acceptodds
Under review as a conference paper at ICLR 2027

VidForensics-M1: AI-Generated Video Detection and Forensics via Verifiable Meta-Detection Reinforcement Learning

Abstract

Recent advances in video generation models have dramatically improved the realism of synthetic videos, blurring the boundary between generated and authentic content and raising significant concerns about misinformation. Existing MLLM-based detectors predominantly rely on supervised fine-tuning or label-level reinforcement learning, where predefined or coarse-grained supervision signals limit their generalization to out-of-domain scenarios and emerging video generators. To overcome these limitations, we are the first to introduce **meta-detection** into AI-generated video detection, jointly supervising label correctness and evidence correctness during reinforcement learning to enable more reliable forgery detection. This paradigm presents two core challenges: (1) identifying evidence forms that provide reliable and scalable supervision for meta-detection, and (2) developing effective mechanisms to integrate such evidence into label-level reinforcement learning for trustworthy and generalizable synthetic video detection. Textual rationales offer semantically rich descriptions of forgery artifacts, yet their generation and verification rely heavily on external reference models, making the resulting supervision vulnerable to hallucinations and semantic biases. This motivates us to seek an evidence representation with verifiable ground truth to support rule-based meta-detection supervision. Temporal grounding meets this requirement because manipulation intervals can be precisely controlled and recorded during data construction. Building on this insight, we adopt temporal grounding as a more reliable meta-detection signal and propose an automated data construction pipeline that generates paired real and fake videos by reconstructing and replacing temporal segments using boundary-frame-conditioned video generation models. Furthermore, we propose **Evidence-Guided Reward Redistribution**, which performs evidence-aware credit assignment by redistributing rewards among label-correct responses according to their evidence quality, thereby preserving reliable label supervision while progressively encouraging the detector to acquire fine-grained and verifiable forgery localization capabilities. Extensive experiments demonstrate that **VidForensics-M1** effectively leverages verifiable temporal evidence to achieve more robust and generalizable AI-generated video detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MLLM-Based AI-Generated Video Detection

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_MSa1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes verifiable meta-detection for AI-generated video detection: beyond label-level rewards, the VLM detector (built on Qwen3.5-9B) outputs evidence in the form of temporal grounding, i.e., intervals where forgery artifacts occur, and is trained with reinforcement learning under EGRR. Evidence verification is rule-based computation (tIoU ≥ 0.7 against ground-truth intervals created by controlled construction); no reference model is needed at reward time, although the textual-evidence training subset is filtered by Gemini-3.1-Pro. Experiments cover held-out tests on five in-domain generators (ViF-Bench) and the external GenBuster-Bench. On ViF-Bench, Label-Tem-RL improves over Label-RL by 10.38% accuracy and over the textual-evidence route (Label-Exp-RL) by 7.06% accuracy; temporal grounding success rate rises e.g. from 26.5 to 79.5 on the LTX-Video-2B column (a 3.0–3.7x ratio across generators, range 61.5–79.5); on the fake-only GenBuster-Bench subset the paper reports Recall improvements of 7.0% (OOD) and 11.4% (Wild).

Critical strengths
  1. The problem framing is clear and valuable. Many prior explanation-producing detectors rely on teacher- or annotator-derived evidence that cannot be verified by rule; recasting detection supervision as "evidence that can be mechanically verified" is directly relevant to AI-generated content governance, where auditability of detection reports is a practical requirement.
  2. The supervision signal is cleanly designed and formally supported. Ground-truth intervals come from the controlled construction process and verification is pure rule computation (tIoU), so the signal is auditable at reward time. EGRR's redistribution design carries conservation, strict ordering, and anti-multiplicative-collapse properties (Appendix A), and the paper is careful to note these do not assert policy-gradient preservation. The temporal-grounding gains are large and consistent across generators.
Critical weaknesses
  1. EGRR lacks an isolated ablation. The +10.38% accuracy gain is a bundled comparison (Label-RL vs. Label-Tem-RL with EGRR, on ViF-Bench); the control variant "temporal grounding evidence + naive multiplicative reward (without EGRR)" is absent — the paper argues against multiplicative coupling only theoretically. Since EGRR is one of the four claimed contributions, its independent contribution over the evidence signal itself is not isolated empirically.
  2. The closest peer baselines are absent, and reported variance is incomplete. VidGuard-R1 (RL with specialized temporal-artifact rewards) appears in neither the main tables nor anywhere in the experiments, and the explanation-based line (DAVID-XR1, DeepTraceReward, VideoVeritas) is not compared under the same protocol either. Results are averaged over three random seeds, but no variance or significance tests are reported, so whether the 7.06% accuracy gain of Label-Tem-RL over Label-Exp-RL exceeds seed noise cannot be verified from the paper.
Questions for authors
  1. Can you add the control variant "temporal grounding evidence + naive weighted/multiplicative reward (without EGRR)" to isolate EGRR's independent contribution?
  2. How do VidGuard-R1 / VideoVeritas perform under your evaluation setting? If not reproducible, please explain why.
  3. Can the key comparisons (the 7-point gap at ladder step 3→4, and the bundled EGRR gain) be reported as multi-seed mean ± variance?
  4. At inference time on wild fake videos there is no ground-truth interval. How does evidence verifiability degrade? Could a small-scale human audit (e.g., plausibility of predicted evidence intervals on 200 external samples) support this?
  5. The controlled construction replaces segments with content aligned in duration, resolution, and frame rate on 5-second clips, so duration shortcuts appear controlled — but could splice-boundary or re-encoding discontinuities introduced by the replacement act as shortcuts? Was a shortcut audit performed?
3 comments
AuthorsOpenReview ↗

Response to Reviewer MSa1 (Part 1/2)

Dear Reviewer MSa1:

We greatly appreciate your thoughtful review and encouraging assessment of our work. Your constructive suggestions have provided valuable insights into further improving the quality and rigor of our paper. Below, we provide detailed responses to each of your questions and suggestions, and all corresponding revisions to the manuscript are highlighted in blue for ease of reference. If you have any further questions, please feel free to let us know.


Q1:Can you add the control variant "temporal grounding evidence + naive weighted/multiplicative reward (without EGRR)" to isolate EGRR's independent contribution?:

A1: Thank you for this valuable suggestion. We have included this important experiment in our revised manuscript to further validate the effectiveness of EGRR. Specifically, we compare EGRR with two naive reward strategies: Additive : 0.1 × Format Reward + 0.45 × Evidence Reward + 0.45 × Label Reward. Multiplicative : 0.1 × Format Reward + 0.9 × Evidence Reward × Label Reward. The experimental results are presented in Table 1 below:

Table 1: Effectiveness of EGRR Compared with Naive Reward Strategies.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓In-Domain Temporal Localization ([email protected]) ↑GenBuster-Bench Acc. ↑GenBuster-Bench Fake-Video Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
Additive81.5872.6079.769.4468.092.992.692.46.7
Multiplicative80.0172.1378.3012.1167.592.492.191.87.4
EGRR (Ours)83.1872.8681.246.5074.595.593.395.02.5

As shown in Table 1, EGRR consistently outperforms both naive reward strategies in detection performance and temporal localization accuracy. These results demonstrate that EGRR effectively addresses the inherent asymmetry in evidence supervision between real and fake videos, where evidence rewards are only applicable to fake samples. By redistributing rewards among label-correct responses based on evidence quality, EGRR better integrates verifiable evidence feedback into label-level reinforcement learning, leading to more accurate detection and more reliable forensic evidence.

**Q2: How do VidGuard-R1 / VideoVeritas perform under your evaluation setting? If not reproducible, please explain why. **

A2: Thank you for your valuable suggestion. We agree that comparing our method with VidGuard-R1 is important for highlighting the advantages of meta-detection. To ensure a fair comparison, we use the same 10K real videos and generators to construct 10K fake videos with each method's respective pipeline, yielding 20K training videos per method. We then train both models using their respective algorithms and evaluate them on ViF-Bench and GenBuster-Bench. The results are presented in Table 1 below:

Table 2: Comparison of VidGuard-R1 and VidForensics-M1 on ViF-Bench and GenBuster-Bench.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc. ↑GenBuster-Bench Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
VidGuard-R176.9067.1074.3913.3088.3886.4787.2410.00
VidForensics-M181.9072.1079.938.3092.490.091.65.5

As shown in Table 2, VidForensics-M1 consistently outperforms VidGuard-R1 across all evaluation metrics under the same training data sources and scale. These results further demonstrate the importance of verifiable meta-detection feedback in improving the generalization ability of AI-generated video detectors.

For VideoVeritas, we evaluate its publicly released model weights on ViF-Bench and GenBuster-Bench using the same evaluation protocol as our method. The results are presented in the following table 3:

Table 3: Comparison of VideoVeritas and VidForensics-M1 on ViF-Bench and GenBuster-Bench.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc. ↑GenBuster-Bench Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
VideoVeritas71.3566.9867.7124.2868.391.972.751.7
VidForensics-M1 (Ours)83.1872.8681.246.5095.593.395.02.5

As shown in the table, the consistent improvements over VideoVeritas across both benchmarks highlight the competitive performance of VidForensics-M1, further establishing its effectiveness in detecting OOD AI-generated videos.

AuthorsOpenReview ↗

Response to Reviewer MSa1 (Part 2/2)

Q3: Can the key comparisons (the 7-point gap at ladder step 3→4, and the bundled EGRR gain) be reported as multi-seed mean ± variance?

A3: Thank you for your valuable suggestion. To further validate the robustness of our method, we conduct three independent trials with different random seeds. The results are presented in the following table:

Table 4: Performance on ViF-Bench and GenBuster-Bench under Different Random Seeds.

MethodViF-Bench Acc ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc ↑GenBuster-Bench Fake Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
Qwen3.5-9B46.257.1711.1114.6769.956.762.416.8
+ Label-RL72.8063.8970.1518.2986.283.084.510.4
+ Label-Exp-RL76.1267.5873.9015.3489.986.388.66.2
+ Label-Tem-RL (Ours)83.18 ± 0.1272.86 ± 0.1481.24 ± 0.196.50 ± 0.2895.5 ± 0.2893.3 ± 0.3095.0 ± 0.332.5 ± 0.39

As shown in Table 4, our method consistently achieves improvements with low variance across different random seeds, demonstrating the stability and reproducibility of our approach.

Q4:At inference time on wild fake videos there is no ground-truth interval. How does evidence verifiability degrade? Could a small-scale human audit (e.g., plausibility of predicted evidence intervals on 200 external samples) support this?

A4: Thank you for your constructive suggestion. To further evaluate the forensic localization capability of VidForensics-M1 in real-world scenarios, we randomly select 200 fake videos from ViF-Bench and predict their forgery intervals. Four authors then manually inspect 50 intervals each for visible AI-generated artifacts. The results are presented in the following table:

Table 5: Human Verification of Visible AI-Generated Artifacts in Predicted Forgery Intervals.

MethodAuthor 1Author 2Author 3Author 4
+ Label-Tem-RL (VidForensics-M1)82%78%86%84%

These human-verified results further demonstrate the effectiveness of VidForensics-M1 in OOD scenarios, showing that our method can identify meaningful AI-generated artifacts in real-world videos. This provides additional evidence of the reliability of our forensic localization capability and highlights the practical value of VidForensics-M1 in real-world applications.

Q5:The controlled construction replaces segments with content aligned in duration, resolution, and frame rate on 5-second clips, so duration shortcuts appear controlled — but could splice-boundary or re-encoding discontinuities introduced by the replacement act as shortcuts? Was a shortcut audit performed?

A5: Thank you for your insightful question, and we sincerely apologize for the confusion. We would like to clarify that we standardize not only the resolution, frame rate, and duration of both real and fake videos, but also their encoding formats. To further investigate the impact of stitching artifacts on detection performance, we select 200 fake videos and evaluate VidForensics-M1 under two settings: using the complete stitched videos and extracting only the manipulated segments. The results are presented in Table 2 below.

Table 6: Performance of VidForensics-M1 With and Without Stitching Boundaries.

Input SettingRecall ↑
Complete Stitched Video91.0%
Manipulated Segment Only89.5%

The results show that VidForensics-M1 maintains effective detection performance even after removing the stitching boundaries, demonstrating that our method does not rely solely on stitching-specific shortcuts but instead learns meaningful AI-generated forensic cues for reliable detection and evidence localization.


Thank you again for your valuable time and encouraging feedback. We sincerely welcome any further questions or suggestions and would be pleased to engage in additional discussions to improve our manuscript.

Dear Reviewer MSa1,

With the discussion period nearing its end, we would greatly appreciate your feedback on whether our clarifications and supplementary experiments have adequately addressed your concerns. We welcome any further questions and sincerely appreciate your continued engagement.

Warm regards,

The Authors

Reviewer_979fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses AI-generated video detection with MLLM-based detectors, arguing that existing paradigms, that SFT on teacher-generated rationales and label-level RL, are respectively overfit-prone and too coarse-grained, encouraging detectors to exploit spurious cues. The authors propose meta-detection, which jointly supervises label correctness and evidence correctness during reinforcement learning, and adopt rule-based temporal grounding as the evidence form: an automated pipeline removes a temporal segment from real videos, regenerates it from boundary frames using video generators, and reinserts it, yielding paired real/fake videos with ground-truth manipulation intervals derived from the construction process itself. Evidence-Guided Reward Redistribution (EGRR) then reallocates reward among label-correct responses according to evidence quality while preserving the aggregate label reward. Experiments show the effectiveness of the proposed method on ViF-Bench and GenBuster-Bench.

Critical strengths
  1. The problem framing is clear and well-argued. The critique of label-level RL and of model-generated textual rationales is concrete, and each criticism is tied to a specific design decision.
  2. The controlled data-construction pipeline is the strongest asset. Ground-truth manipulation intervals come from the construction process itself, are auditable and rule-based, and require no teacher-model supervision, which is different from existing evidence-oriented datasets, which provide rationales but not exact intervals from controlled segment replacement.
  3. EGRR is simple and principled. Reward-preserving redistribution with two checkable guarantees, and the empirical ladder is clean and consistent, with striking temporal localization gains.
Critical weaknesses
  1. Related work already couples label supervision with temporal/localization supervision in RL for forgery detection, including AvatarShield (GRPO with an accuracy reward plus a temporal compensation reward), OmniVL-Guard (balanced RL for joint detection and grounding), and the temporal-forgery-localization line (e.g., ActivityForensics, EVAS, WAFL), none of which is cited even though temporal localization is this paper’s core evidence signal. VidGuard-R1 (cited as “VidGuard-RI”) is described but not evaluated as a baseline, which is the most direct omission.

  2. EGRR is applied only to fake-video samples, so for real videos the reward degenerates to label-only supervision and the headline claim of “jointly supervising label correctness and evidence correctness” is only partially implemented. The paper never specifies how the detector should behave on real videos or whether a hallucinated forged interval on real content is penalized, leaving open the risk that the model learns to always assert a forged interval on real videos.

  3. The paper motivates temporal grounding as fine-grained and verifiable, yet binarizes IoU at 0.7 “for a fair comparison with the binary semantic consistency scores,” while the theoretical treatment keeps the evidence score continuous-valued; this discards exactly the graded signal that distinguishes temporal grounding from textual rationales. A continuous-IoU EGRR variant should be reported as an ablation.

  4. On ViF-Bench, DeepTraceReward’s average recall (≈95.2%) far exceeds Label-Tem-RL’s (≈72.9%), yet the paper emphasizes Accuracy/F1 and claims to “consistently outperform”; this gap deserves explicit discussion, e.g., a precision–recall analysis, given the paper’s own argument that recall is critical for reducing real-world false negatives. The proprietary zero-shot models (Qwen3.7-Plus, GPT-5.5) should also be framed more carefully since they are not trained for this task.

  5. There is no study of the Gemini-3.1-Pro filter’s effect and rate, segment length, number or type of generators, the 5-s clip length, or the risk that paired construction teaches a source-video shortcut, nor any analysis of how segment-manipulated training transfers to the fully-synthetic OOD video that ViF-Bench and GenBuster-Bench actually evaluate.

Questions for authors

Please see weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer 979f (Part 1/3)

Dear Reviewer 979f:

We sincerely thank you for the time and effort you have devoted to reviewing our paper. Your constructive feedback is invaluable in helping us improve the clarity and rigor of our work. Below, we provide point-by-point responses to your comments. We have also incorporated all corresponding improvements into the revised manuscript, with all changes highlighted in blue for ease of reference. Please feel free to let us know if you have any further concerns or questions.


Q1.1:Related work already couples label supervision with temporal/localization supervision in RL for forgery detection, including AvatarShield (GRPO with an accuracy reward plus a temporal compensation reward), OmniVL-Guard (balanced RL for joint detection and grounding), and the temporal-forgery-localization line (e.g., ActivityForensics, EVAS, WAFL), none of which is cited even though temporal localization is this paper’s core evidence signal.

A1.1: Thank you for bringing these relevant studies to our attention. We sincerely apologize for overlooking these valuable contributions to the field. We have incorporated them into the Related Work section of our revised manuscript, with all additions highlighted in blue for ease of reference. Furthermore, we would like to clarify the key contributions that distinguish VidForensics-M1 from these existing approaches.

  1. Meta-Detection Paradigm. We are the first to bring meta-detection to AI-generated video detection as evidence-aware feedback for reinforcement learning.

  2. Temporal Grounding over Textual Rationales. We show that rule-based temporal grounding provides more reliable, verifiable feedback than model-based textual rationales, making it a better evidence source for meta-detection.

  3. Scalable Paired Data Construction. We develop an automated, scalable pipeline for boundary-frame-conditioned temporal segment reconstruction and replacement, producing paired real and fake videos with ground-truth manipulated intervals.

  4. Evidence-Guided Reward Redistribution (EGRR). EGRR redistributes rewards among label-correct responses by evidence quality, integrating meta-detection feedback into label-level reinforcement learning.

Q1.2: VidGuard-R1 (cited as “VidGuard-RI”) is described but not evaluated as a baseline, which is the most direct omission.

A1.2: Thank you for your valuable suggestion. We agree that comparing our method with VidGuard-R1 is important for highlighting the advantages of meta-detection. To ensure a fair comparison, we use the same 10K real videos and generators to construct 10K fake videos with each method's respective pipeline, yielding 20K training videos per method. We then train both models using their respective algorithms and evaluate them on ViF-Bench and GenBuster-Bench. The results are presented in Table 1 below:

Table 1: Comparison of VidGuard-R1 and VidForensics-M1 on ViF-Bench and GenBuster-Bench.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc. ↑GenBuster-Bench Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
VidGuard-R176.9067.1074.3913.3088.486.587.210.0
VidForensics-M181.9072.1079.938.3092.490.091.65.5

As shown in Table 1, VidForensics-M1 consistently outperforms VidGuard-R1 across all evaluation metrics under the same training data sources and scale. These results further demonstrate the importance of verifiable meta-detection feedback in improving the generalization ability of AI-generated video detectors.

Q2:EGRR is applied only to fake-video samples, so for real videos the reward degenerates to label-only supervision and the headline claim of “jointly supervising label correctness and evidence correctness” is only partially implemented. The paper never specifies how the detector should behave on real videos or whether a hallucinated forged interval on real content is penalized, leaving open the risk that the model learns to always assert a forged interval on real videos.

A2: We sincerely apologize for the confusion and appreciate your careful observation. We would like to clarify that textual explanations and temporal intervals are not applicable to real videos. Therefore, our prompt explicitly requires an empty string for the explanation and an empty list for the temporal interval when a video is predicted as real. Consequently, real videos only require correct classification, whereas fake videos require both correct classification and credible evidence. This inherent asymmetry in feedback signals between real and fake videos motivates our proposed EGRR algorithm, which redistributes rewards among label-correct responses for fake videos based on evidence quality, thereby avoiding unintended reward bias arising from asymmetric evidence supervision.

AuthorsOpenReview ↗

Response to Reviewer 979f (Part 2/3)

Q3:Binarizing temporal IoU at 0.7 discards fine-grained evidence information despite the continuous theoretical formulation. A continuous-IoU EGRR ablation is needed to assess this design choice.

A3: Thank you for your insightful question. We apologize for the confusion and clarify that the theoretical advantages of EGRR apply to both continuous and gated evidence feedback, with gated feedback being a special case. To evaluate different temporal grounding reward formulations, we conduct a continuous-IoU ablation, with results presented in Table 2 below:

Table 2: Comparison of Gated and Continuous IoU as Verifiable Temporal Grounding Feedback for Meta-Detection.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓In-Domain Temporal Localization ([email protected]) ↑GenBuster-Bench Acc. ↑GenBuster-Bench Fake-Video Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
EGRR (Continuous IoU)82.4372.2080.437.347393.492.992.86.2
EGRR (Gated IoU)83.1872.8681.246.5074.595.593.395.02.5

We would like to share a counterintuitive observation: gated IoU outperforms continuous IoU. Continuous IoU is more susceptible to reward hacking, as even poorly localized temporal intervals can receive positive rewards. In contrast, gated IoU with a threshold of 0.7 rewards only highly overlapping intervals. This stricter reward criterion encourages the detector to identify high-quality forensic evidence, thereby improving its generalization performance.

Q4:DeepTraceReward achieves substantially higher recall than Label-Tem-RL on ViF-Bench (95.2% vs. 72.9%), raising concerns about the emphasis on Accuracy/F1 and the claim of consistent superiority. A precision–recall analysis and clearer framing of untrained zero-shot baselines (Qwen3.7-Plus and GPT-5.5) are needed.

A4: Thank you for your valuable observation. We would like to clarify that DeepTraceReward's high recall on ViF-Bench primarily stems from its tendency to classify videos as fake, resulting in a substantially higher false positive rate (FPR), as shown in Table 3. Although this prediction bias improves recall, it also leads to frequent misclassification of real videos, limiting its practical utility and explaining why its Accuracy and F1 scores are lower than those of our method.

Table 3: False Positive Rate (FPR) Comparison Between DeepTrace-Reward and VidForensics-M1 on ViF-Bench.

MethodCogV-X1.5Hunyuan T2VHunyuan I2VGen4-TurboHailuo-02Pika-V2PixVerse-V4.5Kling-V1Sora-2Avg. FPR ↓
DeepTrace-Reward93.8793.8293.8792.8092.6593.3393.9992.8693.9293.46
VidForensics-M1 (Ours)3.605.926.444.006.178.427.198.468.376.50

We have incorporated the false positive rate (FPR) into Table 1 of our revised manuscript and revised the claim of "consistently outperforming" existing methods. We have also explicitly clarified that Qwen3.7-Plus and GPT-5.5 are evaluated in a zero-shot setting without any task-specific training. Thank you again for your valuable suggestions, which have helped us improve the rigor and clarity of our manuscript.

Q5.1: Reliability of Textual Evidence Verification.

A5.1: Thank you for your constructive question. To further validate the reliability of our textual evidence supervision, we analyze the filtering performance of Gemini-3.1-Pro and find that 28.7% of the textual explanations generated by GPT-5.5 are rejected. To assess the filtering quality, four authors manually inspect 200 rejected explanations (50 per author) and confirm that 86.5% indeed contain errors. We further examine 200 retained explanations and find that 93% are reliable. These results demonstrate the effectiveness of Gemini-3.1-Pro in filtering unreliable textual evidence, with an acceptable level of judgment error.

Table 4: Reliability of Textual Explanation Filtering and Semantic Consistency Verification.

MetricFiltering Rate (Gemini)Rejection Precision (Gemini)Retention Reliability (Gemini)Semantic Consistency Recall (Qwen)
Rate (%)28.786.593.094.5

Furthermore, we evaluate the reliability of the semantic consistency judge used during training. Specifically, we collect 200 pairs of detector-generated explanations and corresponding GPT-5.5 reference explanations, all manually verified as semantically consistent. Qwen3.5-4B correctly identifies 94.5% of these consistent pairs, indicating that its semantic consistency judgments are reasonably reliable, with a 5.5% false rejection rate on these samples.

AuthorsOpenReview ↗

Response to Reviewer 979f (Part 3/3)

Q5.2:Sensitivity to Manipulated Segment Length.

A5.2: To investigate the impact of manipulated segment duration on detection performance, we evaluate VidForensics-M1 on three held-out test sets, each containing 200 fake videos with manipulated durations of 1 s, 3 s, and 5 s, respectively. The results are presented in Table 5 below:

Table 5: Impact of Manipulated Segment Length on Detection Performance.

Manipulated Segment Length1 s3 s5 s
Recall ↑62.5%78%86.5%

As shown in Table 5, detection recall increases with manipulated segment duration, suggesting that longer manipulations provide more detectable forensic cues. In contrast, shorter manipulations are more challenging to detect, as their forensic artifacts tend to be subtle and temporally localized.

Q5.3:Impact of Generator Number and Diversity.

A5.3: To investigate the impact of generator number and diversity on detection performance, we consider three training settings: (1) one generator (LTX-Video-2B), (2) three generators (LTX-Video-2B, Wan2.2, and SkyReels-V2), and (3) five generators (LTX-Video-2B, Wan2.2, SkyReels-V2, Wan2.7, and Seedance1.0). To ensure a fair comparison, we use 20K training samples for each setting. The results are presented in Table 6 below:

Table 6: Impact of Training Generator Number and Diversity on Detection Performance.

Generator SetViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc. ↑GenBuster-Bench Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
1 Generator (LTX-Video-2B)76.4572.1575.3919.2590.991.490.39.4
3 Generators (LTX, Wan2.2, SkyReels-V2)80.4973.0578.9212.0792.792.292.16.9
5 Generators (LTX, Wan2.2, SkyReels-V2, Wan2.7, Seedance1.0)82.8374.3681.248.7093.793.393.25.9

As shown in Table 6, increasing the number and diversity of training generators consistently improves detection performance. Furthermore, incorporating more challenging samples from advanced generators, including proprietary models, may help the detector develop more generalizable detection capabilities. These findings suggest that diverse training generators contribute to stronger OOD generalization.

Q5.4: Choice of Video Clip Duration.

A5.4: Thank you for your insightful question. We would like to clarify that we adopt a fixed video duration of 5 seconds to systematically evaluate our proposed meta-detection feedback mechanism and EGRR algorithm. We focus on this controlled setting to minimize the additional complexity introduced by varying video lengths and better isolate the contributions of our proposed methods. In future work, we plan to extend training to videos of diverse durations and investigate the effectiveness and generalization of our approach in more complex, long-horizon scenarios.

Q5.5: Potential Shortcuts in Paired Construction.

A5.5: To address the concern that paired construction may introduce shortcuts through shared source videos, we ensure that each real video and its corresponding fake counterpart are assigned to the same data split (either training or evaluation), thereby preventing information leakage across the two splits.

Q5.6:Generalization to Fully Synthetic OOD Videos.

A5.6: As illustrated in Figure 4(a) of our manuscript, label-only reinforcement learning rewards correct binary predictions even when the detector relies on spurious cues rather than genuine forensic evidence. Such shortcuts can easily lead to reward hacking and undermine generalization to unseen scenarios. Our proposed VidForensics-M1 addresses this limitation by introducing meta-detection feedback based on verifiable temporal grounding, encouraging the detector to predict not only the correct authenticity label but also the manipulated temporal intervals. This rule-based meta-detection feedback is scalable and effectively mitigates reward hacking. By encouraging the detector to learn genuine forgery artifacts rather than exploit spurious correlations, VidForensics-M1 achieves stronger generalization to out-of-distribution (OOD) scenarios.


Thank you again for your valuable time and constructive suggestions. If you have any further questions or concerns, please do not hesitate to contact us. We would be delighted to continue our discussion with you to further improve the rigor and quality of our manuscript.

Comments on the paper

8 comments
AuthorsOpenReview ↗

Response to Reviewer dFWC (Part 1/2)

Dear Reviewer dFWC:

Thank you for your insightful comments and positive feedback, which are highly encouraging to us. We sincerely appreciate your valuable suggestions and questions. Below, we provide point-by-point responses to your concerns, and all corresponding revisions have been incorporated into the manuscript and highlighted in blue for ease of reference. If you have any further questions, please do not hesitate to contact us.


Q1:The experiments do not isolate EGRR itself. Tables 1–3 compare label, textual, and temporal supervision, but do not compare redistribution with additive or multiplicative rewards using the same temporal labels. The specific benefit of the proposed reward remains unclear.

A1: Thank you for this valuable suggestion. We have included this important experiment in our revised manuscript to further validate the effectiveness of EGRR. Specifically, we compare EGRR with two naive reward strategies: Additive : 0.1 × Format Reward + 0.45 × Evidence Reward + 0.45 × Label Reward. Multiplicative : 0.1 × Format Reward + 0.9 × Evidence Reward × Label Reward. The experimental results are presented in Table 1 below:

Table 1: Effectiveness of EGRR Compared with Naive Reward Strategies.

MethodViF-Bench Acc. ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓In-Domain Temporal Localization ([email protected]) ↑GenBuster-Bench Acc. ↑GenBuster-Bench Fake-Video Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
Additive81.5872.6079.769.4468.092.992.692.46.7
Multiplicative80.0172.1378.3012.1167.592.492.191.87.4
EGRR (Ours)83.1872.8681.246.5074.595.593.395.02.5

As shown in Table 1, EGRR consistently outperforms both naive reward strategies in detection performance and temporal localization accuracy. These results demonstrate that EGRR effectively addresses the inherent asymmetry in evidence supervision between real and fake videos, where evidence rewards are only applicable to fake samples. By redistributing rewards among label-correct responses based on evidence quality, EGRR better integrates verifiable evidence feedback into label-level reinforcement learning, leading to more accurate detection and more reliable forensic evidence.

Q2:Segment insertion may create shortcut cues. Matching duration, resolution, and frame rate does not rule out boundary or encoding differences. No control establishes whether the detector relies on generated content or the insertion process.

A2: Thank you for your insightful question, and we sincerely apologize for the confusion. We would like to clarify that we standardize not only the resolution, frame rate, and duration of both real and fake videos, but also their encoding formats. To further investigate the impact of stitching artifacts on detection performance, we select 200 fake videos and evaluate VidForensics-M1 under two settings: using the complete stitched videos and extracting only the manipulated segments. The results are presented in Table 2 below.

Table 2: Performance of VidForensics-M1 With and Without Stitching Boundaries.

Input SettingRecall ↑
Complete Stitched Video91.0%
Manipulated Segment Only89.5%

The results show that VidForensics-M1 maintains effective detection performance even after removing the stitching boundaries, demonstrating that our method does not rely solely on stitching-specific shortcuts but instead learns meaningful AI-generated forensic cues for reliable detection and evidence localization.

Q3:Localization generalization is only tested in-domain. Table 3 uses the same five generator families as data construction. The cross-generator detection results do not establish that predicted evidence remains accurate for unseen generators.

A3: Thank you for your valuable suggestion. To further evaluate the OOD forgery localization capability of VidForensics-M1, we construct 800 held-out samples from four unseen generators (200 per generator). The results are presented in Table 3 below:

Table 3: Evaluation of OOD Temporal Localization Capability Across Four Unseen Generators ([email protected]).

MethodCogVideoX-3Kling-Video-V4ViduQ3-ProHailuo-02
Qwen3.5-9B8.5%7.5%6.0%5.5%
+ Label-RL19.0%16.0%15.5%12.5%
+ Label-Exp-RL22.5%22.0%19.5%16.0%
+ Label-Tem-RL (VidForensics-M1)66.5%62.0%58.5%54.0%

These results further demonstrate the effectiveness and generalization ability of VidForensics-M1, highlighting its capability to accurately localize AI-generated forgery artifacts in videos produced by generators unseen during training.

AuthorsOpenReview ↗

Response to Reviewer dFWC (Part 2/2)

Q4:The comparison between evidence types depends on model judgments. GPT-5.5 and Gemini-3.1-Pro construct and filter textual evidence, and Qwen3.5-4B evaluates semantic consistency. Binary scoring alone does not make this verifier as reliable as exact temporal labels. A human audit is missing.

A4: Thank you for your constructive question. To further validate the reliability of our textual evidence supervision, we analyze the filtering performance of Gemini-3.1-Pro and find that 28.7% of the textual explanations generated by GPT-5.5 are rejected. To assess the filtering quality, four authors manually inspect 200 rejected explanations (50 per author) and confirm that 86.5% indeed contain errors. We further examine 200 retained explanations and find that 93% are reliable. These results demonstrate the effectiveness of Gemini-3.1-Pro in filtering unreliable textual evidence, with an acceptable level of judgment error.

Table 4: Reliability of Textual Explanation Filtering and Semantic Consistency Verification.

MetricFiltering Rate (Gemini)Rejection Precision (Gemini)Retention Reliability (Gemini)Semantic Consistency Recall (Qwen)
Rate (%)28.786.593.094.5

Furthermore, we evaluate the reliability of the semantic consistency judge used during training. Specifically, we collect 200 pairs of detector-generated explanations and corresponding GPT-5.5 reference explanations, all manually verified as semantically consistent. Qwen3.5-4B correctly identifies 94.5% of these consistent pairs, indicating that its semantic consistency judgments are reasonably reliable, with a 5.5% false rejection rate on these samples.

Q5:Uncertainty and false alarms are incompletely reported. Section 5.1 averages three runs without reporting their variation. Table 2 reports recall on fake videos only. Although ViF-Bench includes broader metrics, false-positive behavior on the second benchmark remains unknown.

A5: Thank you for your valuable suggestion. To further validate the robustness of our method, we conduct three independent trials with different random seeds and report the corresponding performance variations. We also extend our evaluation on GenBuster-Bench to include Accuracy, F1, and false positive rate (FPR) in addition to Recall, providing a more comprehensive assessment of detection performance. The results are presented in the following table 5:

Table 5: Performance on ViF-Bench and GenBuster-Bench under Different Random Seeds.

MethodViF-Bench Acc ↑ViF-Bench Recall ↑ViF-Bench F1 ↑ViF-Bench FPR ↓GenBuster-Bench Acc ↑GenBuster-Bench Fake Recall ↑GenBuster-Bench F1 ↑GenBuster-Bench FPR ↓
Qwen3.5-9B46.257.1711.1114.6769.956.762.416.8
+ Label-RL72.8063.8970.1518.2986.283.084.510.4
+ Label-Exp-RL76.1267.5873.9015.3489.986.388.66.2
+ Label-Tem-RL (Ours)83.18 ± 0.1272.86 ± 0.1481.24 ± 0.196.50 ± 0.2895.5 ± 0.2893.3 ± 0.3095.0 ± 0.332.5 ± 0.39

As shown in Table 5, our method consistently achieves improvements across a comprehensive set of evaluation metrics with low variance, demonstrating the stability and reproducibility of our approach.


Thank you once again for your valuable time and positive feedback. We would be happy to address any additional questions or concerns you may have and welcome further discussion to strengthen the clarity and rigor of our work.

AuthorsOpenReview ↗

Sincere Appreciation and Looking Forward to Further Feedback and Discussion

Dear Reviewers 979f, dFWC, and MSa1,

Thank you very much for the valuable time and effort you have devoted to reviewing our paper. Your constructive suggestions have helped us further improve the quality and rigor of our work.

As the discussion period will end in about three days, we would like to kindly ask whether our rebuttal has adequately addressed your concerns. We would be delighted to discuss and clarify any remaining questions or concerns you may have.

Warm regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.