- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper proposes BiTRA to address the performance imbalance of MLLM-based AI-generated video detectors, particularly their weakness on I2V videos. It combines SHAI for artifact-guided feature injection and TAS for temporal attention regularization. Experiments on three benchmarks, including comparisons, ablations, robustness tests, and visualizations, support its effectiveness.
Overall, despite some issues with its motivation, the paper maintains strong logical coherence and novelty. I would raise my score if the authors provide reasonable explanations or clarifications.
Critical strengths
S1 - The method is well-motivated, and the overall narrative is logically coherent. Addressing the performance imbalance of MLLMs, especially their weakness on I2V videos, is novel and interesting.
S2 - The experiments are solid. The authors demonstrate the effectiveness of the method through comprehensive comparisons, ablation studies, robustness tests, and visualizations.
Critical weaknesses
W1 - There are some writing and formatting issues. For example, some sentences in the abstract are confusing (e.g., the use of "However" on line 25), the introduction (too long paragraph) is hard to follow (lines 75–98), and there are issues with the quotation marks (line 35).
W2 - The authors' hypothesis is that existing LLMs allocate more attention weight to earlier frames. However, this alone does not clearly demonstrate that the LLMs fail to make full use of features from later frames. The relationship between attention distribution and information contribution has not been directly verified, and lower attention weights do not necessarily mean that detection cues in later frames are not used. From the perspective of attention weights alone, this motivation is not entirely rigorous.
Questions for authors
Q1 - The authors state, "However, as the representations propagate through the LLM decoder, the performance on I2V videos degrades substantially, whereas that on T2V videos improves markedly, thereby further exacerbating the cross-paradigm imbalance, as shown in Fig. 2(a)." However, I do not observe a substantial decline in I2V detection performance across successive LLM decoder layers in Fig. 2(a). Do the authors mean that detection performance on I2V videos drops substantially when the representations produced by the ViT are passed into the decoder, whereas the opposite occurs for T2V videos?
Q2 - How are the probes for the ViT and the LLM decoder in Fig. 2(a) configured? Could the authors describe the complete setup, including where features are extracted, how spatial and temporal aggregation is performed, and whether predictions are made at the frame or video level? Furthermore, I would like to know why this attention bias arises only in the decoder (could the attention analysis in Appendix C.2 and Fig. 7 also be performed on the ViT? ).
Q3 - Based on an estimate from the average attention distribution shown in Fig. 7, the normalized KL divergence appears to be far below the margin (\delta=0.4) used in the paper. I am therefore concerned that TAS may not be providing sufficient regularization. Could the authors provide sensitivity or ablation experiments for this threshold and report the activation rate of the TAS loss during training?
(By the way, the statement “Darker colors indicate a larger proportion of final-token attention allocated to the corresponding frame.” in Figure 7 (Line 1146) is clearly incorrect. The authors should recheck it.)
5 comments
Response to Reviewer c9Be (Part 2/3)
Response to Reviewer c9Be (Part 2/3)
Response to Questions:
Q1: Clarification on the layer-wise probing trends in Fig. 2(a).
A1: Thank you for pointing out the ambiguity. Your interpretation is correct: Fig. 2(a) shows a substantial decrease in probing performance for I2V when comparing the ViT output with early LLM decoder representations, whereas probing performance for T2V increases over the same transition. We have revised the corresponding statement in the Introduction to clarify this distinction (Lines 86–90, highlighted in blue in the revised manuscript).
Q2: Probe configuration in Fig. 2(a) & Why early-frame attention bias arises only in the LLM decoder.
A2: Thank you for requesting these clarifications. We specify the probing protocol and the attention analyses below, covering the Qwen3-VL-Embedding-2B decoder, CLIP ViT-L/14, and the frozen Qwen3-VL-ViT.
-
Probe Configuration in Fig. 2(a).
- Feature Extraction: At each probed ViT layer, we apply spatial mean pooling over the patch tokens within each frame, followed by temporal mean pooling across all sampled frames to obtain a video-level representation. At each probed LLM decoder layer, following [8], we extract the hidden state of the last non-padding token.
- Prediction Level: We train a separate linear probe for each layer on the ViF-Bench training split and evaluate it at the video level on the T2V and I2V test splits.
-
Why Early-Frame Bias Arises Only in the Decoder & ViT Attention Analysis.
- Why in the LLM decoder: The LLM uses a decoder-only architecture with unidirectional causal self-attention, where tokens can only attend to preceding positions. As established in prior studies [9, 10, 11], causal masking structurally induces a strong position bias / attention sink toward initial tokens, causing the decoder to favor early-frame visual tokens while under-attending to later frames.
- ViT Attention Analysis (Appendix G, Fig. 12, highlighted in blue): Because the native ViT in MLLMs (
Qwen3-VL) encodes each frame independently without cross-frame attention or a[CLS]aggregation anchor, the same frame-level attention analysis cannot be applied directly. We therefore design two complementary ViT probes: (i) on a frozen CLIP ViT-L/14, we concatenate the patch tokens of all 16 frames with the[CLS]token and measure normalized[CLS]-to-frame attention across all 24 layers; and (ii) on the frozen Qwen3-VL-ViT, we train a learnable aggregation query with a single cross-attention layer and an MLP classifier on ViF-CoT, and measure the query's normalized attention to each frame across ViF-Bench. As shown in Appendix Fig. 12, while the MLLM decoder exhibits a clear position bias toward early frames due to unidirectional causal attention, the ViT employs bidirectional self-attention, whose attention fluctuates naturally across frames without such position bias.
Response to Reviewer c9Be (Part 3/3)
Response to Reviewer c9Be (Part 3/3)
Q3: Clarification on margin δ, sensitivity ablation, activation rate, and Fig. 7 caption typo.
A3: Thank you for your suggestions. We provide more details below.
-
Correction of the Reported δ Value: Thank you for pointing out this inconsistency. The value
0.4in Section 4.2 was a typographical error; the experiments reported in the original submission used δ = 0.04. We have corrected the value in Line 291 (highlighted in blue). This correction concerns only the text and does not affect the previously reported results. -
Sensitivity Ablation for δ: We ablate δ ∈ {0.02, 0.04, 0.08, 0.1} on ViF-Bench, reporting the average balanced accuracy (bAcc, %), Recall (%), and F1 (%) across all generators.
| Margin δ | bAcc | Recall | F1 |
|---|---|---|---|
| 0.02 | 97.7 | 97.4 | 97.5 |
| 0.04 (Default) | 98.3 | 98.6 | 98.2 |
| 0.08 | 97.3 | 97.7 | 97.1 |
| 0.10 | 96.0 | 96.5 | 95.7 |
-
Training Activation Rate and Distribution Shift of TAS (Appendix E, Figs. 9 & 11):
- Activation Rate (Appendix Fig. 9): Under δ = 0.04,
L_tasstarts with a 96.3% non-zero activation rate in Epoch 1 and smoothly decays to 6.8% by Epoch 10 (averaging 31.7% across epochs) as the distribution enters the target margin zone. - Distribution with and without TAS (Appendix Fig. 11): Without
L_tas, the normalized KL divergence remains large due to inherent attention bias in the LLM decoder; withL_tas, it becomes substantially smaller and is compressed below δ = 0.04, confirming sufficient regularization of our proposedL_tas.
- Activation Rate (Appendix Fig. 9): Under δ = 0.04,
-
Correction of Fig. 7 Caption (Line 1146): We have corrected the caption of Fig. 7 in the revised manuscript (highlighted in blue) to accurately describe the line curves.
References
[1] Li et al., "Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning," CVPR, 2026.
[2] Kay et al., "The Kinetics Human Action Video Dataset," arXiv:1705.06950, 2017.
[3] Wan et al., "Wan: Open and Advanced Large-Scale Video Generative Models," arXiv:2503.20314, 2025.
[4] Jiang et al., "VACE: All-in-One Video Creation and Editing," ICCV, 2025.
[5] HaCohen et al., "LTX-Video: Realtime Video Latent Diffusion," arXiv:2501.00103, 2024.
[6] Zhuang et al., "Myopia Rectification: KV Cache Pruning for MLLMs via Dynamic Attention Subsidy and Token Reclamation," CVPR, 2026.
[7] Li et al., "MIHBench: Benchmarking and Mitigating Multi-Image Hallucinations in Multimodal Large Language Models," ACM MM, 2025.
[8] Azaria and Mitchell, "The Internal State of an LLM Knows When It's Lying," Findings of EMNLP, 2023.
[9] Tian et al., "Identifying and Mitigating Position Bias of Multi-Image Vision-Language Models," CVPR, 2025.
[10] Liu et al., "Lost in the Middle: How Language Models Use Long Contexts," arXiv:2307.03172, 2023.
[11] Li et al., "The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity," ICML, 2026.
Response to Reviewer c9Be (Part 1/3)
Response to Reviewer c9Be (Part 1/3)
We sincerely thank you for your positive assessment of our method as “well-motivated” and “novel and interesting”, as well as for your constructive feedback. We address your comments point by point below.
Response to Critical Weakness:
W1: Writing and formatting issues in the Abstract and Introduction.
A1: Thank you for pointing out these writing and formatting issues. We have addressed all three issues in the revised manuscript, with the changes highlighted in blue:
W2: Insufficient experimental evidence that lower attention weights for later frames imply underutilization of their forgery cues.
A2: To directly verify whether the MLLM fails to make full use of forgery cues in later frames, we conduct a controlled zero-shot comparison between First-Frame-to-Video (F2V / standard I2V) and Last-Frame-to-Video (L2V) samples:
Wan2.1-VACE-1.3B[3, 4] andLTX-Video-0.9.8-13B-distilled[5]) and matched real source videos, we synthesize 1,000 F2V videos (500 + 500) and 1,000 paired L2V videos (500 + 500). Between F2V and L2V, the generators and settings are strictly identical, and the sole difference is whether the conditioning real frame is placed at the first frame or the last frame. We evaluate zero-shot fake recall (%) using pretrained Qwen3-VL-2B, Qwen3-VL-8B, and Qwen3.5-8B: