 75/76/4 um
    Hand-chosen in Algorithm 1; 2 of 20 runs violate constraints, so the values affect whether placements are valid.
assumptions (4)
  • domain assumption The adiabatic evolution on Aquila prepares a state whose measurement outcomes solve or sample near the UD-MIS instance.
    Invoked in Section II-B citing [26], a hardware assumption about Rydberg blockade and adiabatic state preparation.
  • ad hoc to paper The UD-MIS graph constructed by force-directed placement yields a sampling distribution close to the trained QUBO Boltzmann distribution.
    Section III-A reinterprets Q entries as edge probabilities but gives no proof or numerical check; the authors postpone theoretical analysis to future work.
  • domain assumption The class of pairwise QUBO Boltzmann distributions is expressive enough to approximate the token password distribution.
    The paper chooses this model family and fits it via KL divergence; no model capacity analysis is provided (Section III).
  • domain assumption The BPE tokenization built on the full RockYou dataset remains valid for held-out folds.
    Section IV-A builds the tokenizer before the 5-fold split, so token vocabulary is informed by evaluation data; the paper implicitly assumes this does not invalidate holdout comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Adiabatic Generation of Human-Like Passwords." pith.science (2026). https://pith.science/paper/Y2XADQVI

@misc{pith2026250608917,
  author       = {Pith},
  title        = {Pith review of: Quantum Adiabatic Generation of Human-Like Passwords},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y2XADQVI}},
  note         = {Machine review of arXiv:2506.08917}
}
read the original abstract

Generative Artificial Intelligence (GenAI) for Natural Language Processing (NLP) is the predominant AI technology to date. An important perspective for Quantum Computing (QC) is the question whether QC has the potential to reduce the vast resource requirements for training and operating GenAI models. While large-scale generative NLP tasks are currently out of reach for practical quantum computers, the generation of short semantic structures such as passwords is not. Generating passwords that mimic real user behavior has many applications, for example to test an authentication system against realistic threat models. Classical password generation via deep learning have recently been investigated with significant progress in their ability to generate novel, realistic password candidates. In the present work we investigate the utility of adiabatic quantum computers for this task. More precisely, we study different encodings of token strings and propose novel approaches based on the Quadratic Unconstrained Binary Optimization (QUBO) and the Unit-Disk Maximum Independent Set (UD-MIS) problems. Our approach allows us to estimate the token distribution from data and adiabatically prepare a quantum state from which we eventually sample the generated passwords via measurements. Our results show that relatively small samples of 128 passwords, generated on the QuEra Aquila 256-qubit neutral atom quantum computer, contain human-like passwords such as "Tunas200992" or "teedem28iglove".

Figures

Figures reproduced from arXiv: 2506.08917 by the authors.

Figure 1
Figure 1. Based on historical password data, we learn the Boltzmann distribution [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Exemplary graphical structure derived from the atom placement and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Atom placements computed for the encodings listed in Table I. From top to bottom: binarysplit(8), stacked(16), stacked(20), stacked(24). Columns [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Box plot for the mean MED over all splits for the four encoding methods listed in Table I. For comparison, the mean MED and one standard deviation for tokens sampled uniformly from T is shown. using the fifth for evaluation (i. e., as a holdout set for checking if our …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 18 canonical work pages

  1. [1]

    Adiabatic quan- tum computation

    Tameem Albash and Daniel A. Lidar. “Adiabatic quan- tum computation”. In:Reviews of Modern Physics 90.1 (Jan. 2018).ISSN: 1539-0756.DOI: 10 . 1103 / revmodphys . 90 . 015002.URL: http : / / dx . doi . org / 10 . 1103/RevModPhys.90.015002

  2. [2]

    Advances in Pass- word Recovery Using Generative Deep Learning Tech- niques

    David Biesner, Kostadin Cvejoski, Bogdan Georgiev, Rafet Sifa, and Erik Krupicka. “Advances in Pass- word Recovery Using Generative Deep Learning Tech- niques”. In:International Conference on Artificial Neu- ral Networks. V ol. 12893. Lecture Notes in Computer Science. Springer, 2021, pp. 15–27.DOI: 10.1007/978- 3-030-86365-4\ 2

  3. [3]

    Combining Variational Autoencoders and Transformer Language Models for Improved Password Generation

    David Biesner, Kostadin Cvejoski, and Rafet Sifa. “Combining Variational Autoencoders and Transformer Language Models for Improved Password Generation”. In:International Conference on Availability, Reliability and Security. ACM, 2022, 37:1–37:6.DOI: 10.1145/ 3538969.3539000

  4. [4]

    Some approaches to best-match file searching

    W. A. Burkhard and R. M. Keller. “Some approaches to best-match file searching”. In:Commun. ACM16.4 (1973), pp. 230–236.DOI: 10.1145/362003.362025

  5. [5]

    Kalyan Dasgupta and Binoy Paine.Loading Probability Distributions in a Quantum circuit. 2022. arXiv: 2208. 13372[quant-ph]

  6. [6]

    DeepSeek-AI et al.DeepSeek-V3 Technical Report

  7. [7]

    Quantum optimization of maximum independent set using Rydberg atom arrays

    S. Ebadi et al. “Quantum optimization of maximum independent set using Rydberg atom arrays”. In:Science 376.6598 (2022), pp. 1209–1215.DOI: 10.1126/science. abo6587

  8. [8]

    LLMCarbon: Modeling the End-to- End Carbon Footprint of Large Language Models

    Ahmad Faiz et al. “LLMCarbon: Modeling the End-to- End Carbon Footprint of Large Language Models”. In: International Conference on Learning Representations. OpenReview.net, 2024.URL: https : / / openreview. net / forum?id=aIok3ZD9to

Show all 28 references
  1. [9]

    Graph Drawing by Force-directed Placement

    Thomas M. J. Fruchterman and Edward M. Reingold. “Graph Drawing by Force-directed Placement”. In: Softw. Pract. Exp.21.11 (1991), pp. 1129–1164.DOI: 10.1002/SPE.4380211102

  2. [10]

    Stochastic Relax- ation, Gibbs Distributions, and the Bayesian Restoration of Images

    Stuart Geman and Donald Geman. “Stochastic Relax- ation, Gibbs Distributions, and the Bayesian Restoration of Images”. In:IEEE Transactions on Pattern Analysis and Machine Intelligence6.6 (1984), pp. 721–741.DOI: 10.1109/TPAMI.1984.4767596

  3. [11]

    Monte Carlo Sampling Meth- ods Using Markov Chains and Their Applications

    Wilfred Keith Hastings. “Monte Carlo Sampling Meth- ods Using Markov Chains and Their Applications”. In: Biometrika57.1 (1970), pp. 97–109.DOI: 10 . 2307 / 2334940

  4. [12]

    Quantum Wave Function Collapse for Procedural Content Generation

    Raoul Heese. “Quantum Wave Function Collapse for Procedural Content Generation”. In:IEEE Computer Graphics and Applications44.5 (Sept. 2024), pp. 54– 66.ISSN: 1558-1756.DOI: 10.1109/mcg.2024.3447775. URL: http://dx.doi.org/10.1109/MCG.2024.3447775

  5. [13]

    On the effects of biased quantum random numbers on the initialization of arti- ficial neural networks

    Raoul Heese, Moritz Wolter, Sascha M ¨ucke, Lukas Franken, and Nico Piatkowski. “On the effects of biased quantum random numbers on the initialization of arti- ficial neural networks”. In:Mach. Learn.113.3 (2024), pp. 1189–1217.DOI: 10.1007/S10994-023-06490-Y

  6. [14]

    Quantum computing with neutral atoms

    Lo ¨ıc Henriet et al. “Quantum computing with neutral atoms”. In:Quantum4 (2020), p. 327.ISSN: 2521- 327X.DOI: 10.22331/q- 2020- 09-21- 327.URL: http: //dx.doi.org/10.22331/q-2020-09-21-327

  7. [15]

    Beyond-classical computation in quantum simulation

    Andrew D. King et al. “Beyond-classical computation in quantum simulation”. In:Science388.6743 (2025), pp. 199–204.DOI: 10.1126/science.ado6285

  8. [16]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. “Adam: A Method for Stochastic Optimization”. In:International Con- ference on Learning Representations. Ed. by Yoshua Bengio and Yann LeCun. 2015.URL: http://arxiv.org/ abs/1412.6980

  9. [17]

    A comprehensive review of quantum ran- dom number generators: concepts, classification and the origin of randomness

    Vaisakh Mannalatha, Sandeep Mishra, and Anirban Pathak. “A comprehensive review of quantum ran- dom number generators: concepts, classification and the origin of randomness”. In:Quantum Information Processing22.12 (2023).ISSN: 1573-1332.DOI: 10 . 1007/s11128-023-04175-y

  10. [18]

    2021.DOI: 10.21227/ gzcg- yc14.URL: https://dx.doi.org/10.21227/gzcg- yc14

    Rohit Mutalik, Dhairya Chheda, Zeeshan Shaikh, and Dhanashree Toradmalle.Rockyou. 2021.DOI: 10.21227/ gzcg- yc14.URL: https://dx.doi.org/10.21227/gzcg- yc14

  11. [19]

    Complexity of uniqueness and local search in quadratic 0-1 program- ming

    Panos M. Pardalos and Somesh Jha. “Complexity of uniqueness and local search in quadratic 0-1 program- ming”. In:Oper. Res. Lett.11.2 (1992), pp. 119–123. DOI: 10.1016/0167-6377(92)90043-3

  12. [20]

    Exponential families on resource- constrained systems

    Nico Piatkowski. “Exponential families on resource- constrained systems”. PhD thesis. Technical University of Dortmund, Germany, 2018.URL: http://hdl.handle. net/2003/36877

  13. [21]

    On the challenges of using D-Wave computers to sample Boltzmann Random Variables

    Thomas Pochart, Paulin Jacquot, and Joseph Mikael. “On the challenges of using D-Wave computers to sample Boltzmann Random Variables”. In:Companion Proceedings of the IEEE 19th International Conference on Software Architecture. IEEE, 2022, pp. 137–140. DOI: 10.1109/ICSA-C54293...

  14. [22]

    Assessment of image generation by quantum an- nealer

    Takehito Sato, Masayuki Ohzeki, and Kazuyuki Tanaka. “Assessment of image generation by quantum an- nealer”. In:Scientific reports11.1 (2021), p. 13523

  15. [23]

    Neural Machine Translation of Rare Words with Sub- word Units

    Rico Sennrich, Barry Haddow, and Alexandra Birch. “Neural Machine Translation of Rare Words with Sub- word Units”. In:Annual Meeting of the Association for Computational Linguistics. The Association for Com- puter Linguistics, 2016.DOI: 10.18653/V1/P16-1162

  16. [24]

    Boltzmann sampling with quantum annealers via fast Stein correction

    Ryosuke Shibukawa, Ryo Tamura, and Koji Tsuda. “Boltzmann sampling with quantum annealers via fast Stein correction”. In:Physical Review Research6.4 (2024), p. 043050

  17. [25]

    Attention is All you Need

    Ashish Vaswani et al. “Attention is All you Need”. In:Advances in Neural Information Processing Systems

  18. [26]

    Jonathan Wurtz et al.Aquila: QuEra’s 256-qubit neutral-atom quantum computer. 2023. arXiv: 2306 . 11727[quant-ph].URL: https://arxiv.org/abs/2306. 11727

  19. [30]

    by Isabelle Guyon et al

    Ed. by Isabelle Guyon et al. 2017, pp. 5998–6008. URL: https://proceedings.neurips.cc/paper/2017/hash/ 3f5ee243547dee91fbd053c1c4a845aa-Abstract.html

  20. [2025]

    org/abs/2412.19437

    arXiv: 2412.19437[cs.CL].URL: https://arxiv. org/abs/2412.19437

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.