acceptodds
Under review as a conference paper at ICLR 2027

When to Flag, Where to Localize: Risk-Controlled Detection with Authorship Localization in Mixed Human–LLM Text

24 views

Abstract

Documents containing both human-written and LLM-generated text require both deciding whether LLM-generated text exists and identifying where it occurs. Localization methods can in principle address both by treating any predicted LLM segment as evidence that the document contains generated text, but sentence-level errors can then produce false alarms on fully human-written documents. We therefore propose CoMa-LD, a training-free framework for sequential LLM detection and post-alarm localization. CoMa-LD introduces a two-state Human/LLM Markov model into a conformal test martingale: summing evidence over candidate authorship paths determines when to flag, while selecting the highest-weight path identifies where the generated text occurs. The resulting detector provides statistical guarantees that control the probability of ever falsely flagging a fully human-written document. Across multiple domains and authorship transition patterns, CoMa-LD improves both detection and localization over the evaluated baselines while maintaining low false-alarm rates. Our code is available at https://anonymous.4open.science/r/CoMa-LD-B486/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-generated text detection, Authorship localization, Conformal test martingales, False-alarm control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zMV1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes CoMa-LD for document-level detection and sentence-level localization of LLM-generated text. It combines conformal evidence with a two-state Human/LLM switching model. Forward aggregation provides document-level alarms, while post-alarm Viterbi decoding estimates sentence labels. The paper provides false-alarm guarantees under stated exchangeability assumptions and introduces document-level calibration to accommodate dependence within documents.

Critical strengths

The shared Human/LLM sequence model provides a principled way to combine document-level detection with sentence-level localization, and the ablations support the contribution of both components. Document-level calibration is a further strength: it allows dependence among sentences while controlling false alarms under exchangeability across documents, an important consideration for authorship assessment.

Critical weaknesses

Limited external validity of the mixed-authorship evaluation. All evaluated mixed documents are constructed by replacing alternating human-written blocks with LLM-generated continuations (Appendix F.1). This leaves an important gap between the evaluation and the motivating applications of revision, rewriting, and human AI collaboration: the results do not establish whether the gains persist when humans select and edit AI suggestions or contributions are intertwined within sentences. Actual collaborative-writing data exists: for example, CoAuthor, although its mixed-authorship sentences require an explicit labeling policy for this binary framework. Without such evaluation, the practical scope of the claimed detection and localization improvements remains uncertain.

Questions for authors
  1. Clarification of calibration data. What human reference/calibration data and threshold-selection procedures are used for MCP and Online Betting? In particular, do these baselines receive access to a comparable calibration-data budget to CoMa-LD_DC, which uses 500 additional human documents? This affects whether the reported gains reflect the method or differences in calibration resources.

  2. Interpretation of the detection improvement. In Table 5, CoMa-LD reduces missed detections on WritingPrompts while increasing FPR from 0.063 to 0.080. Does the detection advantage persist at comparable false-alarm rates, using thresholds selected on held-out calibration data? This would clarify how much improvement comes from the switching detector versus the operating point.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer zMV1 (1/2)

Thank you for your encouraging comments and constructive feedback. Your comments helped us clarify the calibration budgets, compare detection at similar false-alarm rates, and extend the evaluation to real collaborative writing. We respond to each point below.

  1. (W1) “All evaluated mixed documents are constructed by replacing alternating human-written blocks with LLM-generated continuations (Appendix F.1). This leaves an important gap between the evaluation and the motivating applications of revision, rewriting, and human AI collaboration: the results do not establish whether the gains persist when humans select and edit AI suggestions or contributions are intertwined within sentences. Actual collaborative-writing data exists: for example, CoAuthor, although its mixed-authorship sentences require an explicit labeling policy for this binary framework. Without such evaluation, the practical scope of the claimed detection and localization improvements remains uncertain.”

Thank you for highlighting this concern and suggesting CoAuthor. We agree that evaluating real human–AI collaborative writing is important beyond our controlled experiments.

Following your suggestion, we evaluated CoMa-LD on CoAuthor collaborative-writing documents, using 100100 human reference sentences and Gemma-3-1B-PT for scoring. For mixed-authorship sentences, we adopt a binary labeling policy: a sentence is labeled as AI-positive if it contains any AI-generated contribution, including sentences with both human and AI contributions.

MethodDocument TPR ↑WindowDiff ↓
Fast-DetectGPT0.1770.177—
Online Betting0.096±0.0090.096 \pm 0.009—
VCP—0.365±0.0000.365 \pm 0.000
WCP$_{\mathrm{len}}$—0.373±0.0000.373 \pm 0.000
CoMa-LD0.332±0.0150.332 \pm 0.0150.351±0.0000.351 \pm 0.000

CoMa-LD achieves a higher document-level TPR than Fast-DetectGPT and Online Betting, and a lower WindowDiff than VCP and WCP_len. These results provide further evidence that CoMa-LD's advantages over the evaluated baselines extend to this collaborative-writing setting, although the overall detection sensitivity remains limited.

One limitation of this evaluation is the small number of human-only documents. After filtering, only 3737 documents remain, all from a single writer. This limits our ability to reliably assess false positives across different human authors, so we do not report FPR for this experiment.

We have also added experiments on multi-editor Wikipedia text (Section 4.4, Table 5) and LLM-polished human text (Appendix G.6). These experiments further examine the effects of multiple human authors and local AI edits to human-written text. Finally, we have clarified the limitations of binary sentence-level labeling in Appendix H, as it does not distinguish between different amounts of human and AI contributions within a sentence. Finer-grained detection and localization remain directions for future work.

  1. (Q1) “What human reference/calibration data and threshold-selection procedures are used for MCP and Online Betting? In particular, do these baselines receive access to a comparable calibration-data budget to CoMa-LD_DC, which uses 500500 additional human documents? This affects whether the reported gains reflect the method or differences in calibration resources.”

Thank you for asking us to clarify the calibration budgets. MCP uses 500500 human documents for threshold calibration, matching CoMa-LD$_{\mathrm{DC}}$. Original Online Betting shares CoMa-LD's 100 human reference sentences and uses its theoretical threshold. Calibration and test sources are disjoint (Appendix F.2).

To ensure a more comparable calibration budget, we have additionally evaluated Online Betting + DC, which uses 500500 human documents to calibrate the threshold on the maximum wealth over sentence prefixes at α=0.05\alpha=0.05. The table below reports TPR at K=1K=1 and human-only FPR, while Appendix G.5, Table 16, provides the complete results for K∈{1,2,3,5}K\in\{1,2,3,5\}.

DatasetMethodTPR, K=1K=1 ↑\uparrowFPR, K=0K=0 ↓\downarrow
WPMCP0.323±0.0120.323 \pm 0.0120.050±0.0170.050 \pm 0.017
WPOnline Betting + DC0.336±0.0420.336 \pm 0.0420.067±0.0150.067 \pm 0.015
WPCoMa-LD$_{\mathrm{DC}}$0.597±0.0300.597 \pm 0.0300.040±0.0170.040 \pm 0.017
CNNMCP0.699±0.0860.699 \pm 0.0860.177±0.0950.177 \pm 0.095
CNNOnline Betting + DC0.637±0.0170.637 \pm 0.0170.073±0.0290.073 \pm 0.029
CNNCoMa-LD$_{\mathrm{DC}}$0.795±0.0450.795 \pm 0.0450.017±0.0210.017 \pm 0.021

Across the full comparison, CoMa-LD$_{\mathrm{DC}}$ has the highest TPR in seven of eight settings and the lowest empirical FPR on both datasets. The only exception in TPR is CNN/DailyMail at K=5K=5, where MCP performs best. The gains therefore persist under equal document-calibration budgets.

AuthorsOpenReview ↗

Reply to Reviewer zMV1 (2/2)

  1. (Q2) “In Table 5, CoMa-LD reduces missed detections on WritingPrompts while increasing FPR from 0.063 to 0.080. Does the detection advantage persist at comparable false-alarm rates, using thresholds selected on held-out calibration data? This would clarify how much improvement comes from the switching detector versus the operating point.”

Thank you for pointing out this important distinction. We have replaced the original detector-ablation table with Figure 7, which compares CTM and CoMa-LD across α∈{0.025,0.05,0.10}\alpha\in\{0.025,0.05,0.10\}, using the same 100 human reference sentences and Viterbi localizer. As shown in Figure 7, CoMa-LD achieves lower miss rates and WindowDiff than CTM at comparable empirical FPRs on both datasets. These results suggest that the detection improvement is not solely due to a higher false-alarm operating point. For example, on WritingPrompts, CoMa-LD at α=0.025\alpha=0.025 achieves a lower miss rate (0.3340.334 vs. 0.5610.561) and lower WindowDiff (0.2970.297 vs. 0.3420.342) than CTM at α=0.05\alpha=0.05, while operating at a slightly lower empirical FPR (0.0570.057 vs. 0.0630.063).

The human reference sources are disjoint from the test sources (Appendix F.2). The thresholds are determined by the nominal significance level α\alpha, rather than fitted using a separate held-out calibration set. We have updated the discussion in Section 4.5 accordingly.

Thank you for the additional experiments, especially on CoAuthor. I appreciate the authors taking the time to address my concern. I still feel that the connection to real-world use is a bit unclear, since people might edit or paraphrase AI suggestions and mix them with their own writing in different ways. The CoAuthor results help, but the detection rate is still quite low, and it is hard to tell much about false positives from the limited human-only data. I appreciate the authors acknowledging these limitations. I will maintain my score.

Thank you for your follow-up and for taking the time to consider our additional experiments. We understand your concerns regarding real-world applicability and false-positive control. Importantly, CoMa-LD achieves the best performance among the evaluated methods, with a document-level TPR of 0.332 (compared with 0.177 for the strongest baseline) and the lowest WindowDiff of 0.351. We believe these results demonstrate meaningful improvements in this challenging collaborative-writing setting, while acknowledging the remaining limitations. Thank you again for your constructive feedback, which helped us strengthen the paper!

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

The paper tries to localize and detect mixed LLM-Human text/document. In this work, each sentence's detection score is converted into a conformal p-value, and then a two-state Markov model weights all possible label paths to indicate which sentence/segment is written by LLM. The experiments show that while retaining low false alarm, the proposed method shows better performance than the baselines.

Critical strengths
  1. The formulation is well-motivated, separating two importants problems in LLM text detection and localization and solve them accordingly.

  2. Compared to other baselines, no additional model training is needed, and showing good generalization cross-domain/dataset.

Critical weaknesses
  1. The localization advantage seems to rest on a manually set smoothing parameter.
    • a) Combining Fig. 3 with Table 1, CoMa-LD without Viterbi (WindowDiff 0.70–0.78 on WritingPrompts) is far worse than the change-point baselines (0.28–0.48).
    • b) Labelling each sentence independently is equivalent to Viterbi with ρ = 0.5. The gain therefore comes from sequence smoothing whose strength is set by ρ = 0.15 but this value is never ablated.
Questions for authors
  1. Since Fig. 3 shows that joint decoding drives the localization gains, how much of CoMa-LD's advantage remains against a simple baseline that smooths per-sentence Fast-DetectGPT scores with an HMM/Viterbi decoder (same ρ), without conformal p-values? Is the independent rule in Fig. 3 a fair comparison point?

  2. How were ρ01 = ρ10 = 0.15 and the initial state weights chosen, and were they selected using the evaluated documents? How sensitive are the results to ρ (e.g., 0.01–0.5) and to the calibration size? Could ρ instead be estimated on held-out mixed documents, or mixed over a grid of values, while keeping validity? How does CoMa-LD perform on short LLM segments (1–3 sentences) and on paraphrased or polished human text rather than continuations?

4 comments
AuthorsOpenReview ↗

Reply to reviewer JpTw (1/2)

Thank you for your encouraging comments and constructive feedback. Your questions helped us improve the localization comparison, examine parameter sensitivity, and evaluate LLM-polished text. We respond to each point below.

  1. (W1) “Combining Fig. 3 with Table 1, CoMa-LD without Viterbi (WindowDiff 0.70–0.78 on WritingPrompts) is far worse than the change-point baselines (0.28–0.48). Labelling each sentence independently is equivalent to Viterbi with ρ=0.5\rho = 0.5. The gain therefore comes from sequence smoothing whose strength is set by ρ=0.15\rho = 0.15 but this value is never ablated.”

Thank you for highlighting the role of ρ\rho. We agree that Viterbi decoding plays an important role in CoMa-LD's localization performance. In our two-state Markov model, ρ\rho controls the probability of switching between the Human and LLM states, and 1−ρ1-\rho is the probability of staying in the current state. When ρ=0.5\rho=0.5, switching and staying are equally likely, so the Viterbi decoder labels each sentence independently based on its own betting evidence. The forward detector also reduces to a standard conformal test martingale with the fixed betting function g(p)=(1−log⁡p)/2g(p)=(1-\log p)/2. This independent-labeling variant yielded high WindowDiff on WritingPrompts in the original ablation.

At smaller values of ρ\rho, such as our default ρ=0.15\rho=0.15, Eq. (8) penalizes switching between the Human and LLM states relative to staying in the same state. This favors contiguous spans with consistent labels and reduces sensitivity to weak evidence from isolated sentences. For a short LLM span embedded in human text, the decoder requires sufficient evidence to justify both entering and leaving the LLM state, which can make very short segments harder to localize.

Following your suggestion, we have included ρ=0.5\rho=0.5 in the revised Figure 5 as a reference without transition-based smoothing. Compared with ρ=0.15\rho=0.15, this setting yields higher WindowDiff and lower LLM-sentence F1 on both datasets. We further examine the contribution of conformal betting evidence under the same Viterbi smoothing in Figure 3 and the sensitivity to ρ\rho in Figure 5, as discussed in our responses to Q1 and Q2, respectively.

  1. (Q1) “Since Fig. 3 shows that joint decoding drives the localization gains, how much of CoMa-LD's advantage remains against a simple baseline that smooths per-sentence Fast-DetectGPT scores with an HMM/Viterbi decoder (same ρ\rho), without conformal p-values? Is the independent rule in Fig. 3 a fair comparison point?”

Thank you for suggesting this comparison. We designed a new ablation to assess the contribution of conformal betting evidence to localization. We compare CoMa-LD with Fast-DetectGPT + Viterbi (FD + HMM/Viterbi). Both methods use the same Viterbi decoder and transition probabilities ρ01=ρ10=0.15\rho_{01}=\rho_{10}=0.15, with the baseline using raw sentence-level Fast-DetectGPT scores as evidence. CoMa-LD achieves lower WindowDiff on both datasets at every evaluated transition count, with larger gains on documents containing multiple transitions. These results show that conformal betting evidence improves localization under the same sequence smoothing. The full results are presented in Section 4.5 and Figure 3.

  1. (Q2) “How were ρ01=ρ10=0.15\rho_{01} = \rho_{10} = 0.15 and the initial state weights chosen, and were they selected using the evaluated documents? How sensitive are the results to ρ\rho (e.g., 0.01–0.5) and to the calibration size?”

Thank you for raising these questions. We set ρ01=ρ10=0.15\rho_{01}=\rho_{10}=0.15 before the final evaluation, based on a small preliminary experiment on a separate set of documents. We use equal initial-state probabilities, π0=π1=0.5\pi_0=\pi_1=0.5, for both forward detection and Viterbi decoding. The initial weights set the prior over the authorship state of the first sentence. In documents of typical length, the influence of this prior fades quickly as evidence from subsequent sentences accumulates.

We have added a sensitivity analysis over ρ=0.01\rho=0.01 to 0.50.5 in Figure 5. Very small values of ρ\rho strongly discourage state transitions, while larger values reduce the switching penalty and can lead to fragmented predictions. Localization performance remains relatively good across intermediate values of ρ\rho, particularly around 0.100.10 to 0.150.15, rather than depending on the exact choice of 0.150.15. Figure 6 examines initial human reference sizes from 5050 to 500500 sentences. WindowDiff remains relatively stable across these sizes, while larger reference sets improve sentence-level F1 more on WritingPrompts than on CNN/DailyMail. These results are presented in Section 4.5.

AuthorsOpenReview ↗

Reply to Reviewer JpTw (2/2)

  1. (Q2) “Could ρ\rho instead be estimated on held-out mixed documents, or mixed over a grid of values, while keeping validity?”

Thank you for suggesting these alternatives. One possibility is to estimate ρ\rho using separate mixed development documents and fix it before calibration and evaluation. Alternatively, a fixed convex mixture Mtmix=∑jwjMt(ρj)M_t^{\mathrm{mix}}=\sum_j w_jM_t^{(\rho_j)}, with wj≥0w_j\geq0 and ∑jwj=1\sum_jw_j=1, also preserves the Ville bound under the conditional uniformity assumptions of Proposition 3. For CoMa-LD$_{\mathrm{DC}}$, calibrating the running maximum of the final selected or mixed process on held-out human documents preserves the guarantee in Proposition 4 under document-level exchangeability.

  1. (Q2) “How does CoMa-LD perform on short LLM segments (1–3 sentences) and on paraphrased or polished human text rather than continuations?”

Thank you for asking about short segments and AI-assisted editing. Very short LLM segments, such as a single LLM-generated sentence surrounded by human-written text, remain challenging to localize. With Viterbi smoothing, the evidence for an embedded LLM segment must be strong enough to support both the switch into the LLM state and the switch back to the human state. Evidence can accumulate across adjacent sentences in a longer segment. An isolated sentence must provide enough evidence for both switches on its own.

To evaluate AI-assisted editing, we added experiments on LLM-polished human text using WritingPrompts and CNN/DailyMail. We use Qwen3-8B to polish selected paragraphs with access to the original document as context, and NFT-finetuned Gemma-3-1B-PT for Fast-DetectGPT scoring.

The table below shows K=1K=1, where KK is the number of paragraphs selected for polishing.

DatasetMethodTPR ↑\uparrowWD ↓\downarrowHuman-only FPR ↓\downarrow
WPFast-DetectGPT0.643±0.0230.643\pm0.023—0.050±0.0100.050\pm0.010
WPVCP—0.546±0.0240.546\pm0.0240.820±0.0200.820\pm0.020
WPCoMa-LD0.673±0.0290.673\pm0.0290.293±0.0110.293\pm0.0110.090±0.0100.090\pm0.010
WPCoMa-LD$_{\mathrm{DC}}$0.619±0.0340.619\pm0.0340.291±0.0070.291\pm0.0070.053±0.0230.053\pm0.023
CNNFast-DetectGPT0.767±0.1340.767\pm0.134—0.077±0.0310.077\pm0.031
CNNVCP—0.528±0.0280.528\pm0.0280.833±0.0400.833\pm0.040
CNNCoMa-LD0.778±0.1300.778\pm0.1300.261±0.0260.261\pm0.0260.073±0.0290.073\pm0.029
CNNCoMa-LD$_{\mathrm{DC}}$0.701±0.0460.701\pm0.0460.244±0.0350.244\pm0.0350.050±0.0360.050\pm0.036

In the complete result, across K∈{1,2,3,5}K\in\{1,2,3,5\}, CoMa-LD has the highest TPR on both datasets. CoMa-LD and CoMa-LD_DC have lower WD than VCP and WCP$_{\mathrm{len}}$. Document calibration reduces false alarms, at the cost of detection sensitivity.

We added the complete results, labeling policy, and prompt to Appendix G.6, “LLM-Assisted Text Polishing.” We label any sentence containing LLM-polished content as AI-positive, including partially polished sentences. These labels indicate AI involvement at the sentence level. We discuss this labeling limitation in Appendix H and identify token-level detection and localization as directions for future work.

Thanks for author's rebuttal, I believe these additional ablation studies would further strengthen the paper. I keep my rating.

Thanks for taking the time to read our response. Your suggestions helped us improve the paper.

Reviewer_AMx3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel method called CoMa-LD (Conformal Martingales for LLM Detection and Localization). The authors claim that their method addresses two key challenges in documents co-authored by humans and large language models (LLMs): detecting whether a document contains LLM-generated content (when to flag) and precisely identifying which sentences were generated by an LLM (where to flag).

Critical strengths

Detecting text written by both humans and machines is an important research problem, and the authors' conformal martingale approach is interesting.

Critical weaknesses
  1. The datasets used in the paper are based on relatively old models and may not be suitable for text generated by newer LLMs.

  2. The synthetic data consist of alternating human-written and AI-generated paragraphs, which align almost perfectly with the two-state Markov switching assumption. In contrast, real-world mixed-content scenarios are much more complex.

Questions for authors

Could the detection results be affected if an article is written by different authors?

4 comments
AuthorsOpenReview ↗

Reply to reviewer AMx3

Thanks for the thorough and insightful comments. These suggestions have helped us extend the evaluation to recent models and multi-editor text and clarify the scope of our results. We address each point in detail below. All changes to the manuscript are highlighted in blue for ease of reference.

  1. (W1) “The datasets used in the paper are based on relatively old models and may not be suitable for text generated by newer LLMs.”

Thank you for highlighting the need to evaluate our method on more recent LLMs. To address this concern, we have added experiments using Gemma4-12B (June 2026) for generation and Qwen3.5-9B (March 2026) for scoring on both WritingPrompts and CNN/DailyMail. The table below reports document-level TPR and sentence-level WindowDiff (WD) for K=3K=3, along with FPR on human-only documents (K=0K=0).

DatasetMethodTPR, K=3K=3 ↑\uparrowWD, K=3K=3 ↓\downarrowHuman-only FPR ↓\downarrow
WPFast-DetectGPT0.109±0.0610.109\pm0.061—0.033±0.0120.033\pm0.012
WPOnline Betting0.288±0.0370.288\pm0.037—0.073±0.0120.073\pm0.012
WPVCP—0.371±0.0040.371\pm0.0040.180±0.0170.180\pm0.017
WPCoMa-LD0.616±0.0290.616\pm0.0290.334±0.0070.334\pm0.0070.093±0.0120.093\pm0.012
WPCoMa-LD$_{\mathrm{DC}}$0.611±0.0320.611\pm0.0320.335±0.0110.335\pm0.0110.050±0.0100.050\pm0.010
CNNFast-DetectGPT0.064±0.0140.064\pm0.014—0.077±0.0250.077\pm0.025
CNNOnline Betting0.267±0.0860.267\pm0.086—0.077±0.0250.077\pm0.025
CNNVCP—0.370±0.0130.370\pm0.0130.237±0.0230.237\pm0.023
CNNCoMa-LD0.403±0.2480.403\pm0.2480.354±0.0410.354\pm0.0410.070±0.0440.070\pm0.044
CNNCoMa-LD$_{\mathrm{DC}}$0.371±0.0800.371\pm0.0800.381±0.0080.381\pm0.0080.023±0.0210.023\pm0.021

Across all evaluated values of KK, CoMa-LD achieves the highest TPR. It also achieves the lowest WindowDiff on WritingPrompts for all values of KK, and on CNN/DailyMail for K=2,3,5K=2,3,5, while VCP performs better at K=1K=1 on CNN/DailyMail. Document-level calibration further reduces empirical FPR, at the cost of lower TPR and slightly higher WindowDiff. The complete results for K∈1,2,3,5K\in{1,2,3,5} are reported in Appendix G.3, Tables 12–13.

  1. (W2, Q1) “The synthetic data consist of alternating human-written and AI-generated paragraphs, which align almost perfectly with the two-state Markov switching assumption. In contrast, real-world mixed-content scenarios are much more complex. Could the detection results be affected if an article is written by different authors?”

Thank you for highlighting this limitation of our synthetic data. We agree that real-world mixed-content documents can involve more complex authorship patterns than alternating human-written and AI-generated paragraphs.

To examine the effect of multiple human authors, we have added an evaluation using multi-editor Wikipedia text with authorship provenance identified by WikiWho. We retain the original hyperparameters, including α=0.05\alpha=0.05 and 100100 human reference sentences, and calibrate CoMa-LD$_{\mathrm{DC}}$ using 500500 held-out human Wikipedia documents. Selected results are shown below.

MethodTPR ↑\uparrowWD ↓\downarrowHuman-only FPR ↓\downarrow
Fast-DetectGPT0.259±0.0250.259\pm0.025—0.033±0.0210.033\pm0.021
WCP$_{\mathrm{len}}$—0.327±0.0020.327\pm0.002—
CoMa-LD0.716±0.0090.716\pm0.0090.284±0.0120.284\pm0.0120.053±0.0250.053\pm0.025
CoMa-LD$_{\mathrm{DC}}$0.638±0.0130.638\pm0.0130.283±0.0110.283\pm0.0110.027±0.0060.027\pm0.006

Both CoMa-LD variants achieve higher TPR than Fast-DetectGPT and lower WindowDiff than WCP$_{\mathrm{len}}$, with relatively low empirical FPR. These results provide additional evidence that the performance gains extend to text involving multiple human authors. The two states in our Markov model represent human-written and AI-generated text, rather than individual human authors. Using reference and calibration data from multiple authors helps account for variation in writing styles. We have added the multi-editor Wikipedia evaluation to Section 4.4 and report the results in Table 5.

In mixed documents with more complex authorship patterns, the two-state Markov assumption can affect how well CoMa-LD detects and localizes LLM-generated text. Even in these settings, the false-alarm guarantee for human-only documents holds. We have expanded Appendix H to discuss cases that may call for finer-grained authorship labels, including human edits to LLM-generated text, varying degrees of AI involvement, and sentences with contributions from both humans and LLMs.

AuthorsOpenReview ↗

Summary of Paper Revisions for Reviewer AMx3

Thank you again for your constructive feedback. We have carefully addressed the concerns raised and provide the following new supplementary results and analyses during the rebuttal:

  • (W1) We evaluated CoMa-LD with Gemma 4-12B for generation and Qwen3.5-9B for scoring. CoMa-LD achieved the highest TPR across all tested settings and the lowest WindowDiff in most settings (Appendix G.3, Tables 12–13).
  • (W2/Q1) We evaluated mixed documents built from multi-editor Wikipedia text. CoMa-LD achieved higher TPR than Fast-DetectGPT and lower WindowDiff than WCP$_{\mathrm{len}}$ (Section 4.4, Table 5).
  • (W2) We evaluated LLM-polished text, where edits can affect only part of a sentence. CoMa-LD achieved the highest TPR, and both variants had lower WindowDiff than the localization baselines across all tested settings (Appendix G.6, Tables 17–18).

We appreciate your constructive suggestions and hope these responses address your concerns. As the discussion period draws to a close, we welcome further discussion and look forward to the re-evaluation of our paper.

Thank you for the authors' response. I am considering raising my rating.

Thank you for taking the time to read our response and reconsider your rating. We appreciate your feedback, which helped us improve the paper!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author-reviewer Discussion

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and reviewers for their time and constructive feedback. We are encouraged by the positive feedback from Reviewers JpTw and zMV1, who recognized the contributions of our proposed framework and acknowledged our additional experiments. In particular, Reviewer JpTw confirmed that our new ablation studies further strengthen the paper. We are also pleased that Reviewer AMx3 thanked us for our response and indicated that they are considering raising their rating. We appreciate the reviewers' constructive engagement throughout the rebuttal.

Our key rebuttal updates are summarized below.

Reviewer AMx3

  1. Newer LLMs: Additional evaluations using Gemma 4-12B and Qwen3.5-9B demonstrate strong generalization (Appendix G.3).
  2. Multiple human authors: New WikiWho experiments demonstrate improved detection and localization (Section 4.4, Table 5).
  3. Realistic writing scenarios: Additional LLM-polished text experiments examine more complex AI-assisted editing (Appendix G.6).

Reviewer JpTw

  1. Conformal evidence: A new Fast-DetectGPT + Viterbi baseline confirms the contribution of conformal evidence (Figure 3).
  2. Parameter sensitivity: Additional analyses examine switching probabilities and reference-set sizes (Figures 5–6).
  3. AI-assisted editing: New experiments evaluate LLM-polished text and clarify short-segment limitations (Appendix G.6).

Reviewer zMV1

  1. Collaborative writing: CoAuthor experiments demonstrate improved detection and localization over the evaluated baselines.
  2. Fair calibration: An equally calibrated Online Betting baseline confirms consistent improvements (Appendix G.5).
  3. False-positive control: Additional comparisons demonstrate better detection and localization at comparable false-positive rates (Figure 7).

We believe these additional results have substantially strengthened the paper and addressed the reviewers' concerns. We sincerely appreciate the Area Chair's consideration of our revised submission.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.