acceptodds
Under review as a conference paper at ICLR 2027

LA-CPD: Local-Evidence-Aware Change-Point Detection for Human–LLM Authorship Segmentation

17 views

Abstract

As LLM-generated text becomes increasingly human-like, failing to localize LLM-authored spans in human–LLM co-authored documents may weaken our ability to establish accountability for copyright infringement, fraud, and other harmful uses of AI-generated content. Existing sentence-level detection scores provide local evidence for these transitions, but content differences can cause score fluctuations even among sentences from the same source, potentially intro- ducing spurious boundaries. Identifying genuine authorship transitions and delin- eating the corresponding document partition therefore remain challenging when both boundary locations and counts are unknown. We propose **L**ocal-**E**vidence- **A**ware **C**hange-**P**oint **D**etection (**LA-CPD**), a structured method that converts noisy score sequences into coherent authorship segments. Given scores from a frozen local detector, LA-CPD combines a length-weighted within-segment resid- ual with a windowed two-mean contrast to capture both segment consistency and sustained changes across candidate cuts. Dynamic programming optimizes cut locations for each candidate count, while an AIC-style criterion selects the fi- nal partition for delineating sentence labels, authorship boundaries, and maximal LLM-authored spans. Experiments on a held-out human–LLM co-authored test set show that LA-CPD outperforms WCP+AIC, improving sentence-level accu- racy from 0.747 to 0.796 while achieving better boundary localization and LLM- span delineation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
llm-generated text detection, change-point detection, ai-generated span localization, dynamic programming

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 3 reviews, above 85% of the others.
Reviewer_7H8hOpenReview ↗
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:

  1. 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.

  2. 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 hh, 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 hh.

  • 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
AuthorsOpenReview ↗

[Response to Weaknesses 1–3] Clarifying the Methodological Relationship Between LA-CPD and WCP

Summary of Weaknesses 1–3: Please clarify which components of LA-CPD were established in Li et al. (2026) and what methodological contributions this paper adds.

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:

  1. We compute local transition evidence within a fixed window at each candidate input-unit boundary and precompute it as an additive boundary reward.
  2. We incorporate this reward into the global partitioning objective, allowing it to change the optimal boundary configuration for a fixed number of cuts while retaining exact DP optimization.
  3. We apply this joint objective to sequences of complete sentences and source-consistent sentence fragments, supporting the localization of both inter-sentence and intra-sentence authorship transitions in human–LLM coauthored text.

Summary of Weakness 1: How does the fixed-window evidence relate to WCP’s weighted CUSUM?

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)E(c;h) at every candidate input-unit boundary cc 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)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.

Summary of Weakness 2: How does the boundary reward change weighted DP?

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)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),J(B)=\beta\sum_{j=0}^{k}V(b_j,b_{j+1})-\alpha\sum_{c\in B}E(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)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 kk, 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.

Summary of Weakness 3: The role of the AIC-style criterion in selecting an unknown number of cuts

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 kk, LA-CPD first obtains a candidate partition BkB_k 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.

AuthorsOpenReview ↗

[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.

MethodDecoherence ACC ↑WD ↓AI-F1 ↑Paraphrase ACC ↑WD ↓AI-F1 ↑
SenPred0.45130.93280.02310.53150.91010.0530
Voting0.47870.73920.09080.60230.68040.2042
TextTiling0.48260.51100.16940.62310.48130.3583
PaLD-scores0.59260.87740.00280.59180.93800.0327
VCP0.61330.29080.06370.64390.29340.2327
WCP0.63190.29200.13560.66150.28660.2424
WCP+AIC0.60850.32790.25220.66450.32540.3620
LA-CPD+AIC0.64910.29530.25020.69490.28450.3520

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 granularityConditionACC ↑WD ↓AI-F1 ↑
Sentence levelB20.81750.20020.8130
Token levelB20.71530.40750.6308
Sentence levelB3-20.71440.33960.3381
Token levelB3-20.51250.41620.2371
Sentence levelB3-30.75030.30900.5619
Token levelB3-30.70720.38370.4762

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.

MethodACC ↑WD ↓
SenPred0.52340.5554
Voting0.51710.4869
TextTiling0.50710.4506
PaLD-scores0.65990.4989
VCP0.69180.3955
WCP0.68690.3967
WCP+AIC0.65980.4006
LA-CPD+AIC0.70060.3951

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.


AuthorsOpenReview ↗

[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 E(c;h)E(c;h) 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 Δ\Delta at boundary τ\tau. Assume unit weights, independent zero-mean noise with variance σ2\sigma^2, and full windows of length hh containing no other transitions. For an integer offset d satisfying ∣d∣<h|d|<h,

E[E(τ+d;h)]=σ2+h2(1−∣d∣h)2Δ2.\mathbb E[E(\tau+d;h)] =\sigma^2+\frac h2 \left(1-\frac{|d|}{h}\right)^2\Delta^2.

Thus, the expected evidence peaks at the true boundary and decreases with displacement. The noise variance of the window-mean contrast is 2σ2/h2\sigma^2/h, explaining how aggregation helps distinguish a sustained change from local fluctuations.

(2) The reward strengthens the preference for supported boundaries. Our objective,

J(B)=βR(B)−α∑c∈BE(c;h),J(B)=\beta R(B)-\alpha\sum_{c\in B}E(c;h),

where R(B)R(B) 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 τ\tau versus τ+d\tau+d:

E[J(Bd)−J(Bτ)]=β E[R(Bd)−R(Bτ)]+α(∣d∣−d22h)Δ2.\mathbb E[J(B_d)-J(B_\tau)] = \beta\,\mathbb E[R(B_d)-R(B_\tau)] +\alpha\left(|d|-\frac{d^2}{2h}\right)\Delta^2.

For α>0\alpha>0 and 0<∣d∣<h0<|d|<h, 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.

AuthorsOpenReview ↗

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

Reviewer_WTLcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Authors propos LA-CPD which is a two stage CPD method for sentence level human-LLM authorship segmentation when both boundary locations and counts are unknown.

Given sentence scores from a frozen LoRA adapted local detector, the method evaluates candidate partitions by combining a length weighted within-segment residual with a fixed-window mean contrast around candidate cuts. DP algorithm finds the optimal cut locations for each candidate cut count, and then an AIC style criterion selects the final number of cuts before a two-group clustering step assigns Human or LLM labels and merges adjacent same-label segments.

Experiments on a GPT-5.5 co-authored dataset and five transfer benchmarks show that LA-CPD improves overall accuracy, WindowDiff, and AI-span F1 over existing change-point baselines on the main test set.

Critical strengths
  • The partition objective cleanly combines fixed window transition evidence with length weighted segment residuals in a dynamic program that optimizes cut placement before selecting the cut count.
  • Freezing the local detector and sharing the exact same sentence-score trajectories across all baselines cleanly isolates the contribution of the structured segmentation stage.
  • The evaluation reports multiple complementary metrics, uses an oracle-K baseline to separate cut placement from count selection, and shows consistent gains on the primary GPT-5.5 benchmark.
Critical weaknesses
  • Line 277 enforces that the minimum segment length L >= h only when alpha > 0. So, setting alpha = 0 in the Table 2 ablation also removes the minimum length constraint? Does it confound the ablation and window-size sensitivity while L itself is omitted from Section 5?
  • DP optimizes the combined objective with windowed rewards in equation 5, whereas the AIC selector in equation 7 evaluates only the unpenalized residual, creating a mismatch between cut placement and count selection. How will this affect?
  • The AIC-style criterion in is heuristic rather than derived from a valid weighted likelihood model, and its parameter-count convention and penalty multiplier r = 2.5 lack statistical justification.
  • The deterministic labeler (lines 311-315) labels the entire document as Human whenever zero cuts are selected (making all-LLM detection impossible and giving oracle-K a trivial 1 accuracy on H0), while forcing at least one LLM segment whenever one or more cuts are selected.
  • Weighting is inconsistent across stage: Segmentation uses uncapped squared sentence lengths without inverse variance estimation, whereas label assignment reverts to unweighted sentence means and weights unique segment scores equally regardless of segment length.
  • Capping the search at K_max = 3 while requiring at least 6 sentences per segment restricts performance on shorter or multi-boundary texts, where even oracle-K drops from 0.945 AI-F1 on B1 to 0.456 on B3-2.
  • The experiments omit several relevant baselines, including standard penalized dynamic programming (PELT, FPOP), robust weighted change-point detection (RWCP), adaptive smoothing, and recent span detectors such as GigaCheck and SenDetEX. Could you please add these baselines?
  • Cross-distribution transfer in Figure 3 is mixed, with LA-CPD trailing simpler baselines like Voting and TextTiling on several datasets, and the main text reports only sentence accuracy for transfer without WindowDiff or AI-F1.
  • The main paper omits boundary count error metrics, boundary F1, and runtime scaling analyses, and uses experimental conditions (B1 to B3-3) and the MAD penalty in Tables 1 and 2 without defining them clearly in the main text. I found these very confusing and challenging while reading the draft.
Questions for authors

See strengths and weaknesses.

8 comments
AuthorsOpenReview ↗

[Response to Weaknesses 1 and 2] Segment-Length Coupling and DP–AIC Objective Mismatch

Summary of Weakness 1: Does the local-evidence ablation and window-size sensitivity analysis also change the segment-length constraint?

Response: Thank you for pointing out the interaction between these parameters. We apologize that the original manuscript did not explain this clearly. In the current implementation, removing the local evidence returns the effective minimum segment length to its base value; when local evidence is enabled, changing the window size may also change the effective minimum segment length.

To align the observation range of the local evidence with the partition structure, we use a window-dependent minimum segment length. The local evidence E(c;h)E(c;h) aggregates scores from the hh input units on each side of a candidate cut point. When each segment is at least hh units long, the windows on either side of the cut point can remain within the adjacent statistical segments. This constraint is a design choice for matching scales; it also reduces the set of feasible partitions and may limit the segmentation of short spans. The effective minimum segment length is defined as

L_{\mathrm{eff}}(h,\alpha)= \left\\{ \begin{array}{ll} L_{\min}, & \alpha=0,\\\\ \max(L_{\min},h), & \alpha>0. \end{array} \right. \\$$ Relevant results from Table 2: the full configuration and the configuration without local evidence. | Configuration | ACC ↑ | WD ↓ | AI-F1 ↑ | |-|-:|-:|-:| | LA-CPD (Full) | 0.796 | 0.231 | 0.600 | | LA-CPD w/o $E(c;h)$ | 0.715 | 0.279 | 0.519 | **The full configuration outperforms the version without local evidence on all reported metrics.** This suggests that combining neighborhood transition evidence with its associated segment-length constraint helps produce more coherent partitions and recover complex authorship transitions. Therefore, Table 2 reflects the joint effect of removing local evidence and its associated segment-length constraint. We will clarify this rule in the revised manuscript and appendix, and revise the title and discussion of Table 2 to specify the meaning and scope of each comparison. > **Summary of Weakness 2: How does using different objectives for DP and cut-count selection affect the results?** **Response:** Thank you for pointing out the difference between the criteria used for location optimization and cut-count selection. We apologize for the insufficient explanation in the original manuscript. Since DP generates candidate partitions using a joint objective that includes local rewards, while the AIC-style criterion compares their residuals and complexity, **local rewards may affect the final cut count by changing candidate locations and their residuals.** This division of roles is intended to guide cut-point localization using local transition evidence, then select the cut count based on overall fit improvement and a complexity penalty. **(1) The design combines two levels of assessment.** For each fixed cut count $k$, DP optimizes the following objective over partitions satisfying the segment-length constraint: $$J(B)=\beta R(B)-\alpha\sum_{c\in B}E(c;h).$$ Let $\widehat B_k$ denote the resulting candidate partition. We then recompute its raw weighted residual, $R_k=R(\widehat B_k)$, and evaluate the AIC-style criterion: $$IC_r(k)=n\log\left(\max\left\\{\frac{R_k}{n},10^{-12}\right\\}\right) +4r(k+1).$$ Local transition evidence guides cut-point locations, while more complex candidate partitions must improve the overall fit sufficiently to offset the complexity penalty. **The local reward does not enter the count-selection criterion directly;** instead, it acts through the following path: $$E(c;h)\longrightarrow\widehat B_k\longrightarrow R_k\longrightarrow IC_r(k).$$ **(2) Effects on count selection. The joint objective may favor a partition with a slightly higher residual in exchange for stronger local evidence at the same cut count.** Thus, even when a candidate with more cut points has strong local support, the AIC-style criterion may favor fewer cut points if the residual improvement does not offset the complexity penalty. Because candidates with different cut counts may be affected differently, the final cut count may increase, decrease, or remain unchanged. **(3) Empirical performance of the complete procedure.** Under the unknown-cut-count setting: |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 classification, boundary consistency, and AI-span recovery, providing empirical support for the complete two-stage configuration. We will clarify these roles and their implications in the revision, explicitly stating that **DP’s exact optimization guarantee applies to the joint objective at each fixed cut count,** while the AIC-style criterion selects the count among the resulting candidate partitions.
AuthorsOpenReview ↗

[Response to Weaknesses 3 and 4] Clarifying the AIC-Style Criterion and Deterministic Labeler

Summary of Weakness 3: Statistical Justification for the AIC-Style Criterion and Its Complexity Penalty

Response: Thank you for requesting clarification of our criterion’s statistical basis. Our complete cut-count selector is an AIC-style heuristic, rather than a direct derivation from standard AIC theory. Previous studies have used AIC and MDL for unknown model-order selection; for example, Izacard et al. (2019) used them as baselines for estimating the number of sinusoidal frequencies. This illustrates the applicability of information criteria to unknown-count selection.

(1) A working-likelihood interpretation of the fit term. For a fixed candidate partition BB and fixed positive weights wiw_i, assume that the scores are conditionally independent and satisfy

qi∣B,wi∼N(μs(i),σ2wi).q_i\mid B,w_i\sim\mathcal{N}\left(\mu_{s(i)},\frac{\sigma^2}{w_i}\right).

Maximizing over segment means and the common variance yields the partition-dependent term nlog⁡(R(B)/n)n\log(R(B)/n) in the profiled negative twice log-likelihood; weight-dependent terms remain constant across partitions of the same document. This interpretation relies on a working assumption, not a verified distributional property of detector scores.

(2) The complete criterion remains heuristic. For each fixed cut count, the DP generates a candidate partition using a joint objective comprising within-segment residuals and local boundary rewards. The AIC-style criterion then evaluates that partition’s residual and penalizes complexity. Because the partition selected by the DP need not minimize the residual for that cut count, the complete selector is not standard AIC derived from the maximum-likelihood solution for each cut count.

(3) Parameter counting and the penalty coefficient. pk=2(k+1)p_k=2(k+1) is a nominal complexity convention, accounting for k+1k+1 segment means, kk discrete boundary locations, and one common variance parameter. It is not a standard derivation of the effective degrees of freedom of a change-point model. The value r=2.5r=2.5 is not a theoretically derived constant, but an empirical setting selected on the validation set by prioritizing ACC, followed by AI-F1.

The paper reports the following sensitivity results:

rrACCAI-F1WD
1.00.7750.6140.321
2.00.7950.6070.267
2.50.7960.6000.231
3.00.7860.5440.244

The metrics vary non-monotonically with rr: among the tested values, r=2.5r=2.5 achieves the best ACC and WD, but not the best AI-F1. This reflects a trade-off among classification, boundary localization, and segment recovery. We therefore regard r=2.5r=2.5 as an empirical choice under the current validation protocol.

We will clarify the working assumptions, nominal parameter count, validation procedure, and the criterion’s heuristic nature and theoretical scope in the revision.


Reference:

[1] Shaw et al. Bridging Kolmogorov Complexity and Deep Learning: Asymptotically Optimal Description Length Objectives for Transformers. ICLR 2026.

[2] Izacard et al. Data-driven Estimation of Sinusoid Frequencies. NeurIPS 2019.


Summary of Weakness 4: How does the deterministic labeler constrain source predictions and affect H0/oracle-K results?

Response: Thank you for pointing out how the deterministic labeling rule affects the interpretation of our predictions. We added an H0 false-positive analysis.

With zero cuts, the labeler assigns the entire document to Human and therefore cannot identify fully LLM-authored documents. When one or more cuts are selected and the segment means are distinct, the two-cluster procedure assigns segments to both Human and LLM. This design is intended for the human–LLM co-authored setting studied in this paper, but a predicted cut is not necessarily a true authorship boundary; spurious cuts in H0 documents may therefore lead to false LLM predictions. The oracle-K H0 accuracy of 1.0 follows from the known K=0 and labeling rule; it is not independent evidence of source-identification ability. We measured document-level and input-unit-level false-positive rates on the H0 test documents. The input-unit-level rate is computed over all H0 input units.

MethodDocument-level false-positive rate ↓Input-unit-level false-positive rate ↓
TextTiling98.0%48.37%
SenPred100.0%61.27%
PaLD-scores100.0%22.89%
VCP15.0%9.68%
WCP17.5%15.99%
WCP+AIC32.0%31.76%
LA-CPD+AIC14.0%16.84%

The results show that LA-CPD+AIC has the lowest document-level false-positive rate, indicating that false alarms affect fewer H0 documents. However, VCP has a lower input-unit-level false-positive rate. Thus, we conclude that LA-CPD+AIC performs well in limiting the incidence of document-level false alarms.

We will add this false-positive analysis to the revision and clarify the zero-cut rule, the interpretation of the oracle-K result, and the scope of our method.

AuthorsOpenReview ↗

[Response to Weaknesses 5 and 6] Clarifying Weighting and Segmentation Constraints

Summary of Weakness 5: Why are different weighting schemes used for segmentation and label assignment?

Response: Thank you for pointing out the difference in weighting between segmentation and label assignment.. This design allows length information to influence the segmentation structure, while source labels are determined by relative differences in segment scores.

(1) Roles of weighting in the two stages.

  • Segmentation. We use normalized weights wi=nℓi2∑jℓj2,w_i=\frac{n\ell_i^2}{\sum_j\ell_j^2}, where ℓi\ell_i is the token count of input unit ii. These weights enter both the within-segment residuals and local transition evidence, giving longer units a greater relative contribution to candidate partitions. They are predefined length-importance weights, not estimated inverse variances. Normalization fixes the total weight, but no additional clipping is applied.

  • Label assignment. Labeling computes ordinary segment means and clusters their distinct values with equal weights, regardless of segment length or repeated means. The complete procedure is therefore an empirical design, not a unified weighted-likelihood estimator.

(2) Experimental analysis.

(a) Length weighting in segmentation.

Segmentation weightsACC ↑WD ↓AI-F1 ↑
Uniform weights0.7810.2230.516
Squared-length weights0.7960.2310.600

Squared-length weighting improves ACC and AI-F1, while uniform weighting yields lower WD. This supports the empirical role of length information in the current segmentation procedure and reflects a trade-off between boundary consistency and span recovery.

(b) Weighting configurations in label assignment.

ConfigurationACC ↑WD ↓AI-F1 ↑
Current rule0.7680.2670.613
Adjust means only0.7580.2740.600
Adjust clustering weights only0.7760.2620.622
Adjust both0.7730.2650.614

Ordinary segment means with weighted clustering perform best in this comparison. Weighting both steps offers no further improvement, suggesting that score aggregation and threshold determination should be assessed separately.

We will clarify rules and report the comparisons in the revision.

Summary of Weakness 6: Do the search limit and minimum segment length restrict short-text and multi-boundary detection?

Response: Thank you for pointing out the potential limitations of the search configuration for short-span and multi-boundary detection. We acknowledge that the cut-count cap and minimum segment length restrict the searchable partitions. To examine their effects, we provide additional span-length statistics, segment-length comparisons, and labeling diagnostics using ground-truth boundaries.

(1) Scope of the configuration. The main experiments contain zero to three true boundaries. B3-2 has two and therefore falls within Kmax⁡=3K_{\max}=3. Compared with B1, B3-2 has shorter median AI spans and a higher proportion of documents containing true source segments shorter than six input units. Thus, the default segment-length constraint can exclude some reference partitions. Tasks with more boundaries require a larger search range.

(2) Relaxing the segment-length constraint did not improve recovery. Under oracle-K, we held all other settings fixed and compared two effective minimum segment lengths:

Condition and subgroupL=6L=6: AI-F1 ↑L=1L=1: AI-F1 ↑
B1: All documents0.9450.885
B3-2: All documents0.4560.324
B3-2: All true source segments ≥60.4720.332
B3-2: At least one true source segment <60.3180.255

Relaxing the constraint did not improve results in either condition or in the B3-2 short-segment subset. Therefore, expanding the feasible partition set alone is insufficient to improve detection, and the performance drop on B3-2 cannot be attributed simply to the segment-length constraint.

(3) Knowing the count still requires predicting locations and labels. B3-2 requires recovering both endpoints of an internal AI span, whereas oracle-K supplies only the count. We further supplied ground-truth boundaries while retaining the original labeler:

ConditionConfigurationACC ↑WD ↓AI-F1 ↑
B1Oracle-K0.9100.1690.945
B1Ground-truth boundaries + original labeler0.9950.0000.995
B3-2Oracle-K0.7140.2950.456
B3-2Ground-truth boundaries + original labeler0.8880.0520.778

The substantial improvement with ground-truth locations indicates that boundary localization is an important factor. The remaining errors on B3-2 indicate that source-label assignment is also a challenge. These diagnostics support interpreting the oracle-K gap in terms of both location recovery and label assignment, rather than attributing it solely to the search configuration.

We will include these diagnostics in the revised manuscript and clarify the scope and limitations of the default configuration.

Reviewer_wmaPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Local-Evidence-Aware Change-Point Detection (LA-CPD) for sentence-level segmentation of human–LLM co-authored texts when the locations of authorship transitions are unknown, and also how many are unknown. A LoRA local detector with Gemma-3-1B backbone gives sentence-level authorship scores. LA-CPD then combines a sentence-length-weighted within-segment residual with a windowed contrast measuring differences between neighboring score regions. DP identifies the minimum-cost partition for each candidate cut count, and an AIC-style criterion selects among these partitions. Segment-level mean-score clustering assigns Human/LLM labels and produces contiguous LLM-authored spans. The test for trained model is a balanced held-out GPT-5.5-generated texts, LA-CPD improves sentence accuracy, WindowDiff, and AI-fragment F1. Ablations shows that the windowed contrast contributes to the improvements, and the additional cross-model and cross-datasets on [HybridEssays, RoFT, SQuAD, Writing and XSum] experiments gives variable generalization performance.

Critical strengths
  1. Framework and comprehensive evaluation: The proposed approach thoughtfully combines within-segment consistency, local transition evidence, dynamic programming for finding optimal boundaries, and AIC-style selection for determining the number of boundaries, the combination is relatively new and adaptive for this segmentation task. The paper also provides ablation for the components to know the importance or contribution of each [Tables 2 and 9], and not just one dataset as test, the paper has done the evaluations across different datasets and generation models [Section 5.2 and Appendix B.5].

  2. The method shows improvements over WCP+AIC on the main test set, increasing sentence-level accuracy and AI-fragment F1, and also reducing WD [Table 1], in which it is balancing every aspect of the evaluation. Importantly, the score-based baselines use the same frozen detector and sentence-level scores, making the comparisons more controlled and so that the controlled experiments show the benefits of the proposed approach and the proposed do have the meaningful improvements over the baselines in several aspects like human-only documents, group by source family across train/validation/test splits, and external datasets.

Critical weaknesses
  1. Unknown boundary-count selection/prediction: The paper also addresses a scenario where the no. of boundaries are not revealed, the prediction of boundary count case, the AI-F1 is 0.040 and in the oracle case, revaling the count, it has a huge improvement to 0.914 [Appendix B.5], this shows that the unknown-count is responsible for a large portion of the observed error. Also, zero selected cuts produce all-human labels deterministically and the labeling rule cannot represent fully LLM-authored documents [From appendix A.5]. These issues undermine confidence in the scope and reliability of the inference mechanism, which is the utmost realistic inference testing.

  2. Data: One explicit mention about the data is required, which is that, the paper talks about reference boundaries between sentences, B2 and some B3 contain authorship transitions inside sentences [Appendix B.1], the paper does not have a clear explanation of how such sentences receive a unique binary reference label.

Questions for authors
  1. What are the distributions of predicted statistical cut counts under the unknown-count setting, especially the proportion of mixed documents assigned zero cuts? Could the authors provide these statistics together with document-length-stratified results and cut-count errors?

  2. How are character-level authorship spans converted into binary sentence-level labels? How are these cases reflected in the reference boundary set, WindowDiff, and fragment IoU evaluation?

7 comments
AuthorsOpenReview ↗

[Response to Weakness 1 and Question 1] Clarifying unknown-count inference and the zero-cut labeling rule

Summary of Weakness 1 and Question 1: Inference reliability with unknown boundary counts and the zero-cut labeling rule

Response: Thank you for noting that unknown-count inference and the zero-cut rule were insufficiently explained. We analyze cut-count distributions, length effects, and short-text adaptation to examine how segment-length constraints and clarify the oracle-K and zero-cut results.

(1) Distribution of statistical cut counts in the unknown-count setting. Following the reviewer’s helpful suggestion, we compare each method’s predicted number of statistical cut points, k^\hat{k}, with the true number of authorship boundaries, K⋆K^\star. For each document, we compute the absolute count error, AbsCE=∣K⋆−k^∣\mathrm{AbsCE}=|K^\star-\hat{k}|, and determine whether the predicted count matches the true count exactly. We then report the cut-count distribution, mean AbsCE, and the proportion of documents with an exact count match.

Method0 cuts1 cut2 cuts3 cutsAbsCE ↓Count accuracy ↑
Ground truth–50.0%25.0%25.0%––
TextTiling0.0%0.0%0.4%12.1%3.86130.0%
VCP51.4%34.8%8.2%5.6%1.231229.2%
WCP41.4%38.6%12.9%7.1%1.105032.8%
WCP+AIC15.4%29.2%21.9%33.5%0.870039.5%
LA-CPD+AIC27.1%41.6%16.8%14.5%0.878843.2%

The results show that LA-CPD+AIC achieves the highest exact cut-count accuracy and a near-best mean absolute cut-count error. This is because it combines local boundary evidence with global within-segment fit to optimize cut-point locations, then uses an AIC-style criterion to compare the resulting partitions. This enables it to select among candidate partitions of varying complexity and identify the true number of boundaries to some extent.

(2) Analysis stratified by text length. To further investigate the low AI-F1 score, we stratify mixed-authorship documents by the number of input units.

Input unitsZero-cut rateACC ↑WD ↓AI-F1 ↑AbsCE ↓
12–2361.3%0.69800.23750.55780.6321
≥2421.9%0.77900.27120.61870.9164

The minimum segment-length constraint leaves fewer feasible partitions for short documents, making them more likely to be assigned zero cuts. Given that shorter documents are more likely to be assigned zero cut points, we further examine whether partition constraints and count penalties affect cut-point selection in the short-text setting.

(3) Short-text configuration adjustment and cross-dataset validation. To examine SQuAD’s low AI-F1, we first relax the minimum segment length and then reduce the AIC-style penalty under unknown-count inference; we test the adjusted configuration on short RoFT documents.

(a) Short-text configuration analysis on SQuAD.

ConfigurationL/rL/rACC↑AI-F1↑
Original6/2.50.55200.0400
Relaxed segment length1/2.50.64560.3710
Further reduced penalty1/1.00.77930.6848

These results indicate that the poor performance mainly stems from the combined effects of the constrained candidate space for short texts and the count penalty, rather than from an inability of the local evidence to identify boundaries. This sensitivity analysis is specific to the short-text setting and does not imply that the same configuration is optimal for all datasets. To test whether the improvement with the short-text configuration is limited to SQuAD, we apply it to short texts from the public RoFT dataset.

(b) Evaluation on the RoFT short-text subset. We compare the adapted configuration with WCP+AIC on RoFT documents containing 10 input units each.

MethodACC↑WD↓AI-F1↑
WCP+AIC0.59880.25470.2996
LA-CPD+AIC (short-text configuration)0.61310.23590.3529

This indicates that short-text adaptation is effective beyond SQuAD and transfers to other short-text data. Together with the configuration analysis on SQuAD, these results support separately calibrating partition constraints and count penalties for short texts.

(4) Scope of the zero-cut rule. Zero cuts indicate that no authorship transition was detected. Under the current labeling rule, the single segment is deterministically labeled Human, so the method cannot represent documents written entirely by an LLM. Our study focuses on detecting authorship transitions in human–LLM co-authored documents when the number of boundaries is unknown and it does not cover the detection of fully LLM-authored documents.

We will incorporate these experiments and analyses into the revised manuscript and clarify that our study addresses authorship-transition detection in human–LLM co-authored text.


References:

[1] Dugan et al. RoFT: A Tool for Evaluating Human Detection of Machine-Generated Text. EMNLP 2020 System Demonstrations.


AuthorsOpenReview ↗

[Response to Weakness 2 and Question 2] Constructing input units and evaluating intra-sentence authorship transitions

Summary of Weakness 2 and Question 2: How are intra-sentence authorship spans converted into uniquely labeled input units and reflected in reference boundaries, WindowDiff, and fragment-IoU evaluation?

Response: Thank you for pointing out this ambiguity. To clarify how intra-sentence authorship transitions are represented and evaluated, we explain how input units and labels are constructed, how reference boundaries and WindowDiff are defined, and how AI-fragment IoU is computed.

(1) Input-unit construction and label assignment. Preprocessing preserves natural sentence boundaries. For sentences containing an intra-sentence authorship transition, we introduce additional splits at the character offsets marking the start and end of LLM-authored text recorded during data construction. Each resulting input unit is labeled Human HH or LLM LL according to its source.

For example, the two input units below belong to the same natural sentence but receive different labels:

Input unitTextReference label
15“You know that I'm pressed for time”H
16“in the mornings, ...”L

Thus, an input unit may be a complete sentence or a sentence fragment, and an intra-sentence transition is represented as a source change between adjacent input units.

(2) Reference boundaries and WindowDiff. Let the input sequence be (u1,…,un)(u_1,\ldots,u_n), with reference labels (y1,…,yn)(y_1,\ldots,y_n). We define reference authorship boundaries as positions where the labels of adjacent input units differ:

C⋆={i∈{1,…,n−1}:yi≠yi+1}.\mathcal{C}^{\star} = \{i\in\{1,\ldots,n-1\}:y_i\ne y_{i+1}\}.

Here, ii denotes the position between the ii-th and (i+1)(i+1)-th input units. For example, the label sequence (H,L,L,H)(H,L,L,H) gives the reference boundary set {1,3}\{1,3\}. In the example above, the H→L transition between units 15 and 16 corresponds to boundary position 15. A boundary between input units is an authorship boundary only when their source labels differ.

Given predicted labels (y^1,…,y^n)(\hat y_1,\ldots,\hat y_n), we define the predicted final authorship boundaries analogously:

C^={i∈{1,…,n−1}:y^i≠y^i+1}.\hat{\mathcal{C}} = \{i\in\{1,\ldots,n-1\}:\hat y_i\ne\hat y_{i+1}\}.

WindowDiff measures the proportion of sliding windows in which the reference and predicted boundary counts differ. The window width is

\qquad K^\star = \left|\mathcal{C}^{\star}\right|.$$ This evaluation uses final authorship boundaries: a statistical cut between adjacent segments is excluded if the segments receive the same label. An intra-sentence transition is evaluated as a boundary between the resulting input units. **(3) AI fragments and IoU. Reference and predicted AI fragments are defined as maximal contiguous runs of input units labeled L.** Let $P$ and $G$ denote the sets of input units in a predicted fragment and a reference fragment, respectively. Their IoU is $$\mathrm{IoU}(P,G)=\frac{|P\cap G|}{|P\cup G|}.$$ **IoU is computed over input units, with each complete sentence or sentence fragment counted as one unit.** Predicted fragments are ranked by the mean detection score of their constituent input units, from highest to lowest, and greedily matched one-to-one with the unmatched reference fragment having the highest IoU. A match is accepted only if $\mathrm{IoU}>0.5$. We then aggregate TP, FP, and FN across documents to compute micro-averaged AI-F1. Reference labels, boundaries, WindowDiff, and fragment IoU all use the same preprocessed input-unit sequence. Character-level source annotations are used to construct these units; accordingly, our evaluation concerns cut and source-label predictions over the preprocessed units, rather than direct localization of arbitrary character-level boundaries in unsplit text. We will clarify the input-unit construction and evaluation scope in the task definition and add the reference-boundary and evaluation protocols to Appendices B.1 and B.3.
AuthorsOpenReview ↗

Response for Reviewer wmaP

Dear Reviewer wmaP

Thank you again for your thoughtful feedback of our work. We have added analyses of boundary-count prediction and short texts, and clarified the input units and labeling limitations. When you have a chance, we would be grateful to hear whether these responses address your concerns and whether anything needs further clarification.

Best regards,

on behalf of the authors of Paper 22163

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers, ACs, SACs, and PCs

We sincerely thank all reviewers as well as ACs, SACs, and PCs for their careful handling of our submission. We appreciate the reviewers’ positive feedback on research motivation and practical relevance (Reviewers 7H8h), model design (Reviewers wmaP, WTLc, and 7H8h), and extensive experiments and generalization evaluation (Reviewers wmaP and WTLc). Based on their constructive suggestions, we provide point-by-point responses:

(1) Methodological Contributions and Comparison with WCP (Reviewers WTLc and 7H8h)

  • Overall Differences in Research Motivation. (Reviewers WTLc and 7H8h) WCP and our proposed LA-CPD both study authorship localization in human–LLM co-authored text, but we are the first to solve the problem of intra-sentence AI-authored span localization in this field. Considering that authorship transitions can occur both between and within sentences, our method further focuses on distinguishing score fluctuations within segments from the same source from genuine authorship transitions. It combines local evidence around boundaries with within-segment consistency to guide document partitioning and explore finer-grained intra-sentence boundary localization.

  • Methodological Comparison with WCP. (Reviewer 7H8h) To allow evidence of changes around boundaries to directly inform partition selection, LA-CPD aggregates neighborhood information within a fixed window at each candidate position and incorporates the precomputed local evidence into the joint objective as an additive boundary reward. Compared with the WCP dynamic programming baseline, which optimizes cut locations based on weighted within-segment residuals, this design allows within-segment consistency and local boundary support to jointly determine the complete partition. It can change the optimal boundary configuration for a fixed number of cuts while retaining exact dynamic programming optimization.

(2) Additional Comparative, Generalization, and Robustness Experiments (Reviewers WTLc and 7H8h)

  • Expanded Baseline Coverage. (Reviewer WTLc) We added five baselines: PELT, FPOP, RWCP, Adaptive Smoothing, and GigaCheck. LA-CPD outperforms these additional baselines on all three metrics in the main test setting. These results support combining local boundary evidence, within-segment fit, and complexity control.

  • Evaluation on Real Collaborative Writing. (Reviewer 7H8h) On CoAuthor, LA-CPD achieves the highest accuracy and lowest WD among the compared methods, supporting its potential to handle irregular authorship transitions in real collaborative writing.

  • Performance under Text Perturbations. (Reviewer 7H8h) LA-CPD achieves the highest accuracy under adjacent-word swapping and paraphrasing, and the lowest WD under paraphrasing. These results are consistent with windowed evidence aggregation and global fitting reducing sensitivity to isolated score fluctuations.

(3) Clarification of Methodological Details (Reviewers 7H8h, WTLc, and wmaP)

  • Theoretical Interpretation of Local Evidence. (Reviewer 7H8h) The evidence equals the weighted residual reduction from fitting two window means instead of one and is equivalent to squared weighted CUSUM. Under equal weights, independent homoscedastic noise, and one true change point within the window, its expectation peaks at the true boundary. The additive reward preserves optimal substructure for fixed-count dynamic programming.
  • Input Units. (Reviewer wmaP) We use complete sentences and sentence fragments split at character offsets marking authorship transitions recorded during data construction, and evaluate labels, boundaries, and fragments over these preprocessed units. We distinguish statistical cuts from final authorship boundaries and separately provide experiments that search for token-level boundaries in complete original texts, together with an analysis of their limitations.
  • Implementation Rules for Weighting and Minimum Segment Length. (Reviewer WTLc) Segmentation uses normalized squared-length weights, while label assignment uses ordinary segment means and equally weighted clustering of distinct segment means. Removing local evidence also changes the associated segment-length constraint; therefore, the corresponding ablation reflects their joint effect.
  • Complexity and Runtime. (Reviewer WTLc) We provide additional complexity and runtime analyses excluding detector forward passes. With a fixed cut-count cap, segmentation costs grow approximately quadratically. We also explain the trade-off between the additional computational cost relative to WCP+AIC and localization quality.

We appreciate your time and careful consideration of our submission. We hope this summary assists you in navigating our rebuttal situation. The revised PDF containing all these improvements has been uploaded.

Best regards,

on behalf of the authors of Paper 22163

AuthorsOpenReview ↗

Response for Reviewers, ACs, and SACs

Dear Reviewers and Area Chair,

Thank you for your time and thoughtful feedback. We have responded to all reviewer comments and uploaded an updated manuscript incorporating the clarifications, additional analyses, and experiments discussed in our responses. Your suggestions have helped us present the contributions, evaluation, and limitations of LA-CPD more clearly.

When you have a chance, we would be grateful for your thoughts on our responses and the updated manuscript. We would be happy to clarify any remaining questions.

Best regards,

on behalf of the authors of Paper 22163

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.