acceptodds
Under review as a conference paper at ICLR 2027

WaveHumanize: Localized Text Humanization with Wavelet Energy

Abstract

Text humanization aims to make LLM-generated writing more human-like while preserving its meaning. Achieving this with minimal changes to the original text requires identifying which passages to revise and choosing suitable replacements. We propose **WaveHumanize**, a training-free framework that uses a shared wavelet-energy signal to guide both decisions. A pretrained language model scores the tokens in a document based on its next-token predictions. Wavelet analysis converts these scores into an energy profile aligned with text positions. We use this profile to select short spans for rewriting, and a second language model generates candidate replacements. Each candidate is inserted into the document and rescored, and the replacement with the highest average energy over its span is retained. The procedure requires no feedback from the target detector. In the primary evaluation across fifteen domains, WaveHumanize achieves greater detection reduction with less textual change than the compared limited-edit baselines. At thresholds calibrated to a 1% false-positive rate, an average of 80.8% of initially detected texts evade detection after rewriting. Further evaluation shows transfer to several additional detector families. Code is available at https://anonymous.4open.science/r/WaveHumanize-ICLR-C6FE.

Primary area:
generative models
Keywords:
AI-Generated Text, Text Humanization, Wavelet Analysis, Localized Rewriting, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_uPvyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors use a shared local energy profile to make targeted edits to a piece of text for the purposes of adversarially "humanizing" the text to evade AI detection. They evaluate their humanizer on detectability reduction as well as token sequence dissimilarity and embedding similarity. They achieve stronger AI detection reduction with a claimed "lower textual edit cost".

Critical strengths

The strengths of this paper lie in the novelty of the approach, as I have not seen wavelet energy approaches to guide adversarial generation in prior work. The methodology seems sound and correct.

Critical weaknesses

I have concerns about the metrics the author uses. First, the detectors used to measure detectability reduction are several years old and not representative of the current state of the art in AI detection. Additionally, while embedding similarity remains a proxy for meaning preservation, no metric measures the stylistic preservation of the source text or that the proposed method preserves text coherence.

Questions for authors
  1. What additional detectors can you use to establish this method continues to work even on detectors specifically designed for adversarial robustness?
  2. What additional metrics can be used to establish stylistic features are preserved?
5 comments

Thank you for your careful and valuable comments and questions. We will explain your concern and questions point by point.

Q1 & W1: Detector Age and Robustness

A: WaveHumanize lowers TPR on both corpora for both recent detectors. The largest drops are 33.9 percentage points for WaveDetect on HC3 and 27.7 points for desklib on HC3.

We use the same evaluation setup as in the paper. To assess transfer beyond the likelihood-based evaluator, we additionally rescore the frozen outputs using WaveDetect (ACL Findings 2026), which is designed to resist eleven RAID attacks, and desklib (2025), which was trained on RAID. Together with the evaluations reported in the paper, these results cover five detector families and detectors published between 2019 and 2026.

DetectorMAGE ΔAMAGE TPR@1%HC3 ΔAHC3 TPR@1%
desklib (2025, RAID-trained)0.07547.3→23.60.02298.6→70.9
WaveDetect (2026, attack-resistant)0.05316.6→6.40.04583.3→49.4

Here, ΔA denotes the AUROC drop, and TPR@1% is the detection rate at the fixed threshold.

Q2 & W2: The problem of preserving stylistic features

A: With 11.6% token edits, WaveHumanize leads on three style measures, ties DIPPER on cosine similarity, and keeps the coherence-metric change below 1%. Human ratings show 77.4% of evaluated outputs are coherent and 74.2% natural.

MethodWMD↓fw-JS↓cos↑rank ρ↑Coherence |ΔAUC| (Change) ↓
WaveHumanize (Ours)0.12450.03800.975890.97030.0039 (<1%)
TempParaphraser (N=1)0.16810.07650.973720.95200.0063 (<1%)
TempParaphraser (N=7)0.14670.06760.970970.95100.0048 (<1%)
DIPPER0.12700.04770.975890.96570.0019 (<1%)
Adversarial-Paraphrasing0.26190.09980.910800.92360.0038 (<1%)
Simple paraphrase0.48900.25150.750400.77560.0361 (5–6%)
Recursive paraphrase0.53540.28070.699500.73340.0361 (5–6%)

Following Evaluating Style Transfer for Text [1], WMD measures lexical distance; fw-JS compares function-word distributions. Using eight features from the Stylometric AI Detector toolkit [2], cos and rank ρ measure stylometric-profile similarity and feature-order consistency, respectively. WaveHumanize leads on WMD, fw-JS, and rank ρ, and ties DIPPER on cosine. Its coherence metric changes by 0.0039; DIPPER has the smallest change (0.0019). These automatic results indicate low measured drift.

In our human evaluation, raters see one human-written text and one WaveHumanize rewrite and select the human-written text. Their choices are nearly evenly split: 52.7% choose the human-written text and 47.3% the rewrite, suggesting they have difficulty distinguishing the two (185 valid responses per item). Raters also judge whether texts are coherent and natural; 77.4% judge them coherent and 74.2% natural (201 valid responses per item). Overall, WaveHumanize shows low measured style drift and favorable human ratings for coherence and naturalness. We will report the survey data and protocol in the final appendix, in accordance with ICLR’s ethical principles.

We sincerely thank you for your valuable feedback and hope these additional results adequately address your concerns.


[1] Mir, Remi, et al. "Evaluating style transfer for text." Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.

[2] Aityan, Sergey K., et al. "A lightweight approach to detection of AI-Generated texts using stylometric features." arXiv preprint arXiv:2511.21744 (2025).

AuthorsOpenReview ↗

Looking Forward to Your Reply

Thank you very much for your comments. As the rebuttal period draws to a close, we would be sincerely grateful if you could let us know whether our responses have adequately addressed your concerns. We would greatly appreciate any further feedback, and we would be more than happy to provide additional clarifications.🙏

Thank you again for your time and consideration.

AuthorsOpenReview ↗

Looking Forward to Your Reply Again

As the author-reviewer discussion comes to a close, we'd like to hear any final thoughts you might have. Please let us know if you have remaining questions or need further clarification on any points. We hope our responses have adequately addressed your concerns.

Thank you again for your time and consideration.

Reviewer_gq3AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies selective rewriting of generated text without querying the evaluation detector. A frozen proxy LM supplies token log-ranks, which are transformed into a position-aligned Haar approximation-energy profile. The method excludes high-energy windows, samples short spans, and ranks contextual replacements using the same energy signal. Evaluation covers fifteen domains, several detector families, component ablations, and calibrated operating thresholds.

Critical strengths
  1. Clear threat model and access assumptions. The method operates without querying the evaluation detector during rewriting, using a frozen proxy language model for guidance. This distinction from methods such as RAFT is clearly explained. Evaluation across multiple detector families provides evidence of transfer beyond the primary evaluator, although transfer varies by detector.

  2. Substantial empirical improvements under the evaluated setting. Across fifteen domains, WaveHumanize achieves larger primary-detector AUROC reductions with lower measured textual dissimilarity than the compared local-edit baselines.

Critical weaknesses
  1. Incomplete comparison with closely related selective-editing methods. Although the paper includes detector-guided methods such as Adversarial Paraphrasing and TempParaphraser with oracle selection, it omits RAFT, a particularly relevant selective-editing baseline discussed in the paper. Detector-guided approaches may also transfer to independent evaluators, making a surrogate-detector comparison informative despite the difference in access assumptions.

  2. Limited evaluation of the broader humanization objective. The experiments primarily demonstrate reduced AI detectability, while meaning preservation is assessed through embedding similarity. They do not directly establish improvements in perceived naturalness or writing quality. Although the paper acknowledges this limitation, the humanization framing would benefit from stronger quality evaluation or narrower claims.

Questions for authors
  1. Could you compare against RAFT using a surrogate detector, with all rewritten outputs evaluated on independent detectors?

  2. Is the intended objective primarily detector evasion or improved perceived naturalness? Could you include evaluation of naturalness, fluency, and meaning preservation, potentially supplemented by a human or LM judge?

  3. Could you discuss the implications for improving detector robustness?

4 comments

We sincerely appreciate your thoughtful comments. Our responses are as follows.

W1 & Q1: Missing some important baselines and methods

A: We add a comparison with the original RAFT implementation. WaveHumanize scores higher on all four independent detectors, while RAFT scores higher on its GPT-Neo surrogate. These results indicate that WaveHumanize transfers better to the independent detectors in this evaluation.

RAFT uses detector feedback to guide rewriting, whereas WaveHumanize relies solely on wavelet-energy analysis. Thus, the two methods differ in their detector-access assumptions. Despite this difference, we use GPT-Neo as RAFT’s surrogate detector and evaluate both methods on GPT-Neo and four independent detectors.

DetectorEscape (WaveHumanize)Escape (RAFT)
GPT-Neo (surrogate)0.2240.333
chatgpt-detector0.1040.075
llm-content0.2010.140
openai-base0.061−0.147
openai-large0.018−0.106

RAFT achieves a higher escape score than WaveHumanize only on GPT-Neo, its surrogate detector. On all four independent detectors, WaveHumanize scores higher than RAFT. Its scores are positive across the board, whereas RAFT receives negative scores on the two OpenAI detectors.

W2 & Q2: Humanization Quality Evaluation

A: Following prior relevant work, we evaluate detectability reduction, edit cost, and semantic similarity. Following your suggestion, we further validate rewriting quality through complementary classifier-based and human evaluations. Classifier scores increase (P(human)P(\mathrm{human}): 0.119→0.349; P(natural)P(\mathrm{natural}): 0.176→0.224), while human ratings show 77.4% coherence, 74.2% naturalness, and nearly balanced human-vs-rewrite identification (52.7% vs 47.3%).

On matched MAGE and HC3 texts, classifier scores increase from 0.119 to 0.349 for P(human)P(\mathrm{human}) and from 0.176 to 0.224 for P(natural)P(\mathrm{natural}). Following the adversarial-classification evaluation framework of Evaluating Style Transfer for Text [1], we compute these scores using classifiers for human-likeness and naturalness, respectively.

MetricOriginalWaveHumanize
P(human)P(\mathrm{human})0.1190.349
P(natural)P(\mathrm{natural})0.1760.224

In our human evaluation, raters see one human-written text and one WaveHumanize rewrite and select the human-written text. Their choices are nearly evenly split: 52.7% choose the human-written text and 47.3% the rewrite, suggesting they have difficulty distinguishing the two (185 valid responses per item). Raters also judge whether texts are coherent and natural; 77.4% judge them coherent and 74.2% natural (201 valid responses per item).

We will report the survey data and protocol in the final appendix, in accordance with ICLR’s ethical principles.

Q3: Implications for Detector Robustness

A: Thank you for this insightful question. Our findings suggest that small, localized changes in textual statistical signals can substantially weaken detection without querying the target detector. Our RAFT comparison further highlights the importance of evaluating robustness on independent detectors, particularly those from different model families, rather than relying solely on surrogate-detector performance. To improve robustness, future detectors could incorporate meaning-preserving local rewrites into adversarial training and encourage prediction consistency under such perturbations.

We sincerely thank you for your valuable feedback and hope these additional results adequately address your concerns.


[1] Mir, Remi, et al. "Evaluating style transfer for text." Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.

AuthorsOpenReview ↗

Looking Forward to Your Reply

Thank you very much for your comments. As the rebuttal deadline is approaching, we would be sincerely grateful if you could let us know whether our responses have adequately addressed your concerns. We would greatly appreciate any further feedback, and we would be more than happy to provide additional clarifications.🙏

Thank you again for your time and consideration.

Thank you for your reply. All my concerns have been addressed, and I will maintain my positive rating.

Many thanks for your thoughtful feedback; we will incorporate it into the next version. 😊

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for your thoughtful comments and questions. Our responses are provided below.

Q1 & W3: Budgets

A: WaveHumanize uses no task-specific training and makes zero target-detector queries. Across MAGE and HC3 machine-generated texts, it edits an average of 11.6% of tokens at an estimated 0.01216 PFLOPs per text, including candidate generation and rescoring.

MethodTraining compute (EFLOPs)Detector mechanismDetector queriesDetector-query compute (PFLOPs)Rewriting strategyRewriting compute (PFLOPs)Token edit rateTotal editing compute (PFLOPs)
WaveHumanize0None00LLM-based span-level token editing0.0121611.6%0.01216
TempParaphraser0.18RoBERTa (neural network)70.0004LLM-based full-document rewriting0.01976.3%0.0194
Adversarial Paraphrasing50RADAR (neural network)2500.016LLM-based full-document rewriting1.0254.0%1.036
Recursive paraphrase0None00LLM-based full-document rewriting (two passes)0.01468.7%0.014
DIPPER62None00LLM-based full-document rewriting0.01449.4%0.014
Simple paraphrase0None00LLM-based full-document rewriting0.00762.8%0.007
AuthorsOpenReview ↗

Response Part 2

Q2 & W4: Parameter sensitivity

A: Thank you for this valuable suggestion. We evaluate five hyperparameters: energy quantile (α), smoothing width (h), edit-window width (W), candidate count (K), and rewriting rounds (R).

1. Energy Quantile (α)

αEscape ↑Δ vs Defaultτ ↓Efficiency ↑
0.150.2205−1.6%0.11421.93
0.30 (default)0.2241/0.11591.93
0.500.2212−1.3%0.11561.91

2. Smoothing Width (h)

hEscape ↑Δ vs Defaultτ ↓Efficiency ↑
10.2242+0.0%0.11521.95
2 (default)0.2241/0.11591.93
30.2227−0.6%0.11541.93

3. Edit-Window Width (W)

WEscape ↑Δ vs Defaultτ ↓Efficiency ↑
30.1588−29.1%0.08321.91
5 (default)0.2241/0.11591.93
70.2801+25.0%0.16291.72

4. Candidate Count (K) × Rewriting Rounds (R)

Escape

K \ RR=1 (τ)R=2 (τ)R=3 (τ)
K=50.0913 (0.0601)0.1666 (0.1142)0.2266 (0.1625)
K=200.1251 (0.0720)0.2241 (0.1159)0.2904 (0.1979)
K=500.1439 (0.0757)0.2517 (0.1463)0.3262 (0.2086)

Efficiency (Escape / τ)

K \ RR=1R=2R=3
K=51.521.461.39
K=201.741.931.47
K=501.901.721.56

The results show that α and h have minimal impact on performance (≤1.6%), indicating stable performance across these settings. Increasing W improves escape performance but also increases textual modification, as expected. The joint K×R analysis further shows that increasing candidate count can partially substitute for additional rewriting rounds. Notably, K=50 with R=2 achieves higher escape with fewer textual modifications than K=5 with R=3 (τ=0.1463 vs 0.1625). Thus, the evidence supports local stability to α/h and quantifies sensitivity to edit width and search budget.

Q3: Overall detection rates

A: Thank you for raising this point. WaveHumanize reduces detection: macro TPR drops from 36.8% to 14.0%, and document-weighted TPR drops from 55.9% to 24.0%.

MeasureAggregationBefore rewritingAfter rewriting
ASRMacro-average (15 domains)/80.8%
TPRMacro-average (15 domains)36.8%14.0%
TPRDocument-weighted ( 49,043 texts)55.9%24.0%

The 80.8% ASR means that, on average across 15 domains, 80.8% of texts detected before rewriting are no longer detected afterward. TPR is the proportion of all machine-generated texts that the detector correctly identifies as machine-generated.

W1: Quality and coherence

A: We retain classifier-based metrics commonly used in prior style-transfer work and, as you suggested, add human ratings. Classifier scores improve (P(human)P(\mathrm{human}): 0.119→0.349; P(natural)P(\mathrm{natural}): 0.176→0.224); human ratings show 77.4% coherence, 74.2% naturalness, and near-even choices (52.7% human text; 47.3% rewrite).

On matched MAGE and HC3 texts, classifier scores increase from 0.119 to 0.349 for P(human)P(\mathrm{human}) and from 0.176 to 0.224 for P(natural)P(\mathrm{natural}). Following the adversarial-classification evaluation framework of Evaluating Style Transfer for Text [1], we compute these scores using classifiers for human-likeness and naturalness, respectively.

MetricOriginalWaveHumanize
P(human)P(\mathrm{human})0.1190.349
P(natural)P(\mathrm{natural})0.1760.224

In our human evaluation, raters see one human-written text and one WaveHumanize rewrite and select the human-written text. Their choices are nearly evenly split: 52.7% choose the human-written text and 47.3% the rewrite, suggesting they have difficulty distinguishing the two (185 valid responses per item). Raters also judge whether texts are coherent and natural; 77.4% judge them coherent and 74.2% natural (201 valid responses per item).

We will report the survey data and protocol in the final appendix, in accordance with ICLR’s ethical principles.

AuthorsOpenReview ↗

Response Part 3

W2: Wavelet contribution

A: We clarify this point through additional experiments: useful localization is not limited to a single estimator, and Haar provides a compact, position-resolved implementation with stronger localization than matched-window STFT.

Estimator robustness

MetriccA₂cD₂cD₁MS₂V_raw2WVRandom selection
Escape rate0.18040.18570.19000.19000.19330.19970.1472

Haar vs STFT localization

MetricHaarSTFTHaar − STFT
Token AUROC0.7740.739+0.035
Token AUPRC0.7580.742+0.016

All six local estimators exceed random selection in Escape (0.180–0.200 vs 0.147). Haar also outperforms matched-window STFT in localization AUROC (+0.035) and AUPRC (+0.016), with positive AUPRC gains in all four settings. These results support local statistical guidance and Haar’s localization advantage in the tested diagnostic settings.

We sincerely thank you for your valuable feedback and hope these additional results adequately address your concerns.


[1] Mir, Remi, et al. "Evaluating style transfer for text." Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.

AuthorsOpenReview ↗

Summary of Reviewer Comments and Author-Reviewer Discussions

We have comprehensively addressed all concerns raised by the reviewers through extensive additional experiments and analyses, and we sincerely thank all reviewers for their positive evaluations and constructive feedback. We are deeply grateful for their recognition of the novelty, methodological soundness, and empirical contributions of WaveHumanize. Specifically, Reviewer uPvy highlighted the novelty of our wavelet-energy-based approach and acknowledged its methodological soundness; Reviewer gq3A recognized our clear threat model, detector-independent design, and substantial empirical improvements; and Reviewer WaLY appreciated our comprehensive evaluation across 15 domains, multiple detector families, and detailed component analyses. We greatly appreciate these positive assessments and the valuable suggestions that have helped us further strengthen our work.

Below, we summarize the key concerns raised by the reviewers and our corresponding responses and new results:

  1. Missing Baselines and Generalization: We added a direct comparison with RAFT under surrogate-detector settings. WaveHumanize demonstrated stronger transfer performance across all four independent detectors, providing additional evidence for its generalization without target-detector feedback.
  2. Detector Age and Robustness: We extended our evaluation to desklib (2025) and WaveDetect (2026), both incorporating adversarial robustness considerations. The results confirmed consistent reductions in detection performance across MAGE and HC3, addressing concerns about applicability to recent detectors.
  3. Humanization Quality and Preservation: We added stylometric, coherence, and human evaluations. WaveHumanize led on three style-preservation metrics, with less than 1% coherence change. Human evaluations showed approximately 76% positive coherence and naturalness ratings and nearly balanced identification choices, supporting textual quality preservation alongside reduced detectability.
  4. Computational Budgets and Parameter Sensitivity: We provided detailed comparisons of computational costs, detector queries, and editing budgets, along with sensitivity analyses of five hyperparameters. These results clarified the computational overhead, demonstrated stability for key settings, and quantified the performance-efficiency trade-offs.
  5. Methodological Contribution: We evaluated alternative local statistical estimators and compared Haar wavelets with STFT. The additional experiments supported the effectiveness of local statistical guidance and Haar's localization advantages, further clarifying our methodological contributions.
  6. Interpretation of Detection Metrics: We supplemented the conditional attack success rate with macro-averaged and document-weighted detection rates. These results clarified the interpretation of the reported improvements and provided a more comprehensive assessment of detection reduction.

We sincerely appreciate the reviewers' positive recognition of our contributions and their constructive feedback. We believe the additional experiments and clarifications have substantially addressed the major concerns and further strengthened the evidence supporting WaveHumanize's novelty, effectiveness, and methodological contributions. We will carefully incorporate these improvements into the revised paper.

We thank all reviewers and chairs again for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.