GravText: A Robust Framework for Detecting LLM-Generated Text Using Triplet Contrastive Learning with Gravitational Factor

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper tackles the important challenge of detecting LLM-generated text under adversarial paraphrasing, a scenario where many existing detectors fail. The authors propose GravText, a detection framework that: Uses triplet contrastive learning with dynamic anchor switching to improve robustness against paraphrased adversarial examples. The idea of introducing a gravitational factor inspired by physics is creative and provides an intuitive metaphor for embedding separation.

• This paper is not well motivated. Why the Contrastive learning is adopted, rather than some advanced representation learning models? The authors should explain more.
• How to measure the performance of Anchor Data Selection?
• Some recent neural methods should be cited in case that readers need more frontiers and advances for building advanced models.

+ Sequence Labeling with Meta-Learning
+ Few-Shot Relation Extraction With Dual Graph Neural Network Interaction
+ Few-shot named entity recognition via meta-learning

The paper would be greatly improved with stronger motivation of approach, more valuable observations from experimental results and a good technical exposition.

Author Response

We sincerely thank you for your valuable time and insightful comments on our manuscript. Your suggestions have been instrumental in helping us improve the quality and clarity of our work. We have carefully addressed each of your concerns in the revised manuscript. Below, we provide a point-by-point response to your comments.

 

[Comment 1: This paper is not well motivated. Why the Contrastive learning is adopted, rather than some advanced representation learning models? The authors should explain more.]

Response 1: We sincerely thank you for this important suggestion. In the revised manuscript, we have significantly strengthened the motivation for using contrastive learning in Section 1 (Introduction) from line 50 to line 62. Specifically, we have added the following explanation:

"The core limitation of existing detectors lies in their reliance on surface statistical or syntactic features, which are easily manipulated by paraphrasing. To achieve true robustness, a model must learn paraphrase-invariant semantic features by focusing on the underlying semantic consistency and structural patterns inherent to LLM-generated texts. Contrastive learning is particularly suited for this purpose, as it aims to pull similar samples (e.g., original and paraphrased LLM texts) closer in the embedding space while pushing dissimilar samples (e.g., human text) further away. This mechanism naturally enhances robustness against minor adversarial perturbations like paraphrasing."

This addition clarifies why contrastive learning is a more suitable paradigm for learning paraphrase-invariant features compared to other representation learning models, directly addressing the limitations of existing approaches.

 

[Comment 2: How to measure the performance of Anchor Data Selection?]

Response 2: Thank you for raising this point. We have now added a detailed explanation of how we evaluate the effectiveness of the Dynamic Anchor Switching Strategy (DASS) in Section 3.2 from line 275 to line 284. Specifically, we added:

"The function of this dynamic anchor switching strategy (DASS) extends beyond task definition; it is engineered for effective hard negative mining. Paraphrasing is inherently an adversarial process designed to make the LLM-generated text (P) semantically closer to human text (N), resulting in hard negative samples. The DASS ensures the triplet structure is optimized to target these ambiguous boundaries... This focused optimization process, which forces the model to learn the subtle difference between the two classes at their closest points, is the true measure of DASS performance, as demonstrated by the overall robustness gains in our experimental results (Section 5)."

We believe this clarifies how DASS contributes to model performance through hard sample mining and embedding space optimization.

 

 [Comment 3: Some recent neural methods should be cited in case that readers need more frontiers and advances for building advanced models.Sequence

• Labeling with Meta-Learning
• Few-Shot Relation Extraction With Dual Graph Neural Network
• InteractionFew-shot named entity recognition via meta-learning]

Response 3: We sincerely appreciate this suggestion. We have now added the following references to the "Introduction" section (Section 1) from line 18 to line 24 to reflect the latest advances in meta-learning and representation learning:

• Sequence Labeling with Meta-Learning
• Few-Shot Relation Extraction with Dual Graph Neural Network Interaction
• MetaNER: Named Entity Recognition with Meta-Learning

These citations enrich the background and connect our work to contemporary advances in meta-learning and graph-based representation learning, providing readers with a more comprehensive understanding of the field's frontiers.

 

[Comment 4: The paper would be greatly improved with stronger motivation of approach, more valuable observations from experimental results and a good technical exposition.]

Response: We thank you for this overarching suggestion. In response, we have:

1. Enhanced the motivationthroughout Section 1, explicitly linking the limitations of existing methods to our design choices.
2. Added comprehensive cross-lingual experimentson the English essay dataset, providing valuable observations about the framework's language-agnostic properties (Section 5.2, Tables 4-5).
3. Conducted rigorous statistical significance testingwith paired t-tests across both datasets (Section 5.3, Tables 6-9), demonstrating that all performance improvements are statistically significant (p < 0.05).
4. Improved technical expositionin Sections 3.2 and 3.3, with clearer explanations of the gravitational factor and its implementation.
5. Added research questions(RQ1-RQ3) in the Introduction to better structure our investigation and motivate the technical approach.

These revisions have significantly strengthened the paper's motivation, provided more valuable experimental insights, and enhanced the technical clarity.

Once again, we deeply appreciate your constructive and insightful comments, which have greatly improved our manuscript. We hope that the revisions satisfactorily address your concerns.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents "GravText," a novel framework for detecting LLM-generated text that is robust to paraphrasing attacks. The idea of using dynamic triplet contrastive learning with archor switching and gravitational factor is well motivated and clearly described. Regarding to the details, I have several suggestions or questions as follows:

1. In line 42, "To address these challenges...", the authors only analyzed the current methods and conclude " their detection effectiveness decreases significantly when the original text is rewritten". The limitations of other methods support the necessity of a reliable detaction method. But how the limitations are related to the enhancement of your method is not clearly explained. More specifically, why constrastive learning and the gravitational factor, not the others?

2. While the gravitational factor is novel, its intuitive explanation could be enhanced. A brief discussion on why cross-attention is a suitable proxy for "mass" in this semantic context, beyond its ability to capture fine-grained alignments, would be beneficial. Is it measuring semantic density, information content, or stylistic consistency?

3. A brief discussion on the added computational cost of the cross-attention mechanism for the gravitational factor, especially compared to a standard triplet loss setup, would be valuable for practitioners considering this approach.

4. The experiments are currently limited to a single dataset (HC3) and one language (Chinese). While the results are strong, a critical question for the reviewers and readers will be about cross-domain and cross-lingual generalization. The authors acknowledge this in the conclusion, but even a small experiment on a different (perhaps English) dataset would significantly strengthen the claims of robustness.

5. It would be helpful to explicitly state the version of RoBERTa used (e.g., RoBERTa-base-chinese) and the fine-tuning details for the baseline to ensure a fair comparison.

6. The results report standard deviations, which is good. However, explicitly stating whether the improvements of GravText over the best baseline (RoBERTa) are statistically significant (e.g., using a paired t-test) would add further rigor.

Author Response

We are deeply grateful for your thorough review and valuable suggestions, which have significantly enhanced the quality of our work. We have carefully addressed each of your comments in the revised manuscript. Below, we provide a detailed point-by-point response.

[Comment 1: In line 42, "To address these challenges...", the authors only analyzed the current methods and conclude "their detection effectiveness decreases significantly when the original text is rewritten". The limitations of other methods support the necessity of a reliable detection method. But how the limitations are related to the enhancement of your method is not clearly explained. More specifically, why contrastive learning and the gravitational factor, not the others?]

Response: We sincerely thank you for this insightful comment. We have now substantially strengthened the connection between the limitations of existing methods and our design choices in Section 1 from line 50 to line 62 of the revised manuscript. We explicitly state:

"The core limitation of existing detectors lies in their reliance on surface statistical or syntactic features, which are easily manipulated by paraphrasing. To achieve true robustness, a model must learn paraphrase-invariant semantic features... Contrastive learning is particularly suited for this purpose, as it aims to pull similar samples (e.g., original and paraphrased LLM texts) closer in the embedding space while pushing dissimilar samples (e.g., human text) further away... We introduce the gravitational factor to further refine the separation by dynamically addressing hard negative examples with higher repulsion. This design directly addresses the limitations of watermarking, statistical, and supervised methods, whose reliance on surface-level cues makes them fragile under paraphrasing, whereas GravText explicitly learns semantic invariances and enforces stronger separation through the gravitational factor."

This addition clarifies that contrastive learning is chosen for its inherent ability to learn invariance to paraphrasing, and the gravitational factor is introduced as a targeted mechanism to handle the hard negatives created by such adversarial rewrites.

[Comment 2: While the gravitational factor is novel, its intuitive explanation could be enhanced. A brief discussion on why cross-attention is a suitable proxy for "mass" in this semantic context, beyond its ability to capture fine-grained alignments, would be beneficial. Is it measuring semantic density, information content, or stylistic consistency?]

Response: This is an excellent point. We have significantly expanded the intuitive explanation and justification for using cross-attention as a proxy for "semantic mass" in Section 3.3 from line 307. We now explain:

"A crucial aspect of our framework is the intuitive motivation for using cross-attention as a proxy for semantic mass. Standard triplet loss relies on a single global distance metric... which is a low-resolution measure... To overcome this, we introduce cross-attention as a high-resolution alignment mechanism... we posit that the aggregated cross-attention score serves as a direct measure of semantic density and shared information content... Therefore, cross-attention is uniquely suited as a proxy for 'mass' because it quantifies the density of shared meaning, allowing our gravitational factor to more accurately model the true semantic attraction and repulsion between text pairs."

This new discussion clarifies that cross-attention primarily measures "semantic density" – the concentration of mutually aligned meaning between texts – which is a more foundational and robust property than surface-level stylistic consistency.

[Comment 3: A brief discussion on the added computational cost of the cross-attention mechanism for the gravitational factor, especially compared to a standard triplet loss setup, would be valuable for practitioners considering this approach.]

Response: We thank you for raising this practical consideration. We have added a dedicated paragraph in Section 3.3 from line 304 to line 351 to discuss the computational cost:

"Integrating the cross-attention mechanism into the gravitational factor introduces additional computational overhead during the training phase, compared to standard triplet loss based solely on embedding distances. Specifically, the cross-attention operation adds an  complexity per triplet, where  is the sequence length, contributing to a longer optimization cycle. However, it is important to emphasize that this computational cost is strictly limited to training. During inference, GravText relies solely on the trained RoBERTa encoder to generate text embeddings, followed by a lightweight distance-based classification. The gravitational factor module, including cross-attention, is not required at inference time. As a result, the inference speed of GravText remains comparable to that of a standard fine-tuned RoBERTa baseline. The additional training cost is thus a necessary and efficient trade-off for achieving significant robustness gains against adversarial paraphrasing."

[Comment 4: The experiments are currently limited to a single dataset (HC3) and one language (Chinese). While the results are strong, a critical question for the reviewers and readers will be about cross-domain and cross-lingual generalization. The authors acknowledge this in the conclusion, but even a small experiment on a different (perhaps English) dataset would significantly strengthen the claims of robustness.]

Response: We completely agree with this critical point. Following your suggestion, we have conducted extensive new experiments on an English essay dataset to validate the cross-lingual generalization of GravText. This new evaluation is now fully integrated into our manuscript:

Section 4.1: Introduces the English essay dataset.(from line 402 to line 409)

Section 5.2: Presents the main cross-lingual results (Tables 4 and 5 in the revised manuscript), showing that GravText maintains superior performance over the RoBERTa baseline in English.

Section 5.3: Includes paired t-tests on the English dataset results (Tables 8 and 9 in the revised manuscript), confirming the statistical significance of the improvements.

Section 5.5 (Discussion): Provides a cross-lingual robustness analysis, concluding: "GravText demonstrates consistent and statistically significant performance gains over RoBERTa-base on both the Chinese HC3 and English essay datasets. These results establish its effectiveness and robustness as a language-agnostic solution for detecting AI-generated text."

These new experiments and analyses substantially strengthen our claims regarding the framework's generalizability.

[Comment 5: It would be helpful to explicitly state the version of RoBERTa used (e.g., RoBERTa-base-chinese) and the fine-tuning details for the baseline to ensure a fair comparison.]

Response: Thank you for highlighting this lack of clarity. We have updated Section 4.3 (Baseline Comparisons) from line 434 to line 440 to specify the exact model and fine-tuning protocol:

"RoBERTa: We fine-tune the standard RoBERTa-base model on a portion of the HC3 dataset to distinguish human-written from model-generated content through supervised learning. All models use the same tokenization and training configuration for fair comparison. Specifically, we adopt a batch size of 32, a learning rate of , and a maximum sequence length of 256 tokens. Training is governed by early stopping based on validation F1-score, with a patience of 5 epochs. The best-performing checkpoint is retained for evaluation."

We have ensured that the same fine-tuning strategy (including the RoBERTa-base architecture and hyperparameters) was applied consistently to both the HC3 and English essay datasets for all baseline and GravText models.

[Comment 6: The results report standard deviations, which is good. However, explicitly stating whether the improvements of GravText over the best baseline (RoBERTa) are statistically significant (e.g., using a paired t-test) would add further rigor.]

Response: We agree that statistical significance testing is crucial. In response, we have added a complete new subsection, Section 5.3: Statistical Significance Analysis from line 545. In this section, we report the results of corrected paired t-tests based on F1 scores and accuracy from five independent runs:

Tables 6 and 7 present the t-statistics and p-values for all comparisons on the HC3 dataset.

Tables 8 and 9 present the same for the new English essay dataset experiments.

The results demonstrate that GravText's performance improvements over RoBERTa are statistically significant (with p-values < 0.05) across both datasets and all task configurations, thereby adding the desired rigor to our claims.

We are truly thankful for your expert guidance, which has been invaluable in improving our manuscript. We hope the revisions and additions meet with your approval.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper proposes a new technique to detect AI-generated text. The proposal is based on a novel contrastive learning approach that incorporates a gravitational measure to create better clusters within the embedding space. The paper is interesting, and I believe it could be valuable for the research community. Therefore, I consider that it can be published after addressing the following concerns:

-The introduction section needs to present a clearer objective behind the research. Consider adding research questions, hypotheses, and objectives after the description of the contributions.

-The introduction should end with a paragraph outlining the structure of the paper.

-The writing of the related works section is somewhat superficial. The division into subsections is helpful and improves understanding, but within each subsection (e.g., statistical methods), only a single sentence is provided for each paper. This gives the impression that the referenced works were not reviewed in sufficient detail. A more robust and analytical related works section is necessary.

-Experiments and results sections, are comprehensive and detailed, with extensive results and implications. However, I am not sure if using only one dataset is enough to draw strong conclusions without doubt. The authors should include at least one additional dataset to make their experimental validation more robust.

-Figure 7 is too small and needs to be reformulated for better readability.

Author Response

We sincerely appreciate your thoughtful review and constructive feedback. Your comments have been invaluable in helping us enhance the clarity, depth, and robustness of our work. We have carefully addressed each of your concerns in the revised manuscript, as detailed below.

 

[Comment 1: The introduction section needs to present a clearer objective behind the research. Consider adding research questions, hypotheses, and objectives after the description of the contributions. The introduction should end with a paragraph outlining the structure of the paper.]

Response: Thank you for this excellent suggestion. We have thoroughly revised the Introduction to provide a clearer research framework:

1. Added Research Questions:Following the contributions, we have now explicitly formulated three research questions (RQs) to guide our investigation from line 91 to line 100:
   • RQ1:To improve robustness against rewriting attacks in LLM-generated text detection, can we employ a contrastive learning framework with dynamic anchor switching to learn paraphrase-invariant representations?
   • RQ2:To improve detection accuracy against paraphrasing attacks, can a physics-inspired gravitational factor in the embedding space enhance cluster separation between human and AI-generated texts?
   • RQ3:To evaluate the generalizability of the proposed GravText framework, can its performance consistency be assessed across different text lengths and between open-source and closed-source LLMs?
2. Added Paper Structure:As recommended, the Introduction now concludes with a new paragraph that outlines the structure of the paper from line 101 to line 107:
   "The remainder of this paper is structured as follows: Section 2 reviews related work in LLM text detection and contrastive learning. Section 3 details the architecture of the proposed GravText framework... Section 4 describes the experimental setup... Section 5 presents the comprehensive results and comparative analysis. Finally, Section 6 concludes the paper and outlines directions for future work."

These additions provide a much clearer roadmap for the reader and firmly establish the objectives of our research.

 

[Comment 2: The writing of the related works section is somewhat superficial. The division into subsections is helpful and improves understanding, but within each subsection (e.g., statistical methods), only a single sentence is provided for each paper. This gives the impression that the referenced works were not reviewed in sufficient detail. A more robust and analytical related works section is necessary.]

Response: We thank you for pointing this out. We have significantly expanded the Related Work section (Section 2 from line 115) to provide a more robust, analytical, and in-depth discussion of the cited literature. Rather than merely listing papers, we now synthesize their contributions, highlight their key methodologies, and critically discuss their limitations to better contextualize our work. For instance:

In Section 2.1, we provide a more detailed analysis of each watermarking, statistical, and supervised method, explaining their core mechanisms and, crucially, why they fail under paraphrasing attacks (e.g., discussing the assumptions behind Golowich's watermark that don't cover adversarial paraphrasing, or the short-text vulnerability of Fagni et al.'s BERT detector).

In Section 2.2, we have expanded the discussion on contrastive learning, providing a clearer narrative of its evolution from vision to NLP and its recent application to AI-text detection, thereby offering a stronger foundation for our methodological choices.

This revised section now demonstrates a comprehensive understanding of the field's landscape.

 

[Comment 3: Experiments and results sections, are comprehensive and detailed, with extensive results and implications. However, I am not sure if using only one dataset is enough to draw strong conclusions without doubt. The authors should include at least one additional dataset to make their experimental validation more robust.]

Response: We agree entirely with this critical point regarding the generalizability of our findings. Following your suggestion, we have conducted a comprehensive set of new experiments on a second dataset. Specifically:

New Dataset: We incorporated the English essay dataset, which consists of human-written and ChatGPT-generated argumentative essays, to test cross-lingual generalization.

Integrated Results: The results on this new dataset are now fully integrated into the manuscript:

Section 4.1 introduces the new English essay dataset.

Section 5.2 presents the main comparative results (Tables 4 and 5 in the revised manuscript), demonstrating GravText's superior performance in a different language and domain.

Section 5.3 includes paired t-tests on the English dataset results (Tables 8 and 9 in the revised manuscript), confirming the statistical significance of the improvements.

Section 5.5 (Discussion) provides a cross-lingual analysis, concluding that the performance advantage of GravText is consistent across both Chinese and English datasets.

This major addition significantly strengthens the validity and robustness of our experimental claims.

 

[Comment 4: Figure 7 is too small and needs to be reformulated for better readability.]

Response: We apologize for the issue with the figure. We have addressed this in the revised manuscript:

The original Figure 7 (log-probability curvature) has been remade and is now presented as Figure 8 in the revised manuscript.

The new figure has been carefully reformatted to be larger and clearer, ensuring all elements are easily readable.

 

We are deeply grateful for your meticulous and constructive comments, which have profoundly improved the quality and impact of our paper. We hope our revisions have adequately addressed all your concerns.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The paper has improved considerably, and with the inclusion of the new dataset, it has robustly demonstrated its potential. The addition of new research questions, as well as expanded background in the Introduction and Related Works sections, provides comprehensive coverage of the topics discussed. With all my comments addressed, I recommend the paper for publication.