acceptodds
Under review as a conference paper at ICLR 2027

Learning Sequential–Structural Discrepancy: Label-Free Pretraining for Domain-Robust LLM-Generated Text Detection

Abstract

The rapid progress of large language models (LLMs) has made it increasingly difficult to distinguish human-written from LLM-generated text, particularly under domain shift. Existing training-based detectors typically learn sequential or syntactic cues under human/LLM provenance supervision, while approaches combining the two often rely on direct feature concatenation, leaving their relationship underexplored and potentially capturing domain-specific artifacts. We empirically find a systematic difference between human-written and LLM-generated text in the discrepancy between sequential and structural representations: human-written text tends to show a larger cross-view discrepancy, whereas LLM-generated text exhibits stronger cross-view consistency. Importantly, this distributional gap persists across domains even when the two views are learned without human/LLM provenance labels. Based on this finding, we propose the **Sequential–Structural Discrepancy (SSD)** framework. SSD first pretrains sequential and structural encoders on unlabeled text through masked representation prediction and subsequently freezes both encoders for detection. The resulting token-level discrepancy plays two complementary roles: it controls the adaptive fusion of the sequential and structural views and is retained as an explicit feature for downstream classification. Supervised training with human/LLM provenance labels updates the downstream fusion and classification modules, while both pretrained view encoders remain frozen. Experiments on in-domain and out-of-domain benchmarks demonstrate that SSD consistently outperforms strong baselines and maintains robust performance under adversarial perturbations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-generated text detection, domain generalization, label-free pretraining, dual-view encoding, sequential–structural discrepancy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FBqHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on discrepancies in text distribution between humans and large language models (LLMs)  and proposes a framework called Sequential-Structural Discrepancy (SSD) to distinguish between human-written and LLM-generated text without relying on provenance supervision, which indicates the text's origin.  As the ability to distinguish between human-written and LLM-generated text becomes more important, this line of research becomes more valuable. From a structural perspective, the SSD framework is intriguing because it distinguishes between the two through a two-stage process. The thorough experiments conducted are also commendable. However, this study raises several concerns.

Critical strengths
  • S1: This paper is written in an easy-to-understand manner.
  • S2: An anonymized implementation and experimental configurations are included in the supplementary material and are also available at https://anonymous.4open.science/r/SSD-BC84/.
  • S3: The effect of the unlabeled pretraining scale is presented in Table 4.
Critical weaknesses
  • W1: Kim et al.'s study highlights the structural differences between human- and LLM-generated text. However, it is not included in the references.

    • Kim, Z. M., Lee, K., Zhu, P., Raheja, V., & Kang, D. (2024, August). Threads of subtlety: Detecting machine-generated texts through discourse motifs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 5449-5474).
  • W2: The paper contains several unclear points.

    • How is context C_v, which is described on page 4, created?
    • Which function is used as the activation function, σ, in equation (8)?
    • What is s_j in the equation to define p_i?
    • Is MHSA MultiHeadSelfAttention?
    • In the experiment shown in Table 1, shouldn’t the authors have discussed what information was missing based on a comparison with Lastde, and how to address those gaps?
    • As shown in Table 5, the "label-free pretraining, then encoder fine-tuning" approach in view learning outperforms the full model in L2R(ID). However, the reason for this is unclear.
  • W3: The technical novelty is not particularly high.

  • W4: The limitations of this research are not clearly stated.

Questions for authors

Please see the "Weaknesses" section.

5 comments
AuthorsOpenReview ↗

Clarifying SSD’s Novelty and Relation to Prior Structural Detection Work

Relation to Kim et al. and technical novelty (W1 and W3)

We thank the reviewer for pointing out Kim et al. (2024), which is an important antecedent to our motivation and should have been discussed. We will add this reference in the revision. Kim et al. demonstrate that human- and machine-generated texts exhibit different discourse-level structures; we do not claim that the existence of structural differences itself is new.

The two works nevertheless address different technical questions. Kim et al. transform RST discourse trees into recursive hypergraphs and compare motif frequencies and depths across human and machine corpora. These discourse motifs are subsequently used as structural features in supervised classifiers. SSD neither extracts discourse motifs nor defines discrepancy by comparing two authorship groups. Instead, it computes a token-aligned discrepancy within each individual text, between representations of the same content learned from sequence-conditioned and dependency-conditioned contexts.

The contribution of SSD is therefore not simply combining a Transformer and an RGCN. Both view encoders independently reconstruct the same representations produced by a frozen language model, providing a common reference space for their outputs. They are pretrained without human/LLM provenance labels and frozen before downstream detection. Thus, provenance supervision can learn how to use the discrepancy but cannot create or reshape the encoders that produce it. The resulting discrepancy then controls adaptive view fusion and is retained as an explicit token-level feature for classification.

Our results directly test whether this formulation contributes beyond standard structural features or heterogeneous architectures:

  • The frozen sentence-level discrepancy alone obtains 74.88%/72.14% AUROC on L2R/M4 without fitting a classifier (Table 3), showing that the discrepancy already carries transferable detection information.

  • To control for architectural differences, we replaced the sequential Transformer with an RGCN over a bidirectional word-order chain while retaining an RGCN over the dependency graph for the structural view. This matched-architecture variant obtains 96.40%/78.14% on L2R/M4, close to the full model's 96.59%/78.39%. The effectiveness of SSD therefore does not primarily arise from combining different encoder architectures.

  • Removing DD reduces M4 AUROC from 78.39% to 66.89%; removing discrepancy-guided gating reduces it to 69.26%; and using DD only for gating yields 75.61% (Table 5). These results verify its complementary roles in fusion and classification.

  • Under the same protocol, the syntax-based PRDetect obtains 89.95%/65.12% on L2R/M4, compared with 96.59%/78.39% for SSD, further distinguishing cross-view discrepancy modeling from using syntactic structure alone.

In short, Kim et al. study structural patterns across authorship groups, whereas SSD studies cross-view disagreement within each text and isolates its formation from provenance supervision. We will revise the introduction and related work to make this distinction explicit.

AuthorsOpenReview ↗

We thank the reviewer for identifying several points that require clarification.

We thank the reviewer for identifying several points that require clarification.

Limitations (W4)

We will add a dedicated limitations section covering four aspects. First, SSD depends on automatically predicted dependency trees. Although the spaCy/Stanza and corruption experiments demonstrate some tolerance to parsing errors, performance may still be affected by inaccurate parses. Second, L2R, the evaluated M4 subset, and the C4 pretraining corpus are English, so cross-lingual transfer remains unverified. Third, SSD is less effective in some specialized domains: on MedicalText, it obtains 87.01% AUROC, below Lastde's 91.15%, because formulaic human clinical records may also exhibit strong sequential-structural consistency. Finally, dependency parsing, dual-view encoding, and proxy-model inference introduce additional cost. SSD requires 1.03 seconds per text under our reported setting, making it faster than generation-intensive detectors but slower than lightweight probability-based methods.

Methodological clarifications (W2)

Context CvC_v.
CvC_v is not an additional tensor; it denotes the information available to view vv when reconstructing masked word ii. Cseq,iC_{\mathrm{seq},i} contains the masked sequence X~\widetilde{X}, word order, and positional information. Cstr,iC_{\mathrm{str},i} contains the same masked node features together with the directed, relation-typed dependency graph G\mathcal{G}; after KK RGCN layers, it covers the corresponding KK-hop syntactic neighborhood. The two views therefore share the same masked node features and reconstruction targets, but use different contextual structures and propagation paths. The conditional-expectation expression describes the squared-loss optimum rather than an additional implementation step.

Equation (8).
The outer activation σ\sigma is the logistic sigmoid, which constrains the gate to (0,1)(0,1); the inner activation is tanh⁡\tanh.

Definitions of sjs_j and pip_i.
sjs_j is the jj-th proxy-model subword and s<js_{<j} denotes its preceding subwords. For word tit_i, AiA_i contains its aligned subword indices, and their log-probabilities are averaged:

pi=1∣Ai∣∑j∈Ailog⁡pϕ(sj∣s<j).p_i=\frac{1}{|A_i|}\sum_{j\in A_i}\log p_{\phi}(s_j\mid s_{<j}).

MHSA.
Yes. MHSA denotes Multi-Head Self-Attention.

Additional experimental analysis (W2)

Comparison with Lastde.
We examined representative MedicalText errors. A formulaic human clinical record has a low mean discrepancy of 0.5847 and is misclassified as machine-generated, whereas a syntactically elaborate LLM text has a higher discrepancy of 0.6748 and is misclassified as human-written. This illustrates that standardized human writing can exhibit strong cross-view consistency, while complex machine writing can produce larger discrepancy. Lastde additionally models local fluctuations and multiscale dynamics in token-probability sequences, providing complementary evidence in this domain. We will add this failure analysis and discuss combining these signals as future work.

Encoder fine-tuning in Table 5.
Fine-tuning the pretrained encoders raises L2R AUROC by only 0.61 points (96.59% to 97.20%) but reduces M4 by 8.39 points (78.39% to 70.00%). Provenance supervision improves in-domain fitting but weakens transfer. Freezing the encoders therefore deliberately trades a small ID gain for a substantially more transferable discrepancy.

Dear Reviewer FBqH,

Thank you for your valuable feedback and for taking the time to review our rebuttal. As the discussion period is nearing its end, we would kindly appreciate it if you could take another look at our responses. We hope our clarifications have addressed your concerns and would be grateful if you could consider them in your final assessment. Thank you again for your time and consideration.

Best regards, The Authors

Reviewer_ySHqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SSD, a detector that uses disagreement between a sequential Transformer and a dependency-based RGCN. Both encoders are independently pretrained on 300,000 unlabeled C4 texts to reconstruct frozen RoBERTa representations at masked positions, then frozen for detection. Their token-level absolute difference controls adaptive fusion and is also supplied to an attention classifier, together with a proxy language-model score. Provenance labels train only the downstream modules. The authors report 96.59% mean AUROC on L2R and 78.39% on the English M4 benchmark. A scalar discrepancy score alone reaches 72.14% on M4. Ablations examine pretraining, freezing and feature integration, alongside rewriting attacks and parser perturbations. The main contribution is an empirically motivated way to preserve and exploit a transferable discrepancy signal.

Critical strengths
  1. The paper identifies a useful signal before supervised detector training. Table 3 reports 74.88% AUROC on L2R and 72.14% on M4 using only the negative mean discrepancy, without fitting a classifier. Reconstructing a common frozen teacher provides a principled way to align the two output spaces, and freezing the encoders prevents downstream provenance labels from redefining their discrepancy. This combination makes the observation more informative than a separation measured only after supervised training.
  2. The empirical study provides meaningful support for the design. SSD has the highest reported average AUROC in Tables 1–2. In Table 5, freezing pretrained encoders yields 78.39% on M4 versus 70.00% when they are fine-tuned; removing discrepancy entirely reduces this to 66.89%. The proxy-free variant still reaches 74.97%, suggesting that the proxy score alone does not explain the full system's success. Parser replacement and controlled graph corruption further test an important dependency of the approach.
Critical weaknesses
  1. Both branches receive contextual RoBERTa representations, but one uses a Transformer and the other an RGCN, making it unclear whether the discrepancy reflects complementary sequential and structural information or merely architectural differences. Matched encoder and graph controls would help clarify this.
  2. Pretraining uses masked inputs, while detection uses unmasked inputs. The conditional-expectation argument therefore does not directly explain the discrepancy observed during detection.
  3. The gains on M4 and Rewrite are 1.67 and 0.64 percentage points, respectively, without uncertainty estimates. M4 also lacks domain- and generator-specific results.
  4. Appendices A–B omit exact split sizes, whether related human texts and rewrites stay in the same split, and the precise M4 evaluation subset.
  5. PRDetect is cited but excluded from the comparisons. A matched comparison would help establish SSD’s advantage over existing syntax-based detectors.
Questions for authors
  1. Does SSD still outperform matched controls using two sequential encoders or randomized dependency graphs?
  2. Can you provide confidence intervals for the M4 and Rewrite gains, along with M4 results by domain and generator?
  3. What are the exact data splits and M4 evaluation subset? Are each human source and its associated rewrites kept in the same split?
3 comments
AuthorsOpenReview ↗

Matched Controls, Pretraining–Inference Gap, and Syntax Baseline

We thank the reviewer for these constructive questions. We conducted two additional controls and evaluated a closely related syntax-based detector.

Architectural control (W1 and Q1)

To determine whether the discrepancy merely reflects the use of a Transformer and an RGCN, we constructed a matched-architecture variant in which both branches use RGCNs. The sequential branch operates on a bidirectional word-order chain with forward and backward relation types, while the structural branch retains the directed, relation-typed dependency graph. The two branches therefore share the same encoder architecture but differ in graph topology and relation types.

This variant obtains 96.40% AUROC on L2R and 78.14% on M4, close to the full model's 96.59% and 78.39%. This control removes the encoder-family confound and shows that SSD remains effective when both views use the same RGCN architecture.

Figure 3 provides a complementary graph control by separately perturbing dependency labels and parent attachments. Performance declines progressively, with attachment corruption having a larger effect because it changes graph connectivity. We use controlled corruption rather than a fully randomized graph because it isolates different sources of structural error. Together, these results indicate that the model uses the distinction between word-order and dependency structures rather than merely exploiting a Transformer–RGCN mismatch. We will add the matched-architecture experiment and clarify the scope of this control.

Masked pretraining versus unmasked detection (W2)

We agree that the conditional-expectation argument applies directly to the masked reconstruction objective and does not theoretically establish the discrepancy observed on unmasked detection inputs. It explains how the encoders learn predictions of the same target from different contextual evidence; the provenance gap on unmasked text remains an empirical observation supported by Table 3.

We additionally tested BERT-style corruption: among 15% selected tokens, 80% are replaced by [MASK], 10% by random tokens, and 10% remain unchanged, while the loss is computed over all selected positions. Downstream training and evaluation use unmasked text. This variant obtains 96.67% on L2R and 78.45% on M4, comparable to the original 96.59%/78.39%. Thus, performance is stable under a standard alternative corruption scheme and is not specific to the all-[MASK] corruption used in the main setting. We will revise the theoretical discussion to state its scope more precisely.

Comparison with PRDetect (W5)

We evaluated PRDetect under the same L2R/M4 protocol. PRDetect obtains 89.95% on L2R and 65.12% on M4, whereas SSD obtains 96.59% and 78.39%, respectively. PRDetect uses syntactic-tree information directly, while SSD models the discrepancy between sequential and dependency-conditioned representations. This comparison supports the benefit of cross-view discrepancy modeling beyond using syntax alone. We will include PRDetect in the revised comparison.

AuthorsOpenReview ↗

Uncertainty, Data Splits, and Evaluation Protocol

We thank the reviewer for requesting uncertainty estimates and more precise documentation of the evaluation protocol.

Uncertainty and M4 subgroup results (W3 and Q2)

The reported SSD results are averages over three random seeds. We agree, however, that seed averaging does not establish whether the 1.67-point M4 improvement over Lastde or the 0.64-point Rewrite improvement over L2D is statistically significant. We therefore interpret these margins descriptively and do not claim statistical significance.

Our current M4 experiment evaluates the complete benchmark subset in aggregate and does not report source- or generator-specific AUROCs. Accordingly, the current evidence supports transfer to the overall M4 distribution, but does not establish that the improvement is uniform across every source or generator. We will state this scope and the absence of paired confidence intervals and subgroup analysis explicitly as limitations.

Exact data splits and evaluation subsets (W4 and Q3)

L2R contains 21 domains. Each domain includes 200 human texts and 200 machine texts from each of four generators. The per-domain train/validation/test counts are 146/18/36 human texts and 146/18/36 machine texts per generator. Across all domains, this yields:

  • Training: 3,066 human + 12,264 machine = 15,330
  • Validation: 378 human + 1,512 machine = 1,890
  • Test: 756 human + 3,024 machine = 3,780

Training batches are balanced between human and machine text, and the four generators are sampled equally. The same fixed split indices are applied across the human and generator files within each domain, keeping aligned human texts and their generated counterparts in the same partition. Following the L2R protocol, evaluation is conducted separately for each generator, and the four AUROCs are averaged.

For OOD evaluation, we use the complete English subtaskA_dev_monolingual.jsonl split from SemEval-2024 Task 8, containing 2,500 human and 2,500 machine texts. This M4 subset is used only for final zero-shot evaluation and is not involved in training, validation, hyperparameter selection, or threshold selection.

For adversarial evaluation, perturbations are applied to the machine-generated side of the matched test subset, while the human texts remain unchanged. Rewrite uses DeepSeek-V4-Pro with a fixed rewrite instruction. Decoherence follows the Fast-DetectGPT setting by selecting a sentence longer than 20 words and randomly repeating one internal word. All detectors are evaluated on the same matched clean and attacked examples. We will add these split and subset details to the appendix.

I appreciate the detailed response and the additional experiments. I have carefully considered both, but my overall assessment remains unchanged. I will therefore retain my original scores.

Reviewer_9Cw8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SSD, a dual-view framework for robustly detecting LLM-generated text by modeling the discrepancy between sequential and syntactic representations. SSD achieves strong performance across multiple datasets and offers a new perspective on enhancing detector robustness through the relationship between complementary views.

Critical strengths
  1. The idea of multi-view fusion is innovative and inspiring.

  2. The experiments are relatively comprehensive.

Critical weaknesses
  1. In the Introduction, the transition from the limitations of existing methods to modeling the relationship between sequential fluency and syntactic structure is abrupt. This transition contains a logical gap and insufficient conceptual preparation. The connection between these two concepts and the preceding discussion is unclear, and the notion of a “view” is introduced without sufficient explanation.

  2. Sequential–structural discrepancy would benefit from further explanation through concrete textual examples. A case study could help readers understand.

Questions for authors
  1. Although the dimensions of several major tensors are provided, some variables and parameters in the equations still lack explicit shape annotations, which may cause confusion.

  2. Some table captions are too brief and should include additional information. And The paper lacks detailed statistics for train/validation/test splits and the subsets used in different experiments.

2 comments
AuthorsOpenReview ↗

Clarifying the Motivation and Sequential–Structural Discrepancy

We thank the reviewer for identifying places where the motivation and interpretation of sequential-structural discrepancy can be made clearer.

Motivation and definition of the two views (W1)

We agree that the transition in the Introduction is too abrupt. Our intended motivation is not to introduce two additional features after discussing domain shift, but to ask whether the relationship between two linguistic organizations is more transferable than either feature alone.

The sequential view represents how each word is interpreted through its linear word-order context, whereas the structural view represents the same word through dependency-mediated context. Autoregressive generation directly optimizes locally compatible word transitions, which may also reinforce recurring syntactic patterns. Human writing is additionally shaped by communicative intent, revision, and non-local planning, which can introduce greater variation between linear flow and hierarchical organization. This motivates examining the disagreement between the two views rather than treating them as independent features.

We will revise the Introduction to define a “view” at its first occurrence as a representation of the same word sequence conditioned on a particular context structure: word order for the sequential view and dependency relations for the structural view. We will also make clear that this argument motivates an empirical hypothesis rather than constituting a deterministic rule for every text.

Concrete examples and failure cases (W2)

We agree that concrete examples would make the discrepancy easier to understand. We analyzed representative samples from Environmental, where SSD performs well, and MedicalText, where its performance is weaker.

In a correctly classified Environmental pair, the human text uses negation and an infinitival construction to organize the relationship between corporate action and its purpose, producing a mean discrepancy of 0.6923. The LLM text follows a more regular enumerative structure and has a lower mean discrepancy of 0.5686. This illustrates the typical pattern underlying SSD: human text exhibits greater disagreement between the sequential and structural views.

We also examined two MedicalText errors. A human clinical record containing formulaic expressions such as “was discharged in stable condition” and “advised to follow up” has a relatively low discrepancy of 0.5847 and is misclassified as machine-generated. Conversely, an LLM text containing a concessive clause and a temporal/conditional constraint has a higher discrepancy of 0.6748 and is misclassified as human-written. These cases show that the discrepancy is a distributional tendency rather than an instance-level rule: standardized human writing can exhibit strong cross-view consistency, while syntactically elaborate machine writing can produce a larger discrepancy.

We will add a compact case study reporting the token-level and sentence-level discrepancies for both successful and failure cases.

AuthorsOpenReview ↗

Tensor Shapes and Evaluation Details

We thank the reviewer for requesting more complete notation and experimental documentation.

Tensor and parameter dimensions (Q1)

We will add a notation table covering the main tensors and parameters. Let LL be the number of words, dd the view-encoder dimension, dgd_g the hidden dimension of the gate, dad_a the attention dimension, and C=2C=2 the number of provenance classes.

The view representations and discrepancy satisfy

Hseq,Hstr,D∈RL×d.H_{\mathrm{seq}},H_{\mathrm{str}},D\in\mathbb{R}^{L\times d}.

For token ii, the gate in Equation (8) uses

W_1\in\mathbb{R}^{d_g\times d},\quad b_1\in\mathbb{R}^{d_g},\quad W_2\in\mathbb{R}^{1\times d_g},\quad b_2\in\mathbb{R},$$ and produces $\alpha_i\in(0,1)$. Collecting all token gates gives $\alpha\in\mathbb{R}^{L\times1}$, which is broadcast across the feature dimension, so $H_{\mathrm{mix}}\in\mathbb{R}^{L\times d}$. The proxy score is $P\in\mathbb{R}^{L\times1}$. Therefore, $$Z_{\mathrm{raw}}=[H_{\mathrm{mix}};\operatorname{sg}(D);P] \in\mathbb{R}^{L\times(2d+1)}.$$ With $W_z\in\mathbb{R}^{(2d+1)\times d_a}$ and $b_z\in\mathbb{R}^{d_a}$, the projected sequence is $Z\in\mathbb{R}^{L\times d_a}$. MHSA preserves this shape, max pooling produces $S\in\mathbb{R}^{d_a}$, and the final classifier uses $W_c\in\mathbb{R}^{C\times d_a}$ and $b_c\in\mathbb{R}^{C}$. ## Table captions and dataset statistics (Q2) We will expand the captions to specify the dataset, metric, ID/OOD setting, averaging protocol, and the meaning of bold and underlined values. L2R contains 21 domains. Each domain has 200 human texts and 200 machine texts from each of four generators. The per-domain train/validation/test counts are **146/18/36 human texts** and **146/18/36 machine texts per generator**. Across all domains, the resulting splits are: - Training: 3,066 human + 12,264 machine = **15,330** - Validation: 378 human + 1,512 machine = **1,890** - Test: 756 human + 3,024 machine = **3,780** Training batches contain equal numbers of human and machine texts, with the four generators sampled equally. Following the L2R protocol, evaluation is conducted separately for each generator and the four AUROCs are averaged. For OOD evaluation, we use the complete English `subtaskA_dev_monolingual.jsonl` split from SemEval-2024 Task 8, containing **2,500 human and 2,500 machine texts**. It is used only for final zero-shot evaluation and not for training, validation, model selection, or threshold selection. The adversarial experiments use matched clean and attacked subsets. Perturbations are applied to machine-generated test texts while human texts remain unchanged. Rewrite uses DeepSeek-V4-Pro with a fixed rewrite instruction. Decoherence follows the Fast-DetectGPT setup by selecting a sentence longer than 20 words and repeating one randomly selected internal word. All detectors are evaluated on the same matched examples. We will add these construction details and the resulting subset sizes to the appendix.

Comments on the paper

1 comment

Dear Reviewers and Area Chair, Thank you for your valuable feedback. We have revised the manuscript and highlighted all additions and changes in red. The main revisions are summarized below:

  • Clarified the motivation, methodological definitions, and implementation details. (Introduction and Methodology)
  • Added the L2R and M4 data splits, sample counts, evaluation protocols, and dataset links. (Experimental Setup and Appendix A)
  • Added PRDetect to the main and adversarial result tables and updated its results throughout the paper.(Experimental Results and Appendix K)
  • Added RGCN–RGCN and BERT masking control experiments. (Out-of-Domain Experiments and Appendix C)
  • Expanded the table captions to clarify the datasets, metrics, evaluation settings, comparison protocols, and result markers. (Main and Appendix Tables)
  • Added detailed textual case studies and presented the results using two separate figures. (Experimental Results and Appendix I)
  • Expanded the related work and limitations. (Related Work, Limitations)

We appreciate the reviewers’ comments, which substantially improved the clarity, completeness, and reproducibility of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.