acceptodds
Under review as a conference paper at ICLR 2027

PEC-Detector: Planning-Execution Coupling Trajectories for Document-Level LLM-Generated Text Detection

Abstract

Reliable AI-generated text detection is increasingly difficult as large language models (LLMs) produce fluent and coherent long texts. Existing detectors typically derive document-level evidence from static textual representations or aggregated statistics, without explicitly modeling dynamic coupling between discourse planning and current writing execution as a document unfolds. We show that remote discourse history provides measurable conditional information about the current sentence beyond local context, and that its contribution follows different trajectories in machine-generated text (MGT) and human-written text (HWT). Motivated by this observation, we propose **PEC-Detector**, a planning-execution coupling framework for document-level machine-generated text detection. At each sentence position, we induce planning and local context states from remote discourse history and local context, respectively, and construct a conditional planning-execution relation with the current writing execution. The resulting position-wise representations form an ordered coupling trajectory for document-level source prediction. A joint conditional semantic prediction objective encourages the planning state to retain information relevant to current writing execution. Experiments on public benchmarks spanning diverse domains, generators, and languages show that our method consistently achieves state-of-the-art detection performance. Furthermore, PEC-Detector remains robust under distribution shifts and adversarial attacks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM-generated text detection, discourse modeling, large language models, content authenticity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2PxvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEC-Detector for document-level machine-generated text detection, which models relations between remote discourse history, local context and current sentence. Its key observation is that conditioned on local context, the contribution of remote history to semantic prediction tends to grow as the machine-generated documents unfold, but remains fairly constant in human-written documents. Motivated by this observation, the proposed method learns position-wise “planning – execution coupling” representations and encodes their sequence to classify the source with an auxiliary semantic prediction objective. Experiments are performed across domains, generators, attack settings, and languages.

Critical strengths
  1. The research question is sound and testable. Investigating how the conditional predictive contribution of remote context changes over the course of a document is an approach to view document-level detection from the most useful perspective. The use of an independent predictor trained without source labels for the first observation also avoids circular explanations based on supervised classification features directly.
  2. Detections results reported are competitive. PEC-Detector is superior to the compared baselines in several evaluation settings, indicating that the designed contextual relation representations are applicable in real detection.
  3. Experiment coverage is relatively wide. Besides the main benchmarks, there are component ablations, analyses of sentence order, text length, training data scale and multilingual performance in this paper.
Critical weaknesses
  1. The core interpretation as “planning – execution coupling” is not sufficiently established. Operationally, the planning state is a representation that is encoded from remote history. The fact that it enhances current-sentence prediction does not automatically prove that it represents discourse planning: topic-continuity, entity-recurrence, semantic-repetition, and formulaic-organization could all yield the same effect. The authors correctly note that this state does not directly reconstruct the author’s original plan. Still, additional justification and evidence should be provided to validate why planning – execution coupling is a more mechanistic explanation than general long-range semantic dependence. The present results more directly bolster the detection value of remote-context relations rather than the postulated mechanistic interpretation.
  2. The link between the central observation, the architecture, and the performance gain is still not complete. Fig. 1 addresses the position-wise evolution of a scalar predictive gain, while the final classifier is fed with high-dimensional representations comprising current-sentence semantics, contextual states, and the multiple interaction features. Good performance of the full model thus does not directly prove that the observed gain trajectory accounts for the gain. The ablations do verify that the individual components are useful, but do not allow isolating the contribution of this specific observation in isolation from richer semantic representations, context modeling, and model capacity. A simple detector on only predictive-gain trajectories with an similarly matched ordinary sentence-sequence classifier in terms of encoder and context coverage and number of parameters would directly help.
  3. Alternative explanations for the original observation are not well controlled. The amount of available remote history naturally increases with document position. Human and machine texts can also differ in the number of sentences, sentence length, topic concentration, and semantic repeats. The progress normalization and document-level aggregation in the appendix are intuitive, but do not entirely control these factors. Further, the central tendency is mainly observed on CMV without reliable evidence of consistent replication across domains and generators. As a result, it is not clear whether the trajectory indicates a stable source-related difference or characteristics of the given dataset and its text distributions.
  4. Some experiments do not entirely support the related mechanistic and generalization claims. Performance degradation under test-time sentence shuffling indicates sensitivity to ordered discourse information, but not uniquely to planning – execution relationships; a general coherence model may indicate a similar sensitivity. Additionally, Multi-Attack experiments are trained on training data from the same settings, thus these results should not be taken to directly indicate robustness to unseen attacks. MAGE experiments should similarly specify which generators or domains are not seen during training, to separate effectiveness on heterogeneous data from transfer to truly unseen data.
Questions for authors
  1. Does the trajectory difference in Figure 1 continue given remote-history length fixed and document length, topic and semantic repetition are equal? Does it reproduce itself across other domains and generators?
  2. How well can a detector do with only the mean or the slope or the full sequence of predictive gains? What added value does PEC have over a capacity matched ordinary sentence-sequence classifier?
  3. During joint training, does the classification loss also update the predictor through the predictive-gain branch? If yes, how is the desired semantic-prediction interpretation of this gain confirmed?
Reviewer_7p6JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEC-Detector, a document-level detector of LLM-generated text that models the relationship between remote discourse history, local context, and the current sentence. Its central idea is to measure the incremental contribution of remote history to semantic prediction after conditioning on local context. The method combines this predictive gain with relational features between remote-history and current-sentence representations, and aggregates the resulting trajectory for document classification. Training combines a source-classification objective with an auxiliary semantic-prediction objective. Experiments on DetectRL and MAGE report strong detection performance, while component ablations, sentence-order perturbations, and supplementary evaluations investigate the usefulness of the proposed representations.

Critical strengths

1-Measuring how much remote history improves the prediction of a sentence beyond its immediate context offers an interpretable way to characterize document structure. The independent analysis in Appendix A is particularly useful: source labels are not used to train or select the conditional predictor, and the resulting gains are subsequently examined across human-written and machine-generated documents. This provides motivation for the proposed features beyond the final supervised classification results.

2-PEC-Detector achieves the highest AUROC in seven of the eight DetectRL Multi-Domain/Multi-LLM configurations and reports strong performance across the evaluated attack settings. Table 4 examines several components across four generator settings, including remote history, conditional predictive gain, auxiliary prediction, and trajectory aggregation. The sentence-level and block-level shuffling experiments further indicate that the detector uses ordered discourse information. Together, these results support the practical usefulness of the proposed representation.

Critical weaknesses

1-Section 4.1 and Table 2 describe the selected MAGE settings as evaluations under distribution shifts. However, Section 5.1 and Table 3 of the original MAGE benchmark explicitly classify Fixed-domain & Arbitrary-models and Arbitrary-domains & Model-specific as in-distribution settings, with the relevant domains and model sets represented during training and testing. Appendix B.1 states that the prepared benchmark splits are retained. Consequently, these results establish performance across diverse sources, but do not establish transfer to unseen domains or generators. MAGE provides separate Unseen Models and Unseen Domains testbeds for that purpose.

The newer-generator experiments in Appendix C.3 could provide relevant additional evidence. However, the manuscript does not clearly identify the trained detector checkpoint or establish whether these generators were excluded from training, validation, and model selection. This protocol needs to be explicit before the results can substantiate unseen-generator generalization.

There is also a concrete configuration inconsistency. Section 4.1 identifies frozen all-mpnet-base-v2 as the default sentence encoder, whereas Appendix C.5 and Table 10 describe replacing the default encoder with that same model and report different results for the replacement and the main system. It is therefore unclear which encoder produced the main results and what the encoder-robustness comparison actually tests. This affects both reproducibility and the interpretation of the reported robustness to encoder choice.

2-Section 3.2 appropriately defines the planning state as a latent representation induced from remote discourse history, rather than a reconstruction of the author’s original plan. Nevertheless, the stronger interpretation of the resulting trajectories would benefit from more targeted validation.

The conditional predictive gain can also reflect topic persistence, entity recurrence, or general discourse coherence. Likewise, component ablations establish that particular information and modules are useful, while sentence and block shuffling disrupt semantic coherence and the distribution of ordered inputs. These findings support the importance of remote context and sentence order, but do not clearly distinguish planning–execution coupling from more general long-range semantic dependencies.

Beyond the existing semantic-only and context-only ablations, a comparison with a generic long-context model using the same sentence encoder and comparable capacity would help assess what is specific to the proposed coupling construction. Alternatively, a controlled analysis involving known planning structure could strengthen the interpretation. Without such evidence, the conceptual claims should be stated more cautiously.

Questions for authors

1-In particular, which trained checkpoint is evaluated in Appendix C.3, and are the newer generators entirely unseen during model development? Results on MAGE’s Unseen Models and Unseen Domains testbeds would directly address the generalization concern.

2-How does the all-mpnet-base-v2 variant in Table 10 differ from the default encoder specified in Section 4.1? Please provide the exact checkpoints and clarify their frozen or fine-tuned status.

3-For example, does the method retain its advantage over a model matched in sentence encoder, context coverage, and capacity? Alternatively, do the predictive-gain trajectories respond systematically to controlled changes in adherence to a known outline?

Reviewer_XQw9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEC-Detector, a detection pipeline that combines context information with writing-trajectory analysis to detect AI-generated text. Specifically, the proposed method considers the remote history, the local context and the current sentence, and designs components to extract their relationships as features. It then analyzes the writing trajectory formed by these context-aware features for reliable AI-generated text detection. Evaluation on several datasets shows that PEC-Detector outperforms several existing baselines.

Critical strengths
  1. The paper proposes a detection pipeline that integrates context information with trajectory dynamics.
  2. Evaluation on several popular datasets demonstrates higher performance of the proposed method compared with existing baselines.
Critical weaknesses
  1. The design of Module 1 causes a natural information asymmetry across sentence positions: later sentences have far more remote history than earlier ones. The authors should add an ablation that uses only the later sentences, e.g., the latter half of the text, and check whether the method keeps similar performance. This may also save computational cost.
  2. The pipeline appears to contain redundancy. For example, in Module 2's Planning-Execution Interaction, it is unclear why the two mappings ApA_p and AeA_e are added. The main text gives no justification, and the supplementary code contains no corresponding implementation.
  3. Several design choices are unjustified. In Module 2's Relation Statistics, the two input vectors pˉt\bar{\mathbf{p}}_t and eˉt\bar{\mathbf{e}}_t are not normalized, which may leave the statistics unnormalized and position dependent. In addition, the relation mapping concatenates pt\mathbf{p}_t and ct\mathbf{c}_t with κt\kappa_t, although κt\kappa_t is computed from pt\mathbf{p}_t and et\mathbf{e}_t rather than ct\mathbf{c}_t; the introduction of ct\mathbf{c}_t at this point is abrupt and confusing.
  4. Lack of evaluation under realistic out-of-distribution settings. Whether the method generalizes to unseen generator LLMs, unseen text domains and unseen adversarial attacks is unknown. The authors should train the method on a clean dataset and test on cross-generator, cross-domain and unseen-attack test sets.
  5. Lack of discussion of more recent context-aware detectors, such as Profiler [1] and Lastde [2].
  6. The writing and presentation could be improved. In Figure 2, the authors should mark which components are frozen and which are trainable. In Figure 2's Module 1, a projection component is missing between the sentence encoder and the execution state et\mathbf{e}_t. The learnable document token qdocq_{doc} of Module 3 appears in the main text but never in the overview figure. These discrepancies between the overview figure and the main text make the paper hard to read.
  7. The reference list has minor inconsistencies. For example, for the HLD (ICLR 2026) paper, the fifth author is omitted.

[1] Guo, Hanxi, et al. "Profiler: Black-box AI-generated Text Origin Detection via Context-aware Inference Pattern Analysis." Conference on Empirical Methods in Natural Language Processing. 2025.

[2] Xu, Yihuai, et al. "Training-free LLM-generated text detection by mining token probability sequences." International Conference on Learning Representations. 2025.

Questions for authors
  1. In Module 1, can FplanF_{\text{plan}} and FlocalF_{\text{local}} be replaced by average pooling of the sentence embeddings instead of a Transformer and a GRU?
  2. In Module 2's conditional predictive gain, do the hidden and visible paths share the same FpredF_{\text{pred}} or use different ones? The two paths take inputs of different sizes, and the released code uses two separate networks.
  3. What is the performance of the proposed method at a low false-positive rate, such as TPR at FPR = 1%? This is important for practical deployment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

For clarity and simplicity, we refer to Reviewers XQw9, 7p6J, and 2Pxv as R1, R2, and R3, respectively.

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged by their recognition of the strengths of our work.

We have carefully addressed each individual comment raised by the reviewers and believe that we have successfully responded to most of their concerns. In the revised manuscript, we have incorporated the suggested experiments, expanded the related discussions, and made the corresponding revisions to further strengthen our work. Below, we summarize the core contributions of our study, the updates to our experiments, and the clarifications included in our revision.


Core Contributions

  1. We revisit document-level machine-generated text detection from a dynamic writing perspective. Given local context, the conditional contribution of remote discourse history to the current writing follows distinct trajectories as documents unfold: it generally increases in MGT, but remains relatively stable in HWT.

  2. Based on this observation, We propose PEC-Detector, a framework that formulates source prediction as identifying ordered planning-execution coupling trajectories to distinguish MGT from HWT.

  3. Extensive experiments on public benchmarks demonstrate that our method delivers superior performance across a variety of complex scenarios, including multi-domain, multi-LLM, cross-setting, out-of-distribution transfer, and adversarial attacks.


Experimental Updates During Rebuttal

  • Main Text (Section 4.2 and Table 3): Added out-of-distribution evaluation on MAGE Unseen Domains and Unseen LLMs, and distinguished these held-out settings from the original MAGE cross-setting evaluation (R1, R2, R3).

  • Main Text (Robustness Against Attacks) and Appendix C.6: Added an evaluation that trains PEC-Detector on clean Multi-Domain data and tests it on unseen attack distributions (R1, R3).

  • Main Text (Section 4.3 and Table 5): Added the AvgPool State ablation to evaluate whether average pooling can replace the planning and local-context encoders (R1).

  • Appendix A.1: Added predictive gain controls using gain mean, gain slope, the full predictive gain trajectory, and the sentence semantic trajectory (R3).

  • Appendix C.2: Added results at TPR@1%FPR and TPR@5%FPR (R1).

  • Appendix C.4: Added detailed F1 results for MAGE Unseen Domains and Unseen LLMs (R1, R2, R3).

  • Appendix C.5: Added a strict unseen newer generator protocol, where the detector is trained and selected using GPT-3.5 only, then evaluated on GPT-4-Turbo, Gemini-2.0-Flash, and Claude-3.7-Sonnet without adaptation (R1, R2).

  • Appendix C.7: Added the Late-Trajectory ablation, which evaluates whether the latter half of the position-wise coupling trajectory is sufficient for detection (R1).

  • Appendix C.8: Added a capacity matched generic long-context control to evaluate the additional value of the explicit planning-execution relation construction beyond general long-range semantic information (R2, R3).


Clarification and Manuscript Updates During Rebuttal

  • Main Text (Related Work): Added discussion of Lastde and Profiler and clarified their relation to PEC-Detector (R1).

  • Main Text (Sections 3.2, Figure 2): Clarified the operational meaning of the planning state, removed the unimplemented mappings ApA_p and AeA_e, distinguished the two conditional prediction heads, explained the role of local context in the relation mapping, and aligned Figure 2 with the implementation (R1, R2, R3).

  • Main Text (Section 4.1): Clarified that the original MAGE Fixed-domain & Arbitrary-models and Arbitrary-domains & Model-specific settings are cross-setting evaluations rather than strict out-of-distribution transfer settings (R2).

  • Main Text (Section 4.2) and Appendix C.5: Clarified the checkpoint selection protocol for newer generator evaluation and explicitly stated that the target generators are excluded from model development (R1, R2, R3).

  • Appendix C.10: Corrected the encoder robustness description and explicitly identified all-mpnet-base-v2 as the frozen default encoder and all-MiniLM-L6-v2 as the frozen replacement encoder (R2).

  • Reference List: Corrected the omitted HLD author and checked the complete reference list for remaining bibliographic inconsistencies (R1).


We believe that these additions and clarifications comprehensively address the reviewers' concerns and strengthen the manuscript. For ease of review, all revisions are highlighted in blue in the uploaded revised manuscript.

We look forward to the reviewers' favorable consideration and remain grateful for their valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.