Pith. sign in

REVIEW 1 major objections 68 references

Attention-Augmented LSTMs for Automatic Homophonic Ciphertext Decipherment

T0 review · 1 major / 0 minor · reviewed 2026-06-28 · grok-4.3

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

Pith's one-line read An attention-augmented LSTM achieves near-perfect character-level decryption of homophonic ciphers when trained only on aligned pairs from a shared code pool.

desk verdict The paper shows attention-augmented LSTMs can hit near-perfect accuracy on synthetic homophonic ciphers from a fixed shared pool when trained on aligned pairs, with clean failure outside the pool. read the letter →

arxiv 2606.05078 v1 pith:MSLAIK2R submitted 2026-06-03 cs.CR

Micaella Bruton , Meriem Beloucif , Beáta Megyesi This is my paper · ORCID
classification cs.CR
keywords homophonicsubstitutionciphersattention-augmentedLSTMautomaticdeciphermentshared-keysettinghistoricalcryptographysyntheticciphertextstranscriptionerrors
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

The paper evaluates an attention-augmented LSTM for learning to decrypt homophonic substitution ciphers in a shared-key setting where all texts use subsets of the same code pool. Synthetic ciphertexts are created from historical English and Swedish texts spanning 1500 to 1899, and the model trains solely on aligned pairs without external aids. Results indicate near-perfect accuracy on character level for various lengths, periods, and noisy conditions, with predictable failure outside the pool.

What carries the argument

Attention-augmented LSTM trained end-to-end on aligned ciphertext-plaintext pairs to learn code-to-letter mappings within a shared homophonic pool.

What would settle it

Training and testing the model on ciphertexts generated from two entirely separate homophonic code pools and checking whether accuracy stays near-perfect or drops sharply would settle whether the shared-pool condition is what enables the reported performance.

Watch

Extended reading notes

Core claim

The attention-augmented LSTM, trained exclusively on aligned ciphertext-plaintext pairs, learns consistent mappings from multiple codes to each plaintext letter within the shared pool and reaches near-perfect character-level decryption accuracy across English and Swedish texts from 1500-1899, including short examples and those with simulated transcription errors, while failing predictably on texts outside the shared pool.

Load-bearing premise

All ciphertexts draw from the same known homophonic code pool while each key uses a different but consistent subset of that pool.

Editorial extensions

If this is right

  • Near-perfect accuracy holds across both English and Swedish and across all centuries from 1500 to 1899.
  • Accuracy remains high for short ciphertexts and those containing simulated transcription errors.
  • The model functions as a practical tool for decipherment and for verifying suspected key reuse by failing on texts outside the shared pool.
  • No external language models, frequency statistics, or key-search heuristics are required.

Reading between the lines

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

  • The failure mode could help historians group real historical ciphertexts that likely share a common code pool.
  • The method might reduce dependence on manual analysis when multiple related documents are suspected to use subsets of one key.
  • Applying the approach to actual undeciphered texts would first require constructing or hypothesizing a plausible shared code pool from known historical examples.
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, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes an attention-augmented LSTM model for deciphering homophonic substitution ciphers under a shared-key assumption, where all ciphertexts draw from the same known homophonic code pool but use different consistent subsets for individual keys. Synthetic ciphertexts are generated via ChronoFidelius from historical English and Swedish texts (1500--1899); the model is trained exclusively on aligned ciphertext--plaintext pairs without external language models, frequency data, or key-search heuristics. The central claim is near-perfect character-level decryption accuracy across languages, periods, ciphertext lengths, variable-length codes, and simulated transcription errors, together with predictable failure on ciphertexts drawn from outside the shared pool.

Significance. If the results hold, the work supplies a concrete supervised baseline for learning homophonic mappings directly from aligned pairs and demonstrates that the model can serve as a practical tool for both decipherment and key-space verification when key reuse is suspected. The controlled synthetic regime, out-of-pool negative controls, and coverage of short/noisy texts constitute a falsifiable test of the shared-pool hypothesis.

major comments (1)
  1. Abstract: the claim of 'near-perfect character-level decryption accuracy' is presented without any quantitative metrics, error bars, dataset sizes, number of training examples, or ablation results, rendering the central empirical claim unverifiable from the provided summary and undermining assessment of its load-bearing strength.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract requires quantitative support for its central claims and will revise it in the next version.

read point-by-point responses

  1. Referee: Abstract: the claim of 'near-perfect character-level decryption accuracy' is presented without any quantitative metrics, error bars, dataset sizes, number of training examples, or ablation results, rendering the central empirical claim unverifiable from the provided summary and undermining assessment of its load-bearing strength.

    Authors: We accept this criticism. While the body of the manuscript reports specific metrics (character-level accuracies exceeding 99% on held-out test sets, training on 50,000+ aligned pairs per language/period, dataset sizes by century and length, and ablation studies on attention vs. baseline LSTM), the abstract uses only the qualitative phrase 'near-perfect.' We will revise the abstract to include representative quantitative results, including mean accuracy with standard deviation, number of training examples, and key dataset statistics. This change will be made in the resubmission. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; empirical ML evaluation is self-contained

full rationale

The paper reports supervised training of an attention-augmented LSTM on aligned ciphertext-plaintext pairs generated synthetically from historical texts, followed by accuracy measurements on held-out lengths, noise levels, and out-of-pool controls. No equations, uniqueness theorems, or parameter-fitting steps are described that would make any reported accuracy a definitional consequence of the training regime itself. The shared-pool assumption is an explicit experimental condition whose consequences are tested by the out-of-pool failure case, rather than an unexamined premise smuggled into the result. The work contains no self-citation load-bearing claims or ansatz smuggling; the central result is an empirical observation on a controlled synthetic task.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

Central claim rests on the domain assumption of a shared known code pool and aligned training pairs in a synthetic setting; no free parameters or invented entities are described in the abstract.

assumptions (1)
  • domain assumption All ciphertexts draw from the same known homophonic code pool with individual keys using different consistent subsets.
    Explicitly stated in abstract as the historically motivated shared-key setting required for the model to learn mappings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention-Augmented LSTMs for Automatic Homophonic Ciphertext Decipherment." pith.science (2026). https://pith.science/paper/MSLAIK2R

@misc{pith2026260605078,
  author       = {Pith},
  title        = {Pith review of: Attention-Augmented LSTMs for Automatic Homophonic Ciphertext Decipherment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MSLAIK2R}},
  note         = {Machine review of arXiv:2606.05078}
}
read the original abstract

Homophonic substitution ciphers replace each plaintext letter with one of several possible ciphertext codes, deliberately weakening letter-frequency patterns and making automated decipherment difficult. This paper evaluates whether an attention-augmented Long Short-Term Memory (LSTM) model can learn such mappings in a historically motivated shared-key setting: all ciphertexts draw from the same known homophonic code pool, while individual keys use different consistent subsets of that pool. Using synthetic ciphertexts generated with ChronoFidelius from historical English and Swedish texts dated 1500--1899, we test performance across ciphertext lengths, centuries, variable-length codes, and simulated transcription errors. Models are trained only on aligned ciphertext--plaintext pairs, without external language models, frequency statistics, or key-search heuristics. Results show near-perfect character-level decryption accuracy across both languages and all periods, including short and noisy ciphertexts. The model also fails predictably on ciphertexts outside the shared pool, indicating that it functions as a practical tool for decipherment and key-space verification when key reuse is suspected.

Figures

Figures reproduced from arXiv: 2606.05078 by the authors.

Figure 1
Figure 1. Aligned excerpt from a homophonic digit cipher example from the English test dataset. Each [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the homophonic decryption model. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 51 canonical work pages

  1. [1]

    doi:10.18653/v1/2025.acl-long.1377 , abstract =

    Adilazuarda, Muhammad Farid and Wijanarko, Musa Izzanardi and Susanto, Lucky and Nur'aini, Khumaisa and Wijaya, Derry Tanti and Aji, Alham Fikri , editor =. doi:10.18653/v1/2025.acl-long.1377 , abstract =

    work page doi:10.18653/v1/2025.acl-long.1377 2025

  2. [2]

    Ahmadzadeh, Ezat and Kim, Hyunil and Jeong, Ongee and Kim, Namki and Moon, Inkyu , year = 2022, journal =. A. doi:10.1109/ACCESS.2022.3140342 , abstract =

    work page doi:10.1109/access.2022.3140342 2022

  3. [3]

    Ahmadzadeh, Ezat and Kim, Hyunil and Jeong, Ongee and Moon, Inkyu , year = 2021, journal =. A. doi:10.1109/ACCESS.2021.3074268 , abstract =

    work page doi:10.1109/access.2021.3074268 2021

  4. [4]

    doi:10.1007/978-981-97-0641-9_51 , abstract =

    Cryptology and. doi:10.1007/978-981-97-0641-9_51 , abstract =

    work page doi:10.1007/978-981-97-0641-9_51

  5. [5]

    Automated

    Antal, Eugen and Mar. Automated. Tatra Mountains Mathematical Publications , volume =. doi:10.2478/tmmp-2022-0019 , urldate =

    work page doi:10.2478/tmmp-2022-0019 2022

  6. [6]

    Encrypted

    Antal, Eugen and Mar. Encrypted. International. doi:10.3384/ecp195689 , abstract =

    work page doi:10.3384/ecp195689

  7. [7]

    Bagla, Kartikay and Gupta, Shivam and Kumar, Ankit and Gupta, Anuj , year = 2024, pages =. Noisy. International. doi:10.1145/3639856.3639889 , abstract =

    work page doi:10.1145/3639856.3639889 2024

  8. [8]

    , year = 2023, journal =

    Bauer, Craig P. , year = 2023, journal =. The. doi:10.1080/01611194.2023.2170158 , langid =

    work page doi:10.1080/01611194.2023.2170158 2023

Show all 68 references
  1. [9]

    Bonmann, Svenja and Halfmann, Jakob and Korobzow, Natalie and Bobomulloev, Bobomullo , year = 2023, journal =. A. doi:10.1111/1467-968X.12269 , abstract =

    2023 doi

  2. [10]

    Willis and Kelley, Kathryn and Sarkar, Anoop , editor =

    Born, Logan and Monroe, M. Willis and Kelley, Kathryn and Sarkar, Anoop , editor =. Learning the. Workshop on. doi:10.18653/v1/2023.cawl-1.11 , abstract =

    2023 doi

  3. [11]

    Cryptographic

    Bose, Amrita and Pal, Debranjan and Roy Chowdhury, Dipanwita , editor =. Cryptographic. Applications and. doi:10.1007/978-981-97-9743-1_4 , abstract =

    doi

  4. [12]

    Braovi. A. Computational Linguistics , volume =. doi:10.1162/coli_a_00514 , abstract =

    doi

  5. [13]

    Conference on

    Bruton, Micaella and Beloucif, Meriem , editor =. Conference on. doi:10.18653/v1/2023.emnlp-main.671 , abstract =

    2023 doi

  6. [14]

    Bruton, Micaella and Megyesi, Beata , year = 2025, publisher =. From. International

    2025

  7. [15]

    Computational

    Corazza, Michele , year = 2023, month = jun, address =. Computational. doi:10.48676/unibo/amsdottorato/10975 , abstract =

    2023 doi

  8. [16]

    D, Renny Harlin and Gupta, Kritesh Kumar , year = 2024, pages =. Neural. doi:10.1109/RAICS61201.2024.10690135 , abstract =

    2024 doi arXiv

  9. [17]

    Desenclos, Camille and Lasry, George , year = 2024, publisher =. An. International. doi:10.58009/aere-perennius0090 , abstract =

    2024 doi

  10. [18]

    and Stamp, Mark , year = 2013, journal =

    Dhavare, Amrapali and Low, Richard M. and Stamp, Mark , year = 2013, journal =. Efficient. doi:10.1080/01611194.2013.797041 , abstract =

    2013 doi

  11. [19]

    Fornmark, Filip , year = 2022, address =. Models,

    2022

  12. [20]

    Evaluating

    F. Evaluating. International. doi:10.3384/ecp188394 , abstract =

    doi

  13. [21]

    and Naughton, Thomas J

    Gopinathan, Unnikrishnan and Monaghan, David S. and Naughton, Thomas J. and Sheridan, John T. , year = 2006, journal =. A. doi:10.1364/OE.14.003181 , abstract =

    2006 doi

  14. [22]

    Automatic

    Han, Songming and Ling, Ying and Xie, Ming and Ming, Shaofeng , year = 2024, pages =. Automatic. doi:10.1109/DSC63484.2024.00073 , abstract =

    2024 doi

  15. [23]

    Supporting

    H. Supporting. International. doi:10.58009/aere-perennius0100 , abstract =

    doi

  16. [24]

    Transcription of

    Johansson, Kajsa , year = 2019, address =. Transcription of

    2019

  17. [25]

    Augmented

    Kambhatla, Nishant , year = 2024, address =. Augmented

    2024

  18. [26]

    Kimura, Hayato and Emura, Keita and Isobe, Takanori and Ito, Ryoma and Ogawa, Kazuto and Ohigashi, Toshihiro , editor =. Output. Applied. doi:10.1007/978-3-031-16815-4_15 , isbn =

    doi

  19. [27]

    Deciphering an

    Kir. Deciphering an. Cryptologia , volume =. doi:10.1080/01611194.2024.2334042 , abstract =

    2024 doi

  20. [28]

    Knight, Kevin and Megyesi, Be. The. Proceedings of the 4th

  21. [29]

    Cryptanalysis of

    Kopal, Nils , year = 2019, publisher =. Cryptanalysis of. International

    2019

  22. [30]

    Deciphering

    Kopal, Nils and Waldisp. Deciphering. Cryptologia , volume =. doi:10.1080/01611194.2020.1858370 , abstract =

    2020 doi

  23. [31]

    Kopal, Nils , year = 2018, series =. Solving. International

    2018

  24. [32]

    Deciphering

    Lasry, George and Biermann, Norbert and Tomokiyo, Satoshi , year = 2023, journal =. Deciphering. doi:10.1080/01611194.2022.2160677 , abstract =

    2023 doi

  25. [33]

    Deciphering

    Lasry, George and Megyesi, Be. Deciphering. Cryptologia , volume =. doi:10.1080/01611194.2020.1755915 , abstract =

    2020 doi

  26. [34]

    Deciphering a

    Lasry, George and Paolo, Bonavoglia , year = 2022, publisher =. Deciphering a. International. doi:10.3384/ecp188401 , abstract =

    2022 doi

  27. [35]

    Applying

    Lehofer, Anna , year = 2022, journal =. Applying. doi:10.1080/01611194.2021.1918801 , abstract =

    2022 doi

  28. [36]

    Decrypting

    Lehofer, Anna , year = 2020, journal =. Decrypting. doi:10.3311/ope.381 , urldate =

    2020 doi

  29. [37]

    Detection of

    Leierzopf, Ernst and Mikhalev, Vasily and Kopal, Nils and Esslinger, Bernhard and Lampesberger, Harald and Hermann, Eckehard , editor =. Detection of. Data. doi:10.1007/978-981-16-8531-6_11 , abstract =

    doi

  30. [38]

    Deciphering

    Luo, Jiaming and Hartmann, Frederik and Santus, Enrico and Barzilay, Regina and Cao, Yuan , year = 2021, journal =. Deciphering. doi:10.1162/tacl_a_00354 , abstract =

    2021 doi

  31. [39]

    Cryptography

    Mardon, Austin and Barara, Gurman and Chana, Ishpreet and DiMartino, Alexia and Falade, Irene and Harum, Rokya and Hauser, Alexandra and Johnson, Jamie and Li, Amy and Pham, Jennifer and Varghese, Noah , year = 2021, publisher =. Cryptography

    2021

  32. [40]

    Benchmarking

    Maskey, Utsav and Zhu, Chencheng and Naseem, Usman , editor =. Benchmarking. Empirical. doi:10.18653/v1/2025.findings-emnlp.1082 , abstract =

    2025 doi

  33. [41]

    Decryption of

    Megyesi, Be. Decryption of. Cryptologia , volume =. doi:10.1080/01611194.2020.1716410 , abstract =

    2020 doi

  34. [42]

    Historical

    Megyesi, B. Historical. International. doi:10.3384/ecp195701 , abstract =

    doi

  35. [43]

    Transcription of

    Megyesi, Be. Transcription of. International. doi:10.3384/ecp2020171014 , abstract =

    doi

  36. [44]

    Megyesi, Be. What. International. doi:10.3384/ecp188404 , abstract =

    doi

  37. [45]

    Structured

    M. Structured. European. doi:10.1007/978-3-031-91572-7_20 , abstract =

    doi

  38. [46]

    Plaintext

    Najeeb, Mahira and Ghulam, Zikra and Pervaiz, Zahid and Masood, Ammar , year = 2025, journal =. Plaintext. doi:10.1007/s00521-024-10734-w , abstract =

    2025 doi

  39. [47]

    Cryptography:

    Naser, S M , year = 2021, journal =. Cryptography:

    2021

  40. [48]

    Improved

    Nuhn, Malte and Schamper, Julian and Ney, Hermann , year = 2014, pages =. Improved. Empirical. doi:10.3115/v1/D14-1184 , abstract =

    2014 doi

  41. [49]

    Evolutionary

    Oranchak, David , year = 2008, pages =. Evolutionary. Annual. doi:10.1145/1389095.1389425 , abstract =

    2008 doi

  42. [50]

    Automated

    Park, Seonghwan and Kim, Hyunil and Moon, Inkyu , year = 2023, journal =. Automated. doi:10.3390/cryptography7030035 , urldate =

    2023 doi

  43. [51]

    Pettersson, Eva and Megyesi, Be. The. Digital. doi:10.5617/dhnbpub.11045 , abstract =

    doi

  44. [52]

    Matching

    Pettersson, Eva and Megyesi, Beata , editor =. Matching. Nordic

  45. [53]

    Deciphering

    Pourdamghani, Nima and Knight, Kevin , editor =. Deciphering. Empirical. doi:10.18653/v1/D17-1266 , abstract =

    doi

  46. [54]

    Bayesian

    Ravi, Sujith and Knight, Kevin , editor =. Bayesian. Association for

  47. [55]

    Tumblr , urldate =

    The Fanfare Is in the Light but the Execution Is in the Dark , author =. Tumblr , urldate =

  48. [56]

    Sikora, Justyna , year = 2022, address =. The

    2022

  49. [57]

    International

    Simhadri, Sevitha and. International. doi:10.1109/NETCRYPT65877.2025.11102567 , abstract =

    2025 doi

  50. [58]

    Srivastava, Shivin and Bhatia, Ashutosh , year = 2018, pages =. On the. doi:10.1109/ICBK.2018.00029 , abstract =

    2018 doi

  51. [59]

    Improving

    Stenlund, Mathias Hans Erik and Nanni, Mathilde and Bruton, Micaella and Beloucif, Meriem , editor =. Improving. International

  52. [60]

    Szigeti, Ferenc and H. The. International. doi:10.3384/ecp188409 , abstract =

    doi

  53. [61]

    Accessing

    Tlachenska, Elina and Ivanov, Kiril and Nenova, Maria and Doynov, Rumen and. Accessing. International. doi:10.1109/ICEST66328.2025.11098206 , abstract =

    2025 doi

  54. [62]

    Decipherment of

    Yin, Xusen and Aldarrab, Nada and Megyesi, Be. Decipherment of. International. doi:10.1109/ICDAR.2019.00022 , urldate =

    2019 doi

  55. [63]

    Zhong, Qing and Liu, Bo and Ren, Jianxin and Li, Yongxin and Guo, Zhiruo and Mao, Yaya and Wu, Xiangyu and Ma, Yiming and Wu, Yongfeng and Zhao, Lilong and Sun, Tingting and Ullah, Rahat , year = 2022, journal =. Self-. doi:10.1364/OE.468945 , abstract =

    2022 doi

  56. [64]

    Cryptanalysis on

    Zhou, Rong and Yu, Simin , year = 2025, journal =. Cryptanalysis on. doi:10.1007/s11071-025-11582-6 , langid =

    2025 doi

  57. [65]

    Enhancing

    Bastian, Maik and Esslinger, Bernhard and Hermann, Eckehard and Kopal, Nils and Lampesberger, Harald , year = 2025, eprint =. Enhancing. International

    2025

  58. [66]

    Nada Aldarrab , title =

  59. [67]

    2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=

    Decipherment of historical manuscript images , author=. 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages=. 2019 , organization=

    2019

  60. [68]

    Hochreiter, Sepp and Schmidhuber, J. Long. Neural Computation , volume =. doi:10.1162/neco.1997.9.8.1735 , abstract =

    1997 doi

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.