acceptodds
Under review as a conference paper at ICLR 2027

Generalizing Machine-Generated Text Detection to Unseen Generators via Harmful Subspace Removal

Abstract

Machine-generated text (MGT) detection is essential for identifying the provenance of online content, yet existing detectors often generalize poorly to unseen generators. We find that this failure is closely tied to generator-specific shortcuts: detectors trained for human–AI classification encode strong generator identity information, and their cross-generator transfer patterns are highly structured rather than uniformly transferable. A natural solution is to identify and remove generator-cue tokens, but our analysis shows that token-level deletion is ineffective because cue-bearing tokens also contain useful human–AI discrimination evidence. Motivated by this observation, we propose Harmful Cue sUbspace Removal with REplay Pool (CURE), a representation-level purification framework for robust MGT detection under unseen-generator shifts. CURE first uses a frozen generator-ID teacher to estimate token-level generator cues, then maintains a three-stream replay pool to construct a low-rank harmful cue subspace by contrasting harmful-cue and clean-reference representations. Instead of dropping tokens, CURE selectively removes generator-specific directions inside token hidden states with confidence-guided strength, while training the detector on both original and purified views. Experiments on 12 unseen generators show that CURE achieves an F1 score of 89.9, outperforming 8 representative detectors by 8.5 points. Further analyses demonstrate that subspace-level removal is substantially more effective than token-level deletion, suppresses generator-specific information while preserving human–AI discrimination signals. We will release our code and datasets upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Machine-Generated Text Detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_w3K9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies AI-text detection on unseen generators and proposes CURE. The main idea is to use a generator classifier to locate source-related cues, then selectively suppress directions in token representations. Experiments on 12 unseen generators, supported by ablations, cross-domain tests, and attack evaluations, show improvements over the tested baselines.

Critical strengths

CURE outperforms the tested baselines on unseen generators.

Critical weaknesses
  1. The paper misses two closely related works. “Robust AI-Generated Text Detection by Restricted Embeddings (Findings of EMNLP, 2024)” already uses subspace removal to improve cross-generator detection. “Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection (ACL, 2026)” separates detection cues from generator-specific features. Without discussing or comparing against them, it is hard to see what CURE adds.
  2. I am not convinced that CURE selectively removes harmful shortcuts. Generator-ID probing alone shows that generator information is present, but not that the detector relies on it. In Table 7, lower AI scores on AI-generated samples could also reflect the removal of useful detection evidence. Moreover, the residual representation in Table 6 remains highly predictive of the human–AI label (94.2% accuracy). So, it is unclear how well CURE separates harmful generator cues from useful detection signals.
Questions for authors
  1. Does CURE outperform Restricted Embeddings and DRGD under the same training and validation setup? What advantage comes specifically from its token-level intervention?
  2. At a fixed human false-positive rate, compared to random removal matched in rank and removed energy, does removing the learned subspace improve unseen-generator recall?
  3. Does +Subspace differ from the full model only in confidence gating? If other components also change, how much of the reported gain can be attributed to the gate alone?
Reviewer_MsHwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the poor generalization of machine-generated text detectors to unseen generators and links this failure to generator-specific shortcuts. It proposes CURE, which identifies tokens that bear generator cues, estimates a low-rank harmful subspace using three replay pools, and removes projections onto this subspace. The detector is trained on both raw and purified representations, while their predictions are fused at inference time. Experiments on 12 unseen generators show substantial improvements over the compared baselines.

Critical strengths
  1. The problem is important, the motivation is clear, and the evaluation is broad. The paper supports its motivation through pairwise transfer and generator-ID probing experiments.
  2. The comparison between token deletion and subspace removal is insightful. The results suggest that cue-bearing tokens contain both generator-specific shortcuts and useful human-AI evidence. Selectively removing representation directions is therefore better motivated than deleting entire tokens.
  3. The generator-shortcut hypothesis is interesting and empirically supported. Pairwise transfer exhibits clear generator-dependent structure, while linear probes recover generator identity from detector representations. These observations provide useful evidence that binary detector training may increse the effection of generator-specific information.
Critical weaknesses
  1. The novelty is insufficiently established. Prior work has already studied generator-specific overfitting and harmful-subspace removal (e.g. Robust AI-Generated Text Detection by Restricted Embeddings). The paper should discuss these works and clarify how its method differs from them.
  2. Many design choices are not adequately ablated. The paper does not separately evaluate the broad-harmful, hard-harmful, and clean-reference pools, even though the first two may overlap. Further ablation experiments are needed to quantify the contribution of each design choice.
  3. Several methodological and experimental details are insufficiently explained. See the Questions for Authors below.
Questions for authors
  1. Could the authors provide the missing implementation details of CURE? In particular, how are full replay pools updated, are they balanced across generators, and are the two harmful pools concatenated or combined as a duplicate-free set? Please also specify the teacher architecture, running-statistics computation, cue-score normalization, and ωmin⁡\omega_{\min}.
  2. Could the authors clarify the dataset construction and evaluation protocol? The paper describes the training data as balanced but appears to use 1,000 human and 4,000 machine-generated samples; how is class balance achieved? Please also provide the generation prompts, decoding parameters, and controls for topic and length, and clarify whether results can be reported on an established public protocol such as MAGE or RAID.
  3. What operations are included in the reported latency of CURE? Does the 5.07 ms include the teacher pass, gradient-based saliency computation, and both raw and purified predictions? Please report end-to-end latency with a component-wise breakdown and the complete timing setup.
Reviewer_xu1tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CURE, a supervised AI-generated text detection framework based on hidden-state subspace purification. Specifically, CURE first employs a generator-identity classifier to identify tokens that potentially contain generator-specific cues. It then constructs a low-rank harmful cue subspace through replay pools and contrastive covariance estimation. The projected components along this subspace are subsequently removed from token hidden representations, and the detector is jointly trained on both original and purified views. The proposed method is evaluated on 12 unseen generators using arXiv abstract-based datasets, as well as additional benchmarks including AGNews, WritingPrompts, ELI5, and HC3.

Critical strengths
  1. This paper presents a relatively novel perspective on improving the generalization of AI-generated text detectors. Specifically, it suggests that supervised detectors may overfit generator-specific identity information, which could negatively affect their generalization to unseen generators.

  2. The proposed method achieves competitive or superior performance against multiple representative baselines across several evaluation datasets, demonstrating its potential effectiveness in improving detection generalization.

Critical weaknesses
  1. Formatting issues. There appear to be several formatting issues in the manuscript. First, two required sections, the AI Use Statement and the Reproducibility Statement, are missing. Second, the paragraph spacing around Lines 361–362 appears to have been modified and may be inconsistent with the rest of the manuscript.

  2. Experimental comprehensiveness. The experimental evaluation could be further strengthened in several aspects:

    (a) More recent and competitive baselines should be considered. For instance, metric-based detectors could include recent methods such as DNA-DetectLLM, while model-based detectors could include representation-based approaches such as RepreGuard.

    (b) Additional evaluations on widely adopted benchmarks in this research area, such as MAGE, RAID, RealDet, and DetectRL, would provide a more comprehensive assessment of the proposed method and strengthen the fairness and persuasiveness of the experimental comparisons.

    (c) It would also be beneficial to evaluate the robustness of CURE under different input sequence lengths, particularly to examine whether its performance remains stable for longer texts.

  3. Insufficient evidence supporting the central hypothesis. The representation probing results in Table 6 demonstrate that CURE can indeed suppress some generator-specific information. However, the removed residual representations still retain strong human–AI discriminative information. Moreover, the AI-vs-Human probing accuracy is largely preserved rather than improved after purification, decreasing slightly from 0.947 to 0.933. These observations raise questions about whether the removed subspace primarily captures harmful generator-specific cues or also contains useful task-relevant information. Could the authors further clarify how these results support the central hypothesis that selectively removing generator-specific information improves detection generalization?

  4. Limited backbone generalizability. Although CURE is presented as a general representation-level purification framework, the experiments are conducted exclusively using ALBERT-base as the detector backbone. Evaluating CURE with additional backbone architectures, such as RoBERTa or other LLMs, would provide stronger evidence of its applicability and effectiveness across different model architectures, thereby further strengthening the paper's contributions.

Questions for authors

See Weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.