Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection Sondos Mahmoud Bsharat1 Jiacheng Liu1 Xiaohan Zhao1 Tianjun Yao1 Xinyi Shang1,2 Yi Tang1 Jiacheng Cui1 Ahmed Elhagry1 Salwa K. Al Khatib1 Hao Li1 Salman Khan1 Zhiqiang Shen1,†
arXiv:2606.06481v1 [cs.CL] 4 Jun 2026
1
Mohamed bin Zayed University of Artificial Intelligence 2 University College London † Correspondence: [email protected]
Abstract As AI writing assistants become increasingly integrated into real-world drafting and revision workflows, many documents are no longer purely human-written or AI-generated, but instead result from progressive human–AI co-editing. However, existing AI-text detection benchmarks largely focus on final outputs and provide limited understanding of how AI authorship signals emerge, accumulate, or disappear throughout the revision process. We introduce OpAI-Bench, an operation-guided benchmark for studying progressive human-to-AI text transformation across document, sentence, token, and span granularities. Starting from human-written documents, OpAI-Bench constructs nine sequentially revised versions for each sample under predefined AI coverage levels and five representative AI edit operations, covering four domains while preserving complete authorship provenance at multiple granularities. The benchmark supports comprehensive evaluation with 8 document-level detectors, 7 sentence-level detectors, and 2 finegrained token/span-level detectors. Experiments reveal that AI-text detectability is governed not only by the proportion of AI-edited content, but also by edit operation, domain, and cumulative revision history. Interestingly, we notice that mixed-authorship intermediate versions are often harder to detect than both fully human and heavily AI-edited endpoints, exposing non-monotonic detection patterns missed by existing benchmarks. OpAI-Bench provides a controlled testbed for analyzing whether, when, and how AI-assisted writing becomes detectable under realistic progressive editing scenarios. Our code and benchmark are available at https://github.com/VILA-Lab/OpAI-Bench.
1
Introduction
AI-assisted writing and editing workflows [17, 23, 25, 6, 7, 14] are increasingly embedded in practical writing workflows, where they are used not only to generate complete passages, but also to revise, polish, expand, compress, and restructure human-written drafts. As a result, many real-world documents are better characterized as products of progressive human–AI co-editing rather than as purely human-written or fully AI-generated text. This emerging writing paradigm challenges the conventional binary framing of AI-text detection, where a document is assumed to belong to one of two endpoint categories. In practice, AI involvement may appear locally, accumulate gradually, and interact with prior human content across multiple rounds of revision. Existing AI-text detection benchmarks [5, 29, 28] provide valuable resources for evaluating whether a completed text is human- or AI-written, but they offer limited support for analyzing how detectability evolves during the transformation from human writing to AI-edited text. Most benchmarks are Preprint.
constructed from static final outputs and do not preserve the intermediate revision states that lead to those outputs. Consequently, they cannot answer several important questions: at what stage does AI involvement become reliably detectable, which types of edits introduce the strongest detection signals, and whether mixed-authorship texts are easier or harder to detect than endpoint cases. These questions are increasingly important as AI-assisted writing becomes more incremental, interactive, and operation-specific. Evaluation should capture not only how much text is AI-edited, but also how it is edited and where the edits occur. Edit operations can leave different signals: polishing may preserve lexical choices while improving fluency, paraphrasing changes form while retaining meaning, expansion adds explanatory detail, and compression removes details or simplifies structure. Collapsing these operations into a single AI-written label obscures their different effects on detection. Similarly, document-level labels support global screening but cannot localize AI involvement, while sentence-level labels may miss mixed human–AI fragments within a sentence. Token- and span-level provenance is therefore needed to evaluate fine-grained localization. To address these limitations, we introduce OpAI-Bench, a novel operation-guided benchmark for progressive human-to-AI text transformation and multi-granularity AI-text detection. Starting from human-written source documents, OpAI-Bench constructs a sequence of progressively revised versions under predefined AI coverage levels and representative edit operations. Each revision trajectory records how human text is transformed by AI edits over time, while preserving authorship provenance at the document, sentence, token, and span levels. This design enables controlled evaluation of detection performance as a function of AI coverage, edit operation, domain, and cumulative revision history, rather than only on isolated final texts. Using OpAI-Bench, we conduct a comprehensive evaluation of document-level, sentence-level, and fine-grained AI-text detectors. Our results show that AI-text detectability is not determined solely by the proportion of AI-edited content. Instead, detector performance varies substantially across edit operations, domains, and revision stages. In particular, intermediate mixed-authorship versions can be more challenging than both purely human and heavily AI-edited endpoints, revealing non-monotonic detection behavior that is overlooked by existing static benchmarks. These findings suggest that reliable AI-text detection requires moving beyond binary endpoint classification toward trajectory-aware and operation-aware evaluation. Our contributions are summarized as follows: • We introduce OpAI-Bench, an operation-guided benchmark for progressive human-to-AI text transformation that preserves intermediate revision states rather than only final outputs. • We construct cumulative revision trajectories with predefined AI coverage levels and five edit operations: polish, paraphrase, style rewrite, compress, and expand, and provide provenance across document, sentence, token, and span levels. • We benchmark diverse detector families across granularities and use OpAI-Bench to assess detector stability across coverage, edit operations, domains, generators, and revision history. Our results reveal non-monotonic detectability, with a critical mixed-authorship region around v4 where intermediate AI coverage and compression coincide.
2
Related Work
AI-text detection benchmarks. Prior work on AI-text detection has introduced increasingly diverse benchmarks spanning human-written, machine-generated, and AI-edited text. Benchmarks such as TuringBench [25], HC3 [6], MGTBench [7], MULTITuDE [14], RAID [5], M4 [29], and M4GTBench [28] broaden evaluation across generators, domains, languages, and robustness conditions. More recent benchmarks also expand the problem formulation to include robustness under attacks, generator attribution, and mixed-authorship boundary detection. DetectRL [30] evaluates detectors under prompt attacks, paraphrasing, perturbations, and data-mixing settings, while M4GT-Bench [28] further incorporates generator attribution and mixed human–machine change-point detection. However, these benchmarks still mainly evaluate the final text obtained after generation or editing, or at most a single authorship transition within it, rather than a full revision process with explicit intermediate stages. 2
(A) Naive Solution: Independent versions generated directly from V0, without a cumulative edit trajectory. Human Source Document (V0) Fully Human-Written Document
n sentences (s1 ... sn) Minimum: 10 sentences per document
AI edit (e.g., paraphrase, polish, style rewrite) is applied to a subset of sentences ------------------------------Each version is edited independently from V0 each time
V1
15%
V2
25%
V3
40%
V8
100%
Limitation of Naive Approach Independent rewrites without cumulative history No explicit trajectory or provenance tracking AI coverage not aligned with a progressive human-to-AI transition
(B) Our Approach: Cumulative versioned Human-to-AI Authorship Trajectory. Each version inherits all previously edited sentences and adds new edits from the fixed shuffled order. (1) Human Source Document (V0) Fully Human-Written Document
(3) Versioned Human-to-AI Authorship Trajectory
(2) Deterministic Sentence Shuffle
Progressively increase AI coverage by deterministically following the shuffled order
Fixed Shuffle Order (seed = hash(doc_id)) Shuffled order
Version
AI Coverage
Operator
S1
Sentence 1
S1
S2
V0
0%
Human (no edit)
S2
Sentence 2
S2
S5
V1
15%
Polish
S3
Sentence 3
S3
S6
V2
25%
Paraphrase
Select {S2, S5, S6}
S4
Sentence 4
S4
S1
V3
40%
Style Rewrite
Select {S2, S5, S6, S1}
Sn-1
Sentence n-1
Sn-1
Sn
V7
90%
Paraphrase
Select {S2, S5, S6, S1, S3, S10, S4, S9, S7}
Sn
Sentence n
Sn
S4 v8
100%
Polish
Document D1
n sentences (s1 ... sn) Minimum: 10 sentences per document
Original order
Deterministic, Reproducible, content-agnostic ordering
Deterministic Cumulative Selection Select first k% of shuffled sentences Select {S2, S5}
Select {S2, S5, S6, S1, S3, S10, S4, S9, S7, S8}
Figure 1: OpAI-Bench construction pipeline. Top: a naive setup creates each version independently from the original human document, so versions do not form a true revision history and provenance is not tracked across stages. Bottom: OpAI-Bench builds a cumulative trajectory by progressively editing a larger prefix of that order. Each version is generated from the previous one. Mixed-authorship and fine-grained attribution. A related line of work moves beyond pure humanversus-AI classification to study hybrid and coauthored text. MixSet [32] highlights the difficulty of detecting mixed human–AI writing in settings such as AI-revised human drafts and humanrevised machine outputs. Localization methods such as AdaLoc [33] identify machine-generated sentences within otherwise human-written documents, while SenDetEX [11], HACo-Det [20], and DAMASHA [22] move toward finer attribution at the sentence, word, or token levels. Beemo [2] and RealBench [8] further show that human post-editing and diverse collaboration patterns substantially complicate detection. These works demonstrate the importance of mixed-authorship and fine-grained detection, but typically focus on a completed hybrid sample or local attribution within it, rather than the revision trajectory through which mixed authorship is formed. Varying degrees of AI involvement. Recent work has begun to move beyond binary AI-text detection by considering text with varying degrees of AI involvement. APT-Eval [17] varies the degree of AI polishing and shows that lightly edited text can be difficult to distinguish from fully human writing. PaLD [13] estimates the proportion and localization of LLM-written content, while EditLens [23] models AI involvement as a continuous edit-magnitude signal relative to an original human draft. HACo-Det [20] similarly motivates numeric notions of AI ratio through fine-grained attribution. Together, these works move beyond binary labels, but they typically focus on final edited samples or single editing outcomes. In contrast, OpAI-Bench studies explicit cumulative revision trajectories with preserved intermediate versions, controlled AI coverage, and diverse edit operations.
3
OpAI-Bench
3.1
Overview and Design Principles
Most prior AI-text detection benchmarks evaluate a completed text and assign it a final authorship label, such as human, AI-generated, or mixed. This formulation has been valuable for static detection, but it is less informative for studying how detectability evolves as AI involvement is introduced through revision. Prior benchmarks generally do not preserve explicit intermediate versions, do not model cumulative version-to-version editing, and do not jointly control AI coverage, edit type, and multi-granularity provenance within a unified setup. Consequently, they provide only limited support for analyzing when detection becomes reliable, whether matched AI coverage yields different difficulty across edit types, and whether detector behavior depends on revision history as well as on the final edited text. 3
Benchmark
Authorship
Trajectory
Mixed AI_Cov Interm Cumul
Diversity
Annotation
Edit_Types Domain LLM
Gran.
DetectRL [30] MixSet [32] RAID [5] M4GT-Bench [28] RealBench [8] Beemo [2] HACo-Det [20] SenDetEX [11] APT-Eval [17] DAMASHA [22] EditLens [23]
✓ ✓ ✗ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗
✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗ ✗ ✗ ✗
✓ ✓ ✓ ✗ ✓ ✓ ✗ ✗ ✗ ✗ ✓
✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
D M D M D D M S D B D
OpAI-Bench (Ours)
✓
✓
✓
✓
✓
✓
✓
M
Table 1: Comparison with prior AI-involved text detection benchmarks. Mixed: mixed human– AI authorship; AI_Cov: controlled AI coverage; Interm: intermediate revision stages; Cumul: cumulative version-to-version revision; Edit_Types: diverse AI edit operations; Gran.: supported annotation/evaluation granularity (D: document, S: sentence, B: boundary, M: multi-granularity). OpAI-Bench is the only benchmark listed that combines all properties. OpAI-Bench is designed to study this setting by representing mixed authorship as a controlled revision trajectory. Given a source document D, we construct T (D) = D(0) , D(1) , . . . , D(8) , where D(0) is the original human-written document and D(V ) denotes version V ∈ {0, . . . , 8}. Rather than generating each version independently from the original source, we construct each D(V ) by editing the previous version D(V −1) . This yields a cumulative sequence in which the target AI coverage increases monotonically from fully human (v0) to fully AI-edited (v8) under predefined coverage ratios and edit operations. Each version is paired with provenance annotations at the token, sentence, and document levels. This structure enables analyses that are difficult to carry out on static benchmarks, including coveragecontrolled comparisons, edit-type-controlled comparisons, and trajectory-controlled comparisons. In particular, it allows us to study whether detectability depends only on the final amount of AI-edited content, or also on the path through which mixed authorship is formed. Table 1 situates OpAI-Bench relative to prior resources, and Figure 1 illustrates the construction pipeline. OpAI-Bench is guided by five design principles: versioned trajectories, expanding each source document into nine ordered versions (v0–v8) from fully human to fully AI-edited text; controlled AI coverage, introducing edits at predefined sentence-level ratios from 0% to 100%; edit-type diversity, covering polish, paraphrase, style rewrite, compress, and expand; cumulative construction, generating each version from the previous one rather than independently from the source; and multi-granularity provenance, preserving AI-revision authorship labels at word-level token/span, sentence, and document levels. 3.2
Benchmark Construction
Source Domains, Representation, and Filtering. OpAI-Bench is constructed from human-written documents across four domains: student essays [12], news articles [15], government reports [10], and scientific abstracts [16]. These domains vary in writing style, discourse structure, and sentence length, enabling evaluation across diverse text types. We use document as the unit of editing. Depending on the source corpus, a document may be a multi-paragraph text or a single paragraph treated as a document. Each document is normalized and segmented into paragraphs and sentences, represented as D = (p1 , . . . , pm ), where each paragraph (i) (i) pi = (s1 , . . . , sni ). Stable paragraph and sentence identifiers are assigned and preserved across all derived versions for sentence selection, reconstruction, and authorship tracking. To support intermediate revision trajectories, we retain only documents with at least 10 sentences, Pm i.e., N (D) = i=1 ni ≥ 10. Each retained document is then expanded into a nine-version trajectory while preserving its paragraph and sentence structure. Summary statistics are reported in Table 2. 4
Domain
# Source Docs
# Traj.
# Versioned Samples
Avg. sents
Avg. tokens
Student essays News articles Government reports Scientific abstracts
3,969 3,998 3,993 3,762
7,906 7,892 8,000 7,291
71,154 71,028 72,000 65,612
21.0 24.0 20.6 11.0
398.8 491.3 563.7 234.3
Total
15,722
31,089
279,794
19.3
426.1
Table 2: Statistics of the OpAI-Bench main split. Source documents are distinct human-written v0 texts, while trajectories are generator-specific revision sequences initialized from v0. Versioned samples count all released versions from v0 to v8 across the train, development, and test splits. Statistics are reported for the three primary generators and exclude ablation splits. Version
v0
v1
v2
v3
v4
v5
v6
v7
v8
Operation Coverage
none 0%
polish 15%
para. 25%
style 40%
compress 50%
expand 60%
style 75%
para. 90%
polish 100%
Table 3: OpAI-Bench version schedule. Each version is defined by an edit operation and target sentence-level AI coverage. “para.” denotes paraphrase. Versioned Trajectories and Cumulative Editing. For each source document D, OpAI-Bench constructs an ordered trajectory T (D) = (D(0) , D(1) , . . . , D(8) ), where D(0) is the original humanwritten document and later versions follow the fixed operation–coverage schedule in Table 3. The trajectory spans the transition from fully human text at v0 to fully AI-edited text at v8, with intermediate versions increasing the amount of edited content while varying the edit operation. To select edited sentences, we define a fixed document-specific ordering π(D) = (π1 , . . . , πN (D) ) using a deterministic shuffle seeded by the document identifier. This ordering is computed once and reused across all versions, ensuring reproducibility and removing positional bias. Given target (t) (t) sentence-level coverage csent at version t, we select k (t) = ⌈csent · N (D)⌉ sentences and define the (t) edited set as S = {π1 , . . . , πk(t) }. Because coverage increases monotonically, the edited sets satisfy S (0) ⊆ S (1) ⊆ · · · ⊆ S (8) . This yields a cumulative editing process: once selected, a sentence remains editable in later versions, so subsequent versions both add newly selected sentences and re-edit previously modified ones under the current operation. This preserves not only where AI edits occur, but also the revision history through which they accumulate. Editing Operations and Coverage Control. OpAI-Bench uses five sentence-level rewrite operations to model different forms of AI-assisted revision: polish, paraphrase, style rewrite, compress, and expand. These operations vary the form of intervention: polish makes fluency edits, paraphrase changes wording while preserving meaning, style rewrite modifies tone, compress removes nonessential phrasing, and expand adds limited clarifying detail. All prompts require the model to rewrite only the target sentence, preserve names, numbers, entities, and factual content, avoid unsupported additions, and return a single sentence. We further enforce sentence-level consistency through automatic validation and manual verification: outputs are discarded if they split the target into multiple sentences, merge across sentence boundaries, or violate the expected format. Full prompts are provided in Appendix J. (t)
For each version D(t) , the target sentence-level AI coverage csent determines the fraction of sentences selected for rewriting according to Table 3. A sentence is labeled AI-edited if selected for rewriting. Token- and span-level annotations are projected from AI-edited sentence regions using whitespace tokenization: words whose character offsets overlap an AI-marked region are labeled AI, and consecutive AI-labeled words are merged into spans. These labels are tokenizer-agnostic and can be aligned to any subword vocabulary via character offsets. This convention is consistent with segment-level mixed-authorship settings that label AI-revised regions as machine-involved [32].
4
Experiments
4.1
Tasks
OpAI-Bench evaluates AI-text detection at three granularities over the same versioned trajectory T (D) = (D(0) , . . . , D(8) ). Document-level detection predicts whether a document version contains any AI-edited content. The source version D(0) is labeled human, while later versions are labeled 5
Detector
Ver.
Essays Acc
Reports
F1-AI
Acc
F1-AI
News
Abstracts
Acc
F1-AI
Acc
F1-AI
Sentence Level
Claude-Haiku
v0 v1 v2 v3 v4 v5 v6 v7 v8
91.2 77.2 -13.9 74.7 -2.5 75.0 +0.3 62.9 -12.1 69.8 +7.0 75.2 +5.4 66.0 -9.2 64.3 -1.7
0.0 22.0 +22.0 45.5 +23.5 65.5 +20.0 52.7 -12.9 69.7 +17.1 81.5 +11.7 77.5 -4.0 78.0 +0.5
96.4 79.9 -16.5 69.4 -10.5 56.2 -13.2 47.8 -8.4 48.7 +0.9 41.1 -7.6 34.4 -6.7 30.7 -3.6
0.0 7.0 +7.0 11.4 +4.4 17.1 +5.6 13.4 -3.6 34.2 +20.8 37.8 +3.6 42.3 +4.5 43.0 +0.7
91.9 78.0 -13.9 70.5 -7.5 61.9 -8.6 51.2 -10.7 52.9 +1.7 46.2 -6.7 33.4 -12.7 27.3 -6.1
0.0 11.2 +11.2 20.3 +9.1 31.3 +11.0 23.0 -8.3 40.3 +17.3 45.2 +4.9 41.7 -3.5 40.4 -1.3
53.2 50.6 -2.6 48.9 -1.6 51.4 +2.5 49.8 -1.6 60.0 +10.2 66.8 +6.9 69.1 +2.3 71.7 +2.5
0.0 25.9 +25.9 37.6 +11.7 51.2 +13.7 52.2 +1.0 67.4 +15.2 77.5 +10.1 80.5 +3.0 83.1 +2.6
AdaLoc [33]
v0 v1 v2 v3 v4 v5 v6 v7 v8
85.1 72.3 -12.8 66.2 -6.1 57.1 -9.1 50.9 -6.2 43.0 -8.0 34.7 -8.3 28.6 -6.1 29.8 +1.2
0.0 20.5 +20.5 23.7 +3.3 30.1 +6.3 30.9 +0.8 27.5 -3.4 30.9 +3.4 37.8 +6.9 44.7 +6.9
94.7 79.0 -15.7 70.5 -8.5 56.9 -13.6 48.1 -8.7 38.6 -9.5 25.4 -13.3 11.6 -13.8 5.0 -6.6
0.0 7.9 +7.9 7.9 -0.0 8.8 +0.9 6.5 -2.3 7.2 +0.7 7.4 +0.3 8.2 +0.8 9.5 +1.3
86.6 73.9 -12.7 67.4 -6.5 56.3 -11.1 49.4 -6.9 42.0 -7.4 30.7 -11.3 21.2 -9.5 16.9 -4.3
0.0 15.8 +15.8 18.6 +2.8 20.8 +2.2 21.1 +0.3 22.5 +1.4 23.0 +0.4 26.3 +3.4 28.8 +2.4
96.7 79.6 -17.1 70.3 -9.3 57.2 -13.1 48.4 -8.8 38.3 -10.2 22.3 -16.0 12.1 -10.2 5.6 -6.5
0.0 6.1 +6.1 5.7 -0.5 5.2 -0.4 5.1 -0.1 4.9 -0.2 5.1 +0.2 7.1 +2.0 10.6 +3.5
RADAR [9]
v0 v1 v2 v3 v4 v5 v6 v7 v8
71.2 33.7 -37.5 36.2 +2.5 45.2 +9.0 42.8 -2.4 49.4 +6.7 53.1 +3.7 52.5 -0.6 55.8 +3.3
0.0 50.4 +50.4 53.1 +2.8 62.2 +9.0 59.9 -2.3 66.1 +6.3 68.9 +2.7 68.7 -0.2 71.5 +2.7
99.9 0.1 -99.9 0.2 +0.2 0.3 +0.1 0.2 -0.1 0.3 +0.1 0.3 +0.1 0.3 -0.1 0.4 +0.2
0.0 0.1 +0.1 0.4 +0.3 0.5 +0.1 0.3 -0.2 0.5 +0.2 0.7 +0.1 0.5 -0.1 0.9 +0.3
99.9 0.1 -99.9 0.3 +0.3 0.6 +0.2 0.2 -0.4 0.8 +0.6 1.0 +0.2 1.4 +0.4 2.0 +0.6
0.0 0.1 +0.1 0.7 +0.6 1.2 +0.5 0.3 -0.8 1.5 +1.2 2.0 +0.5 2.7 +0.7 3.8 +1.2
99.3 0.9 -98.4 1.0 +0.2 1.3 +0.2 0.7 -0.6 1.4 +0.7 1.2 -0.2 0.9 -0.3 0.2 -0.7
0.0 1.7 +1.7 2.1 +0.4 2.6 +0.5 1.5 -1.1 2.8 +1.3 2.3 -0.4 1.8 -0.5 0.5 -1.3
Fast-DetectGPT [3]
v0 v1 v2 v3 v4 v5 v6 v7 v8
83.7 19.7 -64.0 11.1 -8.6 10.1 -0.9 7.9 -2.3 8.5 +0.7 10.1 +1.6 7.0 -3.1 12.1 +5.1
0.0 32.8 +32.8 19.9 -12.9 18.4 -1.6 14.5 -3.8 15.7 +1.1 18.3 +2.6 12.9 -5.4 20.8 +7.9
58.3 44.2 -14.1 28.4 -15.8 32.5 +4.1 19.1 -13.4 20.1 +1.0 19.2 -1.0 10.5 -8.7 18.9 +8.4
0.0 61.3 +61.3 43.9 -17.3 48.4 +4.5 31.8 -16.6 32.6 +0.8 31.0 -1.5 18.5 -12.6 29.1 +10.7
52.3 48.5 -3.8 32.4 -16.1 29.9 -2.5 22.2 -7.7 22.4 +0.2 20.8 -1.6 18.2 -2.6 23.9 +5.6
0.0 65.0 +65.0 48.0 -17.0 44.2 -3.8 35.1 -9.0 34.9 -0.2 32.9 -2.0 29.4 -3.5 35.2 +5.8
82.5 15.9 -66.6 7.8 -8.0 9.8 +2.0 5.8 -4.1 5.8 +0.0 6.1 +0.3 4.4 -1.7 12.2 +7.8
0.0 27.3 +27.3 14.4 -13.0 17.6 +3.2 10.9 -6.8 10.8 -0.0 11.3 +0.4 8.3 -2.9 20.4 +12.1
DAMASHA [22]
v0 v1 v2 v3 v4 v5 v6 v7 v8
89.5 75.8 -13.7 68.7 -7.1 59.0 -9.7 56.2 -2.9 50.4 -5.8 51.2 +0.8 42.3 -8.9 43.2 +0.9
0.0 19.2 +19.2 25.0 +5.8 38.3 +13.3 30.9 -7.4 48.7 +17.8 59.8 +11.1 55.9 -3.9 59.9 +4.0
79.4 68.5 -10.9 62.5 -6.0 52.9 -9.6 51.4 -1.5 45.0 -6.4 41.6 -3.5 32.8 -8.7 32.5 -0.3
0.0 19.4 +19.4 22.5 +3.1 31.7 +9.2 26.3 -5.5 39.2 +13.0 46.8 +7.6 43.5 -3.3 48.2 +4.7
98.6 82.0 -16.6 72.4 -9.6 57.5 -14.9 53.0 -4.5 39.0 -14.0 27.9 -11.1 14.1 -13.9 7.6 -6.4
0.0 2.4 +2.4 4.8 +2.3 9.4 +4.6 5.2 -4.2 13.3 +8.2 17.2 +3.9 12.8 -4.3 13.5 +0.7
83.1 70.2 -12.9 63.6 -6.5 57.2 -6.5 53.8 -3.3 46.1 -7.7 41.9 -4.2 35.1 -6.7 34.9 -0.3
0.0 20.6 +20.6 26.1 +5.5 35.3 +9.2 31.9 -3.4 41.2 +9.3 47.9 +6.6 46.6 -1.3 51.5 +4.9
GigaCheck [24]
v0 v1 v2 v3 v4 v5 v6 v7 v8
98.7 – 82.6 -16.1 74.2 -8.4 66.1 -8.1 58.9 -7.2 54.6 -4.3 56.7 +2.1 41.6 -15.1 51.1 +9.4
0.0 – 7.6 +7.6 19.0 +11.4 39.0 +20.0 26.6 -12.4 44.8 +18.3 62.3 +17.4 53.8 -8.4 66.7 +12.8
98.5 – 81.8 -16.8 71.9 -9.9 56.6 -15.3 52.6 -4.0 38.6 -14.1 28.9 -9.7 17.0 -11.9 11.2 -5.8
0.0 – 2.6 +2.6 3.3 +0.6 5.7 +2.5 8.7 +3.0 11.5 +2.7 18.3 +6.8 17.7 -0.6 18.1 +0.4
99.5 – 82.9 -16.6 72.6 -10.4 57.2 -15.3 51.9 -5.3 38.7 -13.3 30.1 -8.6 19.0 -11.1 15.7 -3.3
0.0 – 1.5 +1.5 3.0 +1.5 8.2 +5.2 6.8 -1.4 12.1 +5.3 20.7 +8.7 20.0 -0.7 23.9 +3.8
97.5 – 79.6 -17.9 69.4 -10.1 58.5 -11.0 53.3 -5.2 45.9 -7.4 43.9 -2.0 32.8 -11.1 30.9 -1.9
0.0 – 6.8 +6.8 12.1 +5.3 22.4 +10.3 21.4 -1.0 29.9 +8.5 46.5 +16.6 40.5 -5.9 45.2 +4.7
Document Level
Token and Span Level
Table 4: Aggregate main results across revision versions and domains, averaged over generators. We report accuracy and F1-AI for zero-shot detectors and LLM-as-detectors at their native evaluation granularity: document, sentence, or fine-grained token/span level. Colored deltas show the change from the previous version. AI-involved. Sentence-level attribution predicts a binary authorship label for each sentence, marking whether the sentence has entered the cumulative AI-edited set by the corresponding version. Tokenand span-level localization evaluates fine-grained provenance by identifying the AI-edited portions of a document at the benchmark token or character-span level. All tasks are evaluated at each version t ∈ {0, . . . , 8} rather than only at the final edited document. This allows us to measure detector behavior throughout the human-to-AI revision trajectory. 6
4.2
Evaluation Protocol
Evaluation regimes. We organize detectors into three categories corresponding to distinct levels of task-specific adaptation. (i) Zero-shot methods: detectors are evaluated in their original settings, using either released checkpoints or models reproduced from the original training scripts, with no exposure to OpAI-Bench data. This measures how well existing detection methods transfer to trajectory-style mixed authorship without adaptation. Document-level methods include Desklib [4], DetectLLM [19], E5-Small [26], Fast-DetectGPT [3], OOD-LLM Detect [31], RADAR [9], RoBERTa OpenAI [18], and GigaCheck [24]; sentence-level methods include AdaLoc [33], GenAI Sentence [21], GL-CLiC [1], and SeqXGPT [27]; token- and span-level methods include DAMASHA [22] and GigaCheck [24]. (ii) LLM-as-detector: three frontier language models: GPT-5.4, Gemini 3 Flash, and Claude Haiku 4.5 are queried as zero-shot classifiers via sentence-level confidence prompting, without any gradient-based adaptation; they serve as prompt-based reference points. (iii) Trained on OpAI-Bench: detectors are fine-tuned on the OpAI-Bench training split using sentence-level provenance labels drawn from three in-distribution generators (GPT-5.4, GPT-5.4nano, Gemini 2.5 Flash). Qwen3-8B is withheld from training and evaluated as a held-out generator, isolating cross-generator transfer while holding the training protocol fixed. Document-level methods include Fast-DetectGPT and GigaCheck; sentence-level methods include AdaLoc, GenAI-Sentence, GL-CLiC, and SeqXGPT; token-level and span-level methods include DAMASHA and GigaCheck. Full implementation details and training hyperparameters are provided in the appendix. Domain and generator splits. We report results by domain and generator to evaluate robustness across writing settings and generation backends. For fine-tuned methods, Qwen3-8B is used only as a held-out generator for cross-generator evaluation. Metrics. For each task and version, we report accuracy and AI-class F1 (F1AI ). Main tables also show changes relative to the previous reported version, and the ablation studies examine coverage control, edit operations, and cumulative construction.
5
Main Results
Detection performance does not increase smoothly as AI coverage grows, but instead varies across revision stages, domains, and detector types. Table 4 summarizes aggregate results across granularities, averaged over generators. Figures 2 and 3 show document- and sentence-level accuracy by domain and generator, with full breakdowns in Appendices F.1 and G. Document-level detectors show domain-specific failure patterns. While aggregate results in Table 4 suggest overall trends, Figure 2 shows that zero-shot document-level detectors exhibit substantial variation across domains and revision stages, rather than following a consistent trend with increasing AI coverage. RADAR is effective mainly on essays: its F1-AI remains near zero on reports, news, and abstracts across the trajectory, but increases on essays from 50.4 at v1 to 71.5 at v8 . Fast-DetectGPT shows a different pattern. It detects early edits in reports and news, reaching 61.3 and 65.0 F1-AI at v1 , but declines as revision progresses, falling to 29.1 and 35.2 at v8 . These results suggest that document-level detectors rely on domain- and operation-sensitive cues, so increasing AI coverage alone does not guarantee easier detection. Sentence-level supervision improves stability, but does not remove revision effects. Sentencelevel results show a clear gap between detector families. LLM-as-detectors often improve through early revisions, drop near the compression step, and only partially recover afterward. For example, Gemini-Flash on essays rises from 9.5 F1-AI at v1 to 71.5 at v3 , drops to 53.6 at v4 , and reaches 57.9 at v8 . Zero-shot sentence-level detectors are less consistent: AdaLoc and GL-CLiC recover in some settings, while SeqXGPT remains near-zero in F1-AI across versions despite high early accuracy. Fine-tuned sentence-level models are generally more stable, with GenAI-Sentence and AdaLoc often recovering at higher AI coverage. However, Figure 3 shows that performance still varies across domains and generators, indicating that sentence-level supervision improves robustness but does not eliminate sensitivity to revision stage or edit operation. Fine-grained localization follows a different trade-off. DAMASHA operates at the token level and GigaCheck at the span level, so we treat them as localization methods rather than directly 7
Essays
Reports
Abstracts
News
75 50 25 0 100
GPT-5.4
75 25
Generators
0 100
GPT-5.4-nano
Accuracy (%)
50
75 50 25 0 100
Qwen3-8b
75 50 25 0
Gemini-2.5-flash
100
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version
DeskLib DetectLLM
E5-Small Fast-DetGPT
OOD-LLM-Detect RADAR
RoBERTa GigaCheck
Figure 2: Document-level accuracy across revision versions, domains, and generators. Each curve represents a document-level detector, illustrating how performance changes along the progressive human-to-AI revision trajectory. comparable document classifiers. Unlike document- and sentence-level detectors, fine-grained methods generally show decreasing accuracy but increasing F1-AI as AI-revision-labeled regions become denser. This reflects class imbalance in early versions: predicting mostly human yields high accuracy when few word-level tokens or spans are labeled AI, while F1-AI becomes more informative at higher AI coverage. Nevertheless, localization remains domain- and generator-dependent, with news consistently harder than essays or abstracts. Overall, the main trajectory shows that AI-text detection depends on more than the amount of AI-edited content. The ablations in Section 5.1 further disentangle this effect, showing that the degradation around v4 is linked to compression as well as intermediate human–AI coverage. 5.1
Controlled Analysis of Coverage, Operation, and Revision History
The main trajectory varies three factors jointly: AI coverage, edit operation, and cumulative revision history. We therefore run controlled analyses to clarify the non-monotonic patterns in the main results, especially around v4 , where intermediate human–AI coverage coincides with compression. Unless otherwise stated, we use Gemini-2.5-Flash, since it gives the clearest separation from human-written text in the main experiments. Varying coverage within each operation. We first fix the edit operation and vary target AI coverage. Figures 12 and 13 show that sentence-level performance does not uniformly collapse at 50% coverage. Instead, trends differ by operation, with compression generally harder to detect than paraphrase or expansion at comparable coverage. Document-level results show a similar but more detectordependent pattern (Figures 14 and 15). Thus, the degradation around v4 is not explained by the human–AI ratio alone. Comparing operations at fixed coverage. We next fix one coverage level at a time (25%, 50%, or 75%) and compare compress, paraphrase, and expand. Figures 16 and 17 show that, for LLM8
Essays
News
Reports
Abstracts Gemini-2.5-flash
100 75 50 25 0 100
25 0 100
GPT-5.4-nano
Accuracy (%)
50
75 50 25 0 100
Qwen3-8B
75 50 25 0
Generators
GPT-5.4
75
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version Claude-Haiku Gemini-Flash
GPT-5.4 GenAI-Sentence (FT)
GL-CLiC (FT) AdaLoc (FT)
SeqXGPT (FT)
Figure 3: Sentence-level accuracy across revision versions, broken down by domain and generator. The figure compares LLM-as-detectors with fine-tuned sentence-level detectors (FT), showing how detector behavior changes across the progressive human-to-AI revision trajectory.
as-detectors, detectability generally increases from compression to expansion, especially at higher coverage levels. Some zero-shot detectors remain weak or nearly flat across operations. Documentlevel results show a similar but less uniform trend; full results are provided in Appendix H.2. These results show that edit operation affects detectability even when AI coverage is fixed. Cumulative versus independent editing. Finally, we compare the main cumulative trajectory with an independent-editing setting, where each version is edited directly from the original human source. This preserves the same version schedule while removing accumulated re-editing. As shown in Appendix H.3, both sentence- and document-level results remain non-monotonic, including a visible drop around the compression step. The independent trajectories are generally smoother and the drop is less pronounced, indicating that cumulative revision can amplify, but does not fully account for, the observed instability.
6
Conclusion
We introduced OpAI-Bench, an operation-guided benchmark for evaluating AI-text detection under progressive human–AI revision. Unlike endpoint-based benchmarks, OpAI-Bench preserves intermediate revision states and multi-granularity provenance, enabling detection to be studied as a trajectory rather than a single binary decision. Our results show that AI-text detectability is not governed by AI coverage alone. Across document, sentence, token, and span granularities, detector behavior depends strongly on edit operation, domain, generator, and detector family. In particular, mixed-authorship intermediate versions can be harder to detect than both human-written and heavily AI-edited endpoints, revealing non-monotonic failure modes that static evaluations miss. Controlled ablations further show that compression is often harder to detect than expansion at matched coverage. These findings suggest that reliable AI-text detection requires moving beyond endpoint human-versusAI classification toward trajectory-aware and operation-aware evaluation. OpAI-Bench provides a controlled testbed for this setting and highlights the need for detectors that can localize and reason about partial, incremental, and operation-specific AI involvement. 9
Acknowledgments This work is supported by the United Al Saqer Group Grant.
References [1] Rizky Adi, Bassamtiano Renaufalgi Irnawan, Yoshimi Suzuki, and Fumiyo Fukumoto. Gl-clic: Global-local coherence and lexical complexity for sentence-level ai-generated text detection. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 3600–3617, 2025. [2] Ekaterina Artemova, Jason S Lucas, Saranya Venkatraman, Joo-Young Lee, Sergei Tilga, Adaku Uchendu, and Vladislav Mikhailov. Beemo: Benchmark of expert-edited machine-generated outputs. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6992–7018, 2025. [3] Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature. arXiv preprint arXiv:2310.05130, 2023. [4] Desklib. Desklib AI Text Detector v1.01. Hugging Face model, 2024. URL https:// huggingface.co/desklib/ai-text-detector-v1.01. Fine-tuned DeBERTa-v3-large for AI-generated text detection. Accessed: 2026-05-04. [5] Liam Dugan, Alyssa Hwang, Filip Trhlík, Andrew Zhu, Josh Magnus Ludan, Hainiu Xu, Daphne Ippolito, and Chris Callison-Burch. Raid: A shared benchmark for robust evaluation of machine-generated text detectors. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12463–12492, 2024. [6] Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. arXiv preprint arXiv:2301.07597, 2023. [7] Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. Mgtbench: Benchmarking machine-generated text detection. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 2251–2265, 2024. [8] Yongxin He, Shan Zhang, Yixuan Cao, Lei Ma, and Ping Luo. Detree: Detecting humanai collaborative texts via tree-structured hierarchical representation learning. arXiv preprint arXiv:2510.17489, 2025. [9] Xiaomeng Hu, Pin-Yu Chen, and Tsung-Yi Ho. Radar: Robust ai-text detection via adversarial learning. Advances in neural information processing systems, 36:15077–15095, 2023. [10] Luyang Huang, Shuyang Cao, Nikolaus Parulian, Heng Ji, and Lu Wang. Efficient attentions for long document summarization. In Proceedings of the 2021 conference of the north American chapter of the association for computational linguistics: Human language technologies, pages 1419–1436, 2021. [11] Lei Jiang, Desheng Wu, and Xiaolong Zheng. Sendetex: Sentence-level ai-generated text detection for human-ai hybrid content via style and context fusion. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5287–5302, 2025. Learning agency lab – automated essay scoring 2.0. [12] Learning Agency Lab. Kaggle competition, 2024. URL https://www.kaggle.com/competitions/ learning-agency-lab-automated-essay-scoring-2. Accessed: 2026-05-07. [13] Eric Lei, Hsiang Hsu, and Chun-Fu Chen. Pald: Detection of text partially written by large language models. In The Thirteenth International Conference on Learning Representations, 2025. 10
[14] Dominik Macko, Robert Moro, Adaku Uchendu, Jason Lucas, Michiharu Yamashita, Matúš Pikuliak, Ivan Srba, Thai Le, Dongwon Lee, Jakub Simko, et al. Multitude: Large-scale multilingual machine-generated text detection benchmark. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9960–9987, 2023. [15] Shashi Narayan, Shay B Cohen, and Mirella Lapata. Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 1797–1807, 2018. [16] Sayak Paul. arxiv paper abstracts. Kaggle dataset, 2021. URL https://www.kaggle.com/ datasets/spsayakpaul/arxiv-paper-abstracts. Accessed: 2026-04-17. [17] Shoumik Saha and Soheil Feizi. Almost ai, almost human: The challenge of detecting aipolished writing. In Findings of the Association for Computational Linguistics: ACL 2025, pages 25414–25431, 2025. [18] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. [19] Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov. Detectllm: Leveraging log rank information for zero-shot detection of machine-generated text. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12395–12412, 2023. [20] Zhixiong Su, Yichen Wang, Herun Wan, Zhaohan Zhang, and Minnan Luo. Haco-det: A study towards fine-grained machine-generated text detection under human-ai coauthoring. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 22015–22036, 2025. [21] LDM S Sai Teja, Annepaka Yadagiri, Partha Pakray, Chukhu Chunka, and Mangadoddi Srikar Vardhan. Fine-grained detection of ai-generated text using sentence-level segmentation. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 814–828, 2025. [22] LDM S Sai Teja, N Siva Gopala Krishna, Ufaq Khan, Muhammad Haris Khan, and Atul Mishra. Damasha: Detecting ai in mixed adversarial texts via segmentation with human-interpretable attribution. In Findings of the Association for Computational Linguistics: EACL 2026, pages 6189–6206, 2026. [23] Katherine Thai, Bradley Emi, Elyas Masrour, and Mohit Iyyer. Editlens: Quantifying the extent of ai editing in text. arXiv preprint arXiv:2510.03154, 2025. [24] Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson, Aleksandr Gordeev, Vladimir Dokholyan, and Maksim Kuprashevich. Gigacheck: Detecting llm-generated content. arXiv preprint arXiv:2410.23728, 2024. [25] Adaku Uchendu, Zeyu Ma, Thai Le, Rui Zhang, and Dongwon Lee. Turingbench: A benchmark environment for turing test in the age of neural text generation. In Findings of the association for computational linguistics: EMNLP 2021, pages 2001–2016, 2021. [26] Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533, 2022. [27] Pengyu Wang, Linyang Li, Ke Ren, Botian Jiang, Dong Zhang, and Xipeng Qiu. Seqxgpt: Sentence-level ai-generated text detection. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1144–1156, 2023. [28] Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Osama Mohammed Afzal, Tarek Mahmoud, Giovanni Puccetti, Thomas Arnold, et al. M4gtbench: Evaluation benchmark for black-box machine-generated text detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3964–3992, 2024. 11
[29] Yuxia Wang, Jonibek Mansurov, Petar Ivanov, Jinyan Su, Artem Shelmanov, Akim Tsvigun, Chenxi Whitehouse, Osama Mohammed Afzal, Tarek Mahmoud, Toru Sasaki, et al. M4: Multi-generator, multi-domain, and multi-lingual black-box machine-generated text detection. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1369–1407, 2024. [30] Junchao Wu, Runzhe Zhan, Derek F Wong, Shu Yang, Xinyi Yang, Yulin Yuan, and Lidia S Chao. Detectrl: Benchmarking llm-generated text detection in real-world scenarios. Advances in Neural Information Processing Systems, 37:100369–100401, 2024. [31] Cong Zeng, Shengkun Tang, Yuanzhou Chen, Zhiqiang Shen, Wenchao Yu, Xujiang Zhao, Haifeng Chen, Wei Cheng, and Zhiqiang Xu. Human texts are outliers: Detecting llm-generated texts via out-of-distribution detection. arXiv preprint arXiv:2510.08602, 2025. [32] Qihui Zhang, Chujie Gao, Dongping Chen, Yue Huang, Yixin Huang, Zhenyang Sun, Shilin Zhang, Weiye Li, Zhengyan Fu, Yao Wan, et al. Llm-as-a-coauthor: Can mixed human-written and machine-generated text be detected? In Findings of the Association for Computational Linguistics: NAACL 2024, pages 409–436, 2024. [33] Zhongping Zhang, Wenda Qin, and Bryan Plummer. Machine-generated text localization. In Findings of the Association for Computational Linguistics: ACL 2024, pages 8357–8371, 2024.
12
Appendix A
Limitations
OpAI-Bench provides a controlled benchmark for studying progressive human-to-AI text transformation, but it still has several limitations. First, the revision trajectories are constructed under predefined AI coverage levels and edit operations, which may not fully capture the diversity and unpredictability of real-world human–AI writing workflows. Second, the benchmark focuses on a fixed set of domains, detectors, and AI editing models, so conclusions may not directly generalize to unseen domains, emerging writing assistants, or future detector architectures. Third, preserving fine-grained authorship provenance requires controlled generation and alignment procedures, which may introduce artifacts that differ from naturally occurring collaborative writing.
B
Societal Impacts
This work can have positive societal impacts by improving the transparency and accountability of AI-assisted writing. By analyzing how AI authorship signals evolve under progressive editing, OpAIBench can support the development of more reliable detection and provenance tools for education, publishing, journalism, and research integrity. Its multi-granularity annotations may also help move AI-text detection beyond coarse document-level judgments toward more interpretable evidence, reducing the risk of unsupported accusations when only small portions of a text are AI-edited. The study may also have negative societal impacts. More capable detection benchmarks could be misused to train stronger evasion strategies, enabling users to deliberately rewrite AI-generated content to bypass detectors. Detection tools developed from such benchmarks may also be overrelied upon in high-stakes settings, despite known risks of false positives, domain bias, and uneven performance across writing styles, languages, and user populations. Therefore, OpAI-Bench should be used as an evaluation resource rather than as a definitive authority for authorship judgment, and any deployment of AI-text detection should include human review and clear limitations.
C
Ethics statement.
This work follows the NeurIPS Code of Ethics. OpAI-Bench is constructed from cited source datasets and model-generated revisions, and does not involve new human-subject experiments or the collection of private user data. The released benchmark and code are intended to support reproducible research on AI-text detection under progressive editing scenarios.
D
Detector descriptions
We evaluate the following AI-text detectors. 1. AdaLoc [33] is a sentence-level classifier over RoBERTa-large with a sliding window of three adjacent sentences: every window position emits an AI/human label, and overlapping scores are averaged into one prediction per sentence. 2. RADAR [9] is a Vicuna-7B classifier trained adversarially against a paraphraser: the paraphraser produces hard negatives during training, pushing the classifier to learn signals that survive paraphrasing. 3. Fast-DetectGPT [3] is zero-shot and scores a candidate by its conditional probability curvature: the likelihood of the observed tokens under a scoring LM is compared to the average likelihood of perturbed alternatives drawn from a sampling LM. AI text has higher curvature than human text and the whole score is computed in a single forward pass. 4. DAMASHA [22] is a token-level CRF tagger over a dual encoder. RoBERTa-base and ModernBERT-base read the same input; their hidden states are fused by an Info-Mask layer driven by simple stylistic features, and the CRF decodes the per-token AI/human tag sequence. 13
5. GigaCheck [24] is a DETR-style span detector on top of Mistral-7B: the LM encodes tokens and a DETR decoder predicts a fixed-size set of character intervals, each labelled AI or human, alongside a coarse document-level head. 6. Desklib [4] is a single-transformer document-level classifier; we use the public Hugging Face release out of the box, with no further training. 7. E5-small [26] is an E5-small encoder with a LoRA adapter trained for AI-text classification by the original authors. We use the public weights directly as a document-level binary classifier. 8. OOD-LLM-Detect [31] treats AI-text detection as one-class classification: a Deep SVDD model is fitted to language-model embeddings of human text only, and a candidate is scored by its distance from the learnt human-text region. 9. RoBERTa-OpenAI [18] is RoBERTa-base fine-tuned by OpenAI on GPT-2 outputs; we use the released document-level binary classifier as is. 10. DetectLLM [19] is zero-shot and combines two ranking statistics under a single reference causal LM: the log-rank ratio (LRR) and the normalised perturbation rank (NPR). Both capture how unusually high-ranked the observed tokens are under the reference distribution. 11. GL-CLiC [1] is a sentence-level classifier whose feature vector concatenates a DeBERTa contextual embedding, per-sentence global–local coherence scores, and per-sentence lexical complexity statistics; the resulting features are passed through a small classification head. 12. SeqXGPT [27] represents each token by its log-probability under four reference LMs (gpt2-xl, gpt-neo-2.7B, gpt-j-6B, llama-7B), yielding a (T, 4) feature matrix. A small CNN + Transformer + CRF stack reads this matrix and emits per-word labels, which are aggregated to sentence level. 13. GPT-5.4 (reasoning level: none) is an API-based judge prompted with the candidate document and asked to return a per-sentence AI/human label list directly. 14. Gemini 3 Flash (thinking level: minimal) is an API-based judge prompted with the candidate document and asked to return a per-sentence AI/human label list directly. 15. Claude Haiku 4.5 (reasoning level: minimal) is an API-based judge prompted with the candidate document and asked to return a per-sentence AI/human label list directly. 16. GenAI-Sentence [21] is a token-level CRF tagger: a DeBERTa backbone feeds a BiGRU encoder, a linear classifier, and a CRF decoder that emits per-token AI/human labels; sentence labels are obtained by aggregation.
E
Implementation Details
E.1
Text Normalization and Segmentation
All source documents are normalized prior to processing: line endings are standardized, consecutive blank lines are collapsed to a single paragraph boundary, and leading and trailing whitespace is removed. Documents are segmented into paragraphs by splitting on blank lines, and each paragraph is assigned a stable identifier preserved across all derived versions. Sentence segmentation is performed using NLTK’s sent_tokenize, applied independently to each paragraph. Sentences receive stable identifiers assigned sequentially across paragraphs, used for sentence selection, provenance tracking, and document reconstruction across all nine versions. E.2
Deterministic Sentence Selection
For each source document D with N sentences, a single fixed ordering π(D) = (π1 , . . . , πN ) is computed once and reused across all versions. The shuffle seed is derived deterministically from the document identifier, ensuring reproducibility and content-agnostic ordering. Given target coverage (t) csent ∈ [0, 1] at version t, the number of sentences selected is: l m (t) k (t) = csent · N . 14
(1)
The edited set is constructed cumulatively: previously selected sentences are always retained, and new sentences are drawn from π(D) to reach k (t) , guaranteeing S (0) ⊆ S (1) ⊆ · · · ⊆ S (8) . E.3
Token- and Span-Level Provenance
Token- and span-level provenance labels are derived deterministically from the editing process, not from post-hoc alignment or model-provided markers, ensuring annotations are exact and reproducible. Span derivation. During construction, each sentence selected for AI rewriting is wrapped in XMLstyle delimiters before being passed to the LLM, and the returned rewrite is stored with those (t) delimiters intact. After each version is constructed, the tagged document Dtagged is parsed to extract (t)
character-level AI spans Achar = {(aℓ , bℓ )}L ℓ=1 , where aℓ and bℓ are character offsets in the clean text D(t) . Span boundaries reflect precisely which characters were produced by the LLM rewriting step. Word-level projection. Word-level labels are projected from character spans via whitespace tokenization. For word wj with character span [sj , ej ]: h i (t) (t) yj = 1 ∃ (aℓ , bℓ ) ∈ Achar s.t. max(sj , aℓ ) < min(ej , bℓ ) .
(2)
Consecutive AI-labeled words are merged into contiguous spans. Because the editing unit is the sentence, each AI span corresponds to one rewritten sentence; boundaries therefore reflect sentence boundaries rather than intra-sentence partial rewrites.
F
Benchmark Statistics Domain
Generator
# Traj.
# Versioned Samples
Avg. sents
Avg. tokens
Student essays
GPT-5.4 GPT-5.4-nano Gemini-2.5-flash
1,969 1,968 3,969
17,721 17,712 35,721
20.9 20.9 21.0
383.6 397.1 407.2
News articles
GPT-5.4 GPT-5.4-nano Gemini-2.5-flash
1,904 1,998 3,990
17,136 17,982 35,910
24.0 24.0 23.9
477.0 479.4 504.1
Government reports
GPT-5.4 GPT-5.4-nano Gemini-2.5-flash
2,000 2,000 4,000
18,000 18,000 36,000
20.5 20.6 20.6
539.4 552.5 581.4
Scientific abstracts
GPT-5.4 GPT-5.4-nano Gemini-2.5-flash
1,764 1,764 3,763
15,876 15,876 33,860
11.0 11.0 11.0
224.1 227.7 242.2
31,089
279,794
19.3
426.1
Total
Table 5: Main split statistics by domain and generator. Trajectories are generator-specific revision sequences initialized from human-written v0 texts, and versioned samples count all versions from v0 to v8.
Domain
Train
Development
Test
Total
Student essays News articles Government reports Scientific abstracts
49,149 50,688 50,049 46,145
11,205 10,566 10,881 9,711
10,800 9,774 11,070 9,756
71,154 71,028 72,000 65,612
Total
196,031
42,363
41,400
279,794
Table 6: Train, development, and test row counts for the primary-generator benchmark subset, broken down by domain. Counts are reported over GPT-5.4, GPT-5.4-nano, and Gemini-2.5-Flash.
15
F.1
Detailed Results by Domain and Generator
This appendix provides the full per-domain and per-generator results that complement the aggregate analyses in the main text. For each evaluation granularity, we report both accuracy and F1-AI. Accuracy captures overall classification behavior, while F1-AI focuses on the detector’s ability to identify AI-edited content, which is especially important under mixed-authorship settings where the class distribution changes across versions. F.1.1
Document-Level Results
Figures 4 and 5 report document-level accuracy and F1-AI, respectively, broken down by domain and generator. These figures show how zero-shot document-level detectors behave across the full revision trajectory. They also expose detector-specific and domain-specific variation that is hidden in the aggregate table, including cases where recovery at high AI coverage occurs only for particular generator–domain combinations. F.1.2
Sentence-Level Results
Figures 6 and 7 report sentence-level accuracy and F1-AI for zero-shot and LLM-as-detector methods. Figures 8 and 9 additionally include fine-tuned sentence-level detectors. These breakdowns allow us to compare how detector families respond to progressive rewriting across domains and generators, and to assess whether sentence-level supervision improves stability under mixed-authorship revisions. F.1.3
Token- and Span-Level Results
Figures 10 and 11 report fine-grained token/span-level performance by domain and generator. DAMASHA operates at the token level, while GigaCheck produces span-level predictions. These results complement the coarser document- and sentence-level analyses by showing how localization performance changes as AI-marked regions become denser across the revision trajectory.
G
Aggregated Results
This appendix provides extended aggregate tables averaged over generators. Unlike Appendix F.1, which shows full domain–generator breakdowns, the tables below aggregate over generators and report results by domain and revision version. These tables provide a compact view of detector behavior across the full trajectory and include additional metrics beyond those shown in the main text. G.1
Sentence Level
Table 8 reports sentence-level accuracy and F1-AI for all sentence-level detectors, averaged over generators. The table includes zero-shot detectors, LLM-as-detectors, and fine-tuned sentence-level models, enabling direct comparison across detector families. Table 9 provides the corresponding extended sentence-level metrics, including Macro-F1 and false negative rate. These additional metrics help characterize whether detectors fail by missing AI-edited content or by over-predicting the AI class. G.2
Token and Span Level
Table 10 reports fine-grained token/span-level accuracy and F1-AI, averaged over generators. Table 11 provides the corresponding extended metrics. DAMASHA is evaluated at the token level, while GigaCheck is evaluated at the span level. These results complement the document- and sentence-level tables by evaluating whether detectors can localize AI-edited regions rather than only assign a global label. Because early revision versions contain relatively few AI-marked words or spans, accuracy can remain high even when the detector misses AI-edited regions. F1-AI is therefore particularly important for interpreting fine-grained performance, as it directly measures recovery of the AI-labeled class. The extended metrics further help distinguish between detectors that remain conservative and miss AI edits, and detectors that recover more AI-marked regions as coverage increases. 16
H
Ablations
We provide three controlled ablations to isolate the factors that vary jointly in the main trajectory. Ablation 1 fixes the edit operation and varies AI coverage, testing whether detector behavior is driven by coverage alone. Ablation 2 fixes AI coverage and varies the edit operation, testing whether different operations remain distinguishable at matched coverage. Ablation 3 compares cumulative editing with independent editing from the original source, testing the effect of accumulated revision history. H.1
Ablation 1: Coverage-controlled Edit Operations
This ablation fixes the edit operation and varies the target AI coverage. Sentence-level accuracy and F1-AI are shown in Figures 12 and 13, respectively. Document-level accuracy and F1-AI are shown in Figures 14 and 15, respectively. H.2
Ablation 2: Fixed-coverage Edit Operations
This ablation isolates the effect of edit operation at a fixed amount of AI editing. For each target AI coverage level separately (25%, 50%, and 75%), we hold coverage constant and vary only the edit operation among compress, paraphrase, and expand. Sentence-level accuracy and F1-AI are shown in Figures 16 and 17, respectively. Document-level accuracy and F1-AI are shown in Figures 18 and 19, respectively. H.3
Ablation 3: Independent versus Cumulative Editing
This ablation compares the main cumulative trajectory with an alternative setting in which each version is edited independently from the original human source. Sentence-level results are shown in Figure 20.
I
Experimental Details
For clarity, we summarize the label notation used across the experimental settings. For a document (t) trajectory T (D) = (D(0) , . . . , D(8) ), document-level labels are ydoc = 1[t > 0]. For sentence-level (t) attribution, the label of sentence i at version t is yi = 1[i ∈ S (t) ], where S (t) is the cumulative set of sentences edited by version t. For token-level localization, tokens refer to word-level units: (t) (t) (t) (t) Ytok = (y1 , . . . , yM ), with yj ∈ {0, 1} indicating whether word j belongs to an AI-edited span. The realized token coverage is M 1 X (t) (t) y . (3) ctok = M j=1 j (t)
Span-level annotations are represented as character spans Achar = {(aℓ , bℓ )}L ℓ=1 , where aℓ and bℓ (t) are character offsets in D . I.1
Zero-Shot Detectors
For zero-shot detectors, we use either released checkpoints or models reproduced from the original training scripts, following the inference settings provided by the original papers or released implementations. These methods are evaluated directly on OpAI-Bench without task-specific training or threshold tuning on the benchmark. Depending on the detector output, we evaluate document-level predictions, sentence-level predictions, or fine-grained token and span predictions under the protocol described in the main text. I.2
Language Models as Detectors
We evaluate frontier language models as prompted sentence-level detectors. For each document version, the text is split into numbered sentences and passed to the model with a shared prompt 17
Detector
Backbone
LoRA targets
AdaLoc GenAI-Sentence SeqXGPT DAMASHA GigaCheck
RoBERTa-large (openai-detector) DeBERTa-v3-base CNN+Transformer+CRF over 4-LM feats. RoBERTa-base + ModernBERT-base Mistral-7B-v0.3
q, v 10−5 32 query_proj, value_proj 10−5 32 −5 — 5×10 32 q, v / Wqkv 2×10−4 16 q_proj, v_proj 3×10−5 4×ga 4
LR
Batch
Epochs Setting 2 2 20 5 5
Adam AdamW, bf16 weight decay 0.1 AdamW, wd 0.01, ext. CRF NLL bf16, DeepSpeed ZeRO-2
Table 7: Training hyperparameters for the OpAI-Bench-trained detector variants. The LoRA targets column lists the modules to which the LoRA adapter is attached; entries marked “–” indicate full fine-tuning of the listed module set, with no LoRA. Rank r=8, α=16, dropout 0.1 throughout.
template. The model returns a binary label and a confidence score for each sentence. We use the returned binary labels for sentence-level evaluation. Sentence-Level LLM-as-Detector Prompt User prompt You are an expert linguist and writing analyst specializing in distinguishing human-written text from AI-generated text. The following text has been split into numbered sentences. For EACH sentence, classify it as human-written (0) or AI-generated (1), and estimate the probability that it is AI-generated. Text: """ [numbered_sentences] """ Respond in JSON format: {"labels": [0, 1, ...], "confidences":
[0.1, 0.9, ...]}
- labels: array of integers (0 = human-written, 1 = AI-generated), one per sentence. - confidences: array of floats (0.0 to 1.0), one per sentence. - Both arrays must contain exactly [num_sentences] elements, in sentence order. - Do not include any other keys or text outside the JSON object.
I.3
Training of OpAI-Bench detector variants
For the trained setting we re-train (or LoRA-fine-tune) the architecture of each method on the OpAI-Bench training split from the three in-distribution generators (GPT-5.4, GPT-5.4-nano, Gemini2.5-Flash). Qwen3-8B is excluded from training and used only for cross-generator evaluation. For GPT-5.4-nano, Gemini-2.5-Flash we sample a subset for data-balance. Sentence-level models consume sentence provenance labels, document-level labels are derived from whether a version contains AI-edited content, and fine-grained models consume the token or span annotations available in the benchmark. Models are selected on the development split and evaluated on the test split. We follow the architecture, optimiser, and trainable-parameter choices of each method’s original implementation; deviations are listed in the Setting column of Table 7. All LoRA-tuned variants share a single adapter configuration: rank r=8, α=16, dropout 0.1 — so the table reports only the target modules. Backbone weights are kept frozen, while the task-specific heads (classifier, BiGRU, CRF, fusion module, etc.) are trained in full. A linear schedule with 10% warmup is used everywhere. GLCLIC. GLCLIC is trained with the official upstream pipeline at its released default configuration; we do not override any of its hyperparameters. Compute resources. All local detector training, inference, and aggregation experiments were run on an internal Linux server equipped with NVIDIA GeForce RTX 4090 GPUs (24GB memory each). Lightweight preprocessing, metric aggregation, and table generation were run on CPU on the same server. API-based dataset generation and LLM-as-detector experiments used external model APIs, 18
Polish Prompt System instruction You are a precise text rewriting assistant. You must output ONLY the rewritten target sentence. Return ONLY the rewritten text –- nothing else. STRICT OUTPUT RULES: - Do NOT start with phrases like ’Here is’, ’Here’s’, ’Sure’, ’Certainly’, ’Of course’, ’Rewritten:’, ’Revised:’, ’Result:’. - Do NOT explain what you did. - Do NOT add quotes around your output. - Do NOT add any prefix or suffix. - Your entire response = the rewritten text only. - Preserve meaning. Do NOT add new facts. - Preserve all names, numbers, dates, locations, and other entities exactly. - Do NOT introduce any new named entities, numbers, or specific claims. - Your task is proofreading –- fix errors and improve fluency only. User prompt Operation:
polish
Constraints: - Rewrite ONLY the target sentence. - Do NOT add new facts. - Preserve names, numbers, and entities. - Keep the same language as the target. - Return ONLY the rewritten text. No explanations, no quotes, no prefixes. - Length constraint: 85%–100% of original word count. - No line breaks. - Output EXACTLY ONE sentence. Guidance: Make light edits to improve grammar, punctuation, fluency, and clarity while preserving meaning. Keep the sentence structure mostly unchanged. The output must not be identical to the input. Keep it as exactly one sentence. Do not add new facts. Paragraph context (for coherence only; do not rewrite it): [context] Target sentence: [sentence]
with local compute used for request orchestration, parsing, and metric computation. Preliminary checks used the same compute environment and are not separately reported.
J
Editing Prompts
The following prompts are used to generate the five editing operations in OpAI-Bench. In all cases, the model is instructed to rewrite only the target sentence, preserve factual content and named entities, and return exactly one sentence.
19
Paraphrase Prompt System instruction You are a precise text rewriting assistant. You must output ONLY the rewritten target sentence. Return ONLY the rewritten text –- nothing else. STRICT OUTPUT RULES: - Do NOT start with phrases like ’Here is’, ’Here’s’, ’Sure’, ’Certainly’, ’Of course’, ’Rewritten:’, ’Revised:’, ’Result:’. - Do NOT explain what you did. - Do NOT add quotes around your output. - Do NOT add any prefix or suffix. - Your entire response = the rewritten text only. - Preserve meaning. Do NOT add new facts. - Preserve all names, numbers, dates, locations, and other entities exactly. - Do NOT introduce any new named entities, numbers, or specific claims. - Your task is paraphrasing –- same meaning, different wording. User prompt Operation:
paraphrase
Constraints: - Rewrite ONLY the target sentence. - Do NOT add new facts. - Preserve names, numbers, and entities. - Keep the same language as the target. - Return ONLY the rewritten text. No explanations, no quotes, no prefixes. - Length constraint: 85%–115% of original word count. - No line breaks. - Output EXACTLY ONE sentence. Guidance: Rewrite the sentence with clearly different wording while preserving meaning. You may restructure the phrasing, but keep the same core content. The output must not be identical to the input. Keep it as exactly one sentence. Do not add new facts. Paragraph context (for coherence only; do not rewrite it): [context] Target sentence: [sentence]
20
Style Rewrite Prompt System instruction You are a precise text rewriting assistant. You must output ONLY the rewritten target sentence. Return ONLY the rewritten text –- nothing else. STRICT OUTPUT RULES: - Do NOT start with phrases like ’Here is’, ’Here’s’, ’Sure’, ’Certainly’, ’Of course’, ’Rewritten:’, ’Revised:’, ’Result:’. - Do NOT explain what you did. - Do NOT add quotes around your output. - Do NOT add any prefix or suffix. - Your entire response = the rewritten text only. - Preserve meaning. Do NOT add new facts. - Preserve all names, numbers, dates, locations, and other entities exactly. - Do NOT introduce any new named entities, numbers, or specific claims. - Your task is style transfer –- same meaning, different tone/voice. User prompt Operation:
style
Constraints: - Rewrite ONLY the target sentence. - Do NOT add new facts. - Preserve names, numbers, and entities. - Keep the same language as the target. - Return ONLY the rewritten text. No explanations, no quotes, no prefixes. - Length constraint: 85%–115% of original word count. - No line breaks. - Output EXACTLY ONE sentence. Guidance: Rewrite the sentence in a formal, natural, human-written style by changing tone, register, and phrasing while preserving meaning. Do not change the underlying facts or add new content. The output must not be identical to the input. Keep it as exactly one sentence. Paragraph context (for coherence only; do not rewrite it): [context] Target sentence: [sentence]
21
Compress Prompt System instruction You are a precise text rewriting assistant. You must output ONLY the rewritten target sentence. Return ONLY the rewritten text –- nothing else. STRICT OUTPUT RULES: - Do NOT start with phrases like ’Here is’, ’Here’s’, ’Sure’, ’Certainly’, ’Of course’, ’Rewritten:’, ’Revised:’, ’Result:’. - Do NOT explain what you did. - Do NOT add quotes around your output. - Do NOT add any prefix or suffix. - Your entire response = the rewritten text only. - Preserve meaning. Do NOT add new facts. - Preserve all names, numbers, dates, locations, and other entities exactly. - Do NOT introduce any new named entities, numbers, or specific claims. - Your task is compression –- shorter text, same meaning, no facts lost. User prompt Operation: compress Style target (if applicable):
formal, natural, human-written
Constraints: - Rewrite ONLY the target sentence. - Do NOT add new facts. - Preserve names, numbers, and entities. - Keep the same language as the target. - Return ONLY the rewritten text. No explanations, no quotes, no prefixes. - Length constraint: 60%–80% of original word count. - No line breaks. - Output EXACTLY ONE sentence. Guidance: Rewrite the sentence to be shorter and more concise while preserving all essential meaning. Remove redundancy and non-essential phrasing, but do not omit important information. The output must not be identical to the input. Keep it as exactly one sentence. Do not add new facts. Paragraph context (for coherence only; do not rewrite it): [context] Target sentence: [sentence]
22
Expand Prompt System instruction You are a precise text rewriting assistant. You must output ONLY the rewritten target sentence. Return ONLY the rewritten text –- nothing else. STRICT OUTPUT RULES: - Do NOT start with phrases like ’Here is’, ’Here’s’, ’Sure’, ’Certainly’, ’Of course’, ’Rewritten:’, ’Revised:’, ’Result:’. - Do NOT explain what you did. - Do NOT add quotes around your output. - Do NOT add any prefix or suffix. - Your entire response = the rewritten text only. - Preserve meaning. Do NOT add new facts. - Preserve all names, numbers, dates, locations, and other entities exactly. - Do NOT introduce any new named entities, numbers, or specific claims. - Your task is expansion –- extend current text. User prompt Operation: expand Style target (if applicable):
formal, natural, human-written
Constraints: - Rewrite ONLY the target sentence. - Do NOT add new facts. - Preserve names, numbers, and entities. - Keep the same language as the target. - Return ONLY the rewritten text. No explanations, no quotes, no prefixes. - Length constraint: 120%–150% of original word count. - No line breaks. - Output EXACTLY ONE sentence. Guidance: Rewrite the sentence to be slightly more detailed using only information already stated or directly implied. You may add brief clarifying or descriptive phrasing, but do not introduce new facts, examples, names, dates, numbers, or claims. The output must not be identical to the input. Keep it as exactly one sentence. Paragraph context (for coherence only; do not rewrite it): [context] Target sentence: [sentence]
23
Essays
Reports
Abstracts
News
75 50 25 0 100
GPT-5.4
75 50
0 100
Generators GPT-5.4-nano
Accuracy (%)
25
75 50 25 0 100
Qwen3-8b
75 50 25 0
Gemini-2.5-flash
100
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version DeskLib DetectLLM
E5-Small Fast-DetGPT
OOD-LLM-Detect RADAR
RoBERTa GigaCheck
Figure 4: Document-level accuracy broken down by domain and generator.
Essays
Reports
Abstracts
News
75 50 25 0 100
GPT-5.4
75 50
0 100
Generators GPT-5.4-nano
F1 (%)
25
75 50 25 0 100
Qwen3-8b
75 50 25 0
Gemini-2.5-flash
100
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version DeskLib DetectLLM
E5-Small Fast-DetGPT
OOD-LLM-Detect RADAR
RoBERTa GigaCheck
Figure 5: Document-level F1-AI across revision versions, domains, and generators. Each curve represents a document-level detector, showing how AI-targeted detection performance changes along the progressive human-to-AI revision trajectory.
24
News
Reports
Abstracts
Accuracy (%)
GPT-5.4-nano
Generators
Qwen3-8b
100 75 50 25 0 100 75 50 25 0
Gpt-5.4
100 75 50 25 0
Gemini-2.5-flash
Essays 100 75 50 25 0
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version
GenAI-Sentence GL-CLiC
AdaLoc SeqXGPT
Claude-Haiku Gemini-Flash
GPT-5.4
Figure 6: Sentence-level accuracy across revision versions, domains, and generators. The figure compares sentence-level detectors across the full revision trajectory, including zero-shot methods and LLM-as-detectors sentence-level models.
News
Reports
Abstracts
F1 (%)
GPT-5.4-nano
Generators
Qwen3-8b
100 75 50 25 0 100 75 50 25 0
Gpt-5.4
100 75 50 25 0
Gemini-2.5-flash
Essays 100 75 50 25 0
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version
GenAI-Sentence GL-CLiC
AdaLoc SeqXGPT
Claude-Haiku Gemini-Flash
GPT-5.4
Figure 7: Sentence-level F1-AI across revision versions, domains, and generators. The figure compares sentence-level detectors across the full revision trajectory, including zero-shot methods and LLM-as-detectors.
25
Essays
News
Reports
Abstracts Gemini-2.5-flash
100 75 50 25 0 100
25 0 100
GPT-5.4-nano
Accuracy (%)
50
75 50 25 0 100
Qwen3-8B
75 50 25 0
Generators
GPT-5.4
75
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version Claude-Haiku Gemini-Flash
GPT-5.4 GenAI-Sentence (FT)
GL-CLiC (FT) AdaLoc (FT)
SeqXGPT (FT)
Figure 8: Sentence-level accuracy across revision versions, domains, and generators, including fine-tuned sentence-level models. The figure highlights how fine-tuned detectors compare with LLM-as-detector methods along the progressive human-to-AI revision trajectory. Essays
News
Reports
Abstracts Gemini-2.5-flash
100 75 50 25 0 100
25 0 100
GPT-5.4-nano
F1 (%)
50
75 50 25 0 100
Qwen3-8B
75 50 25 0
Generators
GPT-5.4
75
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version Claude-Haiku Gemini-Flash
GPT-5.4 GenAI-Sentence (FT)
GL-CLiC (FT) AdaLoc (FT)
SeqXGPT (FT)
Figure 9: Sentence-level F1-AI across revision versions, domains, and generators, including fine-tuned sentence-level models. The figure highlights how fine-tuned detectors compare with LLM-as-detector methods along the progressive human-to-AI revision trajectory.
26
News
Reports
Abstracts
Accuracy (%)
Generators
GPT-5.4-nano Qwen3-8b
100 75 50 25 0 100 75 50 25 0
GPT-5.4
100 75 50 25 0
Gemini-2.5-flash
Essays 100 75 50 25 0
v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version GigaCheck
DAMASHA
Figure 10: Token- and span-level accuracy broken down by domain and generator.
News
Reports
Abstracts
Qwen3-8b v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8 v0 v1 v2 v3 v4 v5 v6 v7 v8
Version GigaCheck
DAMASHA
Figure 11: Token- and span-level F1-AI broken down by domain and generator.
27
F1 (%)
Generators
GPT-5.4-nano
100 75 50 25 0 100 75 50 25 0
GPT-5.4
100 75 50 25 0
Gemini-2.5-flash
Essays 100 75 50 25 0
Detector
Ver.
Essays Acc
F1-AI
Reports Acc
F1-AI
News Acc
F1-AI
Abstracts Acc
F1-AI
LLM-as-detector
Claude-Haiku
v0 v1 v2 v3 v4 v5 v6 v7 v8
91.2 0.0 77.2 -13.9 22.0 +22.0 74.7 -2.5 45.5 +23.5 75.0 +0.3 65.5 +20.0 62.9 -12.1 52.7 -12.9 69.8 +7.0 69.7 +17.1 75.2 +5.4 81.5 +11.7 66.0 -9.2 77.5 -4.0 64.3 -1.7 78.0 +0.5
96.4 79.9 -16.5 69.4 -10.5 56.2 -13.2 47.8 -8.4 48.7 +0.9 41.1 -7.6 34.4 -6.7 30.7 -3.6
Gemini-Flash
v0 v1 v2 v3 v4 v5 v6 v7 v8
99.1 0.0 82.6 -16.5 9.5 +9.5 80.8 -1.8 46.3 +36.8 81.0 +0.2 71.5 +25.2 67.3 -13.7 53.6 -17.9 73.5 +6.3 70.9 +17.2 76.9 +3.4 82.0 +11.1 52.7 -24.2 63.3 -18.6 43.1 -9.6 57.9 -5.4
99.8 0.0 99.8 0.0 83.1 -16.7 0.5 +0.5 83.1 -16.6 1.3 +1.3 73.9 -9.2 8.8 +8.3 76.4 -6.8 20.2 +18.9 63.4 -10.4 19.6 +10.8 68.4 -8.0 30.4 +10.2 51.0 -12.4 7.1 -12.4 53.8 -14.6 18.1 -12.3 57.2 +6.2 37.7 +30.5 58.7 +4.8 43.4 +25.3 42.3 -14.9 30.8 -6.9 49.0 -9.7 45.4 +2.0 30.8 -11.5 33.3 +2.5 33.6 -15.4 39.6 -5.8 22.9 -7.9 30.5 -2.7 25.8 -7.8 37.1 -2.5
86.1 73.7 -12.5 67.4 -6.3 62.7 -4.8 52.3 -10.4 58.8 +6.5 55.1 -3.7 47.3 -7.8 45.1 -2.2
0.0 17.9 +17.9 33.1 +15.2 42.4 +9.2 35.5 -6.9 53.0 +17.5 61.9 +8.9 59.1 -2.9 60.9 +1.8
GPT-5.4
v0 v1 v2 v3 v4 v5 v6 v7 v8
99.7 0.0 82.6 -17.1 4.1 +4.1 77.8 -4.8 29.4 +25.4 75.3 -2.5 55.2 +25.8 59.8 -15.4 34.0 -21.2 66.1 +6.3 58.2 +24.2 67.3 +1.2 71.9 +13.7 45.1 -22.2 56.1 -15.8 38.4 -6.7 54.0 -2.1
97.6 0.0 99.9 0.0 80.7 -16.9 0.0 +0.0 83.1 -16.8 0.2 +0.2 71.7 -9.0 2.3 +2.3 74.7 -8.4 12.5 +12.4 58.3 -13.4 7.4 +5.1 65.6 -9.1 23.9 +11.4 47.8 -10.5 0.5 -6.9 50.9 -14.7 8.1 -15.8 52.3 +4.5 33.0 +32.4 54.8 +3.9 34.0 +26.0 37.6 -14.8 27.9 -5.0 43.5 -11.3 36.0 +1.9 24.5 -13.1 27.5 -0.5 27.3 -16.3 30.4 -5.6 15.3 -9.2 23.5 -4.0 19.5 -7.8 28.7 -1.7
98.6 80.6 -18.0 70.9 -9.6 59.1 -11.8 48.8 -10.3 52.1 +3.2 39.0 -13.0 29.0 -10.0 22.5 -6.5
0.0 2.4 +2.4 8.6 +6.2 13.8 +5.2 7.8 -6.0 32.2 +24.4 33.3 +1.1 33.0 -0.3 33.0 +0.0
AdaLoc [33]
v0 v1 v2 v3 v4 v5 v6 v7 v8
85.1 0.0 72.3 -12.8 20.5 +20.5 66.2 -6.1 23.7 +3.3 57.1 -9.1 30.1 +6.3 50.9 -6.2 30.9 +0.8 43.0 -8.0 27.5 -3.4 34.7 -8.3 30.9 +3.4 28.6 -6.1 37.8 +6.9 29.8 +1.2 44.7 +6.9
94.7 79.0 -15.7 70.5 -8.5 56.9 -13.6 48.1 -8.7 38.6 -9.5 25.4 -13.3 11.6 -13.8 5.0 -6.6
0.0 91.9 0.0 53.2 0.0 7.0 +7.0 78.0 -13.9 11.2 +11.2 50.6 -2.6 25.9 +25.9 11.4 +4.4 70.5 -7.5 20.3 +9.1 48.9 -1.6 37.6 +11.7 17.1 +5.6 61.9 -8.6 31.3 +11.0 51.4 +2.5 51.2 +13.7 13.4 -3.6 51.2 -10.7 23.0 -8.3 49.8 -1.6 52.2 +1.0 34.2 +20.8 52.9 +1.7 40.3 +17.3 60.0 +10.2 67.4 +15.2 37.8 +3.6 46.2 -6.7 45.2 +4.9 66.8 +6.9 77.5 +10.1 42.3 +4.5 33.4 -12.7 41.7 -3.5 69.1 +2.3 80.5 +3.0 43.0 +0.7 27.3 -6.1 40.4 -1.3 71.7 +2.5 83.1 +2.6
Zero-shot
v0 v1 v2 v3 GenAI-Sentence [21] v4 v5 v6 v7 v8
51.7 50.5 -1.2 51.2 +0.8 54.2 +3.0 54.7 +0.5 59.5 +4.8 65.8 +6.3 69.7 +3.9 76.4 +6.7
0.0 7.9 +7.9 7.9 -0.0 8.8 +0.9 6.5 -2.3 7.2 +0.7 7.4 +0.3 8.2 +0.8 9.5 +1.3
0.0 20.6 0.0 28.6 +28.6 31.0 +10.4 29.0 +29.0 40.4 +11.8 36.9 +5.8 40.6 +11.5 55.5 +15.1 46.0 +9.1 55.9 +15.3 57.8 +2.3 50.2 +4.2 62.9 +7.0 68.0 +10.2 57.1 +6.9 69.2 +6.3 76.9 +8.9 66.0 +8.9 78.1 +8.9 81.4 +4.6 74.3 +8.3 84.9 +6.8 86.6 +5.1 80.1 +5.8 88.9 +4.0
86.6 0.0 73.9 -12.7 15.8 +15.8 67.4 -6.5 18.6 +2.8 56.3 -11.1 20.8 +2.2 49.4 -6.9 21.1 +0.3 42.0 -7.4 22.5 +1.4 30.7 -11.3 23.0 +0.4 21.2 -9.5 26.3 +3.4 16.9 -4.3 28.8 +2.4
96.7 0.0 79.6 -17.1 6.1 +6.1 70.3 -9.3 5.7 -0.5 57.2 -13.1 5.2 -0.4 48.4 -8.8 5.1 -0.1 38.3 -10.2 4.9 -0.2 22.3 -16.0 5.1 +0.2 12.1 -10.2 7.1 +2.0 5.6 -6.5 10.6 +3.5
31.4 0.0 25.0 0.0 38.3 +6.9 28.0 +28.0 34.0 +9.0 31.3 +31.3 41.7 +3.4 38.6 +10.6 39.0 +5.0 43.2 +11.9 46.8 +5.1 51.7 +13.1 47.7 +8.7 58.2 +14.9 49.4 +2.6 58.2 +6.5 52.1 +4.4 63.3 +5.1 54.1 +4.7 64.2 +6.0 62.0 +9.9 74.3 +11.1 59.6 +5.5 71.9 +7.7 74.9 +13.0 85.1 +10.8 64.5 +4.9 77.7 +5.8 83.0 +8.0 90.5 +5.4 68.8 +4.2 81.5 +3.8 92.0 +9.0 95.8 +5.3
GL-CLiC [1]
v0 v1 v2 v3 v4 v5 v6 v7 v8
67.6 0.0 62.5 -5.1 28.6 +28.6 58.4 -4.1 30.2 +1.7 53.1 -5.4 37.7 +7.4 48.6 -4.5 37.4 -0.2 44.1 -4.4 39.4 +2.0 37.5 -6.6 40.1 +0.7 32.3 -5.2 44.0 +3.9 33.5 +1.2 50.0 +6.0
49.9 0.0 50.4 +0.5 27.1 +27.1 48.9 -1.5 32.2 +5.1 50.8 +1.9 46.5 +14.4 42.2 -8.6 36.7 -9.9 48.0 +5.8 52.2 +15.5 48.5 +0.6 58.7 +6.6 46.0 -2.5 60.9 +2.1 49.1 +3.1 65.8 +4.9
42.3 0.0 45.4 +3.0 28.0 +28.0 46.1 +0.7 35.5 +7.5 48.7 +2.6 47.7 +12.2 47.8 -0.9 49.5 +1.8 48.3 +0.5 54.4 +4.9 48.4 +0.2 59.5 +5.1 50.7 +2.2 65.5 +6.0 54.6 +3.9 70.4 +4.9
74.2 0.0 66.1 -8.1 26.6 +26.6 60.4 -5.7 27.7 +1.1 53.9 -6.5 33.7 +6.0 46.9 -7.1 29.3 -4.4 43.2 -3.7 35.1 +5.8 34.7 -8.4 37.3 +2.1 30.9 -3.8 40.9 +3.7 29.3 -1.6 45.2 +4.3
SeqXGPT [27]
v0 v1 v2 v3 v4 v5 v6 v7 v8
98.8 81.9 -16.9 73.2 -8.8 58.1 -15.0 48.9 -9.2 38.1 -10.7 23.3 -14.9 8.2 -15.1 0.4 -7.7
100.0 83.2 -16.8 73.8 -9.4 59.1 -14.7 50.1 -9.0 39.7 -10.3 25.2 -14.6 10.3 -14.8 2.6 -7.7
99.5 82.9 -16.6 73.3 -9.6 58.4 -14.9 48.9 -9.5 38.4 -10.5 23.4 -15.0 8.3 -15.1 0.2 -8.2
100.0 81.2 -18.7 71.6 -9.7 57.7 -13.8 48.5 -9.3 37.8 -10.7 20.8 -17.0 9.1 -11.6 0.2 -8.9
0.0 2.1 +2.1 0.6 -1.5 0.2 -0.5 0.3 +0.1 0.3 -0.1 0.1 -0.1 0.4 +0.3 0.9 +0.4
0.0 0.0 +0.0 0.0 +0.0 0.1 +0.1 0.0 -0.1 0.1 +0.0 0.0 -0.1 0.0 +0.0 0.0 +0.0
0.0 1.6 +1.6 0.6 -1.0 0.7 +0.1 0.5 -0.3 0.4 -0.0 0.3 -0.2 0.2 -0.1 0.3 +0.1
0.0 0.1 +0.1 0.2 +0.1 0.3 +0.1 0.2 -0.1 0.3 +0.1 0.1 -0.2 0.1 -0.0 0.4 +0.3
Table 8: Sentence-level Main split results for LLM-as-detectors and zero-shot sentence-level detectors across revision versions and domains, aggregated over the three primary generators. We report accuracy and F1-AI. Colored deltas indicate changes relative to the previous version; no delta is shown for v0 .
28
Detector
Ver.
Essays Macro-F1
Reports FNR
Macro-F1
FNR
News
Abstracts
Macro-F1
FNR
47.9 49.3 +1.4 51.1 +1.8 52.3 +1.2 43.6 -8.7 50.0 +6.5 44.6 -5.4 29.6 -15.0 20.2 -9.4
– 91.8 85.1 -6.6 75.5 -9.6 85.3 +9.8 68.6 -16.7 65.6 -3.0 70.9 +5.4 72.7 +1.7
Macro-F1
FNR
LLM-as-detector
Claude-Haiku
v0 v1 v2 v3 v4 v5 v6 v7 v8
47.7 – 54.3 +6.6 81.2 64.5 +10.2 57.8 -23.4 72.9 +8.4 40.0 -17.8 60.9 -12.0 57.5 +17.5 69.6 +8.7 38.9 -18.6 71.8 +2.1 26.7 -12.2 51.9 -19.9 34.4 +7.6 39.0 -12.8 35.7 +1.3
Gemini-Flash
v0 v1 v2 v3 v4 v5 v6 v7 v8
49.8 – 49.9 – 49.9 – 49.9 +0.2 94.6 45.6 -4.3 99.8 46.0 -3.9 99.3 67.2 +17.3 61.9 -32.7 46.7 +1.1 94.4 -5.4 53.1 +7.1 84.5 -14.8 78.6 +11.3 38.1 -23.8 47.7 +1.0 84.3 -10.0 54.8 +1.7 73.4 -11.1 64.0 -14.6 59.0 +20.9 36.9 -10.8 95.8 +11.5 42.8 -12.0 88.1 +14.7 73.0 +9.0 39.4 -19.6 51.9 +15.0 69.1 -26.7 54.9 +12.1 64.7 -23.4 74.8 +1.7 28.2 -11.2 39.0 -12.9 75.8 +6.8 47.4 -7.5 64.9 +0.2 44.9 -29.9 50.6 +22.4 27.5 -11.4 76.3 +0.4 29.8 -17.6 71.8 +6.9 28.9 -15.9 56.9 +6.2 17.8 -9.7 79.0 +2.8 18.5 -11.2 74.2 +2.5
GPT-5.4
v0 v1 v2 v3 v4 v5 v6 v7 v8
49.9 – 49.4 47.3 -2.7 97.9 44.6 -4.7 58.1 +10.8 78.3 -19.5 42.9 -1.8 69.0 +10.9 55.8 -22.6 40.2 -2.6 52.4 -16.6 76.5 +20.7 32.6 -7.7 64.6 +12.2 52.9 -23.6 47.4 +14.8 65.9 +1.3 41.4 -11.5 35.3 -12.1 39.1 -26.8 58.8 +17.4 22.4 -12.9 27.0 -12.1 61.6 +2.8 11.8 -10.5
AdaLoc [33]
v0 v1 v2 v3 v4 v5 v6 v7 v8
46.0 51.9 +5.9 51.0 -0.8 49.5 -1.5 46.4 -3.1 40.1 -6.3 34.0 -6.1 26.4 -7.7 22.4 -4.0
– 48.6 – 79.3 48.0 -0.6 94.7 80.3 +1.0 45.1 -2.9 95.3 +0.5 77.8 -2.6 40.3 -4.9 95.0 -0.3 78.5 +0.7 35.3 -5.0 96.5 +1.5 82.0 +3.6 30.7 -4.6 96.2 -0.3 79.9 -2.1 22.5 -8.2 96.1 -0.1 75.8 -4.2 11.5 -11.0 95.7 -0.4 70.2 -5.5 4.7 -6.7 95.0 -0.7
v0 v1 v2 v3 GenAI-Sentence [21] v4 v5 v6 v7 v8
34.1 45.4 +11.3 49.6 +4.2 54.2 +4.6 54.4 +0.3 56.3 +1.8 55.4 -0.8 49.4 -6.0 43.3 -6.1
– 17.1 42.4 31.0 +13.9 38.1 -4.3 36.7 +5.7 31.8 -6.2 43.1 +6.5 39.3 +7.4 43.5 +0.4 30.1 -9.1 49.1 +5.5 25.8 -4.3 50.6 +1.5 27.7 +1.9 48.5 -2.0 23.6 -4.1 44.5 -4.1
GL-CLiC [1]
v0 v1 v2 v3 v4 v5 v6 v7 v8
40.3 51.6 +11.2 50.3 -1.2 50.0 -0.4 46.8 -3.1 43.8 -3.1 37.2 -6.6 29.1 -8.1 25.0 -4.1
SeqXGPT [27]
v0 v1 v2 v3 v4 v5 v6 v7 v8
49.7 46.1 -3.6 42.6 -3.5 36.8 -5.7 33.0 -3.9 27.7 -5.3 18.9 -8.8 7.6 -11.3 0.4 -7.2
49.1 – 47.9 -1.2 95.5 46.4 -1.4 92.3 -3.2 43.5 -2.9 87.7 -4.6 38.0 -5.5 91.6 +3.9 44.7 +6.8 70.7 -20.9 38.3 -6.4 69.3 -1.5 28.9 -9.4 68.3 -1.0 21.5 -7.4 68.6 +0.3
34.7 – 44.4 +9.7 53.8 47.2 +2.8 45.8 -8.0 51.2 +4.1 39.0 -6.8 49.6 -1.7 46.5 +7.5 57.4 +7.8 31.6 -15.0 56.3 -1.0 26.3 -5.3 49.0 -7.4 27.3 +1.0 41.5 -7.4 28.3 +1.1 46.3 51.1 +4.8 55.8 +4.7 57.2 +1.5 48.8 -8.4 57.7 +8.9 52.5 -5.2 40.0 -12.6 30.4 -9.5
– 84.6 70.7 -14.0 64.2 -6.5 74.2 +10.1 55.8 -18.4 49.9 -5.9 55.1 +5.2 54.9 -0.2
50.0 – 49.7 45.5 -4.5 99.9 45.8 -3.8 48.9 +3.4 92.1 -7.9 45.6 -0.2 50.7 +1.8 80.4 -11.7 43.4 -2.2 37.2 -13.4 95.4 +15.0 36.2 -7.3 49.2 +11.9 72.1 -23.2 46.9 +10.7 41.3 -7.9 72.9 +0.7 36.8 -10.0 24.6 -16.7 79.0 +6.1 26.0 -10.8 14.3 -10.2 80.5 +1.6 16.5 -9.5
– 98.7 94.9 -3.8 91.3 -3.6 95.8 +4.5 73.6 -22.1 74.6 +0.9 76.9 +2.3 77.5 +0.6
46.4 50.2 +3.8 49.1 -1.1 45.3 -3.8 41.9 -3.3 38.1 -3.9 29.9 -8.1 20.8 -9.2 14.4 -6.4
49.2 47.3 -1.8 44.0 -3.3 38.8 -5.2 34.9 -3.9 29.6 -5.3 19.7 -9.9 11.9 -7.8 5.3 -6.6
– 96.4 96.9 +0.4 97.2 +0.3 97.3 +0.1 97.4 +0.1 97.3 -0.1 96.3 -1.0 94.4 -1.9
– 20.0 28.8 33.9 +13.9 31.0 +2.2 38.6 +4.7 31.6 +0.6 44.1 +5.5 31.1 -0.5 47.2 +3.1 33.1 +2.0 50.3 +3.1 32.5 -0.6 52.5 +2.2 32.6 +0.2 51.8 -0.6 31.2 -1.4 47.9 -3.9
– 19.2 18.2 -1.0 14.0 -4.3 19.9 +6.0 11.4 -8.5 9.4 -2.0 10.2 +0.7 8.0 -2.1
– 33.3 – 29.7 56.5 44.8 +11.5 46.8 42.0 +12.2 66.3 +9.8 45.6 +0.8 54.8 +8.1 44.6 +2.6 65.7 -0.5 50.5 +4.9 49.0 -5.8 48.6 +4.0 69.8 +4.1 41.7 -8.7 67.3 +18.3 47.7 -0.9 70.6 +0.9 47.5 +5.7 53.9 -13.4 47.2 -0.4 72.3 +1.6 45.1 -2.3 52.3 -1.6 43.9 -3.4 71.0 -1.3 36.7 -8.4 54.2 +2.0 39.0 -4.8 66.5 -4.4 32.9 -3.8 50.9 -3.3 35.2 -3.8
– 37.7 44.2 +6.5 43.4 -0.8 49.6 +6.2 49.6 -0.0 50.1 +0.4 48.6 -1.5 45.4 -3.2
– 67.1 73.3 +6.2 72.2 -1.1 78.6 +6.4 75.2 -3.3 75.5 +0.3 73.6 -1.9 70.7 -2.9
– 98.9 99.7 +0.8 99.9 +0.2 99.8 -0.1 99.9 +0.0 99.9 +0.1 99.8 -0.2 99.6 -0.2
– 50.0 – 99.2 44.9 -5.1 100.0 99.7 +0.5 41.8 -3.0 99.9 -0.1 99.6 -0.1 36.8 -5.0 99.8 -0.1 99.8 +0.1 32.7 -4.0 99.9 +0.1 99.8 +0.0 27.5 -5.2 99.9 -0.0 99.9 +0.1 17.2 -10.3 100.0 +0.1 99.9 +0.0 8.4 -8.8 100.0 +0.0 99.8 -0.1 0.2 -8.2 99.8 -0.2
– 100.0 98.8 -1.2 95.7 -3.1 99.7 +4.0 74.0 -25.7 79.6 +5.6 81.3 +1.6 84.5 +3.2
Zero-shot
50.0 45.4 -4.6 42.5 -3.0 37.2 -5.3 33.4 -3.8 28.4 -4.9 20.1 -8.4 9.3 -10.8 2.4 -6.8
– 23.9 18.3 37.0 +13.1 19.8 +1.6 41.6 +4.6 18.6 -1.3 46.2 +4.7 17.6 -1.0 47.1 +0.8 22.0 +4.4 50.0 +2.9 20.8 -1.2 49.8 -0.1 21.2 +0.4 45.5 -4.3 19.9 -1.3 40.7 -4.8
– 100.0 100.0 +0.0 99.9 -0.1 100.0 +0.1 100.0 -0.0 100.0 +0.0 100.0 +0.0 100.0 -0.0
49.9 46.1 -3.8 42.6 -3.5 37.2 -5.4 33.1 -4.1 27.9 -5.1 19.1 -8.9 7.7 -11.3 0.1 -7.6
– 85.7 86.0 +0.3 86.2 +0.2 86.7 +0.5 86.2 -0.5 86.4 +0.2 84.5 -1.9 83.1 -1.4
42.6 52.3 +9.7 50.2 -2.1 49.2 -1.0 43.4 -5.8 42.3 -1.1 34.6 -7.7 28.7 -5.8 22.6 -6.1
Table 9: Extended sentence-level Main split results for LLM-as-detectors and zero-shot sentence-level detectors across revision versions and domains, aggregated over the three primary generators. We report Macro-F1 and FNR. Colored deltas indicate changes relative to the previous version; no delta is shown for v0 .
29
Detector
Ver.
Essays Acc
Reports
F1-AI
Acc
F1-AI
News Acc
Abstracts
F1-AI
Acc
F1-AI
Fine-grained Zero-shot (Token + Span)
GigaCheck [24]
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 DAMASHA [22] v4 v5 v6 v7 v8
98.7 0.0 98.5 82.6 -16.1 7.6 +7.6 81.8 -16.8 74.2 -8.4 19.0 +11.4 71.9 -9.9 66.1 -8.1 39.0 +20.0 56.6 -15.3 58.9 -7.2 26.6 -12.4 52.6 -4.0 54.6 -4.3 44.8 +18.3 38.6 -14.1 56.7 +2.1 62.3 +17.4 28.9 -9.7 41.6 -15.1 53.8 -8.4 17.0 -11.9 51.1 +9.4 66.7 +12.8 11.2 -5.8
0.0 2.6 +2.6 3.3 +0.6 5.7 +2.5 8.7 +3.0 11.5 +2.7 18.3 +6.8 17.7 -0.6 18.1 +0.4
99.5 0.0 97.5 0.0 82.9 -16.6 1.5 +1.5 79.6 -17.9 6.8 +6.8 72.6 -10.4 3.0 +1.5 69.4 -10.1 12.1 +5.3 57.2 -15.3 8.2 +5.2 58.5 -11.0 22.4 +10.3 51.9 -5.3 6.8 -1.4 53.3 -5.2 21.4 -1.0 38.7 -13.3 12.1 +5.3 45.9 -7.4 29.9 +8.5 30.1 -8.6 20.7 +8.7 43.9 -2.0 46.5 +16.6 19.0 -11.1 20.0 -0.7 32.8 -11.1 40.5 -5.9 15.7 -3.3 23.9 +3.8 30.9 -1.9 45.2 +4.7
89.1 0.0 82.9 0.0 98.5 0.0 80.6 0.0 76.6 -12.5 22.3 +22.3 71.5 -11.4 19.0 +19.0 82.2 -16.3 4.2 +4.2 69.4 -11.2 23.4 +23.4 71.7 -4.9 31.8 +9.5 65.8 -5.8 23.1 +4.1 73.3 -9.0 7.6 +3.4 64.6 -4.7 30.9 +7.5 65.5 -6.2 46.7 +14.9 56.5 -9.3 31.7 +8.6 59.5 -13.7 12.6 +5.0 60.3 -4.3 40.8 +9.9 60.5 -5.0 37.9 -8.8 53.0 -3.4 27.3 -4.4 53.6 -5.9 6.7 -5.9 56.0 -4.3 36.7 -4.1 54.9 -5.6 51.7 +13.8 44.9 -8.1 35.0 +7.7 40.0 -13.6 13.3 +6.5 51.1 -4.9 46.4 +9.7 53.9 -0.9 61.1 +9.4 38.4 -6.5 40.0 +5.0 28.1 -12.0 15.9 +2.7 46.2 -4.9 51.7 +5.4 41.9 -12.1 55.1 -6.1 26.8 -11.6 35.5 -4.5 13.4 -14.7 11.4 -4.6 38.1 -8.1 49.4 -2.3 41.3 -0.6 58.1 +3.0 24.6 -2.2 39.2 +3.7 6.3 -7.1 11.5 +0.2 37.2 -1.0 53.6 +4.2
Table 10: Fine-grained Main split results across revision versions and domains, aggregated over the three primary generators. DAMASHA is evaluated at the token level and GigaCheck at the span level. We report accuracy and F1-AI. Colored deltas indicate changes relative to the previous version; no delta is shown for v0 .
Detector
Ver.
Essays Macro-F1
FNR
Reports Macro-F1
FNR
News Macro-F1
Abstracts FNR
Macro-F1
FNR
Fine-grained Zero-shot (Token + Span) v0 v1 v2 v3 v4 v5 v6 v7 v8
49.7 49.0 -0.7 51.8 +2.8 57.5 +5.7 48.9 -8.6 51.6 +2.7 53.6 +2.0 35.2 -18.4 34.3 -0.9
– 95.7 87.9 -7.8 71.3 -16.7 82.8 +11.5 65.6 -17.2 50.5 -15.1 61.3 +10.8 48.9 -12.4
49.6 46.3 -3.4 43.4 -2.9 38.8 -4.6 38.4 -0.4 32.0 -6.4 26.8 -5.1 16.3 -10.6 9.6 -6.7
– 98.6 98.3 -0.3 96.9 -1.3 95.2 -1.8 93.4 -1.7 88.5 -4.9 89.5 +1.0 89.2 -0.3
49.9 46.1 -3.8 43.5 -2.6 40.1 -3.4 37.2 -2.9 32.1 -5.0 28.1 -4.1 17.7 -10.4 12.7 -5.0
– 99.2 98.4 -0.8 95.2 -3.2 96.2 +1.0 92.9 -3.3 86.7 -6.3 87.5 +0.8 84.8 -2.7
49.4 47.7 -1.7 46.8 -0.9 46.8 +0.1 44.0 -2.8 41.6 -2.4 40.6 -1.0 29.1 -11.6 23.2 -5.9
– 96.0 92.5 -3.5 84.7 -7.7 86.1 +1.4 78.7 -7.4 64.7 -14.0 72.1 +7.4 69.4 -2.7
v0 v1 v2 v3 DAMASHA [22] v4 v5 v6 v7 v8
47.1 54.3 +7.1 57.0 +2.7 60.5 +3.6 54.5 -6.1 54.3 -0.2 51.8 -2.5 36.0 -15.8 29.1 -6.9
– 79.4 73.4 -5.9 61.3 -12.1 72.1 +10.9 60.5 -11.7 52.3 -8.2 61.0 +8.8 58.7 -2.3
45.3 50.9 +5.6 50.6 -0.3 49.8 -0.7 46.3 -3.5 43.4 -2.9 38.1 -5.3 25.0 -13.1 19.6 -5.4
– 80.4 80.9 +0.4 76.1 -4.8 81.0 +4.9 76.4 -4.6 73.3 -3.1 77.7 +4.4 75.4 -2.3
49.6 47.2 -2.4 46.0 -1.2 43.1 -2.8 37.9 -5.2 33.5 -4.4 26.3 -7.2 13.2 -13.1 5.8 -7.4
– 97.7 95.7 -1.9 92.7 -3.1 96.4 +3.7 92.4 -4.0 90.8 -1.6 93.8 +3.0 93.7 -0.1
44.6 52.1 +7.5 53.5 +1.4 55.3 +1.8 51.4 -3.9 50.1 -1.3 44.5 -5.6 33.7 -10.8 26.8 -6.9
– 73.5 70.8 -2.7 65.4 -5.4 71.4 +6.0 64.9 -6.5 62.0 -2.9 65.7 +3.7 62.8 -2.8
GigaCheck [24]
Table 11: Extended fine-grained Main split results across revision versions and domains, aggregated over the three primary generators. DAMASHA is evaluated at the token level and GigaCheck at the span level. We report Macro-F1 and FNR. Colored deltas indicate changes relative to the previous version; no delta is shown for v0 .
30
Abstracts
100
Essays
News
Reports
Compress
50
50
AI Edits
Paraphrase
Accuracy (%)
0 100
0 100
Expand
50 0
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
AI coverage (%) Target AI ratio Claude
GPT-5.4
AdaLoc
GL-CLiC
GenAI-Sent.
SeqXGPT
Figure 12: Sentence-level accuracy under coverage-controlled edit operations. Abstracts
100
Essays
News
Reports
Compress
50 0
50 0 100
Expand
50 0
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
AI coverage (%) Target AI ratio Claude
GPT-5.4
AdaLoc
GL-CLiC
GenAI-Sent.
SeqXGPT
Figure 13: Sentence-level F1-AI under coverage-controlled edit operations.
31
AI Edits
Paraphrase
F1 (%)
100
Abstracts
100
Essays
News
Reports
Compress
50
50
AI edits
Paraphrase
Accuracy (%)
0 100
0 100
Expand
50 0
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
AI coverage (%) Target AI ratio DeskLib
DetectLLM E5-Small
Fast-DetectGPT GigaCheck
OOD-LLM
RADAR
RoBERTa
Figure 14: Document-level accuracy under coverage-controlled edit operations.
Abstracts
100
Essays
News
Reports
Compress
50 0
50 0 100
Expand
50 0
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
0
25
50
75
100
AI coverage (%) Target AI ratio DeskLib
DetectLLM E5-Small
Fast-DetectGPT GigaCheck
OOD-LLM
RADAR
RoBERTa
Figure 15: Document-level F1-AI under coverage-controlled edit operations.
32
AI edits
Paraphrase
AI F1 (%)
100
Abstracts
100
Essays
News
Reports 25\%
50
0
50
0
AI coverage
50\%
Accuracy (%)
100
100
75\%
50
0 s pres
Com
phra
Para
se
and
Exp
pres
Com
s
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
Edit operation Claude-Haiku
GPT-5.4
Gemini-Flash
AdaLoc
GL-CLiC
SeqXGPT
Figure 16: Sentence-level accuracy at fixed AI coverage while varying edit operation.
Abstracts
100
Essays
News
Reports 25\%
50
0
50
0 100
75\%
50
0 s pres
Com
phra
Para
se
and
Exp
pres
Com
s
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
Edit operation Claude-Haiku
GPT-5.4
Gemini-Flash
AdaLoc
GL-CLiC
SeqXGPT
Figure 17: Sentence-level F1-AI at fixed AI coverage while varying edit operation.
33
AI coverage
50\%
AI F1 (%)
100
Abstracts
100
Essays
News
Reports 25\%
50
0
50
0
AI coverage
50\%
Accuracy (%)
100
100
75\%
50
0 s pres
Com
phra
Para
se
and
Exp
pres
Com
s
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
Edit operation DeskLib DetectLLM
E5-Small Fast-DetectGPT
GigaCheck
OOD-LLM
RADAR
RoBERTa
Figure 18: Document-level accuracy at fixed AI coverage while varying edit operation.
Abstracts
100
Essays
News
Reports 25\%
50
0
50
0 100
75\%
50
0 s pres
Com
phra
Para
se
and
Exp
pres
Com
s
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
s
pres
Com
se
phra
Para
and
Exp
Edit operation DeskLib DetectLLM
E5-Small Fast-DetectGPT
GigaCheck
OOD-LLM
RADAR
RoBERTa
Figure 19: Document-level F1-AI at fixed AI coverage while varying edit operation.
34
AI coverage
50\%
AI F1 (%)
100
Essays
News
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
Reports
Abstracts
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
Accuracy (%)
100 75 50 25 0
Version Claude-Haiku GPT-5.4
Gemini-Flash AdaLoc
Essays
News
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
GL-CLiC GenAI-Sentence
SeqXGPT
Reports
Abstracts
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
100
AI F1 (%)
75 50 25 0
Version Claude-Haiku GPT-5.4
Gemini-Flash AdaLoc
GL-CLiC GenAI-Sentence
SeqXGPT
Figure 20: Sentence-level performance under independent edits from the source document. Top: accuracy; bottom: F1-AI. Each point corresponds to a version (v0 –v8 ). Compared to the cumulative trajectory, trends are smoother and the drop around v4 is less pronounced, suggesting that the nonmonotonic behavior is more strongly associated with edit type than with accumulated edits.
Essays
News
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
Reports
Abstracts
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
Accuracy (%)
100 75 50 25 0
Version DeskLib DetectLLM
E5-Small Fast-DetectGPT
Essays
News
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
GigaCheck OOD-LLM
RADAR RoBERTa
Reports
Abstracts
v0 v1 v2 v3 v4 v5 v6 v7 v8
v0 v1 v2 v3 v4 v5 v6 v7 v8
100
AI F1 (%)
75 50 25 0
Version DeskLib DetectLLM
E5-Small Fast-DetectGPT
GigaCheck OOD-LLM
RADAR RoBERTa
Figure 21: Document-level performance under independent edits from the source document. Top: accuracy; bottom: F1-AI. Each point corresponds to a version (v0 –v8 ). Compared to the cumulative trajectory, trends are smoother and the drop around v4 is less pronounced, suggesting that the nonmonotonic behavior is more strongly associated with edit type than with accumulated edits.
35