acceptodds
Under review as a conference paper at ICLR 2027

TANGO: Watermarking Masked Diffusion Language Models in Token Pairs

Abstract

Masked-diffusion language models fill in masked positions in parallel and in no fixed order. Most practical text watermarks assume left-to-right generation. They key each token to the tokens before it, and in a diffusion model those tokens may still be masked. A fixed green list needs no such context, but it favors the same tokens at every position, so these tokens appear more often in watermarked text. An attacker who compares token frequencies in watermarked and unwatermarked text can recover the list and forge text that the provider's own detector accepts. We present **TANGO**, a watermark for masked-diffusion language models that keys each new token to a nearby token that is already unmasked. A secret key splits the vocabulary into color classes, and TANGO biases the new token toward a color determined by the key and the nearby token's color. The watermark is therefore embedded in *pairs* of tokens. Because the favored color changes from position to position, token frequencies stay close to those of unwatermarked text. Detection needs only the text and the key, and it does not assume any unmasking order. On two masked-diffusion models, TANGO detects nearly all unedited watermarked texts and most edited ones, and frequency attacks that forge the fixed green list fail against it.

Primary area:
generative models
Keywords:
watermarking, diffusion language models, masked diffusion models, text watermarking, watermark forgery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_BjG4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TANGO, a watermarking method for masked diffusion language models that uses a secret vocabulary coloring and modular checksum relations between token pairs. The method guides token selection using already-unmasked neighbors, requires no left-to-right generation order, and does not need to reconstruct the generation order during detection.

Critical strengths
  • The method naturally accommodates out-of-order generation, and both its generation and detection mechanisms are clearly described.
  • The evaluation covers two masked diffusion language models and includes editing and forgery attacks, with useful calibration diagnostics and ablation studies in the appendix.
  • The authors discuss the method's limitations candidly, including failures under deletion, paraphrasing, and pair-based attacks, which helps clarify its applicability.
Critical weaknesses
  • Missing comparison with the closest baseline. Section 5 identifies LR-DWM [1] as the closest work but omits it from the experiments. Both methods use already-unmasked neighbors and logit biasing, so a direct comparison is needed to establish gains in detection, editing robustness, or forgery resistance.

  • Limited applicability of Theorem 1. Assumption U replaces tap-dependent favored colors with independent uniform draws, removing the pairwise coupling and reducing the detection margin to zero (Appendix A.1). The authors should clarify which frequency-hiding conclusions remain valid under the actual tap dependence and distinguish these guarantees from empirical observations.

  • Unclear scope of resistance to forgery. Pair-chaining forgeries succeed in 74% of cases with 200 collected texts, and the language-model-based attack reaches 55% at the stronger bias (Tables 9–10). Since all pairs share one coloring, the O(V²) pair-counting argument may not capture the difficulty of joint color inference.

Questions for authors
  1. Comparison with prior work: Under matched decoding settings and comparable text quality, how does TANGO compare experimentally with LR-DWM [1] in detectability, robustness to editing, and resistance to forgery? Please also discuss its relationship to [2][3] in design, security objectives, and applicability.

  2. Theorem 1: Which conclusions remain valid when the actual dependence of the favored color on the tap's color is retained? Please clarify the implications for token-frequency shifts in the implemented algorithm and identify which conclusions still rely on empirical evidence.

  3. Scope of resistance to forgery: Considering approaches that jointly recover hidden classes [3], could the shared color structure reduce the learning difficulty? Does the O(V²) argument apply specifically to the evaluated pair-counting attack, or does it also constrain attacks that exploit shared structure? Please clarify the scope of the security conclusions accordingly.

[1] Raban, O., et al. [LR-DWM: Efficient Watermarking for Diffusion Language Models][1]. Findings of the Association for Computational Linguistics: ACL 2026, 43510–43517, 2026.

[2] Shen, H., Huang, B., and Wan, X. [Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks][2]. Advances in Neural Information Processing Systems, 38, 2025.

[3] Zhang, Z., et al. [Large Language Model Watermark Stealing With Mixed Integer Programming][3]. arXiv preprint arXiv:2405.19677, 2024.

6 comments

We thank the reviewer for the helpful suggestions. We are glad that you find that the method "naturally accommodates out-of-order generation" and that its generation and detection are "clearly described". We respond to the weaknesses (W) and the questions (Q) below.


W1 and Q1. "Under matched decoding settings and comparable text quality, how does TANGO compare experimentally with LR-DWM in detectability, robustness to editing, and resistance to forgery?"

We ran this comparison, and it is now in the main text (Table 2\textrm{\color{blue}Table 2} and Sec. 4.4\textrm{\color{blue}Sec. 4.4}, Page 9). Under matched decoding settings and at comparable text quality, TANGO detects more texts than LR-DWM, both unedited and after synonym substitution and insertion.

We implemented LR-DWM from its paper. Both methods run on LLaDA with the same decoding settings and the same 200 prompts, and both detectors are set to a 1% FPR on human text. The table shows the perplexity of the generated text (12.9 without a watermark) and the TPR on unedited text, after replacing 30% of the tokens with synonyms, and after inserting 20% new tokens. δ and β are the watermark strengths. Each pair of rows has similar perplexity.

MethodPerplexityUneditedSynonyms 30%Insertion 20%
LR-DWM, δ = 3.2515.70.8450.3950.610
TANGO, β = 416.50.9500.7150.780
TANGO, β = 621.00.9950.9250.910
LR-DWM, δ = 522.80.9900.7150.880

As the table shows, TANGO is more robust than LR-DWM to both edits. On resistance to forgery, an attacker who counts token pairs forges the two methods at similar rates (Table 28\textrm{\color{blue}Table 28}, Page 44), and forgeries that preserve the quality of the text pass the detector in at most 6% of attempts.


Q1. "Please also discuss its relationship to [2][3] in design, security objectives, and applicability."

The revision now discusses SEEK and the mixed-integer programming attack (Sec. 5\textrm{\color{blue}Sec. 5}, Page 10, and Appendix G\textrm{\color{blue}Appendix G}, Pages 45 and 46). SEEK has the same security objectives as TANGO, which are to resist both scrubbing and spoofing. SEEK is designed for left-to-right generation. It keys each token on the tokens before it, which a masked diffusion model may not have generated yet. TANGO keys each token on one already generated neighbor, so it applies to masked diffusion models. The mixed-integer programming attack recovers the hidden token classes of a watermark jointly. We ran an attack of this kind against TANGO and report it under W3 below.


W2 and Q2. "Which conclusions remain valid when the actual dependence of the favored color on the tap's color is retained? Please clarify the implications for token-frequency shifts in the implemented algorithm and identify which conclusions still rely on empirical evidence."

The main conclusion is unchanged. An attacker cannot forge TANGO's watermark from token frequencies. For the idealized sampler, Theorem 1 concludes that the watermark does not change expected token frequencies and that a token's frequency does not reveal its color. Our measurements show that both conclusions remain approximately valid for the implemented algorithm, even when accounting for the dependence. An attacker who guesses each token's color from its frequency does barely better than chance against TANGO (Table 8\textrm{\color{blue}Table 8}, Page 32):

TANGORed–green list
Attacker AUC (0.5 is chance)0.580.82
Forgeries that pass the detectornoneall

We retitled Theorem 1 "Frequency cancellation under an idealized sampler" and state its scope in Sec. 3.4\textrm{\color{blue}Sec. 3.4} (Page 6). Table 24\textrm{\color{blue}Table 24} (Page 43) lists which results are proved and which are measured.

W3 and Q3. "Considering approaches that jointly recover hidden classes [3], could the shared color structure reduce the learning difficulty? Does the O(V²) argument apply specifically to the evaluated pair-counting attack [...]? Please clarify the scope of the security conclusions accordingly."

Yes, the O(V²) argument applies to the pair-counting attack (Sec. 4.3\textrm{\color{blue}Sec. 4.3}, Page 8). Following your comment, we added a joint-recovery attack. The attack recovers the shared coloring from all observed token pairs at once, without the key. We ran three variants, one of them based on the mixed-integer program of the paper that you mentioned, and we report the most successful one. In the evaluated setting, the attacker uses the recovered coloring to generate text that imitates the watermark. A forgery passes if the detector accepts it at 1% FPR.

This attack is successful when the attacker can collect a large amount of watermarked text under a single key, so rotating keys limits its effectiveness. With as few as eight keys, at most 7% of its forgeries pass.

TANGO's detector checks a text against every key. Here, we assume that the attacker knows which key produced each text, which is the best case for the attacker (Table 22\textrm{\color{blue}Table 22}, Page 41):

Number of keys1248
Forgeries that pass, 38,400 watermarked tokens62%27%9%5%
Forgeries that pass, 115,200 watermarked tokens92%75%42%7%

An attacker who cannot tell which key produced each text does worse. With four keys, at most 5% of its forgeries pass.

With 32 colors instead of 3, TANGO roughly halves this attack's forgery success rate under a single key at the same text quality and is more robust to deletion (Sec. 4.5\textrm{\color{blue}Sec. 4.5}, Page 9, and Table 30\textrm{\color{blue}Table 30}, Page 45).

Regarding the high success rates cited for pair chaining (Table 9), we note that they were achieved only with token chains whose median perplexity is above 1000. The 55% success of the language-model attack (Table 10) is also achieved only at a perplexity 61% above that of unwatermarked text. Within a reasonable perplexity of 1.2× that of unwatermarked text, at most 6% of this attack's forgeries pass (Table 23\textrm{\color{blue}Table 23}, Page 41).

Sec. 6\textrm{\color{blue}Sec. 6} (Page 10) now states the scope more clearly.

Dear Reviewer BjG4,

As the discussion deadline approaches, we wanted to follow up to ensure our responses have adequately addressed your concerns. If there are any additional questions or points for discussion, we would be happy to address them promptly.

We look forward to your feedback and sincerely appreciate the time and effort you put into reviewing our work.

Reviewer_8wBtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TANGO, a watermarking method for masked diffusion language models. It assigns vocabulary tokens to keyed color classes and biases token generation to satisfy a color relation with an already unmasked neighbor. Detection counts matching token pairs using only the final text and secret key, without requiring the generation order. The design aims to reduce leakage through individual token frequencies. Experiments on LLaDA and Dream evaluate detection, robustness to editing, and resistance to forgery.

Critical strengths
  1. The pairwise color constraint supports generation from either direction and allows detection from the final text without recovering the unmasking order.

  2. Semantic coloring provides a simple way to preserve watermark evidence when substitutions replace tokens with others in the same color class.

  3. The implementation is lightweight: coloring is precomputed, and detection requires a single linear pass over the token sequence.

  4. The security evaluation includes both token-frequency and stronger token-pair attacks, helping clarify the scope and limits of the method’s resistance to forgery.

Critical weaknesses
  1. Using already generated tokens from either direction has been explored in LR-DWM and DMark. TANGO introduces a different pair constraint, but lacks direct experimental comparisons with these closely related methods. Its added value is therefore difficult to assess.

  2. The frequency-hiding guarantee assumes uniform, independent favored colors and equal probability mass across color classes. These assumptions do not generally hold for the actual sampler. In particular, equal vocabulary sizes do not guarantee equal probability mass.

  3. In Table 10, an LM-assisted pair attack achieves 55% forgery success after observing 200 watermarked texts, although it also reduces text quality. The results support resistance to simple frequency attacks more clearly than broad resistance to forgery.

  4. The main experiments use only 200 examples for LLaDA and 100 for Dream. Such small negative sets provide limited evidence for performance at 1% FPR. Larger, separate calibration and test sets are needed.

  5. On Dream, filtered perplexity increases from 3.8 to 7.8, while the rate of degenerate outputs rises from 20% to 31%. Deletion robustness also falls substantially below Unigram. Downstream task evaluations would help establish whether the security benefit justifies these costs.

Questions for authors
  1. What advantage does the modular pair constraint provide over the bidirectional designs in LR-DWM and DMark? Can you provide direct comparisons at similar text quality?

  2. How closely does the actual sampler satisfy assumptions (U) and (B)? Can you measure the deviations and show how they affect token-frequency leakage?

  3. Table 10 reports 55% forgery success under the stronger attack setting. How does attack success change with larger observation budgets and different requirements for forged-text quality?

  4. Are separate negative sets used for threshold calibration and FPR evaluation? Can you validate the reported 1% FPR using substantially larger negative sets?

  5. Can you provide downstream task evaluations to assess the quality cost beyond perplexity, especially given the increased rate of degenerate outputs on Dream?

3 comments

We thank the reviewer for the detailed feedback. We are glad that you find the implementation lightweight and that the security evaluation helps "clarify the scope and limits" of the method. We respond to the weaknesses (W) and the questions (Q) below.


W1 and Q1. "What advantage does the modular pair constraint provide over the bidirectional designs in LR-DWM and DMark? Can you provide direct comparisons at similar text quality?"

We ran both comparisons, and they are now in the main text (Table 2\textrm{\color{blue}Table 2} and Sec. 4.4\textrm{\color{blue}Sec. 4.4}, Page 9). At similar text quality, TANGO is more robust than LR-DWM to synonym substitution and insertion. TANGO detects as many texts as DMark and is harder to forge. When the attacker has seen 6,400 watermarked tokens, the forgery passes 51% of the time against TANGO and 79% against DMark (Table 28\textrm{\color{blue}Table 28}, Page 44).

We implemented LR-DWM and DMark from their papers and ran all three methods on LLaDA with the same generation setup and prompts. The table shows the perplexity of the generated text (12.9 without a watermark) and the TPR at 1% FPR on unedited text, after replacing 30% of the tokens with synonyms, and after inserting 20% new tokens. δ and β are the watermark strengths.

MethodPerplexityUneditedSynonyms 30%Insertion 20%
LR-DWM, δ = 3.2515.70.8450.3950.610
TANGO, β = 416.50.9500.7150.780
DMark, δ = 3.2517.80.9700.7850.920
TANGO, β = 5 (default)18.90.9650.8050.850
TANGO, β = 621.00.9950.9250.910
LR-DWM, δ = 522.80.9900.7150.880

W2 and Q2. "How closely does the actual sampler satisfy assumptions (U) and (B)? Can you measure the deviations and show how they affect token-frequency leakage?"

Following your comment, we measured how closely both assumptions hold on the LLaDA model (Appendix F.2\textrm{\color{blue}Appendix F.2}, Pages 42 and 43). The deviations do not make TANGO forgeable from token frequencies. No forgery built from token frequencies passes TANGO's detector, and every one passes the detector of Unigram (red–green list) (Table 8\textrm{\color{blue}Table 8}, Page 32).

Assumption (U) says that the favored color at each position is uniformly random, so the model cannot predict it. The real sampler is very close to this idealized case. Using three colors, the favored color receives 1/3 + 0.019 of the model's probability on average.

Assumption (B) says that the model's probability is split equally among the three colors. As you note, equal vocabulary sizes do not guarantee this. The largest deviation of a color's probability from 1/3 averages 0.51 in absolute probability, because the model often concentrates its probability on a few tokens. Yet the effect on token frequencies is small. An attacker who tries to recover the watermark’s token colors from word frequencies does barely better than chance against TANGO (AUC 0.58), but well above chance against Unigram (AUC 0.82).

To be more accurate, Theorem 1 is now titled "Frequency cancellation under an idealized sampler" (Sec. 3.4\textrm{\color{blue}Sec. 3.4}, Page 6).


W3 and Q3. "Table 10 reports 55% forgery success under the stronger attack setting. How does attack success change with larger observation budgets and different requirements for forged-text quality?"

A larger observation budget does not help this attack. With 900 watermarked texts instead of 200, its success at the same bias is 53% (Table 23\textrm{\color{blue}Table 23}, Page 41). Moreover, requiring text quality reduces success sharply. When the forged text must stay within 1.2× the perplexity of unwatermarked text, at most 6% of the forgeries pass the detection threshold.

The attack in Table 10 biases a language model toward the token pairs that it saw in watermarked text, with a bias βforge\beta_{\mathrm{forge}}. A larger bias passes the detector more often and lowers the text quality. The table gives the results with 900 watermarked texts.

Attacker's bias βforge\beta_{\mathrm{forge}}46810
Perplexity of the forged text, relative to unwatermarked text1.04×1.38×2.49×4.69×
Forgeries that pass the 1% FPR threshold4%14%53%87%
Forgeries that pass the 1% FPR threshold (stay within 1.2× perplexity)1%2%3%6%

For comparison, with 200 watermarked texts and bias 6, at a stricter detection threshold (z > 4), the forgery passes 91% of the time against Unigram and 1% against TANGO.

Yet, we agree that the scope of the claim should be more precise. The revision now specifies resistance to frequency attacks and also addresses stronger attacks (Sec. 4.3\textrm{\color{blue}Sec. 4.3}, Page 8).

W4 and Q4. "Are separate negative sets used for threshold calibration and FPR evaluation? Can you validate the reported 1% FPR using substantially larger negative sets?"

Yes. The revision uses 5,000 human-written texts, split equally between calibration and testing, and maintains the 1% FPR (Sec. 4.2\textrm{\color{blue}Sec. 4.2}, Page 8, and Table 25\textrm{\color{blue}Table 25}, Page 43). For each of five independent keys, we set the threshold on one half and measure the FPR on the other half.

FPR (unseen half)TPR (watermarked text)
LLaDA0.60% to 1.00%0.981
Dream0.77% to 0.97%0.990

W5 and Q5. "[...] Downstream task evaluations would help establish whether the security benefit justifies these costs."

We added GSM8K as a downstream task, and TANGO shows no significant loss of accuracy on either model (Sec. 4.2\textrm{\color{blue}Sec. 4.2}, Page 8, and Table 29\textrm{\color{blue}Table 29}, Page 45). The table shows the accuracy on 200 problems.

GSM8K accuracyNo watermarkTANGOUnigram
LLaDA0.585±0.070.585^{\pm 0.07}0.550±0.070.550^{\pm 0.07}0.580±0.070.580^{\pm 0.07}
Dream0.445±0.070.445^{\pm 0.07}0.610±0.070.610^{\pm 0.07}0.250±0.060.250^{\pm 0.06}

On LLaDA, TANGO performs comparably to the no-watermark baseline. On Dream, TANGO outperforms Unigram significantly.

Dear Reviewer 8wBt,

As the discussion deadline approaches, we wanted to follow up to ensure our responses have adequately addressed your concerns. If there are any additional questions or points for discussion, we would be happy to address them promptly.

We look forward to your feedback and sincerely appreciate the time and effort you put into reviewing our work.

Reviewer_zQQPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TANGO embeds the watermark into the modular relationship between two token colors. In mask diffusion generation, if adjacent positions have been determined, logit bias is applied to color categories that can satisfy this relationship. When detecting, only the token pairs that match the relationship in the final text need to be counted, without restoring the order of unmasking. The paper evaluates the detection, editing robustness, and forgery attacks on LLaDA and Dream.

Critical strengths
  1. The final text can be used to check the token pair relationship; Bidirectional bias application can also utilize the first determined one of the two endpoints. The design is relatively simple, the motivation is clear, and it is not directly transferring the autoregressive watermark to another sampler.
  2. By default, the clean text TPR of LLaDA and Dream are 0.97 and 1.00, respectively, with confusion levels close to the fixed list baseline. The paper also reported the failure and degradation rate of deletion attacks, as well as the success rate of token based attacks.
Critical weaknesses

Theorem 1 explicitly removes its dependence on neighboring colors by using a new and uniform bias category, which is an idealized cancellation result and not a security guarantee for actual key samplers. The actual method only has q=3 color categories, and they come from the same embedded projection. Attackers may not necessarily need to independently estimate the probability of V ^ 2 token pairs, and may also infer the shared color structure from the sparse token pair graph. The existing attacks have not verified this.

Questions for authors
  1. For the default configuration, can a separate and larger watermark free calibration set be used, and FPR and TPR with reserved data reported on multiple independent keys? This helps to determine whether the 1% FPR operating point in the article is stable.
  2. Have you tried to recover shared color categories from observed token pairs in the graph instead of estimating token pairs one by one? If not, what is the basis for discussing the difficulty of O (V ^ 2) estimation when all preferences come from the same set of q-color structures?
2 comments

We thank the reviewer for the thoughtful feedback. We are glad that you find the design relatively simple and its motivation clear. We respond to the weakness (W) and the questions (Q) below.


W. "Theorem 1 [...] is an idealized cancellation result and not a security guarantee for actual key samplers."

We agree, and the revision states this. Theorem 1 is now titled "Frequency cancellation under an idealized sampler", and the text says that it is not a security guarantee for a real key (Sec. 3.4\textrm{\color{blue}Sec. 3.4}, Page 6).

The security claim is strengthened by experiments with the real sampler. An attacker who estimates the key from token frequencies forges the red–green list baseline in every attempt and TANGO in none (Table 8\textrm{\color{blue}Table 8}, Page 32).


W and Q2. "Have you tried to recover shared color categories from observed token pairs in the graph instead of estimating token pairs one by one? If not, what is the basis for discussing the difficulty of O (V ^ 2) estimation when all preferences come from the same set of q-color structures?"

Following your suggestion, we ran this attack. The attack recovers the shared coloring from all the token pairs observed in watermarked text at once, without the key. The attacker then uses the recovered coloring to generate text that imitates the watermark. A forgery passes if the detector accepts it at the 1% FPR threshold.

This attack is successful when the attacker can collect a large amount of watermarked text under a single key, and key rotation limits its effectiveness. With as few as eight keys, at most 7% of its forgeries pass (Sec. 4.3\textrm{\color{blue}Sec. 4.3}, Page 8, and Table 22\textrm{\color{blue}Table 22}, Page 41).

The paper recommends that a provider rotate keys (Sec. 6\textrm{\color{blue}Sec. 6}, Page 10), and TANGO's detector can check a text against every key. The table shows the share of forgeries that pass when the attacker spreads the watermarked text over several keys. In this evaluation, we assume that the attacker knows which key produced each text, which is the best case for the attacker.

Number of keys1248
Forgeries that pass, with access to 38,400 watermarked tokens62%27%9%5%
Forgeries that pass, with access to 115,200 watermarked tokens92%75%42%7%

An attacker who cannot tell which key produced each text does worse. With four keys, at most 5% of its forgeries pass, compared with 9% and 42% in the table.

With 32 colors instead of 3, TANGO roughly halves this attack's forgery success rate under a single key at the same text quality and is more robust to deletion (Sec. 4.5\textrm{\color{blue}Sec. 4.5}, Page 9, and Table 30\textrm{\color{blue}Table 30}, Page 45).

We note that the O(V²) estimate indeed applies only to an attacker who estimates token pairs one by one. We clarified this in the revision and report the joint attack next to it (Sec. 4.3\textrm{\color{blue}Sec. 4.3}, Page 8).


Q1. "For the default configuration, can a separate and larger watermark free calibration set be used, and FPR and TPR with reserved data reported on multiple independent keys? This helps to determine whether the 1% FPR operating point in the article is stable."

Yes, and the 1% FPR operating point is stable (Sec. 4.2\textrm{\color{blue}Sec. 4.2}, Page 8, and Table 25\textrm{\color{blue}Table 25}, Page 43). We use 5000 human-written texts and five independent keys. For each key, we set the threshold on one half of the texts and measure the FPR on the other half.

FPR (unseen half)TPR (watermarked text)
LLaDA0.60% to 1.00%0.981
Dream0.77% to 0.97%0.990

Thank you very much for your reply. It has eased some of my doubts, so I will keep this positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We are encouraged that the reviewers find the design "relatively simple" with a clear motivation (zQQP), a method that "naturally accommodates out-of-order generation" and is "clearly described" (BjG4), and an implementation that is "lightweight" (8wBt). We also appreciate that they value an evaluation that covers two models with editing and forgery attacks (BjG4), which helps "clarify the scope and limits" of the method (8wBt).

We ran every experiment that the reviewers asked for, and we respond to each reviewer individually. The new experiments leave the main result unchanged. The uploaded revision marks all changes in blue. Its three main additions are:

  • Comparison with LR-DWM and DMark at similar text quality (Table 2\textrm{\color{blue}Table 2} and Sec. 4.4\textrm{\color{blue}Sec. 4.4}, Page 9). TANGO offers greater robustness to text modifications than LR-DWM, competitive detection performance against DMark, and stronger resistance to forgery attacks.
  • The stronger forgery attack that the reviewers suggested (Sec. 4.3\textrm{\color{blue}Sec. 4.3}, Page 8). This attack needs a large amount of watermarked text generated under the same key, and becomes ineffective when providers rotate keys.
  • Thresholds tested on unseen text (Sec. 4.2\textrm{\color{blue}Sec. 4.2}, Page 8). Each of five independent keys keeps the false positive rate at or below 1% on unseen human text.

Following the reviewers’ comments, Theorem 1 is now stated for an idealized sampler (Sec. 3.4\textrm{\color{blue}Sec. 3.4}, Page 6). Our experiments show that the result extends to the real model, where no forgery based on token frequencies passes TANGO’s detector (Table 8\textrm{\color{blue}Table 8}, Page 32).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.