acceptodds
Under review as a conference paper at ICLR 2027

Attention Meets Artifacts: Bi-level Temporal Re-Alignment for AIGV Detection

17 views

Abstract

Recent advances in video generation have raised growing security concerns, making AI-generated video (AIGV) detection increasingly essential. Existing methods predominantly rely on Multi-modal Large Language Models (MLLMs), but overlook the discrepancies among samples produced by different generation paradigms. However, we empirically find that MLLM-based detectors exhibit pronounced performance imbalance across different generation paradigms, with a notable weakness on image-to-video (I2V) samples. We attribute this to Temporal Artifact–Attention Misalignment. In particular, unlike entirely synthesized T2V samples, I2V samples are conditioned on a real initial image, and thus the forgery artifacts become progressively more pronounced toward later frames. The LLM decoder, however, attends predominantly to earlier frames owing to its causal attention mechanism; as a result, the most discriminative forgery cues in later frames are left largely unattended. To accommodate diverse generation paradigms, we propose a joint optimization strategy termed Bi-level Temporal Re-Alignment (BiTRA) that operates at both the extrinsic and intrinsic levels. Specifically, at the extrinsic level, we introduce ”Selective Hierarchical Artifact Injection”, an adaptive query mechanism that establishes an additional cross-attention pathway for global forgery information access, leveraging a temporal-artifact prior for discriminative information selection. At the intrinsic level, ”Temporal Attention Smoother” is guided by the Kullback-Leibler margin loss to mitigate the excessive attention bias inherent in the LLM decoder. Extensive experiments demonstrate the effectiveness of our designs on all generation paradigms and confirm our superiority over existing state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Video Detection, Multimodal Large Language Models, Temporal Artifact-Attention Misalignment, Video Forensics, Cross-Attention Injection

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_c9BeOpenReview ↗
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
AuthorsOpenReview ↗

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:

  • Abstract (Line 25): To improve logical coherence, we removed the redundant "However" and smoothed the transition between sentences.
  • Abstract (Line 35): To ensure proper formatting, we corrected LaTeX quotation marks around “Selective Hierarchical Artifact Injection” and “Temporal Attention Smoother”.
  • Introduction (Lines 75–98): To improve readability, we split the long diagnostic paragraph into two focused paragraphs—one analyzing the paradigm-dependent temporal evolution of forgery artifacts (T2V vs. I2V), and the other analyzing how the MLLM visual encoder and causal LLM decoder process these representations.

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:

  • Experimental Design: Following Skyra [1], we extract the first frame of each real video from the real video dataset Kinetics [2] as the shared conditioning anchor for both F2V and L2V. Using the exact same two representative generators (Wan2.1-VACE-1.3B [3, 4] and LTX-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:
Wan2.1-VACE-1.3B⠀LTX-Video-13B⠀Overall
ModelF2VL2V⠀F2VL2V⠀F2VL2V
Qwen3-VL-2B34.645.4⠀39.845.6⠀37.245.5
Qwen3-VL-8B36.847.8⠀42.851.2⠀39.849.5
Qwen3.5-8B37.452.2⠀43.850.2⠀40.651.2
  • Analysis: Under strictly aligned generators and source videos, simply shifting the real anchor from the first frame (F2V) to the last frame (L2V)—which moves the primary forgery artifacts from later frames to earlier frames—consistently boosts zero-shot fake recall across both generators and all three MLLM backbones. This directly proves that MLLMs utilize early-frame features much more effectively while failing to fully exploit forgery cues in later frames, aligning with recent studies [6, 7] showing that MLLMs exhibit an attention sink on earlier images and tend to overlook later visual evidence.
AuthorsOpenReview ↗

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.

  1. 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.
  2. 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.
AuthorsOpenReview ↗

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.

  1. Correction of the Reported δ Value: Thank you for pointing out this inconsistency. The value 0.4 in 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.

  2. 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 δbAccRecallF1
0.0297.797.497.5
0.04 (Default)98.398.698.2
0.0897.397.797.1
0.1096.096.595.7
  1. Training Activation Rate and Distribution Shift of TAS (Appendix E, Figs. 9 & 11):

    • Activation Rate (Appendix Fig. 9): Under δ = 0.04, L_tas starts 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; with L_tas, it becomes substantially smaller and is compressed below δ = 0.04, confirming sufficient regularization of our proposed L_tas.
  2. 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.

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

BiTRA addresses the mismatch between later-emerging artifacts in image-to-video (I2V) generation and early-frame-biased attention in multimodal LLMs. Its SHAI module retrieves multilevel, full-video visual features through cross-attention, while TAS regularizes prediction-token attention using a KL-margin loss. Training combines classification, I2V-specific temporal-barycenter supervision, and attention regularization, with ViT LoRA and full LLM fine-tuning on Qwen3-VL-Embedding-2B. The model demonstrates strong classification performance across multiple AI-generated video detection benchmarks.

Critical strengths
  1. The diagnosis of existing detectors’ weaknesses on I2V is an important contribution. Three distinct lines of evidence support a consistent diagnosis and motivate an actionable intervention.
  2. The ablation studies are extensive and informative. They isolate the contributions of BiTRA’s components and clearly demonstrate its detection gains over conventional methods.
Critical weaknesses
  1. The temporal prior may limit applicability beyond first-frame-conditioned I2V. Mitigating early-frame attention bias is effective on the evaluated I2V datasets. However, favoring later-frame retrieval and regularizing attention against a uniform reference introduce a strong prior over temporal attention allocation. For videos conditioned on both the first and last frames, or videos assembled from multiple generated segments, this prior may divert attention from discriminative forgery evidence. These settings require direct evaluation.
  2. The training protocols are insufficiently specified, limiting the interpretation of generalization results. Although the implementation details suggest dataset-specific training, the paper does not explicitly establish whether Table 2 uses the ViF-Bench-trained checkpoint or a model retrained on GenVideo. The corresponding training and adaptation settings for individual baselines are also unclear. The only explicitly specified cross-dataset OOD protocol among the main benchmarks is GenBuster training followed by GenBuster++ testing. This single transfer setting provides limited evidence of generalization under broader domain shifts. A clearly documented evaluation using one checkpoint trained on a single dataset and tested across multiple benchmarks without target-dataset adaptation would provide a more comprehensive assessment.
Questions for authors
  1. How frequently are Lsel\mathcal L_{sel} and Ltas\mathcal L_{tas} non-zero during training? Please report the distributions of μj\mu_j and normalized KL divergence with and without the corresponding regularizers. How does BiTRA perform when the assumed temporal artifact pattern is absent or reversed, including on temporally reversed videos and FL2V?
  2. Are the BiTRA results in Tables 1 and 2 obtained from the same ViF-Bench-trained checkpoint, or from separate models trained on the respective datasets? For each baseline, was the model retrained on the same data as BiTRA, or evaluated using its originally released checkpoint without dataset-specific adaptation?
  3. Does the NTP setting in Table 6 use next-token prediction of textual class labels for supervision and inference? The table distinguishes NTP-based and embedding-based backbones, but the associated prediction heads and supervision schemes are not explicitly specified. Have you compared this setup with attaching a classification head to the final-layer hidden representation used to predict the answer token, immediately before the LM head, using the same non-embedding backbone?
4 comments
AuthorsOpenReview ↗

Response to Reviewer JEbo (Part 1/4)

Response to Reviewer JEbo (Part 1/4)

We sincerely thank you for recognizing our diagnosis of existing detectors’ weaknesses on I2V as “an important contribution” and for the insightful comments. We address your concerns and questions point by point below.


Response to Critical Weakness:

W1: Applicability to FL2V and multi-segment stitched videos.
A1: Beyond the operation specifically tailored for I2V, BiTRA incorporates general designs for AIGV detection that enhance overall generalization. Specifically:

  1. SHAI provides global access to visual forgery evidence. Causal self-attention limits bidirectional interactions among visual tokens. SHAI provides an additional cross-attention pathway to multi-level visual features across the video, allowing access to forgery cues at different temporal positions and feature levels. For visually conditioned videos, L_sel is a soft constraint applied to emphasize later-emerging forgery cues, facilitating more generalizable forgery representations.

  2. TAS promotes balanced multi-frame evidence aggregation. Our attention visualizations in Appendix Fig. 8 show that the decoder can disproportionately attend to the initial frames. TAS reduces this concentration and encourages the prediction token to aggregate evidence from a broader temporal range. This may benefit detection across different temporal artifact patterns, where informative forgery cues can occur beyond the initial frames.

Following your suggestion, we also evaluate BiTRA on FL2V and multi-segment stitched videos against three state-of-the-art video forgery detectors—Skyra [1], VideoVeritas [2], and CoCoDetect (all trained on the same ViF-Bench training set, ViF-CoT, for fair comparison), as shown below:

1). First-and-Last-Frame Conditioned (FL2V) Videos
Following the generation protocol in Skyra [1], we use the first and last frames of each selected real video from Kinetics [3] as conditioning anchors to generate 960 FL2V videos using three popular video generation models: Wan2.1-FLF2V-14B [4], LTX-Video-13B [5], and Wan2.1-VACE-1.3B [6], with 320 videos per model. We train all detectors on the ViF-Bench training set (ViF-CoT) and evaluate them in terms of balanced accuracy (bAcc, %), fake recall (Recall, %), and F1 (%) by pairing with the full real-video test set of ViF-Bench, with the results shown in the table below.

Wan2.1-FLF2V⠀LTX-Video⠀Wan2.1-VACE⠀Overall
MethodbAccRecallF1⠀bAccRecallF1⠀bAccRecallF1⠀bAccRecallF1
VideoVeritas70.468.469.8⠀72.171.972.1⠀71.270.070.8⠀71.270.170.9
CoCoDetect71.175.972.4⠀72.879.474.5⠀71.977.573.4⠀71.977.673.4
Skyra90.787.590.4⠀91.388.891.1⠀90.587.290.2⠀90.887.890.6
BiTRA (Ours)98.198.198.1⠀98.999.798.9⠀97.897.597.8⠀98.398.498.3

2). Multi-Segment Stitched Videos
Following your suggestion, we construct 1,000 stitched videos using the ViF-Bench test split, each formed by concatenating 2–3 temporal segments randomly sampled from videos produced by different generators. We train all detectors on the ViF-Bench training set (ViF-CoT) and evaluate them in terms of bAcc (%), Recall (%), and F1 (%) by pairing with the full real-video test set of ViF-Bench, with the results shown in the table below.

MethodbAccRecallF1
VideoVeritas75.578.676.2
CoCoDetect73.881.375.6
Skyra89.785.689.3
BiTRA (Ours)98.097.998.0

As shown in the tables above, BiTRA performs strong generalization on both FL2V videos (98.3% bAcc / 98.4% Recall / 98.3% F1) and multi-segment stitched videos (98.0% bAcc / 97.9% Recall / 98.0% F1).

AuthorsOpenReview ↗

Response to Reviewer JEbo (Part 2/4)

Response to Reviewer JEbo (Part 2/4)

W2: Training protocols and generalization across broader domain shifts.
A2: We apologize for the lack of clarity in the original manuscript.

  1. Following the official evaluation protocols of ViF-Bench [1], GenVideo [7], and GenBuster/GenBuster++ [8, 9], we train models on each benchmark's official training split and evaluate on its corresponding test split. All compared methods are reproduced using their publicly released implementations or officially released weights.

  2. To further evaluate generalization ability, we conduct an additional experiment on the newly released GenBuster-Bench [8] (NIPS'26), which comprises three progressive evaluation tracks: ID (2024) on seen generators, OOD (2025) on unseen commercial generators, and Wild (2026) on out-of-domain, in-the-wild videos. Strictly following the benchmark's official training protocol, we train BiTRA and all trainable baseline detectors on the provided training split. We evaluate training-free methods using their publicly released implementations, with all methods evaluated under the same official test protocol. Real recall, fake recall and the corresponding balanced accuracy are reported in the table below for all three tracks. The best and second-best results are shown in bold and italics.

ID (2024)⠀OOD (2025)⠀Wild (2026)
MethodsRealFakebACC⠀FakebACC⠀FakebACC
NSG-VD70.835.653.2⠀33.652.2⠀28.649.7
ReStraV93.870.782.3⠀39.366.6⠀18.055.9
STALL72.660.766.7⠀65.068.8⠀29.351.0
D365.258.461.8⠀60.763.0⠀24.044.6
DeMamba93.870.282.0⠀66.980.4⠀48.070.9
CoCoDetect93.881.387.6⠀62.077.9⠀30.061.9
AIGVDetect92.887.490.1⠀66.779.8⠀60.076.4
BusterX [8]83.090.486.7⠀92.587.8⠀72.077.5
BusterX++72.399.185.7⠀98.686.0⠀85.979.1
VideoVeritas88.599.794.1⠀98.493.8⠀85.386.9
Skyra90.976.383.6⠀59.275.1⠀33.962.4
BiTRA (Ours)95.099.897.4⠀98.796.9⠀86.788.4

As shown in the table above, BiTRA consistently achieves the highest balanced accuracy on ID (97.4%), OOD (96.9%), and the highly challenging out-of-domain Wild (88.4%), further confirming the strong OOD generalization and superiority of our proposed method.

AuthorsOpenReview ↗

Response to Reviewer JEbo (Part 3/4)

Response to Reviewer JEbo (Part 3/4)

Response to Questions:

Q1: Activation rates and distributions of L_sel and L_tas & performance on temporally reversed and FL2V videos.
A1: Thank you for raising these questions. We provide further details below.

  1. Regularizer Activation Rates and Distributions of the Constrained Quantities
  • Activation rates of regularizers: As shown in Appendix Fig. 9, the activation rates of L_sel and L_tas decrease from 100.0% and 96.3% in Epoch 1 to 6.0% and 6.8% in Epoch 10, respectively, with epoch-wise mean activation rates of 39.2% and 31.7%, naturally turning off once the distributions of the two constrained objects move into the target margin zones.
  • Distributions with and without regularization: As shown in Appendix Fig. 10, without L_sel, the temporal barycenter μⱼ on I2V samples remains tightly concentrated around the initial midpoint m=0.5. With L_sel, the distribution shifts clearly to the right of the margin, confirming that L_sel effectively guides cross-attention retrieval toward artifact-rich later frames. As shown in Appendix Fig. 11, without L_tas, the normalized KL divergence spreads widely to the right of the margin δ=0.04, reflecting severe temporal attention concentration. Incorporating L_tas compresses almost the entire distribution below the margin, mitigating excessive early-frame attention bias.
  1. Performance on Temporally Reversed Videos and FL2V
    For the temporal reversal evaluation, we use checkpoints of all compared methods trained on the ViF-Bench training set (ViF-CoT), without additional training or fine-tuning. We evaluate these methods on the full ViF-Bench test set with the frame order reversed in each video and report balanced accuracy (bAcc, %), fake recall (Recall, %), and F1 (%) in the table below.
T2V⠀I2V⠀Overall
MethodbAccRecallF1⠀bAccRecallF1⠀bAccRecallF1
VideoVeritas84.182.483.6⠀55.425.134.9⠀73.561.365.7
CoCoDetect84.295.585.7⠀61.450.356.1⠀75.878.874.8
Skyra86.296.087.4⠀70.063.667.9⠀80.284.180.2
BiTRA (Ours)98.697.998.6⠀93.287.191.6⠀96.693.996.0

As shown above, even when the input frame order is reversed, BiTRA achieves 98.6% bAcc on T2V, 93.2% on I2V, and 96.6% overall, outperforming the second-best Skyra by +12.4%, +23.2%, and +16.4%, respectively.
For FL2V Videos, as reported in the answer of W1 above, BiTRA achieves 98.3% Overall bAcc (98.4% Recall / 98.3% F1) across 960 FL2V videos, outperforming Skyra (90.8% bAcc / 87.8% Recall / 90.6% F1), CoCoDetect (71.9% bAcc / 77.6% Recall / 73.4% F1), and VideoVeritas (71.2% bAcc / 70.1% Recall / 70.9% F1).

AuthorsOpenReview ↗

Response to Reviewer JEbo (Part 4/4)

Response to Reviewer JEbo (Part 4/4)

Q2: Training protocols in Tabs. 1–2 and baseline adaptation settings.
A2: Thank you for highlighting this point. We clarify the benchmark-specific protocols and baseline adaptation settings below.

  1. Benchmark-Specific Training and Evaluation Protocols. For the experiments in Tabs. 1–3, we train a separate BiTRA model on the training split of each benchmark—ViF-Bench [1], GenVideo [7], and GenBuster/GenBuster++ [8, 9]—and evaluate it under the corresponding input and evaluation protocols.

  2. Baseline Adaptation Settings. In our original manuscript, following VideoVeritas [2], conventional baselines were trained on the same benchmark-specific training splits as BiTRA, whereas the MLLM baselines (BusterX++, VideoVeritas, Skyra) were evaluated using their released checkpoints without additional fine-tuning. Following your suggestions, we further adapt these MLLM checkpoints on the corresponding training splits, providing a fairer comparison after benchmark-specific adaptation. As shown in Appendix F (Tabs. 19–21, new results highlighted in blue), BiTRA continues to outperform the adapted baselines. We will update the results in the main text of the revised manuscript.

Q3: NTP supervision details & comparison with non-embedding backbones with classification head.
A3:

  1. NTP and Embedding Setup in Tab. 6. In the NTP setting, the model predicts the textual class label ("real" / "fake") through the language modeling (LM) head, with cross-entropy loss applied only to the class-label tokens. At inference, the model generates the class label. In the Embedding setting, a linear binary classification head takes the final-layer hidden state at the last non-padding position of the input sequence. The classifier is trained with binary cross-entropy loss.

  2. Generative Backbones with a Classification Head. Following your suggestion, we additionally evaluate Qwen3-VL-2B and Qwen3-VL-8B with a classification head on ViF-Bench. For each backbone, we bypass the LM head and attach a linear binary classification head, using the same hidden-state readout and classification-head architecture as in the Embedding setting. We denote these variants as “as Emb.” and report balanced accuracy (bAcc) in the table below.

Backbone⠀Setting⠀T2V⠀I2V⠀Overall
Qwen3-VL-2B (as Emb.)⠀Baseline⠀90.4⠀86.6⠀89.0
Qwen3-VL-2B (as Emb.)⠀+ BiTRA (Ours)⠀96.7 (+6.3)⠀93.3 (+6.7)⠀95.5 (+6.5)
Qwen3-VL-8B (as Emb.)⠀Baseline⠀88.4⠀80.9⠀85.6
Qwen3-VL-8B (as Emb.)⠀+ BiTRA (Ours)⠀95.3 (+6.9)⠀90.2 (+9.3)⠀93.4 (+7.8)

BiTRA consistently improves bAcc on both the T2V and I2V subsets, as well as overall, over the corresponding “as Emb.” baselines, with overall +6.5 and +7.8 gains on Qwen3-VL-2B and Qwen3-VL-8B, respectively. These results support BiTRA’s effectiveness on the evaluated generative backbones with the classification-head architecture held fixed, while showing that BiTRA achieves even better performance with embedding backbones.


References
[1] Li et al., "Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning," CVPR, 2026.
[2] Tan et al., "VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning," ICML, 2026.
[3] Kay et al., "The Kinetics Human Action Video Dataset," arXiv:1705.06950, 2017.
[4] Wan et al., "Wan: Open and Advanced Large-Scale Video Generative Models," arXiv:2503.20314, 2025.
[5] HaCohen et al., "LTX-Video: Realtime Video Latent Diffusion," arXiv:2501.00103, 2024.
[6] Jiang et al., "VACE: All-in-One Video Creation and Editing," ICCV, 2025.
[7] Chen et al., "DeMamba: AI-Generated Video Detection on Million-Scale GenVideo Benchmark," SCIS, 2026.
[8] Wen et al., "BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation," NeurIPS, 2026.
[9] Wen et al., "BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM," arXiv:2507.14632, 2025.

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

The authors identify that for AI generated video detection, MLLM’s don’t perform as well on I2V cases due to temporal artifact misalignment (later frames more likely to have more artifacts). They justify architectural and training improvements to allow artifact attention to propagate through layers ("selective hierarchical artifact injection") and to reduce LLM temporal bias to earlier frames ("temporal attention smoother"). The paper has sufficient and convincing experimental validation demonstrating the state-of-the-art performance of the proposed approach across standard datasets.

Critical strengths

The paper identifies an issue in the space of AI generated video detection and clearly illustrates the performance gaps on the different domains and provides an analysis for why this gap exists.

The authors suggest why a naive solution (bidirectional attention over visual tokens) won't work, which motivates the authors selective hierarchical artifact injection (and prior supervision) and temporal smoothing is more appropriate. And the solution seems novel.

Ablation study and experiments are comprehensive and convincing (SOTA performance in many areas) and cover different backbones, main components of the model design (and hyper-parameters), different tasks, as well as sensitivity to various distortions.

Critical weaknesses

I think I've self-resolved this as being critical, but listing it here anyway. At first glance, the results in Table 1 look too good to be true (in the bottom row there are eleven cells with balanced accuracy of “99.1”). The appendix has other metrics on this test set, and it is clear that next-best algorithm also performs pretty good. So I guess this just means that the dataset (and already established protocol, which the authors are following for the dataset in question) is too easy. That said, the ablation study and results in Figure 4 are convincing (e.g., the method also holds up better than others against various unseen distortions).

Questions for authors

A lot of the focus of the intro is on temporal misalignment on I2V tasks. But the proposed model ends up also giving significant improvement on T2V. Is there any intuitive rationale about why the proposed improvements ends up doing so well in T2V too?

Small things: 187: You say what all the other variables represent except B 187: You say “we select S1 = {ℓi} L1 i=1 from all Vision Transformer layers." Maybe clearer to say "we select a subset S1 ..” 201. In (3) should this be a W_3, since you already have W_1 and W_2

4 comments
AuthorsOpenReview ↗

Response to Reviewer opvP (Part 1/2)

Response to Reviewer opvP (Part 1/2)

We sincerely thank you for the positive assessment (4: Clear acceptance & "sufficient and convincing experimental validation") and constructive comments. Below we address your concerns point by point.


Response to Critical Weakness:

W1: Repeated 99.1% in Tab. 1.
A1: Under the official ViF-Bench protocol, each generator's fake videos are paired by ID with source videos from a common real-video pool, and most of the generators share the exact same set of real videos. Given bAcc = (Recall_real + Recall_fake) / 2, whenever BiTRA reaches 100% fake recall on generators evaluated against the same real videos, bAcc depends only on their shared Recall_real. This explains why multiple generators report the identical bAcc of 99.1%.

W2: Additional evaluation on a more challenging benchmark.
A2: Following your suggestions, we conduct additional experiments on the newly released GenBuster-Bench [1] (NIPS'26), which comprises three evaluation tracks: ID (2024) on seen generators, OOD (2025) on unseen commercial generators, and Wild (2026) on out-of-domain, in-the-wild videos. Following the benchmark's official training protocol, we train BiTRA and all trainable baseline detectors on the provided training split. We evaluate training-free methods using their publicly released implementations, with all methods evaluated under the same official test protocol. Real recall, fake recall and the corresponding balanced accuracy are reported in the table below for all three tracks. The best and second-best results are shown in bold and italics, respectively.

ID (2024)⠀OOD (2025)⠀Wild (2026)
MethodsRealFakebACC⠀FakebACC⠀FakebACC
NSG-VD70.835.653.2⠀33.652.2⠀28.649.7
ReStraV93.870.782.3⠀39.366.6⠀18.055.9
STALL72.660.766.7⠀65.068.8⠀29.351.0
D365.258.461.8⠀60.763.0⠀24.044.6
DeMamba93.870.282.0⠀66.980.4⠀48.070.9
CoCoDetect93.881.387.6⠀62.077.9⠀30.061.9
AIGVDetect92.887.490.1⠀66.779.8⠀60.076.4
BusterX [8]83.090.486.7⠀92.587.8⠀72.077.5
BusterX++72.399.185.7⠀98.686.0⠀85.979.1
VideoVeritas88.599.794.1⠀98.493.8⠀85.386.9
Skyra90.976.383.6⠀59.275.1⠀33.962.4
BiTRA (Ours)95.099.897.4⠀98.796.9⠀86.788.4

As shown in the table above, BiTRA consistently achieves the highest balanced accuracy on ID (97.4%), OOD (96.9%), and the highly challenging out-of-domain Wild (88.4%), further confirming the strong OOD generalization and superiority of our proposed method.


AuthorsOpenReview ↗

Response to Reviewer opvP (Part 2/2)

Response to Reviewer opvP (Part 2/2)

Response to Questions:

Q1: Why does BiTRA also achieve substantial gains on T2V?
A1: Good question. Although motivated by Temporal Artifact–Attention misalignment in I2V videos, BiTRA improves global artifact access and temporal evidence aggregation, both of which can also benefit T2V detection.

  1. SHAI enables global artifact access beyond causal masking. In a standard causally masked MLLM decoder, each visual token can attend only to itself and preceding tokens, restricting access to information from subsequent frames. SHAI introduces an additional cross-attention pathway through GAI over the multi-level visual features fused by HFA. This allows each visual token in the LLM to access forgery information from all frames, which could enhance the detection performance on the T2V samples (+4.4 bAcc on T2V in Tab. 5). Moreover, although applied to I2V samples to emphasize later-emerging forgery cues, L_sel also gives +0.5 T2V bAcc in Tab. 5, suggesting that its benefits extend beyond I2V detection for better forgery learning.

  2. TAS mitigates excessive early-frame attention. The T2V example in Appendix Fig. 8(c) shows that the decoder can disproportionately attend to the initial frames and that TAS reduces this concentration, encouraging the prediction token to aggregate evidence from a broader temporal range. This can benefit T2V detection (+1.8 T2V bAcc in Tab. 4), where informative forgery cues may occur throughout the video.


Response to Minor Comments:

M1: B and S₁ in Line 187 & W in Line 201.
A1: Thank you for pointing these out. We have fixed all three issues in Section 3.2 of the revised manuscript (highlighted in blue): clarified that B denotes the batch size, revised the phrase in Line 187 to "we select a subset S₁ = {ℓᵢ} (i = 1, ..., L₁) from all Vision Transformer (ViT) layers", and renamed the projection matrix W in Eq. (3) to W₃ to avoid notation overlap with W₁ and W₂ in Eq. (2).

References
[1] Wen et al., "BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation," NeurIPS, 2026.

Thanks to the authors for responding to my questions, running additional evaluations, and addressing the minor issues.

We sincerely thank you for your kind follow-up. We are glad that your questions have been addressed. Thank you again for your positive assessment and for the constructive comments that helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.