- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper studies sentence-level authorship localisation: given a passage, the goal is to identify which sentences are written by humans and which are generated by LLMs. To address this problem, the author(s) develops LA-CPD, an algorithm that incorporates local evidence into the change point detection framework. Its main contribution beyond existing approaches lies in the use of a windowed CUSUM-type statistic to quantify local evidence and the incorporation of such a reward term into the dynamic programming objective. Numerical experiments demonstrate the superior performance of the proposed approach over existing state-of-the-art algorithms.
Critical strengths
I summarize the critical strengths of the paper as follows:
-
While there is a huge literature on detecting AI-generated text, much of the existing work formulates the problem as binary classification at the document level, assuming that an entire passage is written either by humans or by LLMs. However, this assumption can be easily violated in practice, as human-written and LLM-generated text can coexist within the same passage. In this regard, the paper addresses a less-explored but practically important problem.
-
The proposed local-evidence-based approach has an intuitive advantage over existing change-point-detection-based methods for LLM text localization. Specifically, by constructing the CUSUM statistic within a local window of size , the method can mitigate the dilution of change point signals caused by averaging over longer and heterogeneous segments. For example, if a segment contains an undetected change point, its overall mean may not accurately represent the underlying signal, potentially obscuring the detection for another change point. Using local windows can alleviate this issue by focusing on observations in the immediate neighbourhood of the candidate boundary. The numerical experiments provide empirical support for this advantage, through the comparison with the WCP+AIC approach of Li et al. (2026).
Critical weaknesses
My biggest concern is that the current presentation does not sufficiently discuss the methodological similarities between the proposed LA-CPD approach and the WCP method of Li et al. (2026), which also studies LLM text localization using CPD. As a reader, I find it difficult to distinguish which components are genuinely new in this paper and which have already been developed in Li et al. (2026).
Specifically, there are several substantial methodological similarities between LA-CPD and Li et al. (2026), which should be discussed more explicitly:
-
CUSUM statistics: Li et al. (2026) already employ weighted CUSUM statistics for CPD. The paper's local evidence statistic introduced in Eq. (4) is equivalent to the squared weighted CUSUM statistic of Li et al. (2026), evaluated within a fixed local window of size .
-
Weighted dynamic programming: The weighted least squares segment cost in Eq. (3) is also used in the WCP implementation of Li et al. (2026), which includes a dynamic programming approach for optimising change point locations, although the main paper of Li et al. focuses on the NOT algorithm. The dynamic programming formulation in Eq. (6) follows essentially the same structure, with the different being the incorporation of an additional reward term, for the local change point evidence.
-
AIC-based model selection: The AIC criterion in Eq. (7), which selects the number of change points, is also present in the implementation of Li et al. (2026).
Based on these connections, my understanding is that the main methodological contributions of the proposal lie in two aspects: (i) evaluating weighted CUSUM statistics within a fixed window, and (ii) incorporating this local evidence as an additional reward term in the objective. The authors should more explicitly acknowledge what has already been established in the literature and focus on highlighting these two new components, their motivations, and their advantages over Li et al. in the main paper.
I also have several suggestions regarding the experiments:
-
Li et al. (2026) investigate the adversarial robustness of WCP by considering different types of attacks, such as paraphrasing. It would be interesting to examine whether the proposed LA-CPD approach is more robust under similar adversarial settings.
-
The current paper focuses exclusively on sentence-level localisation. However, authorship transitions can also occur within sentences, making token-level localisation more informative in practice. I encourage the authors to extend their methodology to this setting.
-
The experiments rely on synthetic human–LLM coauthored documents, which may not fully capture the complexity of real-world human–AI collaboration. I suggest that the authors evaluate their approach on the CoAuthor dataset, which records actual human–AI collaborative writing texts.
Questions for authors
Would it be possible to provide theoretical justifications for the benefits of constructing the CUSUM statistic within a local window and incorporating the resulting term into the objective function?
4 comments
[Response to Weaknesses 4–6] Robustness to Text Perturbations, Token-Level Localization, and Evaluation on CoAuthor
Summary of Weakness 4: Robustness under text perturbations.
Response: Thank you for suggesting an adversarial robustness evaluation following Li et al. We added experiments using the same perturbation types: decoherence and paraphrasing. Decoherence swaps an adjacent word pair in eligible LLM-authored units, while paraphrasing rewrites human-authored units.
| Method | Decoherence ACC ↑ | WD ↓ | AI-F1 ↑ | Paraphrase ACC ↑ | WD ↓ | AI-F1 ↑ |
|---|---|---|---|---|---|---|
| SenPred | 0.4513 | 0.9328 | 0.0231 | 0.5315 | 0.9101 | 0.0530 |
| Voting | 0.4787 | 0.7392 | 0.0908 | 0.6023 | 0.6804 | 0.2042 |
| TextTiling | 0.4826 | 0.5110 | 0.1694 | 0.6231 | 0.4813 | 0.3583 |
| PaLD-scores | 0.5926 | 0.8774 | 0.0028 | 0.5918 | 0.9380 | 0.0327 |
| VCP | 0.6133 | 0.2908 | 0.0637 | 0.6439 | 0.2934 | 0.2327 |
| WCP | 0.6319 | 0.2920 | 0.1356 | 0.6615 | 0.2866 | 0.2424 |
| WCP+AIC | 0.6085 | 0.3279 | 0.2522 | 0.6645 | 0.3254 | 0.3620 |
| LA-CPD+AIC | 0.6491 | 0.2953 | 0.2502 | 0.6949 | 0.2845 | 0.3520 |
LA-CPD+AIC shows competitive performance under both perturbations, achieving the highest ACC in both settings and the lowest WD under paraphrasing. This behavior is consistent with its fixed-window evidence, which aggregates neighboring scores, and its global fitting term, which enforces partition consistency and can reduce the influence of local perturbations. However, when a perturbation weakens the score contrast at a true transition, it may still cause boundary shifts or merge adjacent fragments. We will include the perturbation construction, parameter settings, and complete results in the revision.
Summary of Weakness 5: Extending from Sentence-Level to Token-Level Localization.
Response: Thank you for highlighting the value of token-level localization. Following your suggestion, we conducted an additional experiment that directly searches for token-level boundaries in the complete original texts.
Token-level extension and evaluation. We adapt LA-CPD to token sequences while retaining the joint objective combining local boundary evidence with global within-segment residuals. An AIC-style criterion selects the number of cuts from a predefined range. The results for LA-CPD+AIC are as follows:
| Localization granularity | Condition | ACC ↑ | WD ↓ | AI-F1 ↑ |
|---|---|---|---|---|
| Sentence level | B2 | 0.8175 | 0.2002 | 0.8130 |
| Token level | B2 | 0.7153 | 0.4075 | 0.6308 |
| Sentence level | B3-2 | 0.7144 | 0.3396 | 0.3381 |
| Token level | B3-2 | 0.5125 | 0.4162 | 0.2371 |
| Sentence level | B3-3 | 0.7503 | 0.3090 | 0.5619 |
| Token level | B3-3 | 0.7072 | 0.3837 | 0.4762 |
Transferability and remaining challenges. These experiments demonstrate that LA-CPD can be extended to token-level localization, while precise boundary recovery remains challenging. The extension retains the joint partitioning framework and DP optimization, with token-based scores and candidate boundaries. Contextual aggregation can smooth score changes near transitions, while weak score separation and short source spans can make accurate cut placement difficult. Further improvements therefore concern strengthening transition evidence and calibrating the token-based window size, minimum segment length, and count penalty. We will include the token-level experiments and clarify the preprocessing, scoring, and evaluation protocols in the revision.
Summary of Weakness 6: Generalization to real human–AI collaborative writing.
Response: Thank you for recommending CoAuthor. To assess transfer beyond synthetic coauthored documents, we conducted additional experiments on this dataset of real human–AI collaborative writing.
| Method | ACC ↑ | WD ↓ |
|---|---|---|
| SenPred | 0.5234 | 0.5554 |
| Voting | 0.5171 | 0.4869 |
| TextTiling | 0.5071 | 0.4506 |
| PaLD-scores | 0.6599 | 0.4989 |
| VCP | 0.6918 | 0.3955 |
| WCP | 0.6869 | 0.3967 |
| WCP+AIC | 0.6598 | 0.4006 |
| LA-CPD+AIC | 0.7006 | 0.3951 |
Overall, LA-CPD+AIC shows promising transfer to real human–AI collaborative text, achieving the highest ACC and lowest WD among the compared methods. This transfer is consistent with the fact that LA-CPD selects partitions using local transition evidence, within-segment consistency, and an AIC-style criterion, rather than relying on a fixed alternation pattern in synthetic data; Thus, it can accommodate irregular source transitions in real collaborative writing. We will include the CoAuthor preprocessing, label mapping, parameter settings, and complete experimental results in the revision.
References:
[1] Lee et al. Coauthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities. CHI 2022.
[Response to Question 1] Theoretical Justification for Local-Window Evidence and Its Integration into the Objective
Summary of Question 1: Please justify the benefits of constructing CUSUM within a local window and incorporating the resulting evidence into the partitioning objective.
Response: Thank you for highlighting the need to clarify the theoretical motivation for local-window evidence and its incorporation into the objective. Under the same weights and window, Eq. (4) equals the squared weighted CUSUM statistic; we do not claim the statistic itself as new. Our focus is on how fixed-window evidence induces a local boundary preference and contributes to global partitioning as an additive reward. The justification consists of three aspects:
(1) Local windows aggregate evidence around a transition. Appendix A.3, Eq. (20), shows that exactly equals the weighted residual reduction from fitting two local means instead of one shared mean.
To understand its localization behavior, consider an isolated, nonzero mean shift at boundary . Assume unit weights, independent zero-mean noise with variance , and full windows of length containing no other transitions. For an integer offset d satisfying ,
Thus, the expected evidence peaks at the true boundary and decreases with displacement. The noise variance of the window-mean contrast is , explaining how aggregation helps distinguish a sustained change from local fluctuations.
(2) The reward strengthens the preference for supported boundaries. Our objective,
where is the total weighted within-segment residual. This objective combines within-segment consistency with local transition evidence. Under the same model, compare two feasible partitions differing only in a cut at versus :
For and , the added term is positive: the reward increases the expected objective gap in favor of the true boundary. This explains its intended benefit without guaranteeing improved localization in every finite sample.
(3) Additivity preserves exact optimization. Because the evidence is precomputed and additive over cuts, DP exactly optimizes the joint objective for each feasible fixed cut count without additional state dimensions, subject to the candidate-position and minimum segment-length constraints. Appendix A.4, Eq. (24), provides the recurrence and optimality argument. AIC-style selection then compares the resulting candidates across counts.
We will add this analysis and its assumptions to the revision. The expectation result applies to isolated changes; windows containing multiple transitions require further analysis. The residual identity and fixed-count optimization guarantee hold independently of these noise assumptions.
Response for Reviewer 7H8h
Dear Reviewer 7H8h,
Thank you for your encouraging comments and helpful suggestions. We have clarified the relationship with WCP and added theoretical explanations and experiments on robustness, token-level localization, and CoAuthor. If you have time, we would be grateful for your thoughts on whether these additions address your questions.
Best regards,
on behalf of the authors of Paper 22163
[Response to Weaknesses 1–3] Clarifying the Methodological Relationship Between LA-CPD and WCP
Response: We thank the reviewer for the careful analysis of the relationship between LA-CPD and Li et al. (2026). We hope the following responses help alleviate some of the reviewer's concerns. We clarify the established components and our methodological extensions below.
Li et al. use weighted CUSUM with NOT in the main paper; their public implementation also includes weighted segment costs, DP, and AIC-style cut-count selection. We adopt these established components and do not claim them as novel.
Our work focus on three aspects:
Response: Our extension lies in how this existing statistic is used and integrated into the global segmentation objective.
Construction and use of local evidence. To incorporate local transition evidence directly into the comparison of complete partitions, we compute E(c;h) at every candidate input-unit boundary c using a fixed window rule before dynamic programming. The evidence at each position therefore remains unchanged across candidate partitions and enters the global objective as an additive reward for selected cuts, allowing local boundary support and within-segment fitting to jointly guide partition selection.
Residual interpretation of the local reward. Appendix A.3, Eq. (20), shows that E(c;h) exactly equals the reduction in weighted residual obtained by replacing a single shared mean with separate means for the two sides of the local window. This gives the local reward a residual-based interpretation.
Response: Weighted least-squares segment costs and DP are established foundations that we adopt.
Joint objective. To make local boundary evidence directly influence cut placement, we incorporate the precomputed E(c;h) as an additive reward in the global partitioning objective: J(B)=β∑j=0kV(bj,bj+1)−α∑c∈BE(c;h), where the first term measures within-segment fit across the complete partition, and the second sums the local transition evidence at the selected cuts. Thus, even for a fixed number of cuts, the local evidence can change the optimal boundary configuration selected by the joint objective.
Exact DP optimization. Because E(c;h) is precomputed before DP and the reward is additive over cuts, fixing the location of the final cut determines both its reward and the cost of the final segment, leaving an optimal-prefix subproblem. Therefore, for each feasible fixed k, DP exactly minimizes the joint objective over the specified candidate positions, subject to the minimum segment-length constraint, without requiring additional state dimensions. Appendix A.4, Eq. (24), provides the recurrence and optimality argument.
Response: We use an established AIC-style criterion to control partition complexity; our extension concerns the candidate partitions it evaluates.
Count selection. For each feasible cut count k, LA-CPD first obtains a candidate partition Bk through the joint objective, then recomputes its original weighted residual and combines it with a complexity penalty to select the final count. The local reward does not enter the AIC-style fit term directly, but influences count selection indirectly through candidate boundary locations and their residuals.
Empirical gains. Under unknown boundary counts, with shared detection scores, length weights, and subsequent labeling rules, the two complete configurations achieve the following results: | Method | ACC ↑ | WD ↓ | AI-F1 ↑ | |:-|-:|-:|-:| | WCP+AIC | 0.7472 | 0.2698 | 0.5519 | | LA-CPD+AIC | 0.7960 | 0.2305 | 0.6001 |
LA-CPD+AIC improves all three metrics, supporting the effectiveness of the complete configuration. We will clarify the attribution of established components in the revision and highlight the fixed-window boundary reward, joint partitioning objective, and exact optimization property.