acceptodds
Under review as a conference paper at ICLR 2027

Localize to Detect: Low-Budget Syntactic Span Watermarking for LLMs

Abstract

LLM watermarking provides a practical mechanism for identifying machine-generated text. Conventional methods perturb token probabilities at every decoding step, allowing these interventions to accumulate and degrade generation quality. Sparse watermarking addresses this issue by restricting watermarking to selected tokens. However, existing sparse statistical methods remain token-level, with no shared structural unit for embedding and detection. We introduce SpanWM, a span-based watermarking framework that uses parser-recoverable phrases as the unit of both embedding and detection. A secret key selects syntactic phrases for watermarking, and the detector re-identifies the corresponding spans from the received text for localized detection. This span-level design concentrates watermark evidence within a low token budget while leaving most of the output untouched. Experiments show that existing sparse methods must test substantially more tokens to match SpanWM's detection performance. SpanWM also better preserves the draft's meaning and remains robust to paraphrasing and token-level edits. At matched AUROC on C4, SpanWM requires only 16–45% of the watermark budget needed by the sparse baselines. Our code is available at https://anonymous.4open.science/r/spanwm_11-BCD6.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Natural Language Processing, LLM watermarking, Text provenance, Trustworthy AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iZUcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SpanWM, a watermark scheme placed in a few syntactic phrases of an otherwise unwatermarked draft.

It uses a secret key to select up to 4 noun, verb, or prepositional phrases, and regenerates each selected phrase with a green-list token bias. It samples several candidate replacements and chooses one that fits. The detector parses the received text, selects phrases using the same key, and tests the green-token count within them. A pretty simple idea with solid numbers (if reproducible).

Critical strengths
  1. Rewriting selected phrases after draft generation is a clean way to limit changes to the original text, unlike the other semantic watermarking schemes that pose strong constraints (like SemStamp or the others)
Critical weaknesses
  1. The paper says in its conclusion that SpanWM remains robust to paraphrasing, but presents no real-world practical paraphrase experiment. Its evaluated attacks include token insertions, deletions, and swaps. However, an attacker can simply ask another model to rewrite the text while preserving the meaning. This is a central missing experiment for the paper’s motivation. The reported 0.1% false-positive operating points are also unsupported by only 200 negative examples -- how can 0.1% FPR be possible with 200 examples? Maybe I missed something.

  2. The linked anonymous repository is empty, despite the abstract and reproducibility statement saying code is available. This is a strong red flag for me.

Questions for authors
  1. Can you report detection under actual paraphrasing, rather than only token-level edits, and provide the detailed paraphrasing model / prompts, and results?

  2. Can you provide the detailed implementation and scripts?

3 comments
AuthorsOpenReview ↗

Response to Reviewer iZUc

We thank the reviewer for recognizing SpanWM's clean localized rewriting design and strong empirical results. We have conducted additional experiments to directly address the concerns about practical robustness and statistical reliability.

W1+Q1. Real-world Paraphrasing Robustness & Low-FPR Evaluation

We additionally evaluated LLM-based paraphrasing attacks using GPT-5-mini and GPT-OSS-20B, beyond the token-level perturbations in the original submission. SpanWM achieves the highest AUROC under both attacks, outperforming all four baselines:

ParaphraserSpanWMSpARKSWEETKGWIE
gpt-5-mini0.7500.7370.5720.5840.672
gpt-oss-20b0.7990.7780.6240.6380.698

how can 0.1% FPR be possible with 200 examples? Maybe I missed something.

Regarding the reviewer's question, we clarify an important distinction: empirical FPR resolution does not limit the ability to estimate intermediate ROC operating points. With 200 negatives, [email protected]% FPR can be estimated using standard linear interpolation of the ROC curve, as done in our original evaluation.

More importantly, we have now expanded the negative set to 10k samples, enabling direct empirical evaluation at 0.1% FPR. The results are shown below.

DatasetMethodAUROC (95% CI)TPR@10%TPR@5%TPR@1%[email protected]%z-scoreCos. sim.LLM score
C4KGW0.814 [0.783, 0.843]0.480.350.090.011.700.7133.48
SWEET0.956 [0.938, 0.973]0.890.860.740.593.850.5563.32
SpARK-R0.946 [0.930, 0.960]0.860.760.510.153.490.5673.43
IE0.879 [0.850, 0.907]0.680.580.420.142.160.6933.20
SpanWM (Ours)0.986 [0.979, 0.992]0.960.930.820.664.050.9517.25
CNN/DMKGW0.827 [0.798, 0.854]0.510.350.100.031.630.7903.72
SWEET0.917 [0.891, 0.940]0.800.730.600.423.100.6353.86
SpARK-R0.951 [0.933, 0.966]0.880.780.540.193.230.6303.77
IE0.850 [0.818, 0.880]0.640.530.310.151.790.7953.78
SpanWM (Ours)0.983 [0.971, 0.992]0.970.940.850.663.970.9547.09
WMT16KGW0.784 [0.752, 0.814]0.450.310.090.011.590.6582.80
SWEET0.971 [0.958, 0.982]0.920.860.790.624.280.5102.94
SpARK-R0.955 [0.943, 0.965]0.860.760.440.113.490.5173.02
IE0.879 [0.848, 0.905]0.700.580.410.172.390.6692.86
SpanWM (Ours)0.985 [0.975, 0.993]0.950.940.900.774.470.9356.91

The expanded evaluation confirms our central finding without relying on interpolation: SpanWM achieves the highest [email protected]% FPR and AUROC across all three datasets. We additionally report 95% bootstrap confidence intervals for AUROC, further strengthening the statistical evidence.

W2+Q2. Code Availability & Reproducibility

We apologize for the initial delay in the code release. However, the implementation has been publicly available in our anonymous repository since October 13, more than three weeks before the review was last updated on November 6. Thus, the concern regarding the empty repository no longer reflects its current status. We encourage the reviewer to check the implementation and will further improve the documentation to facilitate reproducibility.


We sincerely thank the reviewer for the constructive feedback, which motivated us to strengthen the empirical evaluation of SpanWM. In particular, we have added LLM-based paraphrasing experiments, expanded the negative sample set to 10k for more reliable low-FPR evaluation, and clarified the availability of our implementation.

We hope that these additional results and clarifications satisfactorily address the reviewer's concerns and would greatly appreciate reconsideration of the initial assessment in light of the new evidence.

If you have any further questions or concerns, we would be happy to address them.

Thank you very much,

Authors

"We apologize for the initial delay in the code release. However, the implementation has been publicly available in our anonymous repository since October 13, more than three weeks before the review was last updated on November 6. Thus, the concern regarding the empty repository no longer reflects its current status. We encourage the reviewer to check the implementation and will further improve the documentation to facilitate reproducibility."

This framing is just misleading, imo. My review was submitted on October 8, well within the official review period. The November 6 timestamp simply reflects when conference administrators released reviews to authors, not when the review was written.

Authors are responsible for ensuring all materials are ready at submission time, not weeks after the paper submission deadline. Reviewers cannot be expected to conduct a fresh audit of a completely newly uploaded repository during the brief rebuttal window.

Thank you for the clarification. We understand that the original review was submitted on October 8. Our intention was simply to clarify that the code had been made available by the November 6 update, not to suggest that the reviewer was expected to revisit the repository. We appreciate your clarification and apologize for any confusion caused by our wording.

Beyond this clarification, we sincerely hope that the additional experimental evidence provided during the rebuttal, including the paraphrasing robustness results and expanded low-FPR evaluation, as well as the methodological contributions of SpanWM's shared span-level embedding and detection framework, will also be taken into consideration in your overall assessment of the paper.

We appreciate your time and consideration.

Reviewer_2yb1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies low-budget sparse watermarking for LLM-generated text and proposes SpanWM, which uses syntactic constituents as the shared unit for both watermark embedding and detection. The method first generates an unwatermarked draft, then uses a secret key to select a few noun, verb, and prepositional phrases from the parsed draft. Each selected phrase is regenerated at its original length under a green-list bias, and the replacement is chosen using a right-context likelihood score and a site reconstruction check, while the rest of the text is left unchanged. At detection, the detector re-parses the received text, locates the selected phrases with the same key, and applies a binomial test to the tokens within them, without requiring the original draft, the prompt, or the generating model. Experiments on C4, CNN/DailyMail, and WMT16, compared with KGW, SWEET, SpARK-R, and IE, report stronger detection at a comparable number of tested tokens, especially at low false-positive rates, higher semantic similarity to the original draft, and better robustness to token insertion, deletion, and swap attacks. On C4, SpanWM requires only 16-45% of the watermark budget needed by the sparse baselines to reach the same AUROC.

Critical strengths

The paper uses syntactic constituents as a shared unit for embedding and detection, achieving strong detection with a very small detection budget. The detector can relocate the watermarked sites using only the parser and the key, without requiring the original draft, the prompt, or the generating model. On Llama-3.2-3B, SpanWM tests only 17.0 tokens on average, which is comparable to or fewer than the sparse baselines, while achieving AUROC between 0.983 and 0.986 across the three datasets and the highest AUROC in all reported generator-dataset settings (Table 1, Table 5 in the appendix).

The experiments are broad and the paper is clearly written. The evaluation covers three datasets and four generators, and includes a detection budget sweep, token insertion, deletion, and swap attacks with edit ratios up to 50%, and an ablation of the syntactic role inventory (Tables 3-6). Semantic fidelity is assessed with both embedding-based similarity and an LLM-based score, and Figures 1 and 2 convey the motivation and pipeline of the method in an intuitive way.

Critical weaknesses

The semantic fidelity comparison inherently favors SpanWM, and generation quality is not directly evaluated. SpanWM is compared with its own unwatermarked draft, of which it edits at most 4 phrases, whereas the baselines are compared with an independently generated unwatermarked continuation (Appendix B.1). The gap in cosine similarity (0.951 versus 0.556-0.713 on C4) and LLM score (7.25 versus about 3.3) in Table 1 therefore mainly reflects this difference in reference rather than better text quality. The LLM judge prompt explicitly instructs the evaluator to ignore grammar and fluency (Appendix F), only the first 100 pairs are scored, and the scores are not validated against human annotations. No perplexity or fluency metric is reported, and the SpanWM example in Table 2, "right around the clock clock in that old.", contains a repeated word and an incomplete sentence, suggesting that local regeneration may harm readability. The generation overhead of producing a draft and then up to 4x4 candidate regenerations per site is also not reported.

Several central claims are not adequately supported by the reported results. The main detection result emphasizes gains of 17, 40, and 8 points at 0.1% FPR (Section 3.1), yet Appendix B.1 acknowledges that with 200 negative samples the empirical FPR resolution is only 0.5%, so results at 0.1% FPR are not validated. The budget comparison in Section 4.2 states that the baselines need 2.2x to 6.2x as many tokens to match SpanWM, but none of the baseline curves in Figure 4 reaches SpanWM's AUROC of 0.983, and it is not explained how these matching budgets were obtained. Robustness is evaluated only under random token insertion, deletion, and swap, and only AUROC is reported, while the paper also claims an advantage at stringent FPRs under editing (Section 1) and robustness to paraphrasing (Section 6) without corresponding experiments; no post-hoc watermarking baseline is compared either. Finally, results for the same default configuration differ across tables, for example AUROC 0.983 versus 0.982 and TPR at 1% FPR 0.82 versus 0.855 in Tables 1 and 3, and no confidence intervals or results over multiple runs are reported.

Questions for authors
  1. Could you report perplexity or fluency metrics and a small human evaluation, and compare all methods against the same unwatermarked reference?
  2. How were the matching budgets in Section 4.2 obtained, given that no baseline curve in Figure 4 reaches an AUROC of 0.983?
  3. How does SpanWM perform under paraphrasing, translation, or targeted rewriting of the selected phrases, and what supports the paraphrasing claim in the conclusion?
  4. Could you calibrate the false-positive rate on a substantially larger set of unwatermarked texts and report confidence intervals for the main results?
  5. Why do results for the same default configuration differ across Tables 1, 3, and 6?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 2yb1 (1/2)

We appreciate the reviewer for the constructive feedback and for recognizing SpanWM's strong detectability under small token budgets. We address the concerns below with additional experiments and clarifications.

W1+Q1. Semantic Fidelity & Generation Quality

For decoding-time baselines, watermarked and unwatermarked outputs are generated using identical prompts and random seeds, with only the watermarking intervention changed. Unlike these methods, SpanWM modifies only a few spans of an existing unwatermarked draft. Thus, comparing against the original draft directly captures SpanWM's inherent design advantage of preserving existing content while embedding a detectable watermark, rather than an arbitrary choice of reference.

To further assess generation quality, we additionally report BERTScore (RoBERTa-large), a widely adopted reference-based metric for evaluating NLG quality. The results are shown below.

DatasetKGWSWEETSpARK-RIESpanWM
C40.8520.8500.8490.8510.978
CNN/DM0.8520.8520.8500.8520.976
WMT160.8490.8470.8480.8500.974

SpanWM consistently achieves the highest scores across all datasets, substantially outperforming the strongest baselines. Furthermore, human evaluation of fluency and overall text quality is underway, and we will report its results during the rebuttal period. The repeated "clock" in Table 2 was a transcription typo. We will correct it and provide a clearer example.

W2. Generation Overhead

We additionally benchmarked the end-to-end computational cost across methods.

Cost / RequirementSpanWM (Ours)SWEETIESpARKKGW
Additional trainingNoneNonePer-model tagger trainingNoneNone
Detection requirementsSyntactic parser (no generating LLM)Full generation model (8B)SimCSE + taggerTokenizerTokenizer
Generation time (s/sample)7.25 (1.84×)4.09 (1.04×)4.94 (1.25×)3.91 (0.99×)4.06 (1.03×)
Detection time (s/sample)0.0520.0520.1730.0010.030

Notably, SpanWM takes only 7.25 s/sample (1.84× unwatermarked generation), including the entire initial draft generation, syntactic span selection, localized regeneration, and candidate verification. Thus, even with its two-stage watermarking procedure, the total generation time remains below 2×. Furthermore, detection takes just 0.052 s/sample, requires no generating LLM or watermark-specific training, and matches SWEET's detection latency. This modest additional generation cost enables SpanWM's substantially stronger detectability while retaining efficient, lightweight detection.

W3+Q4. Rigorous Low-FPR Evaluation

We clarify that the original [email protected]% FPR values were ROC-based estimates obtained through standard linear interpolation, rather than claims of directly observed empirical FPR. Although 200 negative samples limit the empirical FPR resolution to 0.5%, TPR at 0.1% FPR can still be estimated through standard ROC interpolation. Nevertheless, to provide more rigorous empirical evidence, we expanded the negative set to 10k samples and re-evaluated all methods.

DatasetMethodAUROC (95% CI)TPR@10%TPR@5%TPR@1%[email protected]%z-scoreCos. sim.LLM score
C4KGW0.814 [0.783, 0.843]0.480.350.090.011.700.7133.48
SWEET0.956 [0.938, 0.973]0.890.860.740.593.850.5563.32
SpARK-R0.946 [0.930, 0.960]0.860.760.510.153.490.5673.43
IE0.879 [0.850, 0.907]0.680.580.420.142.160.6933.20
SpanWM (Ours)0.986 [0.979, 0.992]0.960.930.820.664.050.9517.25
CNN/DMKGW0.827 [0.798, 0.854]0.510.350.100.031.630.7903.72
SWEET0.917 [0.891, 0.940]0.800.730.600.423.100.6353.86
SpARK-R0.951 [0.933, 0.966]0.880.780.540.193.230.6303.77
IE0.850 [0.818, 0.880]0.640.530.310.151.790.7953.78
SpanWM (Ours)0.983 [0.971, 0.992]0.970.940.850.663.970.9547.09
WMT16KGW0.784 [0.752, 0.814]0.450.310.090.011.590.6582.80
SWEET0.971 [0.958, 0.982]0.920.860.790.624.280.5102.94
SpARK-R0.955 [0.943, 0.965]0.860.760.440.113.490.5173.02
IE0.879 [0.848, 0.905]0.700.580.410.172.390.6692.86
SpanWM (Ours)0.985 [0.975, 0.993]0.950.940.900.774.470.9356.91

At 0.1% FPR, SpanWM outperforms the strongest baseline on every dataset. We additionally report 95% confidence intervals for AUROC. These results independently reinforce SpanWM's superior low-FPR detection performance.

AuthorsOpenReview ↗

Response to Reviewer 2yb1 (2/2)

W4+Q2. Budget Comparison

Figure 4 highlights SpanWM's superior detectability at small token budgets. For readability, higher-budget baseline settings are omitted from the plotted curves. The reported 2.2-6.2× ratios are based on directly measured results, not interpolation or extrapolation. Specifically, we evaluated 38 budget settings on Llama-3.2-3B/C4 using the same 200 prompts and random seeds. For each baseline, we report the smallest measured token budget achieving AUROC ≥ 0.983 (SpanWM's AUROC at N = 17.0). The corresponding measured operating points are shown below.

MethodSettingNAUROCN / 17.0
SpARK-Rmodular 438.10.9922.2×
SWEET (γ=0.25)τ=3.544.70.9922.6×
SWEET (γ=0.5)τ=3.057.90.9913.4×
IEτ=2.2105.10.9956.2×
KGW-spanspan 1601600.9949.4×

W5+Q3. Robustness to Paraphrasing

We additionally evaluated robustness against LLM-based paraphrasing attacks using GPT-5-mini and GPT-OSS-20B to address the reviewer's concern. As shown below, SpanWM achieves the highest AUROC under both attacks, consistently outperforming all four baselines. These results demonstrate that SpanWM's robustness extends beyond the token-level perturbations evaluated in the original submission.

ParaphraserSpanWMSpARKSWEETKGWIE
gpt-5-mini0.7500.7370.5720.5840.672
gpt-oss-20b0.7990.7780.6240.6380.698

Q5. Numerical Consistency

Upon investigating the reported discrepancies, we found that they arise primarily from different evaluation samples, rather than inconsistent watermark configurations. Specifically, Table 3 (R=3) is an independent generation run with different random seeds, whereas Table 6 evaluates only the first 100 prompts selected for the attack experiments, rather than all 200 used in Table 1.

We also identified a table assembly error in Table 3: AUROC was reported against human-written negatives, whereas TPR used unwatermarked-model negatives. We will correct this inconsistency and standardize the evaluation protocol and reporting across tables.

Importantly, the two independent runs demonstrate consistent performance across all four backbone models when evaluated against the same 10,000 unwatermarked negatives:

ModelAUROC (Run 1 / Run 2)
Llama-3.2-3B0.986 / 0.984
Llama-3.1-8B0.978 / 0.978
Qwen3-4B0.965 / 0.957
Qwen3-8B0.963 / 0.967

AUROC varies by at most 0.008 across runs, demonstrating stable detection performance. We will include these results and 95% bootstrap confidence intervals in the revision.

Reviewer_oRCgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SpanWM that embeds statistical watermarks into spans rather than individual tokens.

Critical strengths

The detection budgets, editing robustness and site alignment are well-investigated.

Critical weaknesses

I have some concerns about the novelty of the paper. The span-based watermarking itself is not as novel as the authors claimed in the paper.

The ideas substantially overlap with existing research: [1] already uses keyed syntactic structures for watermarking and parser-based detection. [2] introduces linguistically anchored sparse watermarking. [3] investigate post-hoc watermarking through LLM rewriting and candidate selection. The proposed method appears to combine these techniques with limited novelty.

[1] Zhang et al. PersonaMark: Personalized LLM Watermarking for Model Protection and User Attribution. https://arxiv.org/abs/2409.09739

[2] Hoang et al. SpARK: An Embarrassingly Simple Sparse Watermarking in LLMs with Enhanced Text Quality. Findings of EACL 2026.

[3] Fernandez et al. How Good is Post-Hoc Watermarking With Language Model Rephrasing? ICML 2026.

Questions for authors

How does SpanWM fundamentally differ from the existing syntax-based watermarking methods?

3 comments
AuthorsOpenReview ↗

Response to Reviewer oRCg

We thank the reviewer for raising the novelty concern. We acknowledge that syntactic watermarking, sparse token selection, and post-hoc rewriting have been studied individually. However, SpanWM introduces a distinct embedding–detection mechanism, rather than simply combining these techniques. The key methodological differences are summarized below.

DimensionPersonaMark [1]SpARK [2]Fernandez et al. [3]SpanWM (Ours)
Core design principleStructure-dependent sentence selectionSparse token-level watermarkingWatermarked LLM rephrasingSyntax-guided localization of statistical watermark evidence
Role of syntaxSentence structure is the watermark signalPOS tags select individual token positionsNo syntax-guided localizationSyntactic constituents serve as shared embedding–detection anchors
Watermark carrierHashed dependency structuresSelected green-list tokensStatistical signals in rephrased textGreen-list evidence concentrated within key-selected syntactic spans
Embedding strategySelect entire candidate sentencesBias selected tokens during decodingRewrite text in sentence/paragraph chunksLocally regenerate a few selected phrases while preserving the surrounding draft
Detection localizationHash sentence structuresRecover individual token positionsAggregate evidence from rephrased chunksIndependently reconstruct key-selected constituent ranges from received text
Embedding–detection consistencyShared structural hashingShared token-selection rulesShared watermark statisticsReconstruction-aware regeneration with re-parsing and site-alignment checks
Detection requirementsText, parser, and hash configurationText and watermark configurationText and watermark configuration; optional LM-based filteringReceived text, parser, tokenizer, and keys
Optimization objectivePersonalized attributionSparse token selectionQuality–detectability trade-off in post-hoc rephrasingStrong detectability under a small, explicitly measured detection-token budget

Importantly, SpARK, one of the reviewer-cited methods, is already included as a direct baseline (Table 1 and Appendix Table 5). Across all three datasets (Table 1), SpanWM improves average AUROC from 0.952 (SpARK) to 0.984 (+3.2%p) under virtually identical detection budgets. This substantial improvement demonstrates the practical significance of our constituent-level localization design beyond existing token-level sparse watermarking.


We sincerely thank the reviewer for the valuable feedback, which helped us better articulate the methodological distinctions between SpanWM and prior watermarking approaches. We hope that our detailed comparison and clarification of the shared embedding-detection mechanism have addressed the concerns regarding novelty and highlighted the contributions of our work. In light of these clarifications, we would be grateful if the reviewer could reconsider their initial assessment. If you have any further questions or concerns, we would be happy to address them.

Thank you very much,

Authors

Thanks for the clarifications. I still feel the novelty is not justified. I would maintain my score.

Thank you for your response. We respect your assessment.

In order to help us improve the paper, could you point out which specific component of SpanWM you consider to be already covered by [1-3]?

To our knowledge, none of [1-3] uses key-selected sub-sentential constituents (e.g., NP/VP/PP) as watermark sites, pools green token evidence over the constituent ranges that the detector reconstructs from the received text, or regenerates only those constituents under a green-list bias.

We will add an explicit discussion of [1] and [3] to the related work section in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers and the AC for their time and constructive feedback. We appreciate the reviewers' recognition of SpanWM’s strong detection performance, comprehensive experiments, and localized rewriting design.

Based on the reviewers' feedback, we have substantially strengthened the paper. Each concern is addressed in detail in our individual responses, and all revisions are highlighted in blue in the revised paper. The main improvements are summarized below:

  1. Novelty and positioning (oRCg): We clarified SpanWM’s distinct embedding–detection mechanism and added an explicit discussion of PersonaMark and Fernandez et al. to the Related Work section (§5).
  2. Statistical reliability (2yb1, iZUc): We expanded the evaluation to 10,000 unwatermarked negatives, enabling direct low-FPR evaluation without interpolation, and added 95% confidence intervals for AUROC. SpanWM achieves the highest AUROC and TPR at 0.1% FPR on all datasets (§3.1, Table 1).
  3. Robustness (2yb1, iZUc): We added LLM-based paraphrasing experiments using GPT-5-mini and GPT-OSS-20B, complementing the original token-level attacks. SpanWM achieves the highest AUROC under both paraphrasers (§3.3, Table 2).
  4. Experimental clarity (2yb1): We clarified the budget-matching methodology (directly measured over 38 settings, no interpolation), the evaluation protocols, and the numerical differences across experiments (§4.2, App. B.1, App. E.1).
  5. Semantic fidelity and efficiency (2yb1): We clarified the reference-based evaluation protocol, added BERTScore results, and reported generation and detection costs (App. B.1-B.3).

We believe these revisions meaningfully strengthen the paper's methodological clarity and empirical evidence, and we hope they address the reviewers' principal concerns.

We sincerely appreciate your time and consideration.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.