Pith. sign in

REVIEW 2 major objections 5 minor 55 references

Post-training for the receiver's tokenizer: 100% bit recovery

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 ↗

A margin-supervision plus receiver-realistic preference post-training method achieves 100% receiver-side bit accuracy and 21.6% lower normalized perplexity deviation than the strongest baseline.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A solid, well-ablated empirical paper that makes a real contribution by post-training the generator instead of filtering at inference time, though two gaps—an inconsistent cascade definition and unexamined stale preference labels—need attention before acceptance. the 2 major comments →

arxiv 2608.00382 v1 pith:IN7FOABD submitted 2026-08-01 cs.CR

TI-StegoAlign: Channel-Guided Post-Training for Generative Text Steganography under Tokenization Inconsistency

Jiuan Zhou , Yuhao Xue , Yu Cheng , Yuan Xie , Zhaoxia Yin This is my paper
classification cs.CR
keywords generative text steganographytokenization inconsistencyreceiver-realistic channelpost-trainingdirect preference optimizationLoRAanti-steganalysiscovert communication
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 reading

Generative text steganography lets an LLM agent hide secret bits inside otherwise normal messages, but the paper argues that standard evaluation inflates how well the receiver can actually decode them: it measures recovery on the sender's hidden tokens, not on the surface text the receiver retokenizes. A single retokenization boundary change can desynchronize the shared coding state and cascade into many bit errors. TI-StegoAlign is a post-training framework that instead makes the generator itself channel-aware. A supervised stage (BCSO) increases the distance between realized token positions and coding-interval boundaries, and a preference stage (CCPO) ranks complete stegotexts by receiver-realistic recovery, fluency, semantic preservation, and steganalytic risk. The paper reports 100% receiver bit accuracy and exact message recovery across News, Movie, and Tweet, along with lower perplexity deviation and better average anti-steganalysis than the strongest baselines.

Core claim

The central claim is that receiver-side recoverability can be learned into the generator rather than enforced at inference time. TI-StegoAlign models the communication channel as C = τ∘δ, detokenization followed by retokenization, and defines secret recovery exclusively on the receiver-reconstructed token sequence. BCSO adds a margin loss at realized embedding positions: for every token that carries a secret bit, it measures the distance from the token's cumulative-distribution coordinate to the nearest boundary of the CDF interval assigned by the shared coding rule, and penalizes positions closer than a margin γ. CCPO then generates candidate stegotexts from the BCSO policy, pushes each thr

What carries the argument

The load-bearing objects are the two training objectives and the channel model they run on. BCSO (Bit-Consistent Supervised Objective) supervises the realized sender-side embedding positions: at each position that consumes a secret bit, the token's cumulative-probability coordinate F_j(θ) is pushed toward the interior of its intended CDF interval [ℓ_j, u_j], so small shifts caused by retokenization are less likely to flip the decoded bit. CCPO (Channel-Conditioned Preference Optimization) passes complete candidate stegotexts through the detokenization–retokenization channel C = τ∘δ, scores them by recovery R_i, fluency, semantic similarity, and steganalytic risk A_i, and turns the surviving

Load-bearing premise

The argument assumes receiver extraction behaves the same under the final trained policy as under the BCSO reference policy that generated the preference labels.

What would settle it

Recompute receiver-side extraction labels under the final CCPO policy on a held-out set—without re-running candidate generation under the BCSO reference—and measure exact message recovery; if recovery falls below 100 percent, the trained preferences were misaligned with the deployed decoder.

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

If this is right

  • Oracle (sender-side) bit accuracy is not a valid proxy for communication reliability: in the paper's channel-characterization set, 45.3% of tokenization-inconsistent transmissions produce at least three additional bit errors after the first decoding error.
  • The post-trained policy reaches the same 100% exact receiver recovery as the filtering-based baselines but with a lower tokenization-inconsistency (TI) rate of 0.76% and without any inference-time candidate filtering, verification, or synchronization.
  • The two objectives are complementary rather than interchangeable: BCSO alone yields 99.56% receiver bit accuracy, CCPO alone 88.19%, and only their sequential combination reaches 100%.
  • Improvements transfer across model families: switching the backbone from Llama-3.1-8B to Qwen3-8B preserves 100% receiver recovery and further lowers average steganalysis accuracy to 58.15%.
  • Text quality and anti-steganalysis improve without sacrificing recovery: normalized perplexity deviation drops 21.6% relative to the strongest baseline and average detector accuracy reaches 60.61%, the closest to random guessing among the compared methods.

Where Pith is reading between the lines

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

  • The same channel-guided post-training idea should transfer to other interval-based coding rules and to LLM watermarking, since watermark extraction faces the same retokenization instability; the paper's mechanism is formulated for CDF/interval rules but not tested there.
  • Preference labels are computed under the BCSO reference policy; if the deployed final policy's interval boundaries shift during DPO, the trained preferences may not match the deployed decoder—a re-labeling check would make the 100% claim robust.
  • Because CCPO's security score uses a frozen BERT classifier and fixed stylistic statistics, the security gain is tied to that specific steganalysis model; the paper reports the gain is not uniform across detector architectures, suggesting detectors that exploit different features may erode it.
  • The method's dependence on a shared LoRA adapter and shared coding rule means the sender and receiver must synchronize the adapter before communication; the paper frames this as part of the pre-shared configuration, not an inference-time correction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes TI-StegoAlign, a post-training framework for generative text steganography under tokenization inconsistency. It models the detokenization-retokenization cycle as a receiver-side channel and introduces two objectives: BCSO, which adds local coding-margin supervision to a standard LM loss, and CCPO, which constructs preference pairs from receiver-realistic recovery, fluency, semantic consistency, and steganalytic security, then optimizes them with DPO on shared LoRA adapters. Experiments on News, Movie, and Tweet with Llama-3.1-8B and Qwen3-8B report 100% receiver bit accuracy and exact message recovery at 0.5 bpw, along with improved perplexity deviation, semantic similarity, KLD, and anti-steganalysis accuracy relative to five baselines.

Significance. If the results hold, the paper offers a practical alternative to inference-time tokenization disambiguation by adapting the generation policy itself. The empirical work is extensive: 313,800 benchmark transmissions, separate channel-characterization study, ablations, cross-backbone evaluation, human evaluation, and statistical corrections. The paper also ships detailed reproducibility information. These strengths make the central reliability and quality claims plausible. The main risk is whether the CCPO training signal truly reflects receiver-realistic recovery under the final policy, and one headline statistic is defined inconsistently across the paper.

major comments (2)
  1. [Section 3.3 / Appendix A.1] Eq. (10) updates pi_theta via DPO, while Eq. (3) defines receiver recovery under pi_theta. However, CCPO preference labels are computed once under pi_BC: candidates are sampled and labeled by Ext_Gamma on the channel output using the BCSO policy's conditional distributions, and the paper never states that recovery labels or retokenization outcomes are recomputed as theta changes. The DPO loss only increases log pi_theta(y+)-log pi_theta(y-); it does not constrain CDF interval membership or position selection under the updated policy. Thus the 'receiver-realistic recovery' used to supervise CCPO is that of the reference policy, not the deployed policy. The reported 100% test accuracy is an empirical result, but the mechanism as described does not logically ensure that preference labels match the deployed decoder. Please either recompute labels during training (or at checkpoints) or provid
  2. [Abstract / Section 4.2 / Appendix F.1] Cascade Incidence is defined inconsistently. The abstract and Section 4.2 define it as the proportion of TI cases with 'at least three additional bit errors after the first decoding error'; Appendix F.1 defines it as the proportion of TI cases where 'more than 35% of the remaining secret positions are decoded incorrectly.' These are not equivalent in general (e.g., with 10 remaining positions, 35% corresponds to 4 errors). The 45.3% figure is a headline contribution, so the exact threshold, treatment of the first error, and handling of missing positions must be stated once and applied consistently.
minor comments (5)
  1. [Table 3 / Appendix B.5] Detector accuracies are averaged over five seeds, but Table 3 reports only means without standard deviations or confidence intervals. Please report the spread in the main tables or state where the full statistics appear.
  2. [Section 4.4 / Appendix A.2] CCPO's security score A_i uses a BERT-based classifier, and the evaluation includes BERT-FT. Although the classifiers are trained independently, both are BERT-family; the paper should explicitly discuss the possibility of overfitting to this evaluator family. The gains on BD and SANet are reassuring, and an ablation omitting the security component from CCPO would strengthen the claim.
  3. [Eq. (8)] The pairwise preference condition uses existential and universal quantifiers with epsilon_{l|k}; the dependence of the tolerance on the preference axis k should be made explicit in the notation or in the accompanying text for clarity.
  4. [Figure 2] Panel (b) is dense and some labels ('Requirement gate', 'clear gain · bounded degradation') are not explained in the caption. Direct readers to Appendix A for the exact conditions.
  5. [Section 4.2] The TI-rate advantage over StepVerify is small (0.76% vs 0.78% overall, and 0.60 vs 0.62 on News). Since TI rate is a secondary metric, please indicate whether the differences are statistically significant or discuss them as descriptive only.

Circularity Check

0 steps flagged

No significant circularity: the 100% receiver-bit-accuracy result is an empirical held-out measurement, not a consequence of the training losses by construction.

full rationale

The paper's central claims are empirical: Table 1 reports receiver bit accuracy and exact message recovery on a held-out test set after post-training, and these are not identical to the quantities optimized by BCSO or CCPO. BCSO (Eq. 7) is a supervised margin loss over sender-side realized positions; CCPO (Eq. 10) is a DPO loss over preference pairs whose labels were constructed under the reference policy pi_BC (Appendix A.1: 'Each sender-side sequence is evaluated after passing through the same receiver-realistic channel used in the main experiments: y_i^s --delta--> z_i --tau--> y_i^r --Ext_Gamma--> s_hat_i'). Neither loss contains the final-policy extraction Ext_Gamma(.; pi_theta) from Eq. (3), so the reported 100% recovery is not forced by the training objective. The skeptic concern that preference labels are not recomputed as theta changes identifies a possible train/deploy distribution shift, which is a correctness risk, not circularity: if anything, it means the final result is harder to guarantee a priori. The only self-citation, Zhou et al. (2026) for PPL*, defines a standard normalized perplexity deviation that is restated in Eq. (23), so it is not load-bearing. The CCPO security critic uses a BERT-family classifier, but Section 4.4 and Appendix A.2 state it is 'trained independently' of the BF detector and on training-only splits; evaluation detectors BD, BF, and SANet are held-out and independently trained. The paper's own limitations (Twitter PPL* in Appendix C, cross-backbone generalization in Appendix H) are acknowledged and do not affect circularity. The derivation chain is therefore self-contained with respect to the measured results.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

No new physical or mathematical entities are introduced; the channel model C is a formal composition of existing tokenizer operations. The central claim instead rests on hand-set training hyperparameters and domain assumptions about the tokenizer, preference labels, and quality metrics.

free parameters (9)
  • BCSO margin gamma = 0.2
    Chosen margin width that pushes realized tokens away from interval boundaries; no sensitivity analysis in the main text.
  • BCSO weight lambda_BC = 0.5
    Trade-off between language-modeling loss and margin supervision.
  • CCPO DPO coefficient beta = 0.05 (News), 0.10 (Movie/Tweet)
    Controls deviation from the BCSO reference policy; domain-specific.
  • CCPO preference score weights = 100, -10, 5, 3 in Eq. (16)
    Hand-set weights combining recovery, perplexity deviation, semantics, and security; they determine which candidate pairs are considered.
  • Pairwise retention margins and tolerances = Table 5 values, e.g., recovery margin 0.05, degradation bounds 0.02-0.10
    Decide which candidate pairs become preference labels; central to the CCPO training set.
  • Candidate screening thresholds = R_i >= 0.95, A_i >= 0.30, PPL intervals, Q_sem minima, similarity thresholds 0.55 and 0.70
    Filter candidates before preference construction; affects which stegotexts the policy sees.
  • Candidate counts N per domain = 8 (News), 16 (Movie/Tweet)
    Number of sampled stegotexts per training condition; larger N improves pool diversity at compute cost.
  • LoRA configuration = rank 16, alpha 32, dropout 0.05
    Parameter-efficient adaptation capacity; fixed across experiments.
  • Optimization hyperparameters = learning rates, epochs, batch sizes, warmup steps in Appendix B.7
    Standard training settings that affect convergence of BCSO and CCPO.
axioms (5)
  • standard math CDF-partition coding rule with upper-endpoint token assignment decodes correctly when sender and receiver see identical token sequences.
    Assumed in Eqs. (1)-(3); the paper's problem is that retokenization breaks this assumption.
  • domain assumption The retokenization channel C = tau after delta is deterministic and identical at training and deployment.
    Section 3.1, Eq. (2); if tokenizer or model version differs at deployment, learned margins may not transfer.
  • domain assumption Preference labels extracted under the BCSO policy remain valid for the final CCPO policy.
    Appendix A.1 fixes candidate extraction once; Eq. (3) ties extraction to the current policy, and Eq. (10) changes the policy during DPO.
  • domain assumption GPT-2 perplexity, Sentence-BERT cosine similarity, and frozen steganalysis detectors measure the qualities the paper claims.
    Sections 4.3-4.4 and Appendix B.4; the security objective Ai is a BERT-family classifier while BF is a BERT-FT detector.
  • ad hoc to paper The score and threshold-based preference construction approximates a useful total order over stegotext quality.
    Appendix A.3-A.4; the score S_i and pair-retention conditions are introduced specifically for this framework with no optimality guarantee.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TI-StegoAlign: Channel-Guided Post-Training for Generative Text Steganography under Tokenization Inconsistency." pith.science (2026). https://pith.science/paper/IN7FOABD

@misc{pith2026260800382,
  author       = {Pith},
  title        = {Pith review of: TI-StegoAlign: Channel-Guided Post-Training for Generative Text Steganography under Tokenization Inconsistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IN7FOABD}},
  note         = {Machine review of arXiv:2608.00382}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative text steganography enables LLM agents to exchange secret information through task-relevant messages. Yet most methods evaluate recovery on sender-side tokens, whereas the receiver observes only surface text. Detokenization and receiver-side retokenization can alter token boundaries, desynchronize coding states, and cause such evaluation to overestimate receiver-side recovery. Existing remedies rely on inference-time filtering or verification, correcting individual outputs without adapting the generation policy to the receiver-side channel. To address these limitations, we propose TI-StegoAlign, a channel-guided post-training framework. The Bit-Consistent Supervised Objective (BCSO) enlarges local coding margins at realized sender-side embedding positions. Channel-Conditioned Preference Optimization (CCPO) then aligns complete stegotexts using receiver-realistic recovery, text quality, and anti-steganalysis feedback. TI-StegoAlign updates only LoRA parameters and requires no tokenization-specific correction during communication. Experimental results show 100% receiver bit accuracy. Compared with the strongest baselines, TI-StegoAlign achieves a 21.6% reduction in normalized perplexity deviation and a 6.3% relative improvement in anti-steganalysis performance.

Figures

Figures reproduced from arXiv: 2608.00382 by Jiuan Zhou, Yuan Xie, Yu Cheng, Yuhao Xue, Zhaoxia Yin.

Figure 1
Figure 1. Figure 1: Sender-side oracle evaluation reuses the hidden [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TI-StegoAlign framework: (a) BCSO combines in-domain language modeling with coding-margin supervision. (b) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The results of the human evaluation. bility. In our experiments, we take the logarithm of the KLD values to facilitate a clearer comparison of statistical im￾perceptibility between algorithms. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_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

55 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  5. [5]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  6. [6]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  7. [7]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  8. [8]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  9. [9]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  10. [10]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  11. [11]

    and Burger, Doug and Wang, Chi , title =

    Wu, Qingyun and Bansal, Gagan and Zhang, Jieyu and Wu, Yiran and Li, Beibin and Zhu, Erkang and Jiang, Li and Zhang, Xiaoyun and Zhang, Shaokun and Liu, Jiale and Awadallah, Ahmed Hassan and White, Ryen W. and Burger, Doug and Wang, Chi , title =. Proceedings of the Conference on Language Modeling (

    work page

  12. [12]

    Proceedings of the International Conference on Learning Representations (

    Hong, Sirui and Zhuge, Mingchen and Chen, Jiaqi and Zheng, Xiawu and Cheng, Yuheng and Zhang, Ceyao and Wang, Jinlin and Wang, Zili and Yau, Steven Ka Shing and Lin, Zijuan and Zhou, Liyang and Ran, Chenyu and Xiao, Lingfeng and Wu, Chenglin and Schmidhuber, J. Proceedings of the International Conference on Learning Representations (

    work page

  13. [13]

    Proceedings of the International Conference on Learning Representations (

    Chen, Weize and Su, Yusheng and Zuo, Jingwei and Yang, Cheng and Yuan, Chenfei and Chan, Chi-Min and Yu, Heyang and Lu, Yaxi and Hung, Yi-Hsin and Qian, Chen and Qin, Yujia and Cong, Xin and Xie, Ruobing and Liu, Zhiyuan and Sun, Maosong and Zhou, Jie , title =. Proceedings of the International Conference on Learning Representations (

    work page

  14. [14]

    Secret Collusion among AI Agents: Multi-Agent Deception via Steganography , url =

    Motwani, Sumeet Ramesh and Baranchuk, Mikhail and Strohmeier, Martin and Bolina, Vijay and Torr, Philip and Hammond, Lewis and de Witt, Christian Schroeder , booktitle =. Secret Collusion among AI Agents: Multi-Agent Deception via Steganography , url =. doi:10.52202/079017-2336 , editor =

    doi:10.52202/079017-2336

  15. [15]

    and Matthews, O

    Mathew, Y. and Matthews, O. and McCarthy, R. and Velja, J. and Schroeder de Witt, C. and Cope, D. and Schoots, N. , title =. Proceedings of the International Joint Conference on Natural Language Processing and the Asia-Pacific Chapter of the Association for Computational Linguistics (

    work page

  16. [16]

    and Zhong, F

    Xu, S. and Zhong, F. , title =. Proceedings of the Annual Meeting of the Association for Computational Linguistics (

    work page

  17. [17]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Huang, Kaibo and Wei, Yukun and Wu, Wansheng and Zhang, Tianhua and Yang, Zhongliang and Zhou, Linna , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    work page 2026

  18. [18]

    and Guo, X

    Yang, Z. and Guo, X. and Chen, Z. and Huang, Y. and Zhang, Y.-J. , title =. IEEE Transactions on Information Forensics and Security , volume =

    work page

  19. [19]

    and Deng, Y

    Ziegler, Z. and Deng, Y. and Rush, A. , title =. Proceedings of the Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing (

    work page

  20. [20]

    and Ji, H

    Shen, J. and Ji, H. and Han, J. , title =. Proceedings of the Conference on Empirical Methods in Natural Language Processing (

    work page

  21. [21]

    and Yang, Z

    Zhang, S. and Yang, Z. and Yang, J. and Huang, Y. , title =. Findings of the Association for Computational Linguistics:

    work page

  22. [22]

    and Jois, T

    Kaptchuk, G. and Jois, T. M. and Green, M. and Rubin, A. D. , title =. Proceedings of the ACM SIGSAC Conference on Computer and Communications Security (

    work page

  23. [23]

    and Sokota, S

    Schroeder de Witt, C. and Sokota, S. and Kolter, J. Z. and Foerster, J. N. and Strohmeier, M. , title =. Proceedings of the International Conference on Learning Representations (

    work page

  24. [24]

    and Chen, K

    Ding, J. and Chen, K. and Wang, Y. and Zhao, N. and Zhang, W. and Yu, N. , title =. Proceedings of the IEEE Symposium on Security and Privacy , pages =

    work page

  25. [25]

    34th USENIX Security Symposium (

    Wang, Yaofei and Pei, Gang and Chen, Kejiang and Ding, Jinyang and Pan, Chao and Pang, Weilong and Hu, Donghui and Zhang, Weiming , title =. 34th USENIX Security Symposium (

    work page

  26. [26]

    and Chen, K

    Zhang, X. and Chen, K. and Zhao, N. and Zhang, W. and Yu, N. , title =. IEEE Transactions on Information Forensics and Security , volume =

    work page

  27. [27]

    and Wu, Z

    Wu, J. and Wu, Z. and Xue, Y. and Wen, J. and Peng, W. , title =. Proceedings of the ACM International Conference on Multimedia , year =

    work page

  28. [28]

    and Cheng, Y

    Zhou, J. and Cheng, Y. and Xie, Y. and Yin, Z. , title =. Findings of the Association for Computational Linguistics:

    work page

  29. [29]

    and Just, P

    Huang, Y.-S. and Just, P. and Yin, H. and Narayanan, K. and Huang, R. and Tian, C. , title =. Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics (

    work page

  30. [30]

    and Murawaki, Y

    Nozaki, J. and Murawaki, Y. , title =. Proceedings of the Asia-Pacific Chapter of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing (

    work page

  31. [31]

    and Murawaki, Y

    Yan, R. and Murawaki, Y. , title =. Proceedings of the Conference on Empirical Methods in Natural Language Processing (

    work page

  32. [32]

    ReTokSync: Self-Synchronizing Tokenization Disambiguation for Generative Linguistic Steganography

    Wang, Y. and Wang, R. and Pang, W. and Han, J. and Qi, Y. and Hu, D. and Chen, K. , title =. arXiv preprint arXiv:2604.25486 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    and Chen, K

    Qi, Y. and Chen, K. and Zeng, K. and Zhang, W. and Yu, N. , title =. IEEE Transactions on Dependable and Secure Computing , volume =

    work page

  34. [34]

    and Murawaki, Y

    Yan, R. and Murawaki, Y. , title =. Proceedings of the Annual Meeting of the Association for Computational Linguistics (

    work page

  35. [35]

    and Yang, Y

    Yan, R. and Yang, Y. and Song, T. , title =. IEEE Signal Processing Letters , volume =

    work page

  36. [36]

    and Song, T

    Yan, R. and Song, T. and Yang, Y. , title =. Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics (

    work page

  37. [37]

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

    work page

  38. [38]

    and Ermon, Stefano and Finn, Chelsea , title =

    Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Manning, Christopher D. and Ermon, Stefano and Finn, Chelsea , title =. Advances in Neural Information Processing Systems , volume =

  39. [39]

    and Adeleke, T

    Karpov, A. and Adeleke, T. and Cho, S. H. and P. The Steganographic Potentials of Language Models , journal =

    work page

  40. [40]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =. Proceedings of the International Conference on Learning Representations (

    work page

  41. [41]

    arXiv preprint arXiv:2209.11429 , year =

    Misra, Rishabh , title =. arXiv preprint arXiv:2209.11429 , year =

    Pith/arXiv arXiv

  42. [42]

    and Daly, Raymond E

    Maas, Andrew L. and Daly, Raymond E. and Pham, Peter T. and Huang, Dan and Ng, Andrew Y. and Potts, Christopher , title =. Proceedings of the Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (

    work page

  43. [43]

    Go, Alec and Bhayani, Richa and Huang, Lei , title =

    work page

  44. [44]

    arXiv preprint arXiv:2407.21783 , year =

    Grattafiori, Aaron and others , title =. arXiv preprint arXiv:2407.21783 , year =

    Pith/arXiv arXiv

  45. [45]

    Biometrics Bulletin , volume =

    Wilcoxon, Frank , title =. Biometrics Bulletin , volume =

    work page

  46. [46]

    Scandinavian Journal of Statistics , volume =

    Holm, Sture , title =. Scandinavian Journal of Statistics , volume =

  47. [47]

    Recurrent Neural Network Based Language Model , booktitle =

    Mikolov, Tom. Recurrent Neural Network Based Language Model , booktitle =

    work page

  48. [48]

    Proceedings of the Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing (

    Reimers, Nils and Gurevych, Iryna , title =. Proceedings of the Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing (

    work page

  49. [49]

    arXiv preprint arXiv:1907.11692 , year =

    Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin , title =. arXiv preprint arXiv:1907.11692 , year =

    Pith/arXiv arXiv 1907

  50. [50]

    Information Hiding , pages =

    Cachin, Christian , title =. Information Hiding , pages =

    work page

  51. [51]

    and Wu, J

    Xue, Y. and Wu, J. and Ji, R. and Zhong, P. and Wen, J. and Peng, W. , title =. IEEE Transactions on Information Forensics and Security , volume =

    work page

  52. [52]

    and Bao, Y

    Yang, H. and Bao, Y. and Yang, Z. and Liu, S. and Huang, Y. and Jiao, S. , title =. Proceedings of the ACM Workshop on Information Hiding and Multimedia Security (

    work page

  53. [53]

    and Zhang, J

    Peng, W. and Zhang, J. and Xue, Y. and Yang, Z. , title =. IEEE Signal Processing Letters , volume =

    work page

  54. [54]

    Krippendorff, Klaus , title =

  55. [55]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.