acceptodds
Under review as a conference paper at ICLR 2027

AnchorWrite: Detector-Query-Free Rewriting of LLM-Generated Text via Human-Reference Alignment

Abstract

Large language models (LLMs) generate fluent text that can remain statistically distinguishable from human writing. Existing post-generation rewriting approaches are often evaluated primarily through detector reduction, making it difficult to separate detector-specific optimization from movement toward an independently defined human linguistic reference. We introduce AnchorWrite, a detector-query-free framework that treats rewriting as constrained alignment to a fixed human reference. AnchorWrite constructs cross-order language-model statistics from human-authored text, calibrates them against a length-conditioned human reference, and uses token-level evidence to localize regions of linguistic deviation. A pretrained language model proposes local rewrites, which are accepted only when they reduce human-reference distance while satisfying semantic, linguistic, and edit constraints. Across six domains, mean human-reference distance reduction increases from 0.068 at 5.60% realized editing to 0.363 at 24.32%. On frozen outputs, AUROC decreases by 0.057-0.089 across six external AI-text detection methods, despite no detector feedback during rewriting. Matched-edit comparisons with instruction-based humanization and neural paraphrasing show that detector responses remain transformation-dependent. At the strongest regime, BERTScore remains 0.968 while perplexity increases from 8.88 to 10.71. These results position human-reference alignment as an interpretable control signal for constrained rewriting, with reduced detector separability emerging as an external transfer effect rather than an optimization target.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-generated text, human-reference alignment, text rewriting, AI-text detection, detector robustness, controlled text generation

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_3mVbOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AnchorWrite, a detector-query-free framework for rewriting LLM-generated text toward a fixed statistical reference constructed from human-authored documents. It combines cross-order n-gram statistics, length-conditioned calibration, token-level anomaly localization, and constrained LLM rewriting. Candidate rewrites are accepted based on reference-distance reduction and preservation constraints. Evaluation across 900 documents, six domains, and six external detectors reports reduced reference distance and detector separability while maintaining high semantic similarity. Additional analyses examine matched-edit baselines, human judgments, and transfer across model backbones.

Critical strengths
  1. The framework provides a coherent and interpretable approach to controlled rewriting without downstream detector feedback. Separating token-level anomaly localization from document-level reference-based acceptance makes the optimization process understandable and allows explicit preservation constraints.

  2. The evaluation is broad and considers both detector transfer and text quality. Multiple domains, detectors, rewriting strengths, and backbones, together with matched-edit comparisons and human evaluation, provide meaningful evidence about the method’s behavior. The manuscript also carefully acknowledges important limitations.

Critical weaknesses
  1. The distinctive benefit of reference-guided optimization is not sufficiently isolated. Reference-distance reduction is explicitly required for candidate acceptance, so this reduction alone does not establish an advantage over generic constrained rewriting. At L4, Qwen Rewrite achieves lower AUROC on five of six detectors with comparable editing and semantic similarity. Controlled comparisons separating anomaly localization and reference-based acceptance are therefore important for establishing the proposed mechanism’s additional value.

  2. The interpretation and usefulness of the reference objective require stronger validation. Held-out human documents reportedly have a larger mean reference distance than original AI-generated documents, indicating that lower distance does not necessarily imply closer alignment with human writing in the evaluation domains. The authors appropriately acknowledge this distribution shift, but the benefit of optimizing this particular corpus-specific reference remains insufficiently established beyond reducing the defined objective.

Post-rebuttal update Thank you for the response and the additional analyses. After reviewing the rebuttal, I still find the evidence insufficient to support the paper’s main conclusions. I consider this concern substantial enough to warrant rejection and have revised my rating from 2 to 1.

Questions for authors
  1. How do the reference-distance changes achieved by Qwen Humanize, Pegasus, and Qwen Rewrite compare with AnchorWrite under comparable editing and preservation constraints?

  2. Can you isolate the contributions of anomaly localization and reference-based candidate acceptance while keeping the proposal budget, rewriting backbone, and preservation constraints fixed?

  3. Given that held-out human documents have larger reference distances than original AI-generated documents, what specific writing characteristics does reducing this distance capture? How should the claimed alignment benefit be interpreted?

  4. Can you report inference costs, including candidate counts or generated tokens, and paired uncertainty estimates comparing AnchorWrite directly with the main matched-edit baselines?

Reviewer_3tDdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AnchorWrite, a detector-query-free framework for rewriting LLM-generated text through alignment with human-derived linguistic statistics. The method uses cross-order language model statistics to identify anomalous text regions and iteratively generates local rewrites while preserving semantic meaning. Experiments across six domains demonstrate that AnchorWrite reduces the distance to a fixed human reference and decreases the detectability of AI-generated text across multiple external detectors without relying on detector feedback.

Critical strengths

The paper is well motivated and tackles the problem of LLM-generated text rewriting from a relatively novel perspective, using human-reference alignment instead of relying on detector feedback. The paper provides a thorough experimental evaluation from multiple perspectives, including comparisons with baselines, text quality assessment, human evaluation, cross-domain performance and cross-backbone transfer.

Critical weaknesses

Although Table 2 demonstrates the effectiveness of AnchorWrite compared to the baselines, Table 12 shows that Qwen Rewrite and back-translation achieve better results. This raises concerns about the additional benefits of the proposed human-reference alignment mechanism. The authors should further clarify its advantages over these simpler baselines. According to Table 24, the human evaluation exhibits very low inter-annotator reliability, with consistently low Krippendorff's alpha values across the dimensions. Although exact agreement is approximately 70%, the severe ceiling effect substantially limits the discriminative power of the ratings. Consequently, the human evaluation provides limited evidence for meaningful quality differences between the compared methods.

Questions for authors
  1. Why do the authors choose the cross-order statistic as the human-reference metric? Could the authors provide further justification for its validity? 2 Why do the authors present the baseline comparisons separately in Tables 2 and 12 rather than in a unified table, especially given that the methods in Table 12 achieve better results on several detectors?
Reviewer_VTQrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AnchorWrite reframes post-generation rewriting of LLM text as alignment to a fixed human-derived statistic, with no detector feedback: 2-gram/4-gram KenLM models on UMBC WebBase define a cross-order statistic A₂,₄, standardized by length into a distance D; anomalous tokens are localized, Qwen3-8B proposes sentence rewrites under semantic/NLI/edit constraints, and candidates are accepted only if D decreases. On 900 AI documents across six domains, mean D reduction grows from 0.068 to 0.363 over edit ratios 5.6–24.3%; frozen outputs show AUROC drops of 0.057–0.089 on six external detectors; BERTScore stays 0.968 at L4 while PPL rises 8.88→10.71. Matched-edit baselines, a blinded human study, and cross-backbone transfer are included. The authors' own audit shows held-out human controls have larger D (6.14) than the AI originals (5.81).

Critical strengths
  1. Evaluation rigor and self-critical transparency well above the norm. Outputs frozen before detector evaluation, globally fixed score orientation with a 48/48 AUROC-reproduction audit, matched-edit baselines (L4: 24.32/24.05/23.89%), paired bootstrap CIs, and proactive reporting of evidence against the method itself (App. B.4 audit, Krippendorff's α ≈ 0, ceiling effects, appendix controls that beat it).

  2. The core phenomenon is consistently demonstrated and well scoped. AUROC declines monotonically on all six detectors across L1–L4 (e.g., FastDetect 0.9677→0.9036), in every domain, across 18 backbone×detector transfer cells, corroborated by AUPRC/TPR operating points, with quality costs honestly quantified.

Critical weaknesses
  1. The central construct is invalidated by the paper's own audit, and the framing outruns the evidence. Held-out human text is farther from the "human reference" than the AI originals in all six domains (D: 6.14 vs 5.81; rewrites move to 5.45 — away from actual human text), yet the title and abstract keep the humanization framing. App. H shows the machinery is not even the best route to the transfer effect: at matched L4 edits, generic Qwen Rewrite is lower on 5/6 detectors (FastDetect 0.8777 vs 0.9036; LRR 0.6680 vs 0.7214) at identical BERTScore, and the margin over the trivial Qwen Humanize instruction is only 0.0067–0.0284 AUROC at large unreported compute.

  2. No ablation isolates the proposed mechanisms, and the one superiority claim is unquantified. There is no random-localization or D-blind-acceptance control (ΔD > 0 is an acceptance condition, so Table 1's 99.3% "improved" verifies the loop, not the signal), and Table 2's method comparisons are point estimates even though App. J retains per-sample scores for all three methods — the paired CIs could have been computed and were not. Pegasus, which raises RoBERTa AUROC above the original at L4, is a weak representative of the paraphrasing-attack literature.

Questions for authors
  1. Report paired AnchorWrite-vs-baseline bootstrap intervals from the retained per-sample scores: do the 0.0067-level margins over Qwen Humanize survive?

  2. Add ablations at matched edit ratio: random sentence selection, and acceptance of an arbitrary constraint-satisfying candidate. If detector transfer survives, the reference-guidance framing is untenable.

  3. Recalibrate D on in-domain human text: does the human-vs-AI D ordering flip, and does transfer strengthen?

  4. What does the statistic capture that justifies "human reference" over "UMBC web-crawl center", given the z-sign evidence in App. B.4?

  5. Only two generator families (Moonshot, ChatGPT; versions unrecoverable) — does transfer hold for current-generation LLM text?

  6. Report compute per condition and add DIPPER/MASH-style humanization baselines.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers and the Area Chair

Dear Reviewers and Area Chair,

Thank you so much for your time, thoughtful reviews, and constructive feedback. Your comments have helped us improve the methodological clarity, experimental validation, and presentation of AnchorWrite. We have revised the manuscript and provided detailed individual rebuttals. Below, we summarize the main changes in response to each reviewer's concerns.

Reviewer 3tDd, thank you for highlighting the need to justify the reference statistic, clarify baseline comparisons, and strengthen human evaluation. We clarified the 2–4 cross-order statistic, distinguishing token-level localization from document-level reference-based acceptance and defining D as a corpus-relative control objective, not a universal human-likeness measure (Sections 3.2–3.4; Appendix B). We added a 150-document 2×2 control study (Section 5.6; Appendix J) and a 900-document reference-distance comparison with Pegasus, Qwen Humanize, and Qwen Rewrite (Appendix J.5). We acknowledge that Qwen Rewrite and back-translation obtain lower AUROC on several detectors (Appendix H), so we do not claim universal detector superiority. We also expanded the human-evaluation analysis with Krippendorff's alpha, rating distributions, paired tests, and discussion of low agreement and ceiling effects (Appendix M.7).

Reviewer 3mVb, thank you for urging a clearer assessment of reference guidance and its computational cost. At L4, AnchorWrite reduces the fixed reference distance by +0.3630, versus −0.0426 for Pegasus, −0.0669 for Qwen Humanize, and −0.1242 for Qwen Rewrite, with paired confidence intervals (Appendix J.5). These results demonstrate optimization of the specified reference objective, not independent proof of human-likeness. Our 150-document 2×2 study compares cross-order/random localization and reference-guided/D-blind acceptance (Section 5.6; Appendix J). AnchorWrite attains BERTScore 0.96785 and MiniLM similarity 0.97615 with fewer evaluated candidates, but unequal realized edits and proposal pools prevent strict causal isolation. We clarified the held-out-human distribution shift and the corpus-relative meaning of D (Section 5.1; Appendix B.4), and added candidate counts and direct paired AUROC comparisons (Appendix J.4–J.5). Candidate counts are not GPU-hours or generated-token costs.

Reviewer VTQr, thank you for your detailed questions on statistical uncertainty, mechanism validation, reference validity, transfer, and compute. We added paired AUROC confidence intervals against Qwen Humanize and Pegasus for detectors with retained sample-level scores, including inconclusive and baseline-favoring comparisons (Appendix J.5), and the 150-document 2×2 study with its limitations (Section 5.6; Appendix J). We reported that held-out human texts have larger D than original AI texts under the fixed UMBC calibration (Section 5.1; Appendix B.4); as clarified in our rebuttal, in-domain recalibration was not performed, so its effect remains unknown. We explain that cross-order predictability is a corpus-derived control signal, not a comprehensive human-writing measure (Section 3.2). Three additional rewriting backbones were tested on 300 documents, with positive AUROC reductions in all 18 backbone–detector combinations (Section 5.8; Appendix L); this does not establish transfer to new source generators. We added candidate-level accounting (Appendix J.4), acknowledged the absence of DIPPER/MASH evaluations in our rebuttal, and discussed dual-use risks in the Ethics Statement.

Across these revisions, we distinguish optimization of the fixed reference objective from independently evaluated detector transfer and explicitly state the remaining limitations in reference-corpus sensitivity, mechanism isolation, and computation. We appreciate your time and would welcome any further questions or clarification requests.

Best regards,
Authors

Area_Chair_Tw8qOpenReview ↗

Reminder: Discussion deadline approaching

Dear Reviewer,

Thank you for your time and effort in reviewing Submission 56531. As the discussion deadline is approaching, could you please read the authors’ response and any additional evidence, reply in the discussion to indicate whether your concerns have been addressed, and update your evaluation accordingly as soon as possible? Thank you for your contribution to the review process.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.