acceptodds
Under review as a conference paper at ICLR 2027

RLSpoofer: A Sample-Efficient Black-Box Spoofing Stress-Test for LLM Watermarks

12 views

Abstract

Large language model (LLM) watermarking has emerged as a promising approach for detecting and attributing AI-generated text, yet its robustness to black-box spoofing remains insufficiently evaluated. Existing evaluation methods often demand extensive datasets and white-box access to algorithmic internals, limiting their applicability. In this paper, we study watermark spoofing from a distributional perspective, viewing it as selectively reconstructing watermark-induced probability shifts under semantic constraints. We theoretically show that low-dimensional watermark-induced shifts can be learned more sample-efficiently than the full distribution, while KL-bounded reconstruction inherently favors larger watermark deficits. Guided by this analysis, we propose RLSpoofer, a reinforcement learning-based black-box spoofing attack that requires only 100 paired human–watermarked paraphrases and zero access to watermarking internals or detectors. Despite this weak supervision, it empowers a 4B model to achieve a 62.0% spoof success rate on PF-Watermark while preserving semantic fidelity, compared with below 7% for baselines trained on up to 10,000 samples. Our findings expose weaknesses in the spoofing resistance of current LLM watermarking paradigms and provide a sample-efficient framework for evaluating watermark robustness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Spoofing Attack, RL, LLM Watermarks

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_QC2wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates practical black-box spoofing attacks against LLM watermarks and proposes RLSpoofer, a sample-efficient spoofing method. Based on a theoretical analysis suggesting that low-dimensional watermark-induced probability shifts can be learned more efficiently than the full distribution, the authors propose selectively reconstructing the remaining watermark residuals. To this end, RLSpoofer employs GRPO. Using only 100 training pairs, RLSpoofer achieves substantial improvements over existing baselines across multiple watermarking schemes.

Critical strengths
  • Investigating black-box watermark spoofing attacks is important for understanding the practical security risks and vulnerabilities of LLM watermarking systems.

  • The proposed method achieves substantial improvements over existing baselines, particularly against distortion-free watermarking schemes.

Critical weaknesses
  • The theoretical formulation measures watermark-induced probability shifts under the same context, whereas the implementation estimates them using a reference model conditioned on different paraphrases. Since semantic equivalence does not imply similar next-token distributions, the estimated residual may capture generic paraphrasing variations rather than watermark-specific shifts. Further justification would help clarify whether the proposed residual estimation captures watermark-specific information.

  • The evaluation primarily relies on a P-SP threshold of 0.7, which may be insufficient to capture fine-grained semantic inconsistencies in relatively long texts of 500 tokens. A high similarity score does not necessarily guarantee semantic preservation, particularly for longer texts where subtle differences in details may be obscured by overall semantic similarity. This is a known limitation of cosine similarity-based metrics, and the chosen threshold of 0.7 appears relatively lenient. Moreover, RLSpoofer achieves substantially lower LLM-judge scores than the baselines in several settings. These results raise concerns that the improved spoofing performance may come at the cost of semantic fidelity and output quality, weakening the claim of generating semantically coherent text.

  • Despite emphasizing effectiveness against distortion-free watermarks, the evaluation only includes PF-Watermark and PMark, omitting the widely studied EXP watermark. Evaluating against EXP would provide stronger evidence of generalizability across different distortion-free watermarking mechanisms.

  • On PF-Watermark with Qwen3-4B, SSR increases from 20.8% (50 samples) to 62.0% (100 samples), but drops sharply to 25.3% (200 samples) (Table 7, p. 8). This non-monotonic behavior raises concerns about the reliability of the headline result.

Questions for authors

Please see the four critical weaknesses above. Additionally, I have a question regarding cross-watermark transferability.

  • Transferability between similar logit-based watermarks is understandable, as they share similar green-list mechanisms and the same hash key in the experimental setup. However, the substantial transferability between PF and EWD is difficult to understand, particularly given its asymmetric results (PF → EWD: 50.5% vs. EWD → PF: 7.75%). Could the authors provide further analysis or intuition to explain this asymmetric transferability across different watermarking mechanisms?
7 comments

We sincerely thank you for your valuable comments. We hope our response below could address your concerns.

Q1. Does the surrogate residual capture watermark-specific shifts rather than generic paraphrasing variations?

A: We thank the reviewer for the valuable comment. We acknowledge that our practical surrogate does not exactly recover the theoretical watermark-induced residual, as the true distributions are inaccessible under our black-box setting. Since watermark signals can persist after weak paraphrasing [1], we adopt a distributional view, using a lightweight reference model to approximate these shifts.

Furthermore, to validate this approximation, we conducted additional token-level analysis on EWD using 100 paired samples and 9,600 sampled tokens. Specifically, we compared the initial surrogate residual estimated by Llama-3.2-3B with the actual watermark-induced positive probability shifts computed by Llama-3.1-8B.

MetricSimilarity95% CIUnwatermarked Generation
Pearson0.287[0.159, 0.335]0.071
Cosine0.311[0.200, 0.362]0.106

These results show statistically supported positive alignment with actual watermark-induced shifts, exceeding the unwatermarked generation. While contextual variations may still contribute, our surrogate captures meaningful watermark-related signals. We have included this analysis and clarified its limitations in the Appendix C.10 revised manuscript.

[1] Can Watermarks Survive Translation? On the Cross-lingual Consistency of Text Watermark for Large Language Models ACL2024

Q2. Does RLSpoofer maintain semantic fidelity and output quality under stricter evaluation criteria?

A: We thank the reviewer for the valuable question. Prior studies have shown that watermark-induced distributional perturbations can introduce trade-offs between detectability and text quality, including semantic consistency [2]. Similarly, reconstructing watermark-related signals may compromise semantic fidelity. We therefore acknowledge this trade-off and evaluate Qwen3-4B under stricter P-SP thresholds of 0.80 and 0.90. As shown below, RLSpoofer consistently outperforms all baselines. We have included these results in Appendix C.1 and explicitly discussed the semantic fidelity trade-off in the Limitations section of our revised manuscript.

MethodMetricEWDSWEETPFPMARK
DistilledP-SP> 0.8037.7528.755.7521
DITTOP-SP> 0.802417.75312.5
DPOP-SP> 0.800.250.254.250.25
RLSpooferP-SP> 0.8046.34550.533.5
DistilledP-SP> 0.9010.56.25517.5
DITTOP-SP> 0.901.751.252.58
DPOP-SP> 0.900.250.253.750
RLSpooferP-SP> 0.9016.31335.525

[2] Context-aware Watermark with Semantic Balanced Green-red Lists for Large Language Models. EMNLP2024

Q3. Missing evaluation on EXP-Watermark.

A: We thank the reviewer for the valuable comment. We additionally evaluated RLSpoofer on the distortion-free EXP watermark, with results shown below. RLSpoofer achieves 40.5% SSR with only 100 training pairs, significantly outperforming Distill. These results further support the effectiveness of RLSpoofer across different distortion-free watermarking mechanisms. We have included these additional results in the revised manuscript.

MethodSSRP-SP
Distill8.30.87
DITTO5.20.62
RLSpoofer40.50.79

Q4. How stable are the PF-Watermark results across different training-set sizes?

We thank the reviewer for the valuable comments. To assess the stability of PF-Watermark results, we evaluated two additional random seeds and training subsets for each data size using Qwen3-4B, yielding six additional runs. As shown below, the 100-pair setting consistently achieves 60.5–63.5% SSR, confirming the reproducibility of our results. We hypothesize that the non-monotonic trend arises from PF's distortion-free property in expectation, which may cause watermark-specific distributional shifts to become harder to capture as the dataset grows. We have included these results and discussed this potential explanation in the Appendix C.6 of our revised manuscript.

Method50100200
Untrained3.03.03.0
Seed 120.862.025.3
Seed 222.560.524.8
Seed 321.563.527.0
Mean ± Std21.6 ± 0.962.0 ± 1.525.7 ± 1.2

Q5. What explains the asymmetric cross-watermark transferability between PF and EWD?

A: We thank the reviewer for the valuable comments. We hypothesize that the asymmetric transfer arises from differences in watermarking mechanisms and detector statistics. In particular, PF relies on key-dependent sampling correlations, whereas EWD uses entropy-weighted token statistics. We additionally evaluated genuine watermarked texts and found cross-detection rates of only 0% (PF->EWD) and 2% (EWD->PF). This suggests that the observed transfer is not simply inherited from the original watermark signals, but may arise from detector-specific statistical patterns induced during attack optimization. We have clarified this potential explanation in the Appendix C8 of the revised manuscript.

Watermarked TextEWD DetectorPF Detector
EWD100%2%
PF0%100%

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

AuthorsOpenReview ↗

Gentle Follow-up on the Rebuttal

Dear Reviewer QC2w,

We sincerely appreciate your thoughtful review, valuable suggestions, and the time and effort you have devoted to evaluating our submission. We have provided additional experiments and clarifications in response to your valuable feedback.

We hope our responses and additional experimental results could address your concerns. We would be very grateful if you could share any further feedback or remaining concerns at your convenience. We would be happy to provide additional clarifications if needed.

Best regards,

Paper 11130 Authors

Reviewer_Ucx7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RLSpoofer, a black-box spoofing attack on LLM watermarking methods that requires only a small amount of training data. The approach integrates a token-level residual reward with semantic reinforcement learning and cross-entropy supervision to guide the attack model. Experiments involving five attacker models and six watermarking schemes show that RLSpoofer achieves effective spoofing with just 100 training pairs, with particularly strong results against distortion-free watermarking methods.

Critical strengths
  1. Realistic black-box attack setting. RLSpoofer operates without knowledge of the watermarking mechanism or access to the detector. This makes the attack relevant to practical scenarios where such information is unavailable.
  2. Effective spoofing with limited training data. The method outperforms the evaluated baselines using only 100 training pairs. The improvement is particularly notable for PF-Watermark.
  3. Comprehensive experimental evaluation. The authors test RLSpoofer across different attacker models and watermarking schemes. Additional ablation studies and sensitivity analyses help assess the effects of individual components and experimental settings.
Critical weaknesses
  1. The theoretical watermark-induced residual and the surrogate residual used in practice are not directly equivalent, as the latter compares reference-model probabilities conditioned on different inputs. Such differences may reflect ordinary contextual variations rather than watermark-specific shifts.
  2. Insufficient evaluation of semantic fidelity. The fixed P-SP threshold of 0.7 may obscure the trade-off between spoofing success and text quality. RLSpoofer achieves substantially lower P-SP and GPT-Judge scores than Distill on PF-Watermark.
  3. Lack of key-specific validation. The experiments do not establish whether the attack learns patterns associated with the target watermark key.
Questions for authors
  1. Residual reward contribution: Could the authors provide an ablation without the residual reward (w2=0w_2=0) while retaining the other components?
  2. Surrogate validity: Could the authors validate whether the surrogate residual aligns with actual watermark-induced probability shifts under matched contexts? Additionally, why does the Qwen3-8B self-surrogate perform substantially worse than the 0.6B surrogate on PF-Watermark (29.8% vs. 58.5% SSR)?
  3. Semantic fidelity: How does SSR change under stricter P-SP thresholds (e.g., 0.8, and 0.9)? Does RLSpoofer maintain its advantage over Distill when comparing outputs at similar semantic quality levels?
  4. Key specificity: Could the authors evaluate whether outputs trained to spoof a particular watermark key also trigger detectors configured with independent keys? This would help establish whether the learned signal is genuinely key-specific.
6 comments

We sincerely thank you for your valuable and positive comments. We hope our response below could address your concerns.

Q1. Does the surrogate residual capture genuine watermark-induced probability shifts rather than ordinary contextual variations, and can this be validated under matched contexts?

A: We thank the reviewer for the valuable question. We acknowledge that our practical surrogate does not exactly recover the theoretical watermark-induced residual, as the true distributions are inaccessible under our black-box setting. Since watermark signals can persist after weak paraphrasing [1], we adopt a distributional view, using a lightweight reference model conditioned on the original and watermarked texts to approximate these shifts.

To validate this approximation, we conducted additional token-level analysis on EWD using 100 paired samples. Specifically, we compared the initial surrogate residual estimated by Llama-3.2-3B with the actual watermark-induced positive probability shifts computed by Llama-3.1-8B.

MetricSimilarity95% CIUnwatermarked Generation
Pearson0.287[0.159, 0.335]0.071
Cosine0.311[0.200, 0.362]0.106

As shown above, the surrogate exhibits statistically supported positive alignment with actual watermark-induced shifts, exceeding the unwatermarked generation. While ordinary contextual variations may still contribute, these results support the surrogate's ability to capture meaningful watermark-related signals. We have included this analysis to Appendix C.10 and clarified its limitation in the revised manuscript.

[1] Can Watermarks Survive Translation? On the Cross-lingual Consistency of Text Watermark for Large Language Models ACL2024

Q2. Does RLSpoofer maintain its advantage under stricter P-SP thresholds and comparable semantic quality?

A: We thank the reviewer for the valuable comments. We clarify that the reported P-SP scores are averaged over all generated paraphrases, rather than only successful instances. To further assess semantic fidelity, we evaluate Qwen3-4B under stricter thresholds of 0.80 and 0.9. As shown below, RLSpoofer consistently outperforms all baselines. Nevertheless, we acknowledge the trade-off between spoofing effectiveness and semantic fidelity. We have discussed this limitation and included the additional results in Appendix C.1 of the revised manuscript.

MethodMetricEWDSWEETPFPMARK
DistilledP-SP> 0.8037.7528.755.7521
DITTOP-SP> 0.802417.75312.5
DPOP-SP> 0.800.250.254.250.25
RLSpooferP-SP> 0.8046.34550.533.5
DistilledP-SP> 0.9010.56.25517.5
DITTOP-SP> 0.901.751.252.58
DPOP-SP> 0.900.250.253.750
RLSpooferP-SP> 0.9016.31335.525

Q3. Missing validation of watermark key specificity.

A: We thank the reviewer for the valuable suggestion. We conducted cross-key detection experiments on both genuine watermarked texts and RLSpoofer outputs across three independent keys for EWD, SWEET, and PF and report the average detection rates in the following table. As shown, neither genuine watermarked texts nor RLSpoofer outputs trigger detectors configured with different keys. These results provide direct evidence that RLSpoofer learns key-specific watermark-related signals, rather than merely exploiting generic detector biases. We have include these addtional results to Appendix C.11 of our revised manuscript.

SourceDetector KeyEWDSWEETPF
WatermarkedSame100.0100.0100.0
WatermarkedDifferent0.00.00.0
RLSpooferSame87.064.382.8
RLSpooferDifferent0.00.00.0

Q4. Missing ablation without the residual reward (w2=0w_2=0).

A: We thank the reviewer for the valuable suggestion. We additionally conducted ablations on EWD, SWEET, PF and PMARK using Qwen3-4B by removing the token-level residual reward (w2=0w_2=0). As shown below, removing the residual reward substantilly reduces the SSR of the algorithm, confirming the critical contribution of residual reward to effective spoofing. We have included these results in the revised Appendix C.3.

MethodEWD-SSREWD-PSPSWEET-SSRSWEET-PSPPF-SSRPF-PSPPMARK-SSRPMARK-PSP
w2=0w_2=021.50.6920.30.7117.50.739.000.64
RLSpoofer56.50.7352.30.77620.7736.30.91

Q5. Why does the Qwen3-8B self-surrogate underperform the 0.6B surrogate on PF-Watermark?

A: We thank the reviewer for the valuable comments. We evaluated watermark retention after a single paraphrase on 400 PF-watermarked texts. Qwen3-8B removes 75.5% of watermarks while maintaining a P-SP of 0.91, whereas Qwen3-0.6B removes only 14.5% with a P-SP of 0.89. This suggests that the stronger paraphraser substantially weakens the watermark signal, making it less suitable for constructing the surrogate distribution. In contrast, Qwen3-0.6B better preserves watermark-related signals, explaining its superior surrogate performance (58.5% vs. 29.8% SSR). We have included these additional results and analysis in Appendix C.7 of the revised manuscript.

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

AuthorsOpenReview ↗

Gentle Follow-up on the Rebuttal

Dear Reviewer Ucx7,

We sincerely appreciate your positive and thoughtful review, valuable suggestions, and the time and effort you have devoted to evaluating our submission. We have provided additional experiments and clarifications in response to your valuable feedback.

We hope our responses and additional experimental results could address your concerns. We would be very grateful if you could share any further feedback or remaining concerns at your convenience. We would be happy to provide additional clarifications if needed.

Best regards,

Paper 11130 Authors

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

This paper studies black-box spoofing of LLM text watermarks, where an attacker paraphrases unwatermarked text so that it is detected as carrying a target watermark. It proposes RLSpoofer, an RL-based attack that combines a token-level residual reward, semantic reward, and cross-entropy anchor using 100 paired human-watermarked rewrites. Experiments across six watermarks and several attacker models report strong spoofing, especially for PF-Watermark.

Critical strengths

The empirical result on distortion-free watermarking is consequential. Table 3 reports that, for Qwen3-4B on PF-Watermark, RLSpoofer reaches 62.0% SSR and 82.8% SR, while the three generic baselines remain below 9% SR. Together with the broader results in Table 1, this identifies a practically important spoofing vulnerability that conventional distillation and preference-based attacks fail to expose.

The method is unusually well isolated through targeted ablations. Tables 4, 5, and 6 separately test the residual reward, dual-reference semantic objective, and CE anchor, and show large degradations when key components are removed, particularly on PF. This is much stronger evidence than a monolithic RL attack comparison and supports the claim that selective residual shaping, rather than merely fine-tuning on 100 paired targets, drives the reported gains.

Critical weaknesses

The paper presents RLSpoofer as a general stress test across watermarking paradigms, and I think this is overclaimed. The big result is on PF (62% vs under 9% for the baselines). The text then says PMark shows “a similar trend,” and the conclusion talks about “particularly large gains on distortion-free schemes.” On PMark, though, RLSpoofer gets 29.5–36.3%, while Distill and DPO already reach 20–25%, so the gain is only about 6–15 points. PMark is also the only semantic-level watermark in the paper, and it’s left out of every ablation. Semantic schemes like SemStamp and k-SemStamp, which the paper cites, aren’t evaluated at all. This matters because RLSpoofer’s main reward works at the token level, while semantic watermarks put the signal in meaning, not in word choice. The paper has no limitations section, and this mismatch is never discussed.

I’m also missing a basic baseline: how often does each detector flag the untrained attacker’s normal paraphrases, and the human test texts themselves? Without this, I can’t tell how much of the SSR comes from actually learning the watermark versus learning a writing style that happens to push some detectors over the threshold. For PMark especially, the 20–25% that the baselines reach might just be the background rate. Finally, the semantic threshold for counting a spoof as successful (P-SP > 0.7) is loose, and RLSpoofer’s outputs often sit only slightly above it (around 0.71–0.77 on most schemes). Interestingly, the training reward uses a threshold of 0.85, but evaluation uses 0.7. On PF, Distill keeps P-SP at 0.96 and gets a higher GPT-judge score (8.68 vs 6.80), so part of RLSpoofer’s advantage seems to come from accepting weaker paraphrases. Reporting SSR at stricter thresholds (e.g., 0.8 and 0.85) would make the comparison fairer.

Questions for authors

How stable is the PF result? PF seeds each token with a hash of the previous 8 tokens, so it’s surprising that this can be learned from 100 pairs. The training-size ablation is also not monotonic: 12.3% with 50 samples, 62.0% with 100, and 25.3% with 200. Could the authors report results over several random seeds and training subsets, along with the detection rate of the untrained attacker on PF, to show that the 62% isn’t a lucky run or a side effect of the detector’s behavior?

4 comments

We sincerely thank you for your valuable comments. We hope our response below could address your concerns.

Q1. The claim of RLSpoofer as a general stress test may be overclaimed.

A: We thank the reviewer for the valuable comments. We agree that the magnitude of improvement varies across watermarking schemes and that our wording may be misleading. Our key contribution is achieving superior black-box spoofing performance with only 100 pairs, compared with baselines trained on 7,000–10,000 examples. On Qwen3-4B, RLSpoofer improves SSR from 6.0% to 62.0% on PF and from 21.5% to 36.3% on PMark. Moreover, background detection rates are only 3%/3.5% on PF and 8.5%/8% on PMark for untrained paraphrases/human texts, respectively. We have revised our claims in Sections 1, 5.2, and 6 to better emphasize sample efficiency and acknowledge varying gains across schemes.

Q2. PMark ablations and evaluations on other semantic watermarks are missing and there is no limitation section.

A: We thank the reviewer for the valuable comments. We conducted additional ablations on PMark using Qwen3-4B to validate the effectiveness of our core residual reward. As shown below, removing the residual reward reduces SSR from 36.3% to 9.0%, while alternative residual formulations also underperform the full method.

Residual rewardSSRP-SP
Pwm−PhP_{wm}-P_h26.50.90
[Pwm−Ph]+[P_{wm}-P_h]_+29.50.91
Pwm−PθP_{wm}-P_\theta28.30.90
0 (No reward)9.000.64
RLSpoofer36.30.91

We further evaluated RLSpoofer on SemStamp using Qwen3-4B, achieving 35.8% SSR, compared with 18.5% for the Distill. While these results demonstrate the effectiveness of RLSpoofer on semantic watermarks, its SSR remains lower than on PF-Watermark, potentially due to a mismatch between our token-level residual reward and the meaning-level signals embedded by semantic watermarks. We have added a dedicated Limitations section (Section 6) discussing the results and included the additional experimental results to Appendix C.1 and C.3.

MethodSSRP-SP
Plain rewrite2.250.93
Distill18.50.80
DITTO12.00.67
RLSpoofer35.80.84

Q3. How often does each detector flag the untrained attacker’s normal paraphrases, and the human test texts themselves?

A: We thank the reviewer for the valuable comments. We evaluated the background detection rates of original human texts and untrained paraphrases from five attacker models across all six watermarking schemes. As shown below, the detection rate baselines remain low (≤3.5% on PF and ≤8.5% on PMark), supporting that RLSpoofer's gains are not driven by detector false positives. We have added these results to Appendix C.1 of the revised paper.

RephreasedEWDSWEETKGWUnigramPFPMark
Human Text0%0%1.25%0.25%3.50%8%
Qwen3-4B0%0%0.75%0%3.00%8.50%
Qwen3-1.7B0%0%0.50%0.25%2.50%8.50%
Qwen3-0.6B0%0%0%0%3.00%7.75%
Llama3.2-3B-Instruct0%0%0%0%3.25%8.25%
Qwen2.5-3B-Instruct0.25%0%1.75%0.75%3.00%8.00%

Q4. The semantic threshold for counting a spoof as successful (P-SP > 0.7) is loose.

A: We thank the reviewer for the valuable comments. Following prior work [1], we adopt P-SP > 0.7 as the success criterion. We clarify that the reported P-SP scores are averaged over all generated paraphrases, rather than only successful spoofing instances. To further assess semantic fidelity, we evaluate Qwen3-4B under stricter thresholds of 0.80 and 0.85. As shown below, RLSpoofer consistently outperforms all baselines under stricter semantic constraints. Nevertheless, we acknowledge that improved spoofing effectiveness can come at the cost of reduced semantic fidelity. We have explicitly discussed this trade-off in the Limitations section and included the additional threshold sensitivity results in Appendix C.1 of our revised manuscript.

[1] Watermark stealing in large language models.

MethodMetricEWDSWEETPFPMARK
DistilledP-SP> 0.8037.7528.755.7521
DITTOP-SP> 0.802417.75312.5
DPOP-SP> 0.800.250.254.250.25
RLSpooferP-SP> 0.8046.34550.533.5
DistilledP-SP> 0.8525195.2519.5
DITTOP-SP> 0.859839
DPOP-SP> 0.850.250.2540.25
RLSpooferP-SP> 0.8537.832.344.830.8

Q5. How stable is the PF result?

A: We thank the reviewer for the valuable comments. We would like to clarify that the cited 12.3% SSR refers to Qwen3-0.6B, whereas the corresponding Qwen3-4B result is 20.8%. We evaluated two additional random seeds and training subsets for each data size on Qwen3-4B, yielding six additional runs. The PF-Watermark SSR (%) results are shown below. The 100-pair setting consistently achieves 60.5–63.5% SSR, confirming the reproducibility of our results. We hypothesize that the non-monotonic trend arises from PF's distortion-free property in expectation, which may cause watermark-specific distributional shifts to become harder to capture as the dataset grows. We have included these results and discussed this potential explanation in the Appendix C.6 of our revised manuscript.

Method50100200
Untrained3.03.03.0
Seed 120.862.025.3
Seed 222.560.524.8
Seed 321.563.527.0
Mean ± Std21.6 ± 0.962.0 ± 1.525.7 ± 1.2

We sincerely thank you again for your valuable time and constructive feedback, which have greatly improved the quality and clarity of our manuscript.

AuthorsOpenReview ↗

Gentle Follow-up on the Rebuttal

Dear Reviewer FxnC,

We sincerely appreciate your thoughtful review, valuable suggestions, and the time and effort you have devoted to evaluating our submission. We have provided additional experiments and clarifications in response to your valuable feedback.

We hope our responses and additional experimental results could address your concerns. We would be very grateful if you could share any further feedback or remaining concerns at your convenience. We would be happy to provide additional clarifications if needed.

Best regards,

Paper 11130 Authors

AuthorsOpenReview ↗

Gentle Follow-up Before the Discussion Closes

Dear Reviewer FxnC,

As the discussion period is approaching its end, we would like to kindly follow up on our rebuttal and additional experimental results. We would greatly appreciate it if you could briefly let us know whether our responses have helped address your main concerns, or if any issues remain unresolved.

Thank you again for your time and consideration.

Best regards,

Paper 11130 Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

We sincerely thank all reviewers for their time, thoughtful feedback, and constructive evaluations. Reviewer FxnC described our results against distortion-free watermarking as consequential and our targeted ablations as unusually well isolated. Reviewer Ucx7 highlighted our realistic black-box attack setting, effective spoofing with limited training data, and comprehensive experimental evaluation. Reviewer QC2w recognized the practical importance of black-box watermark spoofing and the substantial improvements over existing baselines, particularly against distortion-free watermarking schemes.

Based on the reviewers' valuable suggestions, we have further strengthened our manuscript through additional experiments and analyses, with revisions marked in blue:

  1. Broader Watermark Evaluation: Extended our evaluation to EXP and SemStamp, covering additional distortion-free and semantic watermarking mechanisms, and conducted further residual reward ablations on PMark (Appendix C.1 and C.3).
  2. Semantic Fidelity: Evaluated spoofing success under stricter P-SP thresholds (0.80, 0.85, and 0.90), demonstrating substantial advantages under stronger semantic constraints while acknowledging the fidelity trade-off (Appendix C.1 and Section 6).
  3. Mechanistic and Key-Specific Validation: Validated surrogate residual alignment against actual watermark-induced probability shifts under matched contexts, and introduced residual reward removal ablations and cross-key detection experiments to examine the effectiveness and key specificity of RLSpoofer (Appendix C.10 and C.11).
  4. Statistical Reliability: Conducted multi-seed and training-subset experiments to assess PF-Watermark stability, and established background detection baselines across six watermarking schemes (Table 4 and Appendices C.1 and C.6).
  5. Further Analysis and Limitations: Investigated surrogate model sensitivity, watermark retention, and asymmetric cross-watermark transferability, refined our generalizability claims, and added a dedicated Limitations section (Sections 1, 5.2, and 6; Appendices C.7 and C.8).

We sincerely thank all reviewers again for their constructive feedback, which has strengthened the empirical validation, mechanistic understanding, and overall clarity of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.