acceptodds
Under review as a conference paper at ICLR 2027

Provenance is not Universal: Discovering and Distilling Forensic Subspaces in Frozen Vision-Language Models

Abstract

AI-generated image detectors often achieve strong in-distribution performance but degrade on unseen generators and datasets, raising the question of whether pretrained vision-language models (VLMs) encode more transferable provenance signals. Although frozen features from foundation models have shown promising performance for synthetic image detection, it remains unclear how such provenance information is organized within VLM representations. We address this gap by studying prompt-conditioned hidden states from seven frozen VLMs. We introduce Behavioral Provenance Signatures (BPS) to characterize separability, generator-conditioned forensic subspaces, cross-domain transferability, and distillability. We observe that frozen representations consistently support strong in-distribution separation but exhibit substantial variation across models and source generators. Generator-conditioned probes have partially aligned yet distinct discriminant directions, showing that strong separability does not imply a universal provenance axis. Distillation is similarly heterogeneous across teachers, target domains, and student architectures. Nevertheless, the resulting image-only models eliminate VLM execution and prompt processing at test time, achieving more than a measured inference speed-up. These results distinguish the availability of provenance information from its robustness and deployability.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AI-generated image detection, image provenance, vision-language models, forensic subspaces, knowledge distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jmqyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether hidden states of frozen VLMs support detecting AI-generated images, whether the signal transfers to unseen generators and datasets, and whether it can be distilled into small image-only models. The method concatenates last-token hidden states from eight descriptive prompts, fits a PCA plus linear-probe teacher on each of five GenImage generators, compares the teachers' discriminant directions, and distills them into MobileNetV3 and ResNet-18 students. All seven VLMs separate real and generated images well in distribution, the five teachers' directions are partially aligned but distinct, and transfer and distillation gains vary widely across generators and test sets, with students that are much faster but noticeably weaker than the teacher on external benchmarks.

Critical strengths

S1. Beyond asking whether frozen VLM features can detect generated images, the paper asks whether detectors trained on different generators rely on the same discriminant direction. This question bears directly on why single-source training transfers so unevenly.

S2. The external evaluation and model-selection protocol is carefully controlled. External benchmarks play no role in fitting or model selection, the deployment configuration is fixed in advance, each student is compared with a label-only student that shares its data, architecture and seed, and the paper also reports its below-chance results.

Critical weaknesses

W1. The direction differences are confounded with image size The title's conclusion draws largely on comparing the teachers' discriminant directions. In Fig. 3b, SD1.5 and Wukong are close to each other (cosine 0.67), ADM, GLIDE and VQDM are close to each other (0.64 to 0.75), and every cross-group pair is below 0.49. This grouping may reflect the generators themselves, but it also coincides exactly with fake-image resolution in GenImage (512×512 for SD1.5 and Wukong, 256×256 for the other three). Grommelt et al. (2024) showed that detectors trained on GenImage exploit this size difference, as well as the JPEG-versus-PNG gap between real and fake images, which affects all five teachers alike. The limitations section acknowledges that teachers may pick up dataset or compression cues, but no experiment, such as a size-matched comparison, separates generator-specific signal from these cues. Without such a control, the current evidence cannot attribute the direction differences to generator provenance.

W2. The contribution of the language layers and prompts beyond visual features is not tested BPS runs the VLM once per prompt, eight times per image, and concatenates the final-layer last-token states. The representation comparison in Sec. 4.2 covers only different aggregations of language-layer hidden states. There is no comparison with Qwen3-VL vision-encoder, CLIP or DINOv2 features under the same training data and linear probe, and no ablation on the number of prompts. Zhou et al. (2026), cited in the paper, show that linear probes on frozen visual features already generalize well. Without these comparisons, I cannot tell whether the separability and direction structure the paper analyzes come from the language layers and prompts or are inherited from the visual features, or whether eight VLM forward passes per image are needed.

Questions for authors
  1. Each teacher is fitted only once. If a teacher is refitted on a different sample of training images from the same generator, how high is the cosine between the two directions? If it is comparable to the within-group values among ADM, GLIDE and VQDM in Fig. 3b, the differences among these three may simply reflect sampling noise.
  2. Sec. 6 states that "Decision-weighted matching addresses this issue", namely that accurate global reconstruction may not preserve the coordinates most relevant to detection. In the component ablation cited in Sec. 4.4 (Table S16), however, adding the weighting changes the mean OOD AUROC by −0.004, logit distillation alone reaches 0.690 against 0.683 for the full method, and Sec. 4.4 itself notes no additional gain from weighting. Could you report the full method with only the weighting removed, so that its contribution can be judged in the configuration actually deployed?
  3. PROBE-DINO-L, PROBE-RN50, HSIC Bottleneck and Effort in Table 1 are not cited. Do these baselines and UnivFD use released weights, or were they trained on the same SD1.5 data as the students, and what is their latency under the protocol of Table 1(b)? PROBE-RN50 scores 0.13 to 0.19 higher than the ResNet-18 student on the three benchmarks. Where do the students stand relative to detectors of comparable compute trained on the same data, for example NPR [r1]?
  4. According to Sec. 3.5, LgL_g in Fig. 4b is computed with label-only students. Would the conclusion in Sec. 4.4 that selecting a teacher for transfer and selecting one for distillation are different decisions still hold with distilled students?
  5. A few minor issues. I could not find the code, and the reproducibility statement mentions robustness_results.json and source-cache hashes without saying where to obtain them. Panel B of Fig. 1 derives the signature r(x)r(x) from "geometry + trajectory", whereas Eq. (5) defines r(x)r(x) as the final-layer concatenation. The retention figures of 81.0% and 77.0% in Sec. 4.4 are raw AUROC ratios. Measured as the fraction of the teacher's margin above chance that the student keeps, they are about 53% and 43%.

[r1] Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection. CVPR, 2024.

4 comments

Thank you.

W1: Resolution and compression confounds

We refit all five teachers under the original processing and a deterministic 224x224 center crop without prior resizing. Both conditions use the same eligible images within each source, with identical labels, splits, prompts, and classifier settings. Cropping changes the learned directions, but within-source refit similarity exceeds between-source similarity, as shown under Q1.

These results show reproducible differences beyond the original 512-versus-256 grouping. Image content, real-image pools, and compression history may also contribute.

W2: Visual-feature and prompt controls

We added frozen visual-feature probes with the same source data, PCA-112, and classifier settings. The table reports external AUROC differences, language minus visual:

Visual representationSD1.5ADM
Qwen VT-0.0033-0.0771
CLIP+0.0429+0.0080
DINOv2+0.1237-0.0520

The language representation does not significantly improve over Qwen's vision tower on SD1.5 and performs worse on ADM. These results show that useful detection information is already accessible through visual features.

At the matched dimension, eight prompts outperform a validation-selected single prompt across three sources and underperform on two. For GLIDE, another prompt tied on validation AUROC also exceeds eight prompts externally. Eight passes are therefore not consistently necessary and take 3317.06 ms versus 865.03 ms for one prompt.

Q1: Refitting variability

We refit PCA, standardization, and the classifier on repeated stratified training halves, matching sizes across sources (1000 images per half in the original-data experiment). Within-source refit cosines range from 0.715 to 0.773. The table reports mean within-source cosine minus between-source cosine for pairs with the same original resolution.

Source pairOriginal within-betweenCropped within-between
ADM-GLIDE+0.149+0.246
ADM-VQDM+0.065+0.110
GLIDE-VQDM+0.135+0.288
SD1.5-Wukong+0.070+0.052

Resampling refit repetitions gives significant differences for all four pairs under both conditions (Holm-adjusted p=0.0016), supporting source differences beyond the measured refitting variability.

Q2: Removing only decision weighting

We conducted ablations with five paired seeds by removing only decision weighting from the full objective, keeping all other loss terms and coefficients fixed. Weighting changes external AUROC by -0.0017 for MobileNetV3 and +0.0131 for ResNet-18. The differences are not statistically significant. Logit-only KD also remains competitive. We will revise Section 6 to state that decision weighting emphasizes classifier-relevant coordinates, but its additional benefit over unweighted matching is not established.

Q3: Baseline citations, training data, latency, and NPR

PROBE, HSIC, Effort, and UnivFD use released weights, without retraining on our split. HSIC and Effort checkpoints are trained on SD1.4 data. We will document checkpoint provenance and add citations for PROBE (Cao et al., ICML 2026), HSIC (Yang et al., ICLR 2026), and Effort (Yan et al., ICML 2025).

The latency results (median end-to-end latency, batch size 1, FP32) are:

DetectorExternal macro AUROCLatency (ms)
PROBE-DINOv2-L0.9739493.35
PROBE-ResNet500.802528.79
HSIC Bottleneck0.678028.70
UnivFD0.632420.91
Effort0.692237.82
NPR, trained on our split0.620911.37
Distilled MobileNetV30.683412.78
Distilled ResNet-180.649111.31

We trained the official NPR architecture on our SD1.5 split with five seeds and source-validation checkpoint selection. NPR uses random initialization and 50 epochs, whereas the students use ImageNet initialization and 150 epochs.

Against label-only students with identical architecture, initialization, data, and training settings, full KD improves AUROC from 0.6604 to 0.6834 for MobileNetV3 and from 0.6246 to 0.6491 for ResNet-18. We will report the offline teacher cost.

Q4: Teacher selection using distilled students

With full-KD students, Wukong exceeds SD1.5 by 0.0157 AUROC averaged across architectures, although SD1.5 gives the stronger teacher. The difference is +0.0203 for ResNet-18 and +0.0111 for MobileNetV3. We will replace the label-only proxy with these results and qualify the conclusion by architecture. No deployment choice was changed using external results.

Q5: Code, Fig1, and retention

We will provide anonymous code and reproduction instructions, including access to or regeneration of robustness_results.json and source-cache hashes.

Fig1 will be revised to show final-layer prompt concatenation as the deployed signature, with geometry and trajectory labeled as alternatives.

Retention above chance, Retention=AUROCS−0.5AUROCT−0.5Retention=\frac{AUROC_{S}-0.5}{AUROC_{T}-0.5}, gives 53.4% for MobileNetV3 and 43.5% for ResNet-18 with the SD1.5 teacher. We will report the absolute AUROCs for teachers and students alongside these values.

Thank you for your comments. The revisions are marked in blue.

W1. Section 4.3 and Table 2 (p.7) compare original processing with common-size cropping on the same eligible images. Appendix C.9 and Tables S19–S20 (pp.40–41) give the refitting results and processing details. We revised the title, Introduction, and Limitations to describe source dependence under the tested conditions. The text acknowledges that content, real-image pools, and compression history may still contribute.

W2. Section 4.2 and Table 1 (p.6) add frozen Qwen vision-tower, CLIP, and DINOv2 comparisons. Appendix C.8 and Tables S17–S18 (pp.39–40) report prompt comparisons at the same final dimension. We no longer imply that language layers or eight prompts are necessary.

Q1. Table 2 (p.7) and Appendix C.9/Table S19 (p.40) compare within-source refit similarity with between-source similarity. The appendix separates the original full-cohort results from the paired processing controls.

Q2. Appendix C.10 and Table S21 (p.41) report the requested ablation on both architectures with five paired seeds. Sections 4.4–6 (pp.8–10) revise the claim about weighting to reflect these results.

Q3. Table 3 (p.9) adds baseline citations and measured latency. Appendix B.3–B.4 (pp.20–23) specifies released-weight usage, NPR training, and the timing protocol. Table S21 (p.41) compares matched conventional students, and Table S23 (p.42) reports benchmark-specific baseline scores.

Q4. Figure 4b and Section 4.4 (p.8) use actual full-KD students in place of the label-only proxy. Appendix C.11 and Table S22 (pp.41–42) give the results and uncertainty for each architecture. The deployment choice remains unchanged.

Q5. Figure 1 (p.3) now shows final-layer prompt concatenation as the deployed signature. Section 4.4 (p.9) and Equation 54 in Appendix A.4 (p.17) report retention above chance. Appendix B.3 (p.22) documents the historical teacher and checkpoint checks. The Reproducibility Statement (p.11) describes the code release and regeneration instructions, and the anonymous repository link has been added.

We hope that our clarifications and revisions address your concerns, and we would be glad to continue the discussion if any concerns remain.

Dear Reviewer jmqy,

Thank you for your supportive assessment and detailed suggestions. We added refitting and cropping controls, visual feature and prompt comparisons, and the requested weighting ablation. We also evaluated teacher selection using distilled students, expanded the baseline comparisons, and corrected Fig.1 and the retention metric.

We hope these additions strengthen the paper and address your concerns. As the discussion deadline approaches, we would be happy to clarify any remaining questions. We appreciate your support and consideration of the revised manuscript.

Regards,
Authors

Reviewer_jmqyOpenReview ↗

Official Comment by Reviewer jmqy

Thanks for the response and the new experiments. My main concerns are addressed. I will keep my positive score.

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

This paper provide an analysis on whether VLM frozen features contain transferable provenance information for detecting AI generated images. The proposed BPS analyzes generator-conditioned discrimination directions across 5 generators and then distill into lightweight image-only detectors. The results show consistently strong ID performance while the OoD performance is varying.

Critical strengths
  • The problem is important
  • The evaluation is large-scale covering 7 VLMs, 5 generators and multiple OoD Benchmarks
  • The analysis on representation separability, transferability and distilability are interesting and informative.
Critical weaknesses
  • The methodology or analysis novelty is limited. There are existing studies demonstrating frozen foundation models can train linear classifier for AI generated images.
  • The practical advantage of proposed method is limited. For example, the students reduce inference time significantly, but their OoD performance is considerably lower. We should have a comparison between student with matched computational budgets to observe its practical contribution.
Questions for authors
  • Does using multiple prompts improve OoD detection compared to a single prompt under same representation dimension and classifier settings?
  • It seems that Tab. S16 show that logit-only KD achieve competitive performance to full objective, what is the additional benegit of decision-weighted subspace matching?
5 comments

Thank you.

Weakness 1: Novelty

Our main contribution is the analysis of source dependence, transfer, and distillation. Prior work establishes that frozen foundation-model features support generated-image detection. We investigate whether detectors trained on different sources use reproducible directions and whether a teacher's external performance predicts the performance of its distilled student.

Our new refitting experiments find differences between sources beyond the measured sampling variability, including among sources with the same original resolution and after common-size cropping. Although SD1.5 yields the stronger teacher, Wukong produces students with 0.0157 higher external AUROC averaged across the two architectures. We will make this empirical contribution clearer in the introduction.

Weakness 2: Practical comparison at similar inference cost

For the supplementary experiment, we added NPR trained on the same SD1.5 split and measured its inference cost alongside the students. The comparison uses the same detector-training images. Results are shown as follows (FP32, batch size 1):

DetectorExternal AUROCParametersEnd-to-end median latency
NPR0.62091.44M11.37 ms
Distilled MobileNetV30.6834approximately 3.1M12.78 ms
Distilled ResNet-180.6491approximately 11.2M11.31 ms

Results show that the students improve over NPR by 0.0626 and 0.0283 AUROC at similar measured end-to-end latency.

To assess distillation itself, we also trained conventional students with the same initialization and training setup. Their AUROCs are 0.6604 and 0.6246, compared with 0.6834 and 0.6491 after distillation. The students are approximately 260x and 293x faster than the eight-prompt teacher under the new end-to-end timing protocol. We will report the accuracy loss relative to the teacher and the offline teacher cost alongside these speedups.

Question 1: Multiple versus single prompts

To test this, we conducted supplementary experiments with 85 fits across the five sources, covering all eight individual prompts, four two-prompt subsets, four four-prompt subsets, and the eight-prompt representation. Every fit uses the same source split, training-only PCA and standardization, final dimension of 112, and classifier setting with C=0.1.

We used two single-prompt baselines: a fixed generic prompt (prompt 1) for all sources, and a prompt selected separately for each source from the eight candidates using source-validation AUROC. The table below reports the second comparison.

SourceSelected single promptEight promptsDifference
SD1.50.81930.8480+0.0287
ADM0.58460.6637+0.0791
GLIDE0.57420.7127+0.1385
Wukong0.80420.7789-0.0253
VQDM0.68170.6130-0.0687

The difference shown in the table is eight-prompt AUROC minus validation-selected single-prompt AUROC. It shows that eight prompts perform better on SD1.5, ADM, and GLIDE, but worse on Wukong and VQDM. Single prompts are selected by source-validation AUROC, with ties broken by prompt index. For GLIDE, another validation-tied prompt outperforms eight prompts.

In the separate comparison with the fixed generic prompt, eight prompts improve AUROC for all five sources by 0.0324-0.1870. This gain is relative to prompt 1 and does not indicate an improvement over the validation-selected single prompt. The benefit therefore depends on the source and the single-prompt baseline. The final dimension is matched, but extraction cost differs, as measured latency is 865.03 ms for one prompt and 3317.06 ms for eight prompts. We will include both prompt sensitivity and cost in the revision.

Question 2: Decision-weighted subspace matching

To isolate the contribution of decision weighting, we conducted an additional ablation. The experiment compares the full objective with unweighted subspace matching, keeping all other loss terms, coefficients, training data, and settings unchanged. We evaluated MobileNetV3 and ResNet-18 with five paired training seeds (42-46). The difference between weighted and unweighted external AUROC is -0.0017 for MobileNetV3 and +0.0131 for ResNet-18. Neither difference is statistically significant (with Holm-adjusted p=0.80 and 0.33).

Logit-only KD reaches 0.6898 and 0.6445, compared with 0.6834 and 0.6491 for the full objective. The full objective does not significantly outperform logit-only KD on either architecture. We will clarify that decision weighting emphasizes classifier-relevant coordinates, but our ablations do not demonstrate an additional OOD AUROC benefit relative to unweighted matching.

Thank you for your comments. The revisions are marked in blue.

W1. The Introduction (p.2) clarifies our contribution to understanding source-dependent directions, transfer, and distillation. Section 4.3 and Table 2 (p.7) add refitting and processing controls. Section 4.4 and Figure 4b (p.8) compare teacher ordering with the performance of their distilled students.

W2. Table 3 (p.9) reports external AUROC and end-to-end latency for the students, NPR, and released detectors. Appendix B.3–B.4 (pp.20–23) gives the training and measurement details, including NPR’s different initialization and training duration. Table S21 (p.41) compares distillation with conventional students under matched training settings.

Q1. Section 4.2 (p.6) summarizes the comparison at the same final dimension. Appendix C.8 and Tables S17–S18 (pp.39–40) report all 85 fits, separating the fixed generic prompt from the validation-selected single prompt. Eight prompts outperform the selected single prompt for three sources and underperform for two. Prompt-selection details are in Appendix C.8 (pp.39–40), and extraction costs are in Appendix B.4 (p.23).

Q2. Appendix C.10 and Table S21 (p.41) report the additional ablation that removes only decision weighting from the full objective, using five paired seeds for each architecture. The table also includes logit-only KD. We revised Sections 4.4-6 (pp.8–10) to state that the experiments do not establish an additional external-AUROC benefit from weighting.

We hope that our clarifications and revisions address your concerns, and we would be glad to continue the discussion if any concerns remain.

Dear Reviewer cBEZ,

We clarified the empirical contribution, added comparisons with detectors at similar inference cost, and tested multiple versus single prompts at the same representation dimension. We also added the requested weighting ablation and revised the claims to reflect its results.

We hope our rebuttal and revisions have addressed your concerns. As the discussion deadline approaches, we would be happy to clarify any remaining questions. If your concerns have been resolved, we would appreciate your reconsideration of the rating and support for our paper.

Regards, Authors

Reviewer_c9aJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work introduced BPS, which derives provenance representations from the prompt-conditioned hidden states of frozen VLMs and distills them into lightweight image encoders. This paper investigates whether frozen vision-language models (VLMs) contain transferable information for distinguishing real images from AI-generated images, how such information is organized across different source generators, and whether it can be distilled into lightweight image-only detectors.

Critical strengths
  1. A strength of this work is that it does not assume strong in-distribution detection automatically implies robust cross-generator generalization. The experiments show that frozen VLM representations can achieve nearly perfect in-distribution AUROC while exhibiting substantial variation on external benchmarks.

  2. The paper examines seven VLMs, multiple source generators, PCA dimensions, discriminant-direction alignment, two student architectures, and several distillation objectives. The appendix also provides additional analyses.

Critical weaknesses
  1. The main detection pipeline extracts frozen VLM hidden states, concatenates features across prompts, applies PCA, fits a linear classifier, and then distills the resulting representation into a smaller model. These are established techniques, and the main novelty appears to lie in their combination and the accompanying empirical analysis.

  2. The experimental results do not clearly demonstrate that the proposed prompt-conditioned or relational signature design is essential. The experiments (especially Section 4.2) show that this representation outperforms the explicitly constructed geometry and cross-layer trajectory signatures in cross-generator detection. So what is the actual contribution of BPS beyond standard frozen-feature probing?

  3. The practical effectiveness of the distilled detectors is not competitive with strong existing alternatives, so why is distillation necessary?

Questions for authors

None.

3 comments

Thank you.

Weakness 1: Novelty

In this paper, our key contribution is to study how the source used to train a detector affects its decision direction, external performance, and suitability for distillation. Near-perfect in-distribution detection does not tell us whether two detectors use the same cues or whether either will transfer well.

The new controls show that detectors refitted on images from the same source have more similar directions than detectors trained on different sources. This holds for all four pairs with the same original image resolution, including after a common 224-pixel crop. We also find that a stronger teacher can produce a weaker student: although SD1.5 gives the stronger teacher, Wukong gives students with 0.0157 higher external AUROC on average across the two architectures. We will revise the introduction to clearly state these contributions.

Weakness 2: Contribution of prompts and relational signatures

The deployed representation concatenates hidden states of the final layer across prompts. Geometry and trajectory signatures are alternatives evaluated in the paper and their lower transfer performance is one of the findings. We will correct Figure 1 to make this distinction clear.

We added single- and multi-prompt comparisons with the same training images, final dimension of 112, and classifier settings. Eight prompts outperform the original generic prompt for all five sources. When the single prompt is selected using source-validation AUROC, eight prompts perform better for three sources and worse for two. For GLIDE, another prompt tied on validation AUROC outperforms eight prompts externally. Multiple prompts can therefore help, but eight prompts are not necessary in every setting.

The new Qwen vision-tower, CLIP, and DINOv2 controls also show that visual features contain useful detection information. Compared with Qwen's vision tower, the language-layer representation gives no significant difference on SD1.5 and lower AUROC on ADM. We will clarify that BPS studies prompt-conditioned representations without claiming that the detection signal originates uniquely in the language layers.

Weakness 3: Why distillation is useful

Distillation improves small image-only detectors trained on the same data. Compared with conventional binary-head students using the same architecture, initialization, and seeds, full distillation raises external AUROC from 0.6604 to 0.6834 for MobileNetV3 and from 0.6246 to 0.6491 for ResNet-18. Logit-only distillation remains competitive with the full objective.

We also trained the official NPR architecture on the same SD1.5 train and validation images for 50 epochs over five seeds using random initialization. Our students use ImageNet initialization and 150 epochs. NPR reaches 0.6209 AUROC at 11.37 ms end-to-end latency, compared with 0.6834 at 12.78 ms for distilled MobileNetV3 and 0.6491 at 11.31 ms for distilled ResNet-18.

Distillation improves detection in small image-only models at low inference cost. The students remain less accurate than the VLM teacher and some released detectors: PROBE-ResNet50 reaches 0.8025 AUROC at 28.79 ms. We will report these accuracy-latency comparisons and the offline teacher cost.

Thank you for your comments. The revisions are marked in blue.

W1. We clarified the main contribution in the Introduction (p.2), which is understanding how the training source affects detector directions, transfer, and distillation. Section 4.3 and Table 2 (p.7) report the new refitting and cropping controls. Section 4.4 and Figure 4b (p.8) show that the stronger teacher does not necessarily produce the stronger distilled student.

W2. Figure 1 and Section 3.2 (pp.3–4) distinguish final-layer prompt concatenation, used by the deployed teacher, from the geometry and trajectory alternatives. Section 4.2 and Table 1 (p.6) add visual-feature and prompt comparisons. Appendix C.8 and Tables S17–S18 (pp.39–40) give the full prompt results. These comparisons do not establish that language layers or eight prompts are necessary.

W3. Section 4.4 and Table 3 (pp.8–9) report accuracy and latency, including NPR trained on our split. Table S21 (p.41) compares distillation with conventional students using the same architecture, initialization, data, and training settings. Appendix B.3–B.4 (pp.20–23) specifies the training differences, timing protocol, and offline cost. The Discussion (p.10) explains the practical benefit and the remaining accuracy loss.

We hope that our clarifications and revisions address your concerns, and we would be glad to continue the discussion if any concerns remain.

Dear Reviewer c9aJ,

We clarified the empirical contribution, added prompt and visual-feature controls, and included comparisons of accuracy and latency. The new results support reproducible source differences and show that distillation improves small detectors over matched label-only training.

We hope our rebuttal and revisions have addressed your concerns. As the discussion deadline approaches, we would be happy to clarify any remaining questions. If your concerns have been resolved, we would appreciate your reconsideration of the rating and support for our paper.

Regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.