PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Qing Yang, Zhenyu Mao, Zixiang Luo, Zezheng Wu, Xinghe Cheng, Qinggang Zhang, Jingwei Zhang, Jiapu Wang

arXiv:2609.33787v1cs.CL

TL;DR

Sentence-level scores can fluctuate within same-source text, complicating coherent segmentation when authorship boundaries and their count are unknown. LA-CPD combines segment-fit and local transition evidence to select a partition, and the paper reports improved boundary localization and LLM-span delineation, with detector and cross-distribution limits.

  • Problem

    Sentence scores can fluctuate among sentences from the same source, while authorship-boundary locations and counts are unknown, making coherent partition selection challenging.

    Source: §1 · pp. 1–2

  • Method

    LA-CPD combines a length-weighted within-segment residual and windowed two-mean contrast, then uses dynamic programming and an AIC-style criterion to select a partition and delineate labels and spans.

    Source: §1 · p. 2

  • Results

    LA-CPD is reported to improve boundary localization and LLM-span delineation, with best overall performance on the GPT-5.5 test set when boundary counts are unknown.

    Source: §5.1 · p. 7, §6 · p. 9

  • Takeaways & Limitations

    The findings identify persistent local transition evidence as a primary factor in boundary localization and length weighting as promoting contiguous AI-authored span delineation.

    Source: §6 · p. 9

    LA-CPD depends on the local detector’s discriminative ability, cross-distribution performance is less consistent under simultaneous shifts, and dynamic programming adds overhead for long documents.

    Source: §6 · p. 9, §5.2 · p. 9

Abstract

from arXiv · show

As LLM-generated text becomes increasingly human-like, accurately localizing LLM-authored spans in human-LLM co-authored documents is important for attribution and accountability in cases involving copyright infringement, fraud, and other harmful uses of AI-generated content. Sentence-level detectors provide local authorship evidence, but content variation can cause score fluctuations even among sentences from the same source, creating spurious boundaries. Recovering a coherent document partition therefore remains challenging when both the number and locations of authorship transitions are unknown. We propose Local-Evidence-Aware Change-Point Detection (LA-CPD), a structured method that transforms noisy sentence-level score sequences into coherent authorship segments. Given scores from a frozen local detector, LA-CPD combines a length-weighted within-segment residual with a windowed two-mean contrast to capture segment consistency and sustained changes around candidate cut points. Dynamic programming optimizes cut locations for each candidate count, while an AIC-style criterion selects the final partition, yielding sentence labels, authorship boundaries, and maximal LLM-authored spans. On a held-out human-LLM co-authored test set, LA-CPD outperforms WCP+AIC, increasing sentence-level accuracy from 0.747 to 0.796 while improving boundary localization and LLM-span delineation.

1 INTRODUCTION

Human–LLM contributions can alternate within a document, so document-level labels cannot show where LLM-authored spans begin and end. LA-CPD addresses noisy sentence scores and unknown transition counts with a structured partition objective and a two-stage inference procedure.

1 INTRODUCTION

  • Sentence-level segmentation is needed because human and LLM contributions may alternate, while a document-level label cannot locate or measure the LLM-generated content.The resulting segments delineate contiguous human- and LLM-authored spans and their transition boundaries.Source: §1 · p. 1
  • Sentence scores can fluctuate among sentences from the same source, causing score-based boundary methods to mistake fluctuations for authorship changes and fragment contiguous LLM spans.Short sentences may provide less stable evidence.Source: §1 · pp. 1–2
  • Because boundary locations and counts are unknown, a credible partition must balance internally consistent regions and persistent cross-boundary differences against model complexity.Optimizing within-segment fit alone can favor overly fine partitions, while adjacent-score contrast alone is sensitive to isolated deviations.Source: §1 · p. 2
  • LA-CPD combines a length-weighted within-segment residual with windowed two-mean contrast to assess segment consistency and sustained change across candidate cuts.The method incorporates fixed-window transition evidence directly into a globally optimized partition objective.Source: §1 · p. 2
  • For each candidate cut count, dynamic programming optimizes cut locations, then an AIC-style criterion selects a partition before segment means determine source labels and contiguous LLM spans.Adjacent segments with the same label are merged to obtain final authorship boundaries and maximal LLM-authored spans.Source: §1 · p. 2

2 RELATED WORK

Prior work studies LLM-text detection and localization in hybrid documents, including authorship roles, editing extent, and fine-grained boundaries. The paper frames persistent score fluctuations and unknown boundary counts as unresolved challenges for complete partitioning.

2 RELATED WORK

  • LLM-text detectors use signals such as token probabilities, probability curvature, regeneration discrepancies, and representation-level statistics, while many produce document-level decisions rather than explicit boundaries.ImBD and MSPO address machine-revised or human–machine hybrid text through stylistic preferences and multilevel style differences.Source: §2 · p. 2
  • Research on human–LLM collaboration examines authorship roles, creator–editor relations, hierarchical representations, machine-written portions, and the extent of AI editing.Beyond Binary, Beyond the Final Actor, DETree, PaLD, and EditLens address these respective perspectives.Source: §2 · pp. 2–3
  • Prior work includes methods for fine-grained localization using probabilistic, stylistic, attribution-based, and character-level span representations, alongside change-point methods using ordered local detection scores.Li et al. (2026a) formulate human–LLM localization through VCP, weighted WCP, and GCP.Source: §2 · p. 3
  • Distinguishing genuine transitions from score fluctuations and selecting a complete partition with an unknown boundary count remain challenging; LA-CPD optimizes candidate partitions and selects one before delineating labels and spans.Its two-stage inference uses dynamic programming followed by an AIC-style criterion.Source: §2 · p. 3

3 PRELIMINARY

The preliminary section defines sentence-level source labels and boundaries, then situates the two optimization stages: dynamic programming chooses cuts for a candidate count, and an AIC-style rule selects among partitions.

3 PRELIMINARY

  • A document is represented as sentences assigned human or LLM source labels; a boundary is a change between adjacent labels, and maximal contiguous LLM-labeled intervals are LLM-authored fragments.At inference, reference labels are unavailable and predicted boundaries follow changes in predicted labels.Source: §3 · p. 3
  • Dynamic programming optimizes cut locations for each candidate cut count, while the choice of cut count is handled separately.It decomposes a global problem into overlapping subproblems and combines their solutions to obtain a global optimum.Source: §3 · p. 3
  • LA-CPD uses an AIC-style criterion to balance residual fit against complexity across candidate partitions, but its residual and user-defined penalty multiplier make it distinct from standard AIC.Source: §3 · p. 3

4 METHODOLOGY

LA-CPD adapts a local sentence scorer, then structures its ordered scores into authorship segments despite unknown boundary counts and locations. It combines within-segment consistency and local transition evidence to select cuts and delineate labels and spans.

4.1 TRAINING

  • Human–LLM sentence pairs adapt a LoRA-augmented scorer against a frozen reference model, after which the scorer is frozen to produce ordered sentence-level scores.The training objective encourages higher scores for LLM-written than human-written sentences and does not use document-level boundary annotations.Source: Figure 1 · p. 4, §4.1 · p. 4

4.2 INFERENCE

  • After selecting statistical cuts, LA-CPD groups segment means to assign Human/LLM labels and delineate authorship boundaries and contiguous LLM-authored spans.The inference diagram distinguishes the statistical cut set from the subsequent source-label and span outputs.Source: Figure 2 · p. 5, §4.2 · p. 6
  • LA-CPD combines sentence-length-weighted segment residuals with windowed two-mean cut evidence, which can reduce sensitivity to isolated score fluctuations.Longer sentences contribute more to fitting; for windows wider than one sentence, neighborhood averaging aggregates evidence of local level changes.Source: §4.2 · pp. 5–6
  • Dynamic programming finds globally optimal cut locations for each fixed cut count, then an AIC-style criterion selects among candidate partitions using weighted residuals and model complexity.The dynamic program optimizes the partition objective at a fixed count; the selection criterion recomputes residuals without the boundary reward.Source: §4.2 · p. 6

5 EXPERIMENTS

LA-CPD performs best overall on the GPT-5.5 test set when boundary counts are unknown, using shared sentence-level scores to structure authorship segments. Ablations, generalization tests, and parameter analyses examine its components and scope.

5 EXPERIMENTS

  • Experiments use GPT-5.5 co-authored test data and compare segmentation methods, sentence-level prediction, voting, TextTiling, and PaLD; evaluation reports accuracy, WindowDiff, and AI-fragment F1.The local detector uses Gemma-3-1B with LoRA; cross-model and cross-dataset/domain generalization is also evaluated.Source: §5 · p. 7, §5.1 · p. 7

5.1 PERFORMANCE COMPARISON

  • LA-CPD achieves the best overall performance on the GPT-5.5 test set when boundary counts are unknown, and performs well overall across all metrics.Compared with WCP, the full framework also improves performance; fixing the cut count to the reference boundary count improves overall detection further.Source: §5.1 · pp. 7–8
  • Fixing the statistical-cut count to the reference number of source boundaries improves overall detection, indicating room to improve the current cut-count selection procedure.Cut locations and source labels are still predicted in this diagnostic setting.Source: §5.1 · p. 8

5.2 ANALYSIS OF THE STRUCTURED SEGMENTATION FRAMEWORK

  • Removing local persistent transition evidence degrades performance across all metrics, while the complete framework jointly contributes to authorship-span delineation and boundary detection.Uniform weights slightly improve WD but reduce AI-F1 and B3-2 F1; replacing AIC-based selection with MAD-based penalties slightly improves AI-F1 but degrades Acc and WD.Source: §5.2 · pp. 8–9
  • LA-CPD+AIC shows its clearest advantage on RoFT, but performance is less consistent when generation models, domains, and document characteristics shift simultaneously.The reported limitations include the quality of local detection scores and boundary-count selection.Source: §5.2 · p. 9
  • On the GPT-5.5 dataset, increasing the persistence window first decreases and then increases WD, while tested changes to α and β cause only small performance fluctuations.The complexity penalty r trades off evaluation objectives; an excessively strong penalty may impair AI-fragment delineation and overall segmentation quality.Source: §5.2 · p. 9

6 CONCLUSION

LA-CPD structures sentence-level evidence to delineate human- and LLM-authored spans when boundary counts and locations are unknown. The conclusion highlights persistent transition evidence and length weighting, while noting detector, distribution, and computational constraints.

6 CONCLUSION

  • LA-CPD delineates contiguous human- and LLM-authored spans in co-authored documents with unknown boundary locations and counts.It introduces structured change-point detection for boundary detection in human–AI collaboration.Source: §6 · p. 9
  • Persistent local transition evidence is identified as a primary factor in boundary localization, while length weighting promotes contiguous AI-span delineation.These are the conclusion's stated findings about the method's components.Source: §6 · p. 9
  • LA-CPD depends on the local detector's discriminative ability, cross-distribution detection remains challenging, and dynamic programming adds overhead for long documents.The authors identify more robust detectors and efficient inference as future work.Source: §6 · p. 9
  • Because performance may vary with generation model, domain, document length, and content, outputs should not be the sole basis for consequential decisions.The authors recommend considering outputs alongside human review and other evidence.Source: ETHICS STATEMENT · p. 10

A METHOD DETAILS AND IMPLEMENTATION

The appendix details LA-CPD's detector adaptation, score construction, weighted segmentation objective, cut-count search, label assignment, and end-to-end inference. Together, these steps turn frozen sentence scores into authorship labels, boundaries, and LLM spans.

A.1 LOCAL DETECTOR ADAPTATION

  • The local detector adapts a Gemma-3-1B scoring model with LoRA-based authorship supervision while keeping the reference model fixed.The main experiment continues adaptation from an existing LoRA checkpoint; only the scoring model's LoRA parameters are updated.Source: §A.1 · p. 14

A.2 COMPLETE DEFINITION OF THE LOCAL SCORE

  • The local score is a standardized deviation of observed-token log probability from its conditional expectation under a frozen reference distribution.Analytical vocabulary summation computes the conditional moments without explicitly sampling replacement tokens; the statistic is used as the sentence-level authorship score.Source: §A.2 · pp. 14–15

A.3 INTERVAL STATISTICS AND WINDOWED TRANSITION EVIDENCE

  • Segmentation combines token-length-weighted within-interval residuals with windowed two-mean evidence, which measures the residual reduction from separate local score levels.The windows aggregate observations on both sides of each candidate cut; normalized weights preserve relative sentence-length contributions.Source: §A.3 · pp. 15–16

A.4 SEARCH SPACE, DYNAMIC PROGRAMMING, AND MODEL SELECTION

  • Dynamic programming finds the globally optimal partition for each fixed cut count, then an AIC-style criterion compares the resulting candidates while accounting for complexity.The criterion uses the original residual without the transition-evidence reward; the main configuration uses an effective minimum segment length of six sentences.Source: §A.4 · pp. 16–17

A.5 DETERMINISTIC AUTHORSHIP LABEL AND SPAN DELINEATION

  • Each selected statistical segment receives its arithmetic mean score, and a two-group split of distinct scores determines Human/LLM labels before adjacent equal-label segments are merged.The resulting boundaries are label changes, and maximal contiguous LLM-labeled spans are returned.Source: §A.5 · pp. 17–18

A.6 FULL LA-CPD INFERENCE ALGORITHM

  • The inference algorithm computes sentence scores and candidate partitions, selects a cut count, assigns labels, and returns sentence labels, boundaries, and predicted AI spans.It uses fixed-count dynamic programming and AIC-style selection before deterministic label and span delineation.Source: §A.6 · pp. 18–19

B.1 DATASET SETTINGS

The appendix describes the constructed co-authored data, evaluation resources, and comparison methods. The experiments cover several transition patterns and external datasets, while excluding an entirely AI-written condition.

B.1 DATASET SETTINGS

  • The GPT-5.5 corpus represents five conditions, including human-only and single- or multiple-transition human–LLM patterns.The conditions include H0, B1, B2, B3-2, and B3-3, with transition patterns H→L, H→L→H, and H→L→H→L among them.Source: §B.1 · p. 19
  • The corpus omits the entirely AI-written A0 condition, so the reported experiments do not evaluate that condition.A0 was considered during corpus design but is absent from both the generated corpus and reported experiments.Source: §B.1 · p. 19
  • GPT-5.5 generates spans in the B1, B2, and B3 conditions from human context, with prompts designed to preserve context and avoid repeating it.The corpus uses human texts from WritingPrompts and XSum; B1 begins at a sentence boundary, whereas B2 continues an unfinished sentence.Source: §B.1 · pp. 19–20
  • Evaluation includes Claude continuations on SQuAD, WritingPrompts, and XSum, plus RoFT and Hybrid Essays as public datasets.The Claude sets primarily represent a single transition, while Hybrid Essays includes patterns with one to three transitions.Source: §B.1 · p. 23

B.2 BASELINES

  • The comparison includes sentence-level and heuristic baselines, score-based partial-text detection, and change-point methods including WCP+AIC.LA-CPD's unknown-count setting optimizes locations for candidate cut counts, then evaluates an AIC-style criterion using residuals without the boundary reward.Source: §B.2 · p. 24

B.3 EXPERIMENTAL SETTINGS AND EVALUATION METRICS

This section describes the evaluation protocol and metrics for sentence labels, boundary agreement, and contiguous AI fragments. It also distinguishes those outputs from statistical cut-count errors.

B.3 EXPERIMENTAL SETTINGS AND EVALUATION METRICS

  • Sentence accuracy, WindowDiff, and AI-fragment evaluation use delineated sentence labels, whereas cut-count error compares reference authorship boundaries with statistical cuts.Statistical cuts and final authorship boundaries can differ when adjacent statistical segments receive the same label.Source: §B.3 · p. 25
  • Sentence-level accuracy is averaged across documents, and the main evaluation uses preprocessed sentence lists with human and LLM labels mapped to 0 and 1.Documents with at most one sentence are skipped in the path that recomputes detector scores.Source: §B.3 · pp. 24–25
  • WindowDiff compares predicted and reference boundary counts within sliding sentence windows, with lower values indicating closer segmentation agreement.The boundaries are derived from changes in sentence labels.Source: §B.3 · p. 25
  • AI-Fragment F1 evaluates localization of maximal contiguous AI-labeled spans, matching predicted and reference fragments at IoU greater than 0.5.Matches are greedy and one-to-one; unmatched predictions and references count as false positives and false negatives.Source: §B.3 · p. 25
  • Figure 4 reports cross-distribution sentence-level accuracy and AI-fragment F1 across Hybrid Essays, RoFT, and three Claude datasets.Its panels separate accuracy and AI-F1; colors identify datasets, and LA-CPD oracle-K is included only as a diagnostic setting.Source: Figure 4 · p. 26

B.4 COMPLETE ABLATION RESULTS

Table 9 reports the complete component ablation results, using WCP+AIC as a reference and identifying LA-CPD (Full) as LA-CPD+AIC.

B.4 COMPLETE ABLATION RESULTS

  • Table 9 reports complete component ablations with WCP+AIC as a reference.Source: §B.4 · p. 26
  • LA-CPD (Full) denotes LA-CPD+AIC.Source: §B.4 · p. 26
  • The ablations set α = 0 without E(c; h), use uniform weights without length weighting, or replace AIC selection with a MAD complexity penalty.Source: §B.4 · p. 26

B.5 COMPLETE CROSS-DISTRIBUTION GENERALIZATION RESULTS

Cross-distribution results vary across datasets: LA-CPD+AIC leads some unknown-count comparisons but performs poorly on Claude–SQuAD. Supplying the true boundary count improves both Acc and AI-F1 on all five datasets.

B.5 COMPLETE CROSS-DISTRIBUTION GENERALIZATION RESULTS

  • On Hybrid Essays, LA-CPD+AIC achieves an AI-F1 of 0.532, higher than VCP, WCP, and WCP+AIC; on RoFT, it has the highest Acc among unknown-count methods at 0.638.Its RoFT WD of 0.210 is close to WCP, while its AI-F1 is lower than Voting and TextTiling.Source: §B.5 · p. 26
  • Claude results vary by dataset: LA-CPD+AIC's AI-F1 is 0.646 on WritingPrompts and 0.040 on SQuAD, where its Acc is 0.552.On Claude–WritingPrompts, its Acc of 0.730 is lower than Voting's.Source: §B.5 · pp. 26–27
  • On Claude–XSum, LA-CPD+AIC reaches an AI-F1 of 0.922, the highest among methods with unknown boundary counts, while its Acc of 0.922 is below Voting's 0.940.Source: §B.5 · pp. 26–27
  • Providing the ground-truth boundary count improves both Acc and AI-F1 on all five datasets, while cut locations and source labels remain predicted.For example, Claude–SQuAD Acc and AI-F1 rise from 0.552 and 0.040 to 0.894 and 0.918, respectively.Source: §B.5 · p. 27

B.6 SUPPLEMENTARY ANALYSIS OF THE REGIME-RETURN DISCOUNT

The supplementary analysis varies the regime-return discount under a MAD-based penalty, not the AIC configuration used in the main experiments. Stronger discounts trade improved B3-2 AI-F1 for lower overall and human-only accuracy.

B.6 SUPPLEMENTARY ANALYSIS OF THE REGIME-RETURN DISCOUNT

  • The regime-return discount ρ reduces penalties for the second and subsequent statistical cuts when ρ < 1; ρ = 1 applies no discount.Source: §B.6 · p. 27
  • The analysis varies ρ from 1.00 to 0.35 with h = 3, α = 0.5, β = 1, and MAD penalty coefficient c = 2.This differs from the AIC-based configuration used in the main experiments.Source: §B.6 · p. 27
  • Reducing ρ from 1.00 to 0.35 raises B3-2 AI-F1 from 0.237 to 0.276, but lowers overall Acc from 0.756 to 0.731 and H0 Acc from 0.852 to 0.810.Overall AI-F1 changes from 0.519 to 0.513; these results do not establish a benefit under AIC-based selection.Source: §B.6 · p. 27
Loading 2609.33787v1…