acceptodds
Under review as a conference paper at ICLR 2027

See, Then Reason: Reflective On-Policy Self-Distillation for Deepfake Detection

Abstract

Multimodal large language models (MLLMs) have enabled deepfake detectors to support forensic reasoning alongside binary real-or-fake predictions. However, it remains unclear whether current MLLMs can preserve and effectively exploit the visual information required for deepfake detection. Our analysis reveals that deepfake detection performance consistently degrades across successive language-model layers. Reasoning-based SFT also underperforms the strongest intermediate-layer probe and exhibits a larger real-fake accuracy gap, indicating that the detection capability available in intermediate visual representations is not fully preserved in the final prediction. To address these limitations, we propose **SeeTR (See, Then Reason)**. Instead of entangling visual judgment with reasoning, SeeTR grounds the initial judgment in the most discriminative intermediate representation. It then builds reflective reasoning on this strong visual detection capability to determine whether the judgment should be kept or revised. Building on this reflective structure, we introduce **Contrastive Teacher On-Policy Self-Distillation (CT-OPSD)** to refine the model's reflective behavior on its own rollouts. SeeTR forms contrastive teacher contexts conditioned on the ground-truth label and its opposite, using their token-level disagreement to strengthen the model's ability to keep correct initial judgments and revise incorrect ones. Under the cross-domain & forgery setting, the initial visual judgment alone already outperforms prior MLLM-based detectors, and SeeTR further improves this visual baseline through reflective reasoning, achieving the highest average balanced accuracy among the evaluated MLLM-based detectors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Media Forensics, Deepfake Detection, Multimodal Learning, Reinforcement Learning, Self-Distillation

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_Nsu3OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper finds that deepfake detection performance generally degrades across successive MLLM layers. Based on this observation, the authors train a visual detector using intermediate-layer features to obtain an initial judgment. They then construct reflective reasoning data with the help of an external MLLM and train the target MLLM through RBI to keep or revise the initial judgment. Finally, they propose CT-OPSD, which uses token-level disagreement between positive and negative teachers to reweight on-policy self-distillation, further improving reflective reasoning and detection accuracy.

Critical strengths
  • I like the contrastive teacher token weighting, which uses disagreement between positive and negative teachers to place more emphasis on important reasoning tokens.
  • I like the design of LCLSL_{\mathrm{CLS}} in RBI. By averaging the MLLM and visual-detector logits, it encourages the MLLM to make stronger corrections when the initial visual judgment is wrong.
Critical weaknesses
  • One advantage of MLLMs over traditional binary detectors is that they can provide forensic reasoning in addition to real-or-fake predictions. However, the paper only evaluates detection accuracy, without assessing the reliability of the generated explanations.
  • The paper focuses on deepfake detection, but some baselines are designed for general AI-generated image detection. How well does this method perform on general AIGC detection benchmarks? Also, comparing general AIGC detectors on deepfake benchmarks may not be entirely fair.
  • Some mathematical symbols are not clearly defined, making the method harder to follow than necessary.
Questions for authors

The proposed method seems applicable to general AI-generated image detection, but the paper focuses on deepfake detection, while some of the baselines are designed for general AIGC detection. Could the authors clarify why the scope is limited to deepfake detection and better explain the relationship between these two tasks in the context of this work?

4 comments

We sincerely appreciate your time and effort in reviewing our paper. We respond to each of your questions one-by-one in what follows. We have made every effort to address your concerns thoroughly. If any concerns remain, we would greatly appreciate hearing them and the opportunity to address them further.


[W1] Reliability of the generated explanations

As you suggested, we supplemented the original qualitative analysis (original Figures 11–12, revised Figures 15–16) with machine and human evaluations. SeeTR ranks first overall and on every criterion in both evaluations. To separate explanation quality from detection accuracy, each evaluation uses only images that all detectors in that comparison classify correctly.

i) Machine evaluation of explanation reliability

All three MLLM judges give SeeTR (CT-OPSD) their highest overall score. We randomly sampled 500 fake images that all four detectors in the table correctly classified as fake in the cross-domain & forgery evaluation. Each judge saw the image and one anonymous rationale at a time, was told that the image was fake, and scored the rationale on three criteria using a five-point scale (revised Figure 8).

MethodQwen3-VL 30B-A3BInternVL3.5 30B-A3BGemma 3 27BAvg.
Veritas3.823.823.403.68
Veritas++3.974.263.804.01
SeeTR (RBI)4.794.204.054.35
SeeTR (CT-OPSD)4.914.294.124.44

Averaged across judges, SeeTR (CT-OPSD) also ranks first in artifact validity and localization, description quality, and clarity (4.36, 4.33, and 4.62; revised Figure 7).

ii) Human evaluation of explanation reliability

Human raters gave SeeTR the highest average score on all three criteria, with an overall score of 3.35 out of 4, compared with 2.60 for Veritas++, the strongest baseline (revised Figure 6). We randomly sampled 10 real and 10 fake images. For each image, 23 participants saw the ground-truth label and ranked four rationales on each of three criteria. We hid the detector names and shuffled the rationale order separately for each image (revised Figure 5).

MethodOverallArtifact localizationFactualityDescriptive depth & reliability
Veritas2.152.122.182.16
Veritas++2.602.672.482.66
Deep-VRM1.891.792.021.86
SeeTR (Ours)3.353.423.313.32

First through fourth place received 4, 3, 2, and 1 points, respectively.

The results, evaluation prompt, and study details are provided in Appendix D (D.1 human evaluation, D.2 machine evaluation) of the revised manuscript.


[W2, Q1] AIGC detection baselines

We deliberately included AIGC detectors in our evaluation because the two tasks share generative artifacts. Both AIGC images and facial deepfakes contain artifacts from generative models such as GANs and diffusion models. AIGC detectors learn these artifacts from diverse generated images. The same artifacts also appear in facial deepfakes created by entire face synthesis (EFS) and facial editing (FE). We mark AIGC detectors with † in Tables 1–2 of the revision. For a fair comparison, all baselines are evaluated with their official code and released checkpoints under the same protocol (Section 4.1).

Our comparison shows that AIGC detectors are much weaker on face swapping (FS) and face reenactment (FR), while remaining competitive on EFS and FE. In Table 2, SeeTR outperforms them by large margins on FS (DeepFaceLab: 72.96 vs. 43.35 and 51.30 for TranX-Adapter and Deep-VRM) and FR (HeyGen: 87.34 vs. 45.27 and 76.46). These manipulations replace the identity of a real face or drive its pose, expression, and mouth while preserving identity. Our artifact taxonomy explicitly covers them, which highlights the importance of fine-grained, face-specific rationale construction. This indicates that SeeTR examines facial forgery cues at a finer granularity than AIGC detectors, which makes it more effective on FS and FR.

We agree with the reviewer that extending our work to AIGC detection is an important future direction. Combining the fine-grained facial forgery details of this work with AIGC detection could lead to a broader deepfake detection framework. Such a framework should also strengthen detection of FS and FR, which existing AIGC detectors struggle with. As SeeTR's taxonomy and prompts are designed for facial forgeries, we have not evaluated it on general AIGC benchmarks in this work.

We have clarified the relationship between the two tasks in Section 4.1 and marked the AIGC detectors in Tables 1–2 of the revised manuscript.


[W3] Clarity of mathematical notation

To address your concern, we have thoroughly revised Section 3 to clarify the mathematical notation throughout. The corresponding changes are specifically highlighted in orange in the revised manuscript.

Dear Reviewer Nsu3,

Thank you again for your time and effort in reviewing our manuscript.

As the discussion period is coming to an end, we would like to kindly ask if there are any remaining points that you would like us to clarify.

We appreciate your thoughtful feedback and hope that our responses have adequately addressed your earlier comments.

Sincerely,
Authors

Thank you for the clarifications. Overall, I think this is a very good paper, and most of my concerns have been addressed. It would be helpful to see additional results on general AIGC detection benchmarks. I lean toward weak acceptance.

Dear Reviewer Nsu3,

Thank you for your encouraging feedback and for carefully considering our responses.

As you point out, general AIGC detection is also an important task in this field. We will continue to work on extending SeeTR to the general AIGC detection task.

Sincerely,
Authors

Reviewer_zB6bOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SeeTR, a framework for facial deepfake detection that separates an initial visual judgment from subsequent reflective reasoning. It derives the initial judgment from intermediate multimodal model representations and trains the model to decide whether to keep or revise it. The training combines visual supervision with contrastive teacher on-policy self-distillation to refine this behavior. The authors evaluate SeeTR under cross-domain and unseen-forgery settings, reporting stronger average balanced accuracy than the compared multimodal detectors and ablations supporting the proposed training components.

Critical strengths
  1. The layer-wise probing analysis identifies useful detection information in intermediate representations across several multimodal backbones. This gives the initial-judgment design a concrete empirical basis.

  2. The paper compares plain supervised fine-tuning with RBI and standard on-policy self-distillation with CT-OPSD. These comparisons help attribute the reported improvements to the proposed training stages.

Critical weaknesses
  1. Table 13 shows that CT-OPSD corrects 1,918 of 4,949 fake images initially predicted as real, but only 25 of 1,777 real images initially predicted as fake. This raises the possibility that part of the aggregate gain comes from a stronger tendency to predict “fake,” rather than a generally reliable ability to reconsider either type of error.

  2. Although the motivating layer-wise analysis covers several backbones, the complete SeeTR training and evaluation use only InternVL3. It is therefore unclear whether the proposed training procedure, rather than just the initial observation, transfers to other multimodal models.

Questions for authors
  1. Can the authors analyze the asymmetric flips in Table 13, including per-class error changes and whether the improvement remains after accounting for a possible shift toward “fake” predictions?
  2. Can the authors evaluate the complete SeeTR pipeline on another multimodal backbone and report variation across training runs? This would clarify the method’s transferability and stability.
  3. Which components are specific to deepfake detection, and if CT-OPSD is intended as a general binary-task method, can the authors provide evidence on another task or clarify the scope of their claim?
  4. Can the authors report inference latency and additional computation from the separate detector forward pass, and provide code or implementation details sufficient to reproduce the detector extraction, prompt construction, and artifact-filtering pipeline?
5 comments

We sincerely appreciate your time and effort in reviewing our paper. We respond to each of your questions one-by-one in what follows. We have made every effort to address your concerns thoroughly. If any concerns remain, we would greatly appreciate hearing them and the opportunity to address them further.


[W1, Q1] Further analysis of reflection behavior

As you note, the flips in reflection are asymmetric (original Table 13, revised Table 15), and a shift toward fake predictions can account for part of the aggregate gain. However, the improvement remains after accounting for this shift. The best global threshold we evaluated reproduces only 1.18 of SeeTR's 2.91-point gain in balanced accuracy (BAcc). In other words, reflection does more than shift predictions toward fake: it selectively revises wrong initial judgments.

i) The improvement remains after accounting for a shift toward fake predictions.
We adjust a global threshold δ\delta on the initial detector's fake-class scores. To favor the detector, we select δ\delta using the evaluation labels. The best threshold reaches 75.45% BAcc, 1.73 points below SeeTR (revised Table 5(a)). In other words, a global threshold shift, even an oracle one, cannot reproduce the gain from SeeTR's reflection.

MethodPredicted fake, pooled (%)Real acc., macro (%)Fake acc., macro (%)BAcc, macro (%)
Initial detector (δ=0.5\delta=0.5)35.0184.5164.0374.27
Oracle threshold (δ=0.4187\delta=0.4187)42.5477.0373.8775.45
SeeTR after reflection44.1877.5676.7977.18

The fake-prediction rate is pooled over all images. Accuracies are macro-averaged over the seven cross-domain & forgery sets.

The oracle threshold predicts fake at a rate similar to SeeTR (42.54% vs. 44.18%). Yet SeeTR is more accurate on both real (+0.53 points) and fake (+2.92 points) images. A global threshold only trades accuracy between the two classes, so no single threshold can match SeeTR on both.

ii) Per-class error changes and asymmetric flips (original Table 13, revised Table 15).
The initial detector is skewed toward real (84.51% real vs. 64.03% fake accuracy), so most of its errors are missed fakes. Reflection mainly reduces this fake-class error. Fake accuracy rises by 12.76 points (64.03 → 76.79), while real accuracy decreases by 6.95 points (84.51 → 77.56). The real–fake gap thus narrows from 20.48 to 0.77 points (original Table 10, revised Table 12). This indicates that the asymmetry reflects which initial judgments SeeTR learns to revise, rather than a uniform tendency to predict fake.

This change follows from the asymmetric flips in Table 15. Of the two types of error, reflection mainly corrects missed fakes, and corrections of real images initially predicted as fake remain rare (25/1,777). Flips toward fake, however, are selective. Among images initially predicted as real, SeeTR flips 38.8% of missed fakes (1,918/4,949) but only 5.8% of correctly identified real images (832/14,288). Reflection thus flips an initial "real" prediction 6.7× as often when it is wrong.

We have added the threshold analysis to Section 4.3 (Table 5) and the per-class analysis to Appendix C.2 of the revised manuscript.

[W2, Q2] Transferability to another backbone and stability across seeds

The complete SeeTR pipeline transfers to another multimodal backbone, and its improvement is stable across training runs. As you suggested, we applied it to Qwen3-VL-8B and repeated training with three seeds on both backbones.

For Qwen3-VL-8B, we use the same setting as for InternVL3-8B under the cross-domain & forgery evaluation. Its visual detector uses the best-probing intermediate layer, selected in the same way. For each backbone, we fix the visual detector, train RBI with three seeds, and train CT-OPSD with three seeds from the same RBI model. The only difference is that input images are resized to 512×512 to match the number of visual tokens of InternVL3-8B.

BackboneStageBAcc (mean ± std)Δ
InternVL3-8BVisual detector74.27–
+ RBI74.98 ± 0.58+0.71
+ CT-OPSD77.16 ± 0.03+2.89
Qwen3-VL-8BVisual detector70.27–
+ RBI72.65 ± 1.30+2.38
+ CT-OPSD74.70 ± 0.11+4.43

Mean ± std over three seeds. Δ is the gain in mean BAcc over the visual detector.

On both backbones, RBI and CT-OPSD each improve BAcc over the visual detector. After CT-OPSD, the standard deviation is only 0.03 points on InternVL3-8B and 0.11 points on Qwen3-VL-8B. The three-seed mean on InternVL3-8B (77.16%) is also consistent with the result in the paper (77.18%). With Qwen3-VL-8B, SeeTR still outperforms all MLLM-based detectors in Table 2 (best: Veritas++, 72.76%). This indicates that the gain from SeeTR is a property of the pipeline rather than of a particular backbone or training run.

The Qwen3-VL-8B detector is also skewed toward real (83.44% real vs. 57.09% fake accuracy), and reflection narrows this gap (26.35 → 2.42 points), as on InternVL3-8B. Reflection therefore corrects the detector's class skew consistently across backbones.

We have added these results to Appendix C.3 (Tables 16 and 17) of the revised manuscript. Table 17 further reports the dataset-wise BAcc before and after reflection for both backbones.

[Q3] Deepfake-specific design of SeeTR

As stated in the paper, SeeTR is a framework specialized for deepfake detection. Motivated by our finding that deepfake evidence is strongest in intermediate MLLM representations (Figure 1), we designed the data construction, the reflective response, and the prompts all specifically for deepfake detection. RBI and CT-OPSD are then built on this deepfake-specific design.

As you note, the CT-OPSD formulation itself seems applicable to other binary tasks, although we evaluate it only on deepfake detection in this work. In that case, we think two points would matter most. The first is clearly defined binary classes. In deepfake detection, real and fake are two clear hypotheses about the same image, so each label offers a meaningful alternative to the other. The second is carefully designed teacher prompts. The positive and negative teacher prompts should be crafted so that the two teachers contradict each other on the target task. If their disagreement is nearly the same at every reasoning position, the token weights become nearly uniform, as in standard OPSD (Eq. 11).

We see extending SeeTR to broader AI-generated content (AIGC) detection as a promising future direction and have clarified this point in the conclusion of the revised manuscript.


[Q4] Inference cost and reproducibility

As you suggested, we compared the inference cost of SeeTR with that of Veritas++, the second-best method in Table 2. SeeTR is faster in both latency and throughput, even though it requires two forward passes, including the visual detector.

Both methods run on the same 100 images, GPU (NVIDIA B200), and vLLM configuration. Latency is measured with one request at a time, and throughput with 64 concurrent requests.

MethodOutput tokens / image ↓Latency (s / image) ↓Throughput (images / s) ↑
Veritas++461.22.701 ± 0.69211.09
SeeTR (Ours)190.01.012 ± 0.15014.81 ± 0.08
↳ Visual detector–0.038 ± 0.00348.56 ± 1.47
↳ Reflection190.00.973 ± 0.15021.33 ± 0.43

SeeTR is faster because decoding dominates inference time, and it generates 59% fewer output tokens than Veritas++ under the same token limit. The detector forward pass is negligible: it takes 0.038 s, only 3.8% of SeeTR's latency. SeeTR is measured end-to-end without feature caching, so this is a conservative result. The additional detector pass therefore adds no practical cost, and the shorter reflective response makes SeeTR faster overall.

Reproducibility. The paper describes the detector extraction (Appendix A.3), the prompt construction (original Figures 5–10, revised Figures 9–14), and the artifact extraction and filtering (Appendix B). We will also release the code for the entire pipeline with the camera-ready version.

We have added Table 6 and the corresponding paragraph "Inference cost" to Section 4.3 of the main text.

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

This paper focuses on MLLM-based deepfake detection and proposes SeeTR. It first trains a lightweight vision detector using intermediate visual representations from a frozen MLLM. The model is then trained to keep or revise the initial judgment through Reflective Behavior Internalization(RBI). Contrastive Teacher On-Policy Self-Distillation(CT-OPSD) is applied after RBI. It uses a teacher under both GT and opposite label contexts to weight token-level distillation. The results on DF40 show that SeeTR outperforms the evaluated MLLM-based detectors.

Critical strengths
  1. The authors find that intermediate visual features perform better for deepfake detection than later-layer features under linear probing, and observe this trend on InternVL3, Qwen3-VL, and LLaVA. The SFT comparison on InternVL3 also suggests that the final predictions may not fully use these visual features. I think this observation is useful and motivates the proposed method.
Critical weaknesses
  1. Although the authors compare plain SFT with the RBI objective, there is still a lack of ablation studies for RBI, since it combines several loss terms, including visual knowledge distillation, autoregressive supervision , and regularized classification. I think separate ablations are important to tell which part of the loss is the key contributor, and whether adding a classification loss on the final decision to plain SFT can bring similar improvements.
  2. The reflection behavior is asymmetric. Reflection is able to correct many missed fake images, but corrects very few real images that are falsely predicted as fake. The authors acknowledge this behavior, but I think they need to clarify if the improvement is from better reflection or is just from the model being more likely to predict fake. The behavior needs more analysis.
Questions for authors
  1. Can the authors provide separate ablations of the RBI objectives? For example, a comparison with only plain SFT and a classification loss on the final decision.
  2. Can the authors provide more analysis to the asymmetric behavior of reflection?
4 comments

We sincerely appreciate your time and effort in reviewing our paper. We respond to each of your questions one-by-one in what follows. We have made every effort to address your concerns thoroughly. If any concerns remain, we would greatly appreciate hearing them and the opportunity to address them further.


[W1, Q1] Further ablation study on the RBI objective (Eq. 6)

We agree that a more detailed ablation study of the RBI supervision components would strengthen our paper. In the original submission, Table 3 compared plain SFT with full RBI. Following your suggestion, we have extended this comparison by evaluating final classification supervision (LCLS\mathcal{L}_{\mathrm{CLS}}) and visual knowledge distillation (LVKD\mathcal{L}_{\mathrm{VKD}}) separately on the same reflective training responses. Revised Table 3 shows that each objective improves on plain SFT, but only their combination in RBI surpasses the initial visual judgment, indicating that the two objectives are complementary.

Stage 1LCLS\mathcal{L}_{\mathrm{CLS}}LVKD\mathcal{L}_{\mathrm{VKD}}Avg. BAcc (%)Δ
Plain SFT✗✗65.56—
w/o LVKD\mathcal{L}_{\mathrm{VKD}}✓✗67.83+2.27
w/o LCLS\mathcal{L}_{\mathrm{CLS}}✗✓69.58+4.02
RBI (ours) (Eq. 6)✓✓75.06+9.50

Adding LCLS\mathcal{L}_{\mathrm{CLS}} or LVKD\mathcal{L}_{\mathrm{VKD}} individually improves BAcc over plain SFT by 2.27 and 4.02 points, but both remain below the initial visual judgment (74.27%). Only full RBI surpasses it, reaching 75.06% (+9.50 points). This indicates that the two objectives are complementary: LVKD\mathcal{L}_{\mathrm{VKD}} grounds the initial judgment in the detector distribution, while LCLS\mathcal{L}_{\mathrm{CLS}} adjusts the final decision according to the detector evidence.

We have added these results to Table 3 and the paragraph "Effect of individual RBI objectives" in Section 4.3 of the revised manuscript.

[W2, Q2] Further analysis of asymmetric reflection behavior

As you note, we discussed the asymmetric behavior of reflection in our original submission (original Table 13, revised Table 15). Relative to the initial detector, SeeTR predicts fake more often and has lower real-image accuracy. However, our additional analysis shows that lowering the detector's threshold reproduces only part of SeeTR's gain in balanced accuracy (BAcc). This suggests that the asymmetry reflects which initial predictions SeeTR learns to revise during reflection: these selective flips improve fake-image accuracy enough to outweigh the loss in real-image accuracy.

i) SeeTR outperforms the best global threshold we evaluated.
We adjust a global threshold δ\delta on the initial detector's fake-class scores. To favor the detector, we select δ\delta using the evaluation labels. The best threshold reaches 75.45% BAcc, 1.73 points below SeeTR (revised Table 5(a)). In other words, a global threshold shift, even an oracle one, cannot reproduce the gain from SeeTR's reflection.

MethodPredicted fake, pooled (%)Real acc. (%)Fake acc. (%)BAcc (%)
Initial detector (δ=0.5\delta=0.5)35.0184.5164.0374.27
Oracle threshold (δ=0.4187\delta=0.4187)42.5477.0373.8775.45
SeeTR after reflection44.1877.5676.7977.18

The fake-prediction rate is pooled over all images. Accuracies are macro-averaged over the seven cross-domain & forgery sets.

The oracle threshold predicts fake at a rate similar to SeeTR (42.54% vs. 44.18%). Yet SeeTR is more accurate on both real (+0.53 points) and fake (+2.92 points) images, and it narrows the real–fake gap from 3.16 to 0.77 points. A global threshold only trades accuracy between the two classes, so no single threshold can match SeeTR on both.

ii) Reflection flips an initial "real" prediction 6.7× as often when it is wrong.
Corrections toward real remain rare. Of the 1,777 real images initially predicted as fake, SeeTR corrects only 25 (1.4%). Flips toward fake, however, are selective. Among images initially predicted as real, SeeTR flips 38.8% of missed fakes (1,918/4,949) but only 5.8% of correctly identified real images (832/14,288) (original Table 13, revised Table 15). Thus, reflection uses image evidence to decide which initial judgments to revise, rather than uniformly favoring fake.

iii) Reflection changes low-confidence predictions more frequently.
We group images by detector confidence, defined as the detector's probability for its predicted class. The flip rate decreases from 25.08% in the lowest bin to 0.15% in the highest (revised Table 5(b)). The oracle threshold changes predictions only in the 0.5–0.6 bin. SeeTR also improves BAcc in the 0.6–0.8 bins (+2.45 and +2.59 points), which this threshold does not affect. Reflection therefore mainly revises uncertain initial judgments and also improves cases that no threshold shift can reach.

Detector confidenceBAcc before reflectionBAcc after reflectionΔ BAccFlip rate (%)
0.5–0.658.2362.74+4.5125.08
0.6–0.769.2771.72+2.4516.08
0.7–0.875.1077.69+2.596.80
0.8–0.984.2184.88+0.670.77
0.9–1.092.7793.11+0.340.15

We have added this analysis to Section 4.3 (Revised Table 5 and the paragraph "Analysis of reflection behavior") and Appendix C.2 of the revised manuscript.

Dear Reviewer jSbP,

Thank you again for your time and effort in reviewing our manuscript.

As the discussion period is coming to an end, we would like to kindly ask if there are any remaining points that you would like us to clarify.

We appreciate your thoughtful feedback and hope that our responses have adequately addressed your earlier comments.

Sincerely,
Authors

Dear Reviewer jSbP,

Thank you once again for the time and care you have devoted to reviewing our work.

Given that the discussion period closes within 9 hours, we would like to send one final reminder, as this may be our last opportunity to respond.

For your convenience, we have briefly summarized below how we have addressed your questions.

  • [W1, Q1] RBI ablation. Neither loss term alone (+2.27 and +4.02 BAcc points) matches full RBI (+9.50 points) (revised Table 3).
  • [W2, Q2] Asymmetric reflection. Even an oracle threshold on the detector remains 1.73 BAcc points below SeeTR (revised Table 5).

We hope that these results have adequately addressed your comments, and we would be happy to clarify any remaining points.

Sincerely,
Authors

Comments on the paper

1 comment

Dear Reviewers,

We sincerely appreciate your time and effort in reviewing our paper. In response to your feedback, we have conducted additional experiments and revised our manuscript as summarized below. Please note that the corresponding changes are highlighted in color.

  • RBI objectives. Separate ablations show that final classification supervision and visual knowledge distillation each improve on plain SFT, with the largest gain when combined (jSbP [W1, Q1]).
  • Asymmetric reflection. Reflection mainly corrects missed fakes and flips an initial "real" prediction 6.7× as often when it is wrong. SeeTR also outperforms the best global detector threshold by 1.73 points at a similar fake-prediction rate (jSbP [W2, Q2], zB6b [W1, Q1]).
  • Backbone transferability and seed stability. On Qwen3-VL-8B, RBI and CT-OPSD each improve BAcc over the visual detector, as on InternVL3-8B, and the improvement is stable across three training seeds (zB6b [W2, Q2]).
  • AIGC task and AIGC baselines. We clarify which components are specific to facial deepfakes and why we evaluate general AI-generated image detectors on facial manipulations. Our empirical claims are limited to facial deepfake detection (zB6b [Q3], Nsu3 [W2, Q1]).
  • Inference cost and reproducibility. We report end-to-end latency and the cost of the additional detector pass, and document detector extraction, prompt construction, and artifact filtering (zB6b [Q4]).
  • Explanation reliability. On images correctly classified by all compared methods, SeeTR ranks first on every criterion in both the human and MLLM evaluations. Each of the three MLLM judges also gives SeeTR the highest overall rationale score (Nsu3 [W1]).
  • Mathematical notation. We clarified the notation in the method section; notation-related revisions are marked in orange (Nsu3 [W3]).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.