acceptodds
Under review as a conference paper at ICLR 2027

Latent Chains for Robust and Tamper-Evident Watermark of Diffusion Large Language Models

Abstract

Diffusion large language models (dLLMs) are emerging as an alternative to autoregressive LLMs, with blockwise generation becoming a prominent design. LLM watermarking provides a promising approach for identifying dLLM-generated text. However, existing dLLM watermarks achieve detectability and robustness but provide limited tamper evidence, leaving them vulnerable to piggyback spoofing attacks that modify watermarked text while preserving its attribution to the targeted model. To provide tamper evidence, we introduce Latent Chain Watermark (LCW), which uses a fixed latent chain within each block and lets each finalized block select the latent chain for the next block. To overcome the tension between robustness and tamper evidence, one favours unaffected by modification, the other favours sensitive to modification. We adopt two detections: (1) robust watermark detection searches over all candidate changes to recover watermark evidence, (2) tamper detection measures how modifications disrupt block boundaries and latent chains as tamper evidence. We further develop efficient methods to accelerate detection. At a 5% false-positive rate, LCW achieves 95.1% watermark detection, 80.2% watermark survival under 10% token-level editing, and 82.2% tamper detection after three token edits. Additional experiments demonstrate that LCW's applicability beyond diffusion generation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM watermark, Diffusion LLM, LLM security

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LCU6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Chain Watermark (LCW), a watermarking framework designed to achieve both robustness and tamper evidence for blockwise diffusion language models. LCW assigns a fixed latent chain to each block to determine position-specific green lists, while using the finalized content of each block to select the latent chain for the subsequent block. This design avoids dependence on provisional tokens during parallel generation while introducing content-dependent relationships across blocks. To address the conflicting requirements of robustness and tamper sensitivity, the authors develop two detection procedures. Robust watermark detection searches over candidate block boundaries, latent chains, and within-block alignments to recover watermark evidence from modified text. Tamper detection instead measures the score differences between flexible and constrained detection procedures to identify structural inconsistencies caused by modifications. Beam search and dynamic programming are employed to improve detection efficiency. Experiments on DreamReasoner and LLaDA evaluate watermark detectability, robustness against text editing, and tamper detection. The authors also extend the framework to autoregressive generation and tree-based speculative decoding.

Critical strengths

1.Meaningful research problem. The paper addresses the challenge of achieving both robust watermark detection and tamper evidence in diffusion language models. These two objectives are inherently conflicting, making the problem relevant and technically challenging. 2.Interesting latent-chain design. The proposed approach uses fixed latent chains within blocks and content-dependent transitions across blocks, allowing watermark embedding to remain compatible with parallel generation while introducing sensitivity to post-generation modifications. 3.Well-structured detection framework. The separation between robust watermark detection and tamper detection is intuitive. The combination of beam search and dynamic programming also provides a practical approach to handling the otherwise expensive search over candidate watermark structures. 4.Relatively comprehensive evaluation. The experiments cover multiple models, text lengths, editing attacks, and generation paradigms. The ablation studies further help illustrate the contributions of the proposed components.

Critical weaknesses
  1. Limited robustness and tamper detection under paraphrasing attacks. Paraphrasing is a realistic and important threat to LLM watermarking, as attackers can substantially modify the wording and structure of watermarked text while preserving its meaning. However, LCW shows considerably weaker robustness under paraphrasing than under random token-level edits. As reported in Appendix C.4, LCW achieves watermark survival rates of only 43.6% and 66.3% under DIPPER paraphrasing at text lengths of 128 and 256, respectively, substantially below Unigram (88.1% and 99.0%) and GSJ (48.5% and 84.2%). Furthermore, the evaluation considers only one DIPPER configuration and does not report tamper detection performance after paraphrasing. Given that robustness and tamper evidence are central objectives of the paper, these results raise concerns about LCW's effectiveness against practical text rewriting attacks.
  2. Limited effectiveness against targeted and localized tampering. LCW relies on within-block alignment and cross-block latent-chain transitions to identify tampering, making it less effective against modifications that do not significantly disrupt these structures. The average tamper detection rate for single-token substitutions is only 30.6%, and Appendix C.6 reports a further decrease to 15.8% for substitutions in the final block at a 5% FPR. This limitation is particularly concerning because an adversary could deliberately modify the final block without knowing the secret key, thereby avoiding subsequent chain-transition evidence. The current evaluation does not sufficiently investigate such targeted modification strategies.
  3. Limited analysis of the reliability and discriminative power of tamper evidence. LCW identifies tampering by measuring differences between watermark scores obtained under different structural constraints. However, such differences may also arise from the stochastic nature of soft watermarking and the maximization over multiple candidate configurations, even for unmodified texts. Although the authors calibrate detection thresholds on unmodified watermarked texts, the extent to which the proposed alignment-based and chain-based scores reliably distinguish genuine tampering from normal statistical variations remains insufficiently analyzed. Furthermore, the effectiveness of these signals may depend on watermark strength, introducing a potential trade-off between tamper sensitivity and text quality. Additional analysis of score distributions and sensitivity to watermark strength would strengthen the justification of the proposed mechanism.
  4. Concerns about the evaluation protocol and baseline comparison. The survival tamper detection rate (STR) is evaluated only on edited texts that remain watermark-positive. While this conditional metric is meaningful, it does not directly reflect the overall probability of successfully detecting tampering. This distinction becomes particularly important under attacks with low watermark survival rates, such as paraphrasing. Reporting the joint success rate of watermark detection and tamper detection would provide a more complete assessment. Additionally, the tamper detection baselines employ different mechanisms: LCW uses its dedicated detector, CW uses its native edit detector, while Unigram and GSJ rely on watermark score decreases. The construction of these baseline scores and the information available during detection should be clarified to ensure a fair comparison.
Questions for authors
  1. Robustness against paraphrasing. LCW performs substantially worse than Unigram and GSJ under DIPPER paraphrasing. What are the main reasons for this performance gap? Could the authors evaluate additional paraphrasing settings with different lexical and structural modification strengths, and report the corresponding tamper detection performance?
  2. Targeted tampering and positional sensitivity. Given the low tamper detection rate for substitutions in the final block, how would LCW perform against an adversary who deliberately selects modification positions? Could the authors provide a position-wise analysis of tamper detection performance and discuss possible ways to mitigate this limitation?
  3. Reliability of tamper evidence. How well do the alignment-based and chain-based tamper scores distinguish modified texts from unmodified watermarked texts? Could the authors provide the score distributions for both cases and analyze how tamper detection performance varies with watermark strength?
  4. End-to-end tamper detection. Could the authors report the joint success rate of watermark detection and tamper detection under different editing attacks, particularly paraphrasing? This would help clarify the practical effectiveness of LCW beyond the conditional STR metric.
  5. Fairness of baseline comparisons. How are the tamper detection scores for Unigram and GSJ constructed? Do these baselines require access to the original watermarked text or its watermark score? Are all methods evaluated under comparable information availability and threshold calibration protocols?
5 comments

Q1 Paraphrasing Robustness Why is LCW less robust than Unigram and GSJ under paraphrasing, and how does its tamper detection perform under different paraphrasing strengths?

Thank you for the comment. We acknowledge that LCW exhibits lower robustness than Unigram and GSJ under paraphrasing. This difference stems from their distinct design objectives. Unigram and GSJ primarily focus on watermark detectability and robustness, aiming to preserve watermark signals under text modifications. However, this robustness also makes tampering difficult to detect, as modifications may leave the watermark signals largely unchanged. As demonstrated in Figure 3, both methods exhibit limited tamper-detection capability.

In contrast, LCW explicitly introduces content-dependent cross-block transitions to provide tamper evidence while maintaining watermark robustness. Importantly, LCW does not simply trade substantial robustness for tamper evidence under localized edits. For example, Table 3 shows that introducing chain transitions improves substitution STR@5 from 7.0% to 55.8%, while substitution WSR@5 decreases only slightly from 99.0% to 97.7%. This demonstrates that LCW can substantially improve tamper detection while maintaining strong robustness in the targeted editing setting.

To further investigate paraphrasing robustness and tamper evidence, we additionally evaluate different DIPPER configurations with varying lexical and order diversity on DreamReasoner (NX=256N_X=256). We report the two metrics already used in our paper: watermark survival rate (WSR) and survival tamper detection rate (STR). Following another reviewer's suggestion, we also include the end-to-end tamper detection rate (E2E-TDR), calculated as E2E-TDR=WSR×STR\mathrm{E2E\text{-}TDR}=\mathrm{WSR}\times\mathrm{STR}.

FPRDIPPER (Lex/Order)LCW WSRLCW STRLCW E2E-TDRUnigram WSRGSJ WSR
5%20/00.6631.0000.6630.9900.842
5%60/600.2481.0000.2480.8470.338
5%80/800.1851.0000.1850.6310.185
1%20/00.4751.0000.4750.9410.663
1%60/600.0891.0000.0890.5990.134
1%80/800.0641.0000.0640.3380.076

Notably, LCW achieves 100% STR across all evaluated paraphrasing configurations, meaning that every watermark-positive paraphrased text is correctly identified as tampered in these experiments. Under stronger DIPPER settings (60/60 and 80/80), extensive rewriting substantially disrupts the original watermark evidence, resulting in lower WSR and consequently lower E2E-TDR. GSJ and Unigram also exhibit reduced watermark survival under these settings.

Q2 Targeted tampering and positional sensitivity How does the position of an edit affect tamper detection, particularly for substitutions in the final block, and can this limitation be mitigated?

Thank you for the comment. We have discussed the difficulty of detecting substitutions in Section 4.2 and the final-block limitation in Appendix C.6. First, substitution is inherently more challenging to detect than insertion or deletion, as it preserves token positions and therefore introduces limited block-alignment evidence. This is not unique to LCW, and the evaluated baselines also exhibit limited substitution tamper-detection capability. Nevertheless, as shown in Figure 3, LCW substantially outperforms these baselines in substitution tamper detection.

Second, we further investigate the impact of modification positions. We conduct a position-wise evaluation on 256-token texts consisting of eight blocks. At 5% FPR, when applying three token deletions to the first, seventh, and final blocks, the STRs are 100.0%, 56.5%, and 42.1%, respectively. For insertion, the corresponding STRs are 100.0%, 65.6%, and 36.8%. For substitution, the corresponding STRs are 25.8%, 19.06% and 15.8%, respectively.

Furthermore, a simple practical mitigation is to append one or two non-essential, naturally generated watermarked blocks after the semantically important content. This allows modifications to the final content block to disrupt subsequent latent-chain transitions, providing additional tamper evidence. These trailing blocks can be discarded after watermark and tamper verification without affecting the substantive content presented to users.

Q3 Reliability of Tamper Evidence How well do the alignment-based and chain-based tamper scores distinguishmodified texts from unmodified watermarked texts? Could the authors provide the score distributions for bothcases and analyze how tamper detection performance varies with watermark strength?

Thank you for the comment. We acknowledge that soft watermarking and the maximization over candidate configurations may introduce score variations even for unmodified watermarked texts. However, LCW measures tamper evidence through differences between watermark scores under different structural constraints, rather than relying on absolute watermark scores.

For unmodified watermarked texts, the original block alignments and chain transitions remain consistent with the embedding process, so enforcing these structures typically causes little watermark score loss. In contrast, text modifications can disrupt these structures. Specifically, SalignS_{\mathrm{align}} measures the score loss caused by enforcing block boundaries and within-block alignments, primarily capturing insertion and deletion, while SchainS_{\mathrm{chain}} measures the score loss caused by enforcing chain transitions, providing evidence for substitution. Together, these complementary scores distinguish structural disruptions caused by tampering from normal score variations.

To further substantiate this mechanism, we analyze the score distributions of unmodified and tampered texts under the default configuration (NX=128N_X=128, γ=0.25\gamma=0.25, δ=2\delta=2).

ScoreTextMedianQ25–Q75P95
SalignS_{\mathrm{align}}Clean0.0000.000–0.0000.367
Deletion3.0191.937–3.5614.925
Insertion2.5031.902–3.5245.046
Substitution0.0000.000–0.0000.367
SchainS_{\mathrm{chain}}Clean0.0000.000–0.0001.225
Deletion1.2490.625–1.8742.603
Insertion1.4021.001–2.0033.204
Substitution2.4501.429–4.0835.920

The results show that alignment-based scores clearly distinguish deletion and insertion from unmodified texts, while chain-based scores provide stronger evidence for substitution. This is consistent with the complementary roles demonstrated in Table 5 of Appendix C.1, further validating the effectiveness of LCW's complementary detection design.

We further evaluate the sensitivity of tamper detection to watermark strength by varying γ\gamma and δ\delta. Each component is independently calibrated to a tamper FPR of at most 5%. Results on DreamReasoner at NX=128N_X=128 under five token edits as follows.

γ\gammaδ\deltalog(PPL) ↓Align STR@5 (Del/Ins)Chain STR@5 (Sub)Combined STR@5 (Sub)
0.25022.0590.987 / 0.9800.7660.747
0.25042.7451.000 / 1.0000.8610.861
0.50021.9880.961 / 0.8990.8280.767
0.50042.3420.984 / 1.0000.8820.890

These results demonstrate that both SalignS_{\mathrm{align}} and SchainS_{\mathrm{chain}} remain effective across different watermark strengths.

Q4 End-to-End Tamper Detection What is the joint success rate of watermark detection and tamper detection under different editing attacks, particularly paraphrasing?

Thank you for the suggestion. We report watermark survival rate (WSR) and survival tamper detection rate (STR) separately to evaluate two distinct security objectives: watermark robustness and tamper evidence. WSR measures the fraction of edited texts that remain watermark-positive, while STR measures the fraction of these watermark-positive texts that are correctly identified as tampered.

For the same set of attacked texts, the end-to-end tamper detection rate (E2E-TDR) can be directly computed as P(DW=1,DT=1)=WSR×STRP(D_W=1,D_T=1)=\mathrm{WSR}\times\mathrm{STR}. To make the overall effectiveness more explicit, we additionally report the joint detection rates under different editing attacks, including paraphrasing.

AttackSettingSTR@5E2E-TDR@5STR@1E2E-TDR@1
DeletionNedit=5N_{\mathrm{edit}}=50.9970.9830.9970.980
InsertionNedit=5N_{\mathrm{edit}}=50.9970.9930.9870.980
SubstitutionNedit=5N_{\mathrm{edit}}=50.8630.8630.6760.676
DIPPER20/01.0000.6631.0000.475
DIPPER60/601.0000.2481.0000.089
DIPPER80/801.0000.1851.0000.064

Results on DreamReasoner (NX=256N_X=256). Token-level editing follows the evaluation protocol in Figure 3 with Nedit=5N_{\mathrm{edit}}=5. DIPPER settings denote lexical/order diversity.

The results demonstrate strong end-to-end tamper detection under token-level edits. For paraphrasing, LCW achieves 1.000 STR across all evaluated configurations, indicating that all watermark-positive paraphrased texts are identified as tampered. Under stronger DIPPER settings (60/60 and 80/80), extensive paraphrasing substantially changes the original token sequence, disrupting watermark evidence and reducing watermark survival. Consequently, the lower E2E-TDR primarily reflects watermark removal rather than a failure to detect tampering in texts where the watermark survives.

Q5 Fairness of baseline comparisons How are the tamper detection scores of Unigram and GSJ constructed, and are all methods compared under equivalent information availability and threshold calibration?

Thank you for the question. We would like to clarify that watermark embedding and detection are inherently coupled, and different watermarking mechanisms require corresponding detection procedures. Therefore, we adopt detection methods suited to each watermarking mechanism rather than applying a single tamper detector to all methods. Specifically, LCW uses its proposed alignment- and chain-based tamper detector, while CW uses its native edit detector, as its watermark structure does not involve blocks or latent chains. Unigram and GSJ are primarily designed for watermark detectability and robustness and do not provide native tamper-evidence detectors. For comparison, we use changes in their original watermark scores as surrogate tamper signals. For a consistent evaluation, tamper detection thresholds are calibrated at the same target FPR using unmodified watermark-positive texts.

Reviewer_f7yeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a watermarking method for blockwise diffusion language models. It fixes the green list at each position within a block and uses the completed block to choose the green-list pattern for the next block. This supports parallel generation while creating links between blocks that may break when the text is edited. During detection, the method checks mismatches in block alignment and green-list patterns to identify possible tampering.

Critical strengths
  1. The paper addresses a meaningful problem: preserving watermark detection while identifying later edits to text generated by diffusion models.

  2. The embedding design fits blockwise parallel generation. Fixed within-block patterns avoid dependence on unfinished tokens, while finalized blocks establish content-dependent links.

  3. The ablations help explain the method’s behavior: block links mainly support substitution detection, while alignment scores are more useful for detecting insertions and deletions.

Critical weaknesses
  1. Appendix C.6 reports only 15.8% STR@5 for a single-token substitution in the final block, where no subsequent chain transition exists. The security evaluation should include attackers that deliberately target these vulnerable positions.

  2. With only four candidate chains, a modified block would select the original next-block chain with probability approximately 1/4 under a uniform-hash assumption. This does not directly imply a 25% overall miss rate, but its effect on chain-based tamper evidence should be quantified.

  3. The reported metrics do not directly measure successful spoofing. STR is computed only over edited texts that remain watermark-positive. The paper should also report the fraction of all attacked texts that retain a detectable watermark while escaping tamper detection, using the same attack settings across methods.

  4. The baseline coverage is incomplete. DMark and LR-DWM are discussed but not evaluated in the diffusion experiments. The autoregressive comparison includes Bileve but omits DualGuard, another directly relevant method discussed in the paper.

  5. Text quality evaluation is limited to log perplexity. There are no downstream task evaluations or broader assessments of semantic preservation and usefulness.

  6. The efficiency claims need stronger evidence. Tamper detection takes approximately 3.1 seconds for a 256-token text on a single CPU. Runtime comparisons with baselines and evaluations on longer documents are needed to assess the cost of the search-based detector.

Questions for authors
  1. What is the spoofing success rate when an attacker deliberately makes meaning-changing edits within the final block? Can you evaluate this targeted setting beyond the random final-block edits reported in Appendix C.6?

  2. With only four candidate chains, how often do modified blocks select the original next-block chain, and how does this affect tamper detection? What is the trade-off between increasing the number of chains and detection cost?

  3. Can you report the fraction of all attacked texts that remain watermark-positive while escaping tamper detection?

  4. How does LCW compare with DMark and LR-DWM on diffusion models, and with DualGuard in the existing autoregressive setting?

  5. Can you provide downstream task results to complement the log-perplexity evaluation?

3 comments

Q1 Final-Block Targeted Spoofing What is the spoofing success rate when an attacker deliberately makes meaning-changing edits within the final block?

Thank you for raising this important point. Following your suggestion, we additionally evaluate semantic piggyback spoofing attacks using the protocol in Appendix C.3, with modifications restricted to the final block. At NX=256N_X=256, LCW achieves an STR@5 of 0.346, demonstrating that LCW retains tamper detection capability even in this challenging setting.

Furthermore, a simple practical mitigation is to append non-essential, naturally generated watermarked tokens after the semantically important content. To reduce the predictability of the trailing watermark structure and make it harder for adaptive attackers to exploit a fixed-length extension, the number of additional tokens is determined by a pseudorandom function based on the latent-chain index of the final content block, with three possible lengths: 32, 40, or 48 tokens. These additional tokens follow LCW's chain transitions, providing further tamper evidence for modifications to the final content block. They can be discarded after watermark and tamper verification.

We evaluate this approach under two semantic piggyback spoofing settings at 5% FPR on 256-token texts:

  • Attacking the apparent final block. The attacker modifies the last 32 tokens of the extended text, assuming them to be the final content block. After removing the appended non-essential tokens, LCW achieves a WSR@5 of 0.956, showing that watermark detectability is largely preserved.
  • Attacking the actual final content block. The attacker instead targets the original final 32-token block, excluding the appended tokens. With the additional tokens retained during verification, LCW achieves an STR@5 of 0.595, compared with 0.346 without the extension.

These results suggest that the length extension preserves watermark detectability after removal of the trailing tokens while providing additional tamper evidence for modifications to the original final content block.

Q2 Chain Collision How often do modified blocks select the original next-block chain, and what is the trade-off between the number of candidate chains and detection cost?

Thank you for the question. We clarify that the number of candidate chains, NlcN_{\mathrm{lc}}, is a configurable hyperparameter rather than a fixed limitation of LCW. Under the uniform-hash assumption, the probability that a modified block selects the original next-block chain is approximately 1/Nlc1/N_{\mathrm{lc}}. We empirically measure this probability by comparing the HMAC-derived next-chain states before and after single-token modifications. The results closely match the theoretical expectation. We further evaluate how NlcN_{\mathrm{lc}} affects tamper detection and computational cost.

NlcN_{\mathrm{lc}}Theoretical CollisionEmpirical CollisionSub STR@5Sub STR@1Time (s)
40.2500.2510.5580.3722.097
60.1670.1690.7600.3543.088
80.1250.1230.7270.4354.077
160.0630.0620.7680.5827.018

Empirical collision rates are measured under single-token substitutions at NX=128N_X=128. Substitution STR is evaluated at NX=128N_X=128 under three token edits. Detection time is reported in seconds per text. Increasing NlcN_{\mathrm{lc}} reduces chain collisions and generally improves substitution tamper detection, at the cost of increased detection time. For example, increasing NlcN_{\mathrm{lc}} from 4 to 16 reduces the empirical collision rate from approximately 0.250 to 0.063, while substitution STR@5 improves from 0.558 to 0.768 and detection time increases from 2.097 to 7.018 seconds.

Importantly, chain collision does not necessarily imply tamper detection failure, as LCW also uses block-alignment evidence, particularly for insertion and deletion.

Q3 End-to-End Spoofing Success What fraction of all attacked texts remain watermark-positive while escaping tamper detection?

Thank you for the suggestion. We report watermark survival rate (WSR) and survival tamper detection rate (STR) separately to evaluate watermark robustness and tamper detection. WSR measures the fraction of edited watermarked texts that remain detectable as watermarked, while STR measures the fraction of these watermark-positive edited texts that are further identified as tampered.

Following the reviewer's suggestion, we additionally introduce the spoofing success rate (SSR), which measures the fraction of all attacked texts that retain a detectable watermark while escaping tamper detection: SSR=P(DW=1,DT=0)=WSR×(1−STR)\mathrm{SSR}=P(D_W=1,D_T=0) = \mathrm{WSR} \times (1-\mathrm{STR}). We report SSR using the same experimental protocol as Figure 3, focusing on five token-level edits (Nedit=5N_{\mathrm{edit}}=5), as shown below.

FPRLengthDel SSR ↓Ins SSR ↓Sub SSR ↓
5%640.0940.0600.401
5%1280.0100.0170.194
5%1920.0070.0030.147
5%2560.0030.0030.137
1%640.1370.0900.509
1%1280.0400.0540.445
1%1920.0270.0300.345
1%2560.0030.0130.324

The results demonstrate that LCW effectively mitigates piggyback spoofing attacks, achieving low spoofing success rates.

Q4 More Baselines How does LCW compare with DMark and LR-DWM on diffusion models, and with DualGuard in the autoregressive setting?

Thank you for the suggestion. Our original evaluation includes Unigram, GSJ, and CW, representing three different watermarking approaches applicable to dLLMs: fixed green-list watermarking, diffusion-specific watermarking, and order-agnostic watermarking with native edit detection, respectively.

Similar to GSJ, DMark and LR-DWM primarily focus on watermark detectability and robustness, without dedicated tamper-evidence mechanisms. We additionally evaluate both methods on DreamReasoner at NX=128N_X=128, following the same experimental protocol and watermark settings described in Appendix B.

Watermark detectability and robustness: Robustness is evaluated under 10% token-level edits, as shown below.

MethodTPR@5Del WSR@5Ins WSR@5Sub WSR@5
LR-DWM0.7460.7020.6640.530
DMark0.8440.7500.7660.742
LCW1.0000.7640.7770.975

Tamper detection: Since DMark and LR-DWM do not provide dedicated tamper detectors, we follow the same approach used for Unigram and GSJ, using their negative watermark scores as surrogate tamper signals. Tamper detection is evaluated under five token-level edits. In addition to STR, we also report the end-to-end tamper detection rate (E2E-TDR), which measures the fraction of all edited texts that remain watermark-positive and are correctly identified as tampered: $ \mathrm{E2E-TDR} = P(D_W=1,D_T=1)=\mathrm{WSR}\times\mathrm{STR}. $

MethodDel STR@5Del E2E-TDR@5Ins STR@5Ins E2E-TDR@5Sub STR@5Sub E2E-TDR@5
LR-DWM0.0410.0300.0200.0150.0000.000
DMark0.0580.0470.0750.0630.0000.000
LCW1.0000.9621.0000.9430.7870.777

The results show that LCW achieves the highest detectability, robustness, and tamper detection rates across all reported metrics.

Following the reviewer's suggestion, we further compare LCW with DualGuard, a tamper-evident watermark designed for autoregressive LLMs. WSR is evaluated under 10% token-level edits, while STR and E2E-TDR are evaluated under five token-level edits, at 5% FPR.

AttackMethodTPR@5WSR@5 (10%)STR@5 (5 edits)E2E-TDR@5 (5 edits)
DeletionLCW0.9900.8120.9900.941
DualGuard0.9930.9800.0440.044
InsertionLCW0.9900.8021.0000.941
DualGuard0.9930.9730.0450.044
SubstitutionLCW0.9900.9600.9000.891
DualGuard0.9930.9630.0410.040

Although DualGuard also aims to achieve both robustness and tamper evidence, its tamper detection primarily targets significant semantic changes. It achieves slightly higher watermark robustness than LCW but shows limited sensitivity to token-level modifications.

Q5 Downstream Task Quality Can downstream task evaluations be provided to complement the log-perplexity analysis?

Thank you for the suggestion. We further evaluate the impact of watermarking on two downstream tasks: text summarization and machine translation (English to Romanian). All experiments use DreamReasoner. All metrics are higher-is-better. ROUGE, BLEU, and chrF are reported on a 0–100 scale, while BERTScore-F1 uses its original scale.

Text Summarization:

MethodROUGE-1ROUGE-2ROUGE-LBERTScore-F1
No-WM31.8010.9922.550.719
GSJ31.5310.3921.610.717
Unigram28.229.3019.710.708
LCW30.189.6620.900.709

Machine Translation (English to Romanian):

MethodBLEUchrFBERTScore-F1
No-WM10.1637.910.755
GSJ7.3930.640.731
Unigram8.0935.370.740
LCW7.9335.530.742

The results show that LCW maintains competitive downstream task performance.

W6 Detection Efficiency How does the detection runtime compare with baselines, and how does LCW scale to longer documents?

Thank you for the comment. LCW introduces additional detection costs because it searches over block boundaries, latent chains, and within-block alignments to achieve both robustness and tamper evidence. We mitigate this cost through block-level beam search and within-block dynamic programming. For 256-token texts, LCW requires 2.10 seconds for watermark detection and 3.06 seconds for tamper detection on a single CPU (Table 9). In comparison, Unigram, GSJ, DMark, and LR-DWM require less than 0.1 seconds for watermark detection in our measurements, as they do not perform such structural optimization. However, these methods do not provide dedicated tamper-evidence detectors. DualGuard, which requires an additional semantic encoder, reports an average detection time of approximately 0.42 seconds.

For longer documents, a scalable extension is to divide the text into 256-token segments and process them in parallel, followed by separate verification of cross-segment alignment and latent-chain consistency. This avoids performing a single global search over the entire document.

Reviewer_XRX8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Latent Chain Watermark (LCW), a watermarking method for diffusion LLMs. The method addresses a specific threat model: robustness while providing tamper evidence. This threat model exists in the literature, and existing solutions propose embedding two different watermarks: one robust and one fragile. In this paper, LCW combines watermark robustness and tamper evidence. To do so, the authors follow the green-list/red-list watermarking paradigm for blockwise diffusion LLMs and add tamper-evidence detection.

Critical strengths

The paper's setting is excellent: studying the piggyback spoofing attack for diffusion LLMs is important and well motivated. The authors also demonstrate in Table 2 that despite its simplicity, their method is quite general and directly applies to autoregressive LLMs.

Critical weaknesses

Overall the paper is held back by two main weakness:

Weak conceptual novelty: The method uses the green-list/red-list approach proposed by Kirchenbauer et al. at the block level. The context of the previous block is used to choose among 4 chains of green lists per block, with the list for each token depending on the position of the token in the block. This construction allows the authors to compute a tamper score by comparing the block's score with the score that would have been optimal using the best list chain. Such an approach is completely incremental and, most importantly, does not provide any theoretical FPR guarantees.

Non-rigorous evaluation protocol: The paper lacks critical information for both replicability and assessment of the experimental results. I refer the authors to Question section for an extensive list. To summarize, no sample size is provided, the number of samples used to compute the FPR thresold of each method is never stated. More importantly, watermarking requires ultra-low false positive rates. Here the authors chose 0.010.01 a 0.050.05. To be of practical use, a watermarking method requires at least 10−610^-6. Obviously, this is difficult to guarantee when the method does not allow to compute a p-value easily -- I would enjoin the authors to make their method follow current good practices, as stated in [1] for example, by casting the watermarking problem as a sound statistical problem. This would allow to guarantee the FPR apriori instead of empirically. I admit that I am unsure if it will be easy due to the mixing of a fragile and robust watermark.

[1] Fernandez, P., Chaffin, A., Tit, K., Chappelier, V., & Furon, T. (2023, December). Three bricks to consolidate watermarks for large language models. In 2023 IEEE international workshop on information forensics and security (WIFS) (pp. 1-6). IEEE.

Questions for authors

For this paper, I tend towards a rejection because of (1) the limited conceptual novelty and (2) the evaluation protocol is not rigorous enough.

However, my score can improve if the authors can clarify or justify the following points:

  • There is no information about the number of samples used to compute the FPR threshold for each method.
  • The results on GSJ in the paper at TPR @ 1% FPR are much lower than in the original paper, which achieves 0.99. Why is there a difference?
  • The impact on text quality is only discussed in the appendix and is not fixed for each method. The log perplexity presented for LCW is always higher than for GSJ, so the comparison is not at equal impact on the text. It would be interesting to plot the results as a function of PPL.
  • The paper provides no details on the sample sizes tested or the variance of the results. Are the results really significant?
  • 5% and 1% are high FPR values. How does the method behave for lower rates, say 10−610^-6?
  • Are the results computed after a decoding re-encoding by the tokenizer to analyze the effects of possible desynchronization?
5 comments

W1 Weak conceptual novelty What is the conceptual novelty of LCW beyond applying conventional green-list watermarking at the block level?

Thank you for the comment. We agree that extending KGW to the block level is a natural approach to introducing tamper evidence in dLLMs. However, this alone is insufficient to achieve both robust and tamper-evident detection. As discussed in Section 1, LCW addresses two key challenges:

    1. Establishing tamper evidence without relying on provisional tokens. A straightforward block-level KGW extension can introduce tamper evidence by using finalized blocks to determine the green lists of subsequent blocks. However, text modifications can disrupt block boundaries and chain transitions, causing watermark detection to fail. Without reliable watermark detection after modifications, it is insufficient to achieve either robustness or tamper detection.
    1. Preserving robustness while providing reliable tamper evidence. LCW addresses this tension through two complementary detection procedures built upon the same latent-chain structure. Robust watermark detection searches over candidate block boundaries, latent chains, and within-block alignments to recover watermark evidence, while tamper detection measures the watermark score losses caused by enforcing block alignment and chain transitions.

The key novelty of LCW is therefore not merely introducing block-level tamper evidence, but enabling robustness and tamper evidence through the same latent-chain structure.

W2 Number of Candidate Chains

We clarify that the number of candidate chains, NlcN_{\mathrm{lc}} is a configurable hyperparameter rather than a limitation of the proposed method. We use Nlc=4N_{\mathrm{lc}}=4 as the default experimental setting, but have already evaluated Nlc=6N_{\mathrm{lc}}=6 in Tables 3 and 4 of Appendix C.1. We additionally consider Nlc=8N_{\mathrm{lc}}=8 and 1616 to to further investigate its impact on watermark performance.

We report watermark survival rate (WSR), which measures the fraction of edited texts that remain watermark-positive, and survival tamper detection rate (STR), which measures the fraction of watermark-positive edited texts identified as tampered. @5 denotes evaluation at a 5% false-positive rate (FPR).

NlcN_{\mathrm{lc}}TPR@5Del WSR@5Ins WSR@5Sub WSR@5Del STR@5Ins STR@5Sub STR@5
40.9830.7990.8230.9770.9510.9690.558
60.9900.7920.7820.9800.9900.9790.760
80.9840.7270.7440.9850.9930.9710.727
160.9860.7150.7280.9820.9960.9730.768
NlcN_{\mathrm{lc}}TPR@1Del WSR@1Ins WSR@1Sub WSR@1Del STR@1Ins STR@1Sub STR@1
40.9500.5990.6050.9160.9020.9280.372
60.9700.5540.4650.9310.8920.7710.354
80.9860.4640.4770.8960.9020.8100.435
160.9970.4360.4620.8330.9350.8720.582

Results on DreamReasoner at 128 tokens. Robustness is evaluated under 10% token-level substitution, and tamper evidence under three token-level edits. The results show that increasing the number of candidate chains generally strengthens tamper detection, particularly for substitution, but may reduce watermark robustness.

W3 Theoretical FPR Guarantee Can LCW provide theoretically justified false-positive-rate guarantees rather than relying solely on empirical threshold calibration?

Thank you for this valuable comment. To address this concern, we formulate LCW watermark detection as a statistical hypothesis test, where the null hypothesis assumes that the input text is independent of the watermark randomness. As defined in Eq. (3)–(4), LCW computes the penalized watermark score ZW(x)Z_W(x) by searching over candidate configurations and declares a watermark when ZW(x)≥ταZ_W(x)\ge\tau_\alpha.

We derive a finite-sample FPR upper bound for LCW's original detection statistic without assuming a Gaussian distribution for the maximized score. Specifically, for any fixed input sequence xx of length nn with pairwise-distinct token IDs, under an ideal random-green-list model, we establish Pr⁡0[ZW(x)≥t∣C]≤Un(t)\Pr_0[Z_W(x)\ge t\mid\mathcal C]\le U_n(t), where C\mathcal C denotes the latent-chain index table, and Un(t)U_n(t) is a computable upper bound on the false-positive probability. The probability is taken over the ideal watermark randomness, assuming independent, uniformly sampled fixed-size green lists. The bound holds uniformly over all admissible latent-chain index tables.

The proof proceeds in three steps. First, we derive necessary optimality conditions for penalized within-block alignments, allowing us to exclude alignments that cannot be optimal. Second, we use finite-state weighted dynamic programming to bound the probability of high-scoring alignments. Third, we account for the search over candidate latent chains, block boundaries, and block-key phases using probability bounds, together with a finite-vocabulary correction for fixed-size green lists.

Importantly, the bound covers the complete search space without assuming independence between candidate configurations. It also applies to beam-search detection because the returned configurations belong to the admissible search space. The analysis leaves LCW's original scoring function and detection algorithm unchanged.

For any target FPR α\alpha, we obtain a theoretically sufficient detection threshold τα\tau_\alpha by requiring Un(τα)≤α.U_n(\tau_\alpha)\le\alpha.

For an illustrative configuration at NX=128N_X=128, using LCW's default watermark parameters, and full-length terminal chains, we obtain τ10−6=7.82\tau_{10^{-6}}=7.82, yielding U128(7.82)≈8.20885×10−7<10−6.U_{128}(7.82) \approx 8.20885 \times10^{-7}<10^{-6}.

Q1 Sample Size and Significance The paper provides no details on the sample sizes tested or the variance of the results. Are the results really significant?

Thank you for the comment. Following existing studies [1,2,3], we used 1,000 non-watermarked samples for FPR threshold calibration and 1000 watermarked samples per setting for evaluation. To further assess statistical reliability, we report Wilson 95% confidence intervals for watermark detectability (TPR), watermark survival rate (WSR), and survival tamper detection rate (STR) across different text lengths and editing attacks, as shown below.

NXN_XFPRTPRDel WSRIns WSRSub WSRDel STRIns STRSub STR
645%0.819 [0.772, 0.859]0.599 [0.542, 0.653]0.512 [0.455, 0.568]0.766 [0.715, 0.810]0.814 [0.756, 0.861]0.811 [0.751, 0.860]0.485 [0.422, 0.549]
1285%0.983 [0.962, 0.993]0.799 [0.750, 0.841]0.823 [0.775, 0.862]0.977 [0.953, 0.989]0.951 [0.919, 0.971]0.969 [0.941, 0.983]0.558 [0.501, 0.614]
2565%1.000 [0.987, 1.000]0.756 [0.704, 0.801]0.880 [0.838, 0.912]0.993 [0.976, 0.998]0.983 [0.961, 0.993]0.993 [0.976, 0.998]0.602 [0.546, 0.656]
641%0.649 [0.593, 0.701]0.388 [0.335, 0.444]0.278 [0.230, 0.331]0.572 [0.515, 0.627]0.627 [0.547, 0.700]0.609 [0.524, 0.688]0.126 [0.086, 0.183]
1281%0.950 [0.919, 0.969]0.599 [0.542, 0.653]0.605 [0.549, 0.659]0.916 [0.880, 0.943]0.902 [0.860, 0.932]0.928 [0.890, 0.953]0.372 [0.318, 0.430]
2561%1.000 [0.987, 1.000]0.622 [0.566, 0.675]0.776 [0.725, 0.820]0.990 [0.971, 0.997]0.970 [0.943, 0.984]0.977 [0.952, 0.989]0.401 [0.347, 0.458]

We additionally conduct two-sided Fisher's exact tests with Holm correction for multiple comparisons. For detectablity, at 5% FPR, LCW achieves significantly higher clean True Positive Rate (TPR) than GSJ (p=3.34×10−18p=3.34\times10^{-18}) and CW (p=4.61×10−51p=4.61\times10^{-51}), with no statistically significant difference from Unigram. For robustness, LCW significantly outperforms GSJ and CW across deletion, insertion, and substitution attacks (all adjusted p<0.001p<0.001), while achieving comparable substitution robustness to Unigram.

Q2 GSJ Reproduction Why is the TPR@1% FPR of GSJ substantially lower than the result reported in the original GSJ paper?

Thank you for pointing out this difference. One important reason is the different watermark strengths used in the two studies. Following a common controlled-comparison setting [1,2,3], we use the same green-list fraction (γ=0.25\gamma=0.25) and logit bias (δ=2\delta=2) across all watermarking methods to ensure a consistent evaluation of different watermark mechanisms. In contrast, the original GSJ paper uses a stronger watermark bias (δ=4\delta=4) to achieve higher detectability. Therefore, the reported TPR values are obtained under different watermark strengths and are not directly comparable.

Q3 PPL Comparison How does LCW compare with other watermarking methods at a comparable impact on text quality, rather than under the same logit bias?

Thank you for the suggestion. Our goal is to achieve both robustness and tamper evidence, whereas GSJ primarily focuses on detectability and robustness, with limited tamper-detection capability, as demonstrated in Figure 3. Nevertheless, we agree that a quality-matched comparison is valuable to verify that LCW's detection performance is not simply achieved at the expense of text quality. To this end, we additionally evaluate LCW and GSJ under different watermark strengths and compare their detectability and robustness.

Methodγ\gammaδ\deltaMean log(PPL) ↓TPR@5WSR@5 (Del/Ins/Sub)STR@5 (Del/Ins/Sub)
LCW0.2522.0591.0000.771 / 0.790 / 0.9751.000 / 0.980 / 0.747
GSJ0.2522.0580.9250.817 / 0.842 / 0.8170.111 / 0.092 / 0.057
LCW0.2542.7451.0001.000 / 1.000 / 1.0001.000 / 0.991 / 0.861
GSJ0.2542.6671.0001.000 / 1.000 / 1.0000.082 / 0.102 / 0.092
LCW0.5021.9880.9370.614 / 0.528 / 0.8740.990 / 0.978 / 0.767
GSJ0.5021.9920.9000.825 / 0.800 / 0.7920.077 / 0.048 / 0.131
LCW0.5042.3420.9920.789 / 0.813 / 0.9920.992 / 1.000 / 0.890
GSJ0.5042.2860.9910.982 / 0.982 / 0.9910.101 / 0.073 / 0.083

Results on DreamReasoner at NX=128N_X=128 and 5% FPR. Robustness is evaluated under 10% token-level edits, and tamper detection under five token-level edits. At comparable text quality, LCW maintains competitive watermark robustness while providing substantially stronger tamper-detection capability than GSJ.

Q4 Lower FPR How does LCW perform at substantially lower false-positive rates, such as 10−610^{-6}?

Thank you for the suggestion. Our original evaluation uses 1% and 5% FPR, following common evaluation protocols in existing LLM watermarking studies [1,2,3]. We agree that ultra-low FPR is important for practical deployment. As established in our response to W3, we derive a theoretical FPR upper bound for LCW's search-based detection procedure. Under the stated assumptions, this bound yields a sufficient detection threshold of τ10−6=7.82\tau_{10^{-6}}=7.82 for 128-token texts using LCW's default configuration. At this threshold, LCW achieves a TPR of 0.980, demonstrating strong watermark detectability even at a theoretically controlled FPR of 10−610^{-6}.

Q5 Tokenizer Desynchronization Are the results computed after a decoding–re-encoding process to evaluate tokenizer desynchronization?

Thank you for the question. Yes, our experiments include the decoding–re-encoding process. Specifically, generated token sequences are decoded into text and then re-encoded using the tokenizer before watermark detection.

[1] J. Kirchenbauer, J. Geiping, Y. Wen, J. Katz, I. Miers, and T. Goldstein, “A Watermark for Large Language Models,” in ICML, Jan. 2023. [2] T. Gloaguen, R. Staab, N. Jovanović, and M. Vechev, “Watermarking Diffusion Language Models,” in ICLR, Sep. 2025. [3] X. Zhao, P. Ananth, L. Li, and Y.-X. Wang, “Provable Robust Watermarking for AI-Generated Text,” in ICLR, Oct. 2023.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Point-by-Point Responses

We sincerely thank all reviewers for their constructive comments and valuable suggestions. We find that these comments provide us with the chance to explain our proposed method more clearly.

Strengths Highlighted by Reviewers

  • Meaningful research problem. Achieving both watermark robustness and tamper evidence in diffusion LLMs is an important and challenging problem (XRX8, f7ye, LCU6).
  • Latent-chain design. Fixed within-block latent chains and content-dependent chain transitions support parallel generation while introducing tamper evidence (f7ye, LCU6).
  • Complementary detection framework. The separation of robust watermark detection and tamper detection, together with beam search and dynamic programming, provides a practical solution (LCU6).
  • Broad applicability. LCW extends beyond diffusion LLMs to autoregressive and speculative decoding (XRX8, LCU6).

Questions and Weaknesses

Reviewer XRX8

  • W1. Conceptual novelty. Clarified the two key challenges and how LCW achieves both robustness and tamper evidence through complementary detection procedures.
  • W2. Number of candidate chains. Extended the ablation to Nlc∈{4,6,8,16}N_{\mathrm{lc}}\in\{4,6,8,16\}, analyzing the trade-off between robustness and tamper detection.
  • W3. Theoretical FPR guarantees. Derived a finite-sample FPR upper bound for LCW's original search-based detector under explicit assumptions.
  • Q1. Sample size and significance. Clarified sample sizes and added Wilson 95% confidence intervals and statistical significance tests.
  • Q2. GSJ reproduction. Explained the differences in watermark strength and experimental settings.
  • Q3. PPL comparison. Added quality–performance comparisons under different watermark strengths.
  • Q4. Ultra-low FPR. Obtained a theoretically sufficient threshold for 10−610^{-6} FPR under the stated assumptions, with TPR = 0.980.
  • Q5. Tokenizer desynchronization. Clarified the decoding–re-encoding procedure used in our evaluation.

Reviewer f7ye

  • Q1. Final-block targeted spoofing. Evaluated targeted semantic piggyback spoofing restricted to the final block and investigated a practical mitigation using additional watermarked tokens.
  • Q2. Chain collision. Measured empirical chain collision rates and analyzed the trade-off between candidate-chain diversity, tamper detection, and computational cost.
  • Q3. End-to-end spoofing success. Added spoofing success rate (SSR) evaluations corresponding to Figure 3.
  • Q4. Additional baselines. Added comparisons with DMark, LR-DWM, and DualGuard.
  • Q5. Downstream task quality. Added text summarization and machine translation evaluations.
  • W6. Detection efficiency. Compared detection runtime with baselines and discussed a segmented detection approach for longer documents.

Reviewer LCU6

  • Q1. Paraphrasing robustness. Evaluated additional DIPPER configurations and reported watermark survival, tamper detection, and end-to-end detection rates.
  • Q2. Targeted tampering. Added position-wise tamper detection analysis and discussed mitigation of final-block vulnerabilities.
  • Q3. Reliability of tamper evidence. Analyzed clean and tampered score distributions and evaluated both tamper signals under different watermark strengths.
  • Q4. End-to-end tamper detection. Added joint detection results under token-level editing and paraphrasing attacks.
  • Q5. Baseline comparison fairness. Clarified the tamper detection procedures and threshold calibration used for different watermarking methods.

We hope these clarifications and additional results address the reviewers' main concerns and provide stronger evidence for LCW's ability to combine watermark robustness with tamper evidence in diffusion LLMs.

We sincerely thank the reviewers for their time and thoughtful feedback. We hope our detailed responses and additional experiments have addressed your concerns. Please feel free to let us know if you have any further questions or concerns. We would be happy to provide additional clarifications.

Dear Reviewers,

As the discussion phase is approaching its end, we would like to sincerely thank you again for your time and constructive feedback. We hope that our detailed responses, additional experiments, and theoretical analyses have satisfactorily addressed your concerns. We would greatly appreciate it if you could let us know whether any questions or concerns remain, and we would be happy to provide further clarifications.

If you find that our responses have adequately addressed your concerns, we would also be grateful if you could consider updating your ratings accordingly.

Thank you again for your careful consideration and valuable feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.