- Rating
- 3
- Confidence
- 3(+1)
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes LesionProbe, a method for detecting AI-generated text. It removes the attention and MLP modules at different layers of a frozen language model and measures how the output predictions change. These changes are used as features to train a small detector. The paper also uses Shapley values to compare the effects of attention and MLP modules on human and AI-generated text. Experiments show that the method performs well across different source models and writing domains, although its performance is less stable in some cross-benchmark settings.
Critical strengths
- The paper is easy to follow. The authors explain the motivation and method clearly.
- The main idea is interesting. The paper detects AI text by removing parts of a language model and observing how its predictions change. This is different from most existing methods.
- The experiments are fairly complete. The method is tested in several settings and is compared with strong baselines as LTD, and the results generally support the main claim of the paper.
Critical weaknesses
-
The paper provides a computational cost analysis for LesionProbe but gives limited information about the cost of the baseline methods. The LesionProbe results in Table 1 use the full 32-layer setting, which requires 97 runs per input. The faster five-layer version shows lower performance in a separate experiment. In some settings, strong baselines such as LTD achieve performance close to LesionProbe. Reporting their computational costs would therefore help readers better assess the practical benefits of LesionProbe and its suitability for deployment.
-
LesionProbe performs well on average, but its performance varies across different benchmark transfers. This suggests that the detected patterns may still contain benchmark-specific information or noise.
Questions for authors
q1. Zeroing a module may create unnatural hidden states. Have the authors tested other intervention methods, such as mean replacement or partial scaling, to verify that the detected patterns are not specific to zero ablation?
q2. What factors do the authors believe cause the large performance differences across benchmark transfers, especially in settings such as MAGE-to-AIGTBench?
2 comments
W2 and Q2. Understanding Performance Differences Across Benchmark Transfers
Thank you for raising this concern. We conducted a feature analysis to investigate the differences in cross-benchmark performance. The results suggest that stronger consistency in human–AI feature patterns between RAID and AIGTBench contributes to better transfer performance than between MAGE and AIGTBench.
We examine this explanation using two measures: class separation within each benchmark and alignment of class-separation directions across benchmarks. We use token-averaged lesion-response features from the MAGE, RAID, and AIGTBench test sets, with a shared standardization fitted only on MAGE and RAID training data.
For each benchmark , let denote the standardized lesion-response feature vector of an individual text, and let and denote the AI and human feature centroids, respectively. We define:
Larger indicates greater separation between the two centroids relative to within-class variability, suggesting better class discriminability within the benchmark. To compare how the classes differ across benchmarks, we define:
\cos\left( \mu_{S,A}-\mu_{S,H}, \mu_{T,A}-\mu_{T,H} \right).$$ Higher $A_{S,T}$ means that the human-to-AI feature direction is more consistent between the source and target benchmarks, potentially facilitating cross-benchmark generalization. The results are summarized in Table 3. **Table 3.** Feature separation and directional alignment, with AIGTBench as the shared target. Values are mean ± standard deviation across five runs. | Source → Target | Source separation J | Target separation J | Direction alignment A ↑ | OOD AUROC (%) ↑ | |---|---:|---:|---:|---:| | MAGE → AIGTBench | 0.113 ± 0.002 | 0.149 ± 0.006 | 0.322 ± 0.014 | 60.11 ± 5.49 | | RAID → AIGTBench | 0.100 ± 0.003 | 0.149 ± 0.006 | 0.765 ± 0.001 | 91.45 ± 1.30 | AIGTBench exhibits relatively strong within-benchmark class separation ($J=0.149$), consistent with its high ID AUROC (95.48%). However, its class-separation direction aligns more closely with RAID ($A=0.765$) than with MAGE ($A=0.322$), corresponding to substantially different OOD AUROC (91.45% vs. 60.11%). A similar pattern emerges in reverse transfers: AIGTBench → RAID (89.81%) vs. AIGTBench → MAGE (72.87%). These consistent bidirectional trends further support directional misalignment as a plausible explanation for the observed performance differences. Thank you again for your careful review and constructive feedback. Have a great day!
We thank you for taking the time to review our submission and for your thoughtful feedback. Below, we address your concerns point by point.
W1. Inference Cost and Performance Trade-off
Thank you for this constructive suggestion. We have added a comparison of detection performance and inference cost. LesionProbe achieves the highest AUROC in this setting while requiring less inference time and GPU memory than LTD, and sparse interventions further improve efficiency. Detailed results are presented in Table 1 below.
Table 1. Detection performance and inference cost on DetectRL. Trainable methods are trained on ChatGPT-generated text and corresponding HWT; evaluation covers all four source LLMs.
Although full-layer interventions incur higher inference costs than RoBERTa and RepreGuard, sparse interventions offer a practical trade-off between performance and efficiency. Compared with the full configuration, Peak-5 achieves a 6.4× speedup with only a 1.09 percentage point decrease in AUROC. Using layer 32 alone increases the speedup to 26.4× while maintaining an AUROC of 95.95%.
Q1. Are the Results Specific to Zero Ablation?
Thank you for this constructive suggestion, which helps us better assess the robustness of LesionProbe to different intervention choices. LesionProbe consistently achieves strong ID and OOD detection performance across different intervention strategies, showing that its effectiveness does not depend on zero ablation. Detailed results are presented in Table 2 below.
Table 2. ID and OOD performance under alternative interventions on texts generated by ChatGPT and Google-PaLM (mean ± std, %). α denotes the branch-output scaling factor; α = 0 is zero ablation.
Partial scaling with α = 0.25, 0.50, and 0.75 maintains stable ID and OOD performance. Mean replacement also performs strongly, with a slight decrease in OOD TPR. These results suggest that LesionProbe’s effectiveness is not specific to zero ablation and that module-intervention responses provide strong detection signals across intervention choices.