Pith. sign in

REVIEW 3 major objections 6 minor 77 references

A multi-agent loop that pairs forensic tools with semantic checks can detect social-risk AI videos that pure artifact models and pure language models both miss.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 05:06 UTC pith:S554IOAE

load-bearing objection Solid multi-agent detector plus a genuinely useful social-risk benchmark; the +18.7% claim is real but partly rides on GenVideo-tuned tools whose OOD transfer is untested. the 3 major comments →

arxiv 2607.03069 v1 pith:S554IOAE submitted 2026-07-03 cs.CV

SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection

Wenlin Wu , Sheng Zhou , Peipei Song , Wenhao Wang , Junbin Xiao , Xun Yang This is my paper
classification cs.CV
keywords AI-generated video detectionsocial-risk forensicsmulti-agent systemsperception-reasoning gapSafeVid benchmarkhierarchical perceptual solverself-reflective verifier
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Modern video generators no longer leave obvious local artifacts; they produce full scenes that look real and can depict violence, riots, misinformation, or other high-stakes social events. Current detectors either hunt low-level statistical glitches and fail when those glitches vanish, or reason about meaning and ignore fine forensic traces. The paper argues that the two capabilities must work together in a closed loop: a hierarchical solver first forms a risk-aware hypothesis, crops the critical region, and runs specialized tools for motion, depth, appearance, and pixels; a self-reflective verifier then checks whether the physical evidence supports the semantic claim and can force another round if the story is inconsistent. To test the idea the authors release SafeVid, twenty thousand videos spanning ten social-risk categories generated by twenty-one models. On that benchmark the combined system raises accuracy by roughly nineteen points over the strongest prior detector and also leads on four existing public suites, while producing an explicit chain of evidence rather than a bare label.

Core claim

The Perception–Reasoning Gap in AI-generated video detection can be closed by a multi-agent architecture that extracts localized forensic cues with a hierarchical perceptual solver and then enforces semantic–physical consistency with a self-reflective verifier, yielding large gains on social-risk content and solid cross-benchmark generalization without end-to-end task fine-tuning.

What carries the argument

SafeGuard: a Hierarchical Perceptual Solver that generates social-risk hypotheses, routes and crops candidate regions, and extracts four-tool forensic scores (optical flow, depth, appearance, pixel), closed by a Self-Reflective Verifier that scores evidence–hypothesis consistency and triggers iterative refinement when reliability falls below threshold.

Load-bearing premise

The four lightweight forensic tools, whose weights were tuned on an older general-purpose training set, still return reliable authenticity signals for videos made by twenty-one newer generators under socially charged content distributions.

What would settle it

Replace the GenVideo-tuned forensic tools with the same architectures left completely untrained (or trained only on real video) and re-run the full SafeVid OOD split; if average accuracy collapses toward the no-tool baseline, the claim that collaborative perception-reasoning is what drives the gain is undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper identifies a Perception–Reasoning Gap in AI-generated video detection: artifact-centric detectors lack semantic inference while VLMs miss fine-grained forensic cues, and existing benchmarks under-represent social-risk content. It proposes SafeGuard, a multi-agent system with a Hierarchical Perceptual Solver (hypothesis generation/routing, language-guided cropping, and a four-tool forensic toolbox producing E_phy_t) and a Self-Reflective Verifier that scores evidence–hypothesis consistency and closes the loop via hypothesis update (Eqs. 1–6). It also introduces SafeVid (≈20K videos, 10 social-risk categories, 21 generators, T2V/TI2V/V2V). Empirically, SafeGuard reports +18.7% Acc on SafeVid over the strongest prior baseline and gains on GenVideo, DVF, LOKI, and GenVidBench, with progressive ablations (Tables 4–5) and a qualitative case study.

Significance. If the results hold under clearer attribution of tool transfer versus multi-agent design, the work is a solid systems contribution: SafeVid is the first social-safety-oriented AIGC-video detection benchmark with multi-paradigm synthesis and a fine-grained taxonomy, and SafeGuard is a training-free (for the overall pipeline) closed-loop perception–reasoning architecture that produces an interpretable evidence chain. Strengths include explicit ID/OOD splits (Table 2), dual proprietary/open-source backbones, progressive module and toolbox ablations, public code/dataset, and performance approaching the reported human baseline on SafeVid. The social-risk framing and multi-agent forensic design are timely for high-stakes video authenticity assessment.

major comments (3)
  1. [Sec. 5.1, Eq. (3), Table 5] Sec. 5.1 and Eq. (3): the four forensic tools that produce E_phy_t are optimized on GenVideo train, yet SafeVid OOD uses 11 generators never seen by those tools (Table 2: LTX, Veo3, Wan, etc.). Table 5 shows that adding the tools lifts overall Acc from 81.8% (no tools) to 89.3%, so a large fraction of the headline gains rides on these scores. The paper never reports tool-only Acc/F1, score calibration, or reliability of fo/fd/fa/fp on the SafeVid OOD split (or on LOKI/GenVidBench). Without that, it is hard to separate residual GenVideo-style artifact sensitivity from the claimed closed-loop multi-agent specialization. Please add tool-only and tool-calibration results on SafeVid OOD (and ideally the other public OOD sets) and discuss how often tool scores are discarded by the verifier.
  2. [Abstract; Sec. 5.2–5.3; Table 4] Table 4: the no-FTE variant (SHGR+VC+SRV) already reaches 74.5 Acc on SafeVid and 81.8 overall, while full SafeGuard reaches 85.5 / 89.3. The abstract and introduction attribute the +18.7% SafeVid gain primarily to collaborative perception–reasoning. Please restate the main claims so that (i) multi-agent reasoning without tools and (ii) the incremental contribution of GenVideo-finetuned tools are reported as separate, first-class results, and so that the +18.7% figure is not read as pure multi-agent gain over artifact-free baselines.
  3. [Sec. 3.2, Eq. (6)] Sec. 3.2 and Eq. (6): the Self-Reflective Verifier is load-bearing for reliability claims (γ=0.5, L=3), but the paper does not report how often S_ver_t < γ, how often refinement changes the final label, or accuracy conditioned on number of refinement cycles. A short analysis (fraction of videos that trigger refinement; Acc before vs after refinement on SafeVid ID/OOD) is needed to show that the closed loop is doing causal work rather than rarely firing.
minor comments (6)
  1. [Table 3; Sec. 5.1] Table 3 caption: evaluation of BusterX++, Skyra-RL, and generalist models on 30% stratified subsets (and humans on 10%) should also be flagged in the main text of Sec. 5.1 so readers do not treat all numbers as full-set.
  2. [Sec. 5.1] Implementation Details: free parameters γ, L, CLIP threshold 0.75, and 8-frame sampling are stated but not sensitivity-tested. A brief appendix sweep (e.g., γ ∈ {0.3,0.5,0.7}, L ∈ {1,3,5}) would strengthen robustness claims.
  3. [Fig. 1; Fig. 2] Fig. 1 and Fig. 2 are dense; the evidence-chain example in Fig. 1(c) is helpful but hard to parse at print size. Consider enlarging the verdict/feedback boxes or moving a cleaned version to the appendix.
  4. [Throughout] Manuscript text has many missing spaces and line-break artifacts (e.g., “highperceptualrealism”, “thatcloselyresemble”, “Percep-tion–Reasoning”). A full copy-edit pass is needed before camera-ready.
  5. [Sec. 2] Related work: LAVID and other agentic AIGC detectors are cited; a short explicit comparison table (training-free vs trainable, perception-only vs perception+reasoning, social-risk coverage) would clarify positioning.
  6. [Sec. 4.1] SafeVid taxonomy (Sec. 4.1): state whether category labels were human-verified after CLIP filtering or only automatically assigned; inter-annotator agreement if any would help.

Circularity Check

0 steps flagged

No circular derivation: empirical multi-agent system with transparent GenVideo-tool pretraining; headline gains rest on independent SafeVid/OOD and external benchmarks.

full rationale

SafeGuard is an empirical systems paper, not a first-principles derivation. Its equations (1–6) define a closed-loop agent pipeline (hypothesis generation, region routing, forensic tool scores E_phy_t, LVLM reasoning, consistency check, optional refinement); none of these equations define the reported accuracy in terms of a fitted target quantity. The only domain-matched component is the forensic toolbox T = {f_o, f_d, f_a, f_p}, whose weights are optimized on GenVideo train (Sec. 5.1). The authors explicitly label the 97.9% GenVideo result an “in-domain reference” rather than a surprising prediction, and the load-bearing claims (+18.7% Acc on SafeVid, gains on DVF/LOKI/GenVidBench) are measured on generators and distributions outside that training split (Table 2 OOD generators; public test sets). Ablations (Tables 4–5) isolate module contributions without renaming a fit as a prediction. No uniqueness theorem, self-citation chain, or ansatz is load-bearing for the central claim. Domain-transfer risk of the GenVideo-tuned tools is a generalization/correctness concern, not circularity of derivation. Score 0; steps empty.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 3 invented entities

The central performance claims rest on a small set of hand-chosen thresholds, the assumption that GenVideo-finetuned forensic tools remain informative out-of-domain, and the reliability of two commercial LVLMs for hypothesis generation and consistency checking. No new physical entities are postulated; the invented components are engineering modules and a dataset.

free parameters (4)
  • reliability threshold γ = 0.5
    Set to 0.5 by hand; controls whether the verifier accepts the initial verdict or triggers refinement (Eq. 6).
  • max refinement cycles L = 3
    Hard upper bound on closed-loop iterations, chosen as 3.
  • CLIP similarity filter threshold = 0.75
    0.75 used to retain real videos matching the social-risk taxonomy during dataset construction.
  • number of sampled frames = 8
    Uniformly sample 8 frames per video for all LVLM and tool inputs.
axioms (3)
  • domain assumption LVLMs (GPT-4o / Gemini-2.5-Pro / Qwen3-VL) can produce useful social-risk hypotheses, grounding queries, and consistency feedback from video frames plus tool scores.
    Invoked throughout Sec. 3.1-3.2; no independent verification that the models are free of systematic bias on forensic tasks.
  • domain assumption The four forensic tools (optical flow, depth, appearance, pixel) return calibrated real/fake probabilities that remain informative for generators and content distributions outside GenVideo.
    Stated in Implementation Details and validated only by the ablation in Table 5; tools themselves were optimized on GenVideo train.
  • ad hoc to paper The 10-category social-risk taxonomy plus CLIP filtering yields a representative and unbiased sample of high-stakes scenarios.
    Defined in Sec. 4.1; no external validation that the taxonomy exhausts real-world social risks or that the 0.75 threshold avoids selection bias.
invented entities (3)
  • Hierarchical Perceptual Solver no independent evidence
    purpose: Coarse-to-fine hypothesis generation, region routing, and multi-tool forensic cue extraction.
    Engineering module introduced in Sec. 3.1; no independent existence outside the framework.
  • Self-Reflective Verifier no independent evidence
    purpose: Evidence-hypothesis consistency checking and closed-loop hypothesis refinement.
    Engineering module introduced in Sec. 3.2; reliability score and feedback are internal constructs.
  • SafeVid benchmark independent evidence
    purpose: Provide 20K videos across 10 social-risk categories for evaluating physical, structural, and social-logic violations.
    New dataset constructed by the authors; value is empirical utility rather than a physical discovery.

pith-pipeline@v1.1.0-grok45 · 22745 in / 3098 out tokens · 33817 ms · 2026-07-12T05:06:16.491411+00:00 · methodology

0 comments
read the original abstract

As video generation paradigms evolve from localized manipulation to full-scene synthesis, AI-generated video detection becomes increasingly challenging, as forgeries exhibit coherent global structure and high perceptual realism. However, existing benchmarks are biased toward perceptual fidelity and primarily evaluate detectors based on perceptual artifacts, providing limited coverage of scenarios that require reasoning about violations of physical laws, structural coherence, or social logic. This dataset bias shapes current approaches and results in a Perception-Reasoning Gap: artifact-centric models capture low-level statistical irregularities yet lack semantic inference, whereas vision-language models perform semantic reasoning but remain insensitive to fine-grained forensic cues. To bridge this gap, we propose SafeGuard, a multi-agent framework that enables collaborative specialization between forensic perception and semantic reasoning. A hierarchical perceptual solver extracts fine-grained forensic evidence, while a self-reflective verifier enforces consistency between semantic inference and physical plausibility, forming an interpretable evidence chain. To support evaluation, we introduce SafeVid, a novel AI-generated video detection benchmark comprising 20K videos spanning 10 social risk categories, designed to evaluate physical plausibility, structural consistency, and the rationality of social behaviors. Extensive experiments demonstrate the generalization of SafeGuard, improving accuracy on SafeVid by +18.7% and consistently outperforming prior methods across four public benchmarks.

Figures

Figures reproduced from arXiv: 2607.03069 by Junbin Xiao, Peipei Song, Sheng Zhou, Wenhao Wang, Wenlin Wu, Xun Yang.

Figure 1
Figure 1. Figure 1: The illustration of social-risk AI-generated video detection. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The SafeGuard Framework. SafeGuard combines a Hierarchical Percep￾tual Solver for coarse-to-fine trace extraction with a Self-Reflective Verifier that checks evidence–hypothesis consistency and performs closed-loop refinement. 2.2 Multi-Agent System for Visual Forensics Multi-Agent Systems (MAS) address complex tasks by decomposing problem￾solving into coordinated interactions among specialized agents, ena… view at source ↗
Figure 3
Figure 3. Figure 3: SafeVid statistics. (a) Video count distribution. (b) Social risk category distribution. (c) Top-10 high-risk terms in video generation prompts (including T2V and TI2V). (d) Video duration distribution. (e) Video resolution distribution. to filter out irrelevant videos. CLIP [41] computes semantic similarity between candidate videos and taxonomy keywords, and videos are retained if the simi￾larity exceeds … view at source ↗
Figure 4
Figure 4. Figure 4: A case study of SafeGuard and other existing models on SafeVid. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

77 extracted references · 20 linked inside Pith

  1. [1]

    In: International Conference on Machine Learning (2026)

    Asad, A., Obadinma, S., Shayanfar, R., Zhu, X.: Reddebate: Safer responses through multi-agent red teaming debates. In: International Conference on Machine Learning (2026)

    2026

  2. [2]

    In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV)

    Bai, J., et al.: Ai-generated video detection via spatial-temporal anomaly learning. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 460–470 (2024)

    2024

  3. [3]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:2404.18930 (2024)

    Bai, Z., et al.: Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930 (2024)

    Pith/arXiv arXiv 2024

  5. [5]

    arXiv preprint arXiv:2311.15127 (2023)

    Blattmann, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    Science China Information Sciences p

    Chen, H., et al.: Demamba: Ai-generated video detection on million-scale genvideo benchmark. Science China Information Sciences p. 162103 (2026)

    2026

  7. [7]

    arXiv preprint arXiv:2506.10975 (2025)

    Chen, W., et al.: Genworld: Towards detecting ai-generated real-world simulation videos. arXiv preprint arXiv:2506.10975 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    In: International Conference on Learning Representations (2026)

    Cheng, Z., et al.: Moca: Modeling object consistency for 3d camera control in video generation. In: International Conference on Learning Representations (2026)

    2026

  9. [9]

    arXiv preprint arXiv:2511.18102 (2025)

    Chinchure, A., et al.: Spotlight: Identifying and localizing video generation errors using vlms. arXiv preprint arXiv:2511.18102 (2025)

    arXiv 2025

  10. [10]

    IEEE Transactions on Information Forensics and Security pp

    Coccomini, D.A., et al.: Mintime: Multi-identity size-invariant video deepfake de- tection. IEEE Transactions on Information Forensics and Security pp. 6084–6096 (2024)

    2024

  11. [11]

    arXiv preprint arXiv:2507.06261 (2025)

    Comanici, G., et al.: Gemini 2.5: Pushing the frontier with advanced reason- ing, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Advances in Neural Information Processing Systems pp

    Corvi, R., et al.: Seeing what matters: Generalizable ai-generated video detection with forensic-oriented augmentation. Advances in Neural Information Processing Systems pp. 26418–26446 (2025)

    2025

  13. [13]

    Advances inNeural Information Processing Systems pp

    Feng,K., et al.:Video-r1: Reinforcingvideo reasoningin mllms. Advances inNeural Information Processing Systems pp. 99114–99137 (2025)

    2025

  14. [14]

    arXiv preprint arXiv:2506.14827 (2025)

    Gao, Y., et al.: David-xr1: Detecting ai-generated videos with explainable reason- ing. arXiv preprint arXiv:2506.14827 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    Model card, Google DeepMind (2025), https : / / storage

    Google DeepMind: Veo 3 Model Card. Model card, Google DeepMind (2025), https : / / storage . googleapis . com / deepmind - media / Model - Cards / Veo - 3 - Model-Card.pdf

    2025

  16. [16]

    In: International Conference on Learning Representations

    Hong, S., et al.: Metagpt: Meta programming for a multi-agent collaborative frame- work. In: International Conference on Learning Representations. pp. 23247–23275 (2024)

    2024

  17. [17]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Huang, B., et al.: Implicit identity driven deepfake face swapping detection. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 4490–4499 (2023)

    2023

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, Q., et al.: Unishield: An adaptive multi-agent framework for unified forgery image detection and localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8121–8132 (2026)

    2026

  19. [19]

    arXiv preprint arXiv:2410.21276 (2024)

    Hurst, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Ad- vances in Neural Information Processing Systems pp

    Internò, C., et al.: Ai-generated video detection via perceptual straightening. Ad- vances in Neural Information Processing Systems pp. 20672–20705 (2025) SafeGuard 17

    2025

  21. [21]

    arXiv preprint arXiv:2405.15343 (2024)

    Ji, L., et al.: Distinguish any fake videos: Unleashing the power of large-scale data and motion features. arXiv preprint arXiv:2405.15343 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Jiang, Z., et al.: Vace: All-in-one video creation and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17191–17202 (2025)

    2025

  23. [23]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lai, Y., et al.: Agent4faceforgery: Multi-agent llm framework for realistic face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14073–14083 (2026)

    2026

  25. [25]

    arXiv preprint arXiv:2504.06958 (2025)

    Li, X., et al.: Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning. arXiv preprint arXiv:2504.06958 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Li, Y., et al.: Unim: A unified any-to-any interleaved multimodal benchmark. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 15902–15911 (2026)

    2026

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, Y., et al.: Skyra: Ai-generated video detection via grounded artifact reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4482–4493 (2026)

    2026

  28. [28]

    arXiv preprint arXiv:2511.00181 (2025)

    Liang, M., et al.: From evidence to verdict: An agent-based forensic framework for ai-generated image detection. arXiv preprint arXiv:2511.00181 (2025)

    Pith/arXiv arXiv 2025

  29. [29]

    arXiv preprint arXiv:2402.00045 (2024)

    Lin, L., et al.: Detecting multimedia generated by large ai models: A survey. arXiv preprint arXiv:2402.00045 (2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Advances in Neural Information Processing Systems pp

    Ling, H., et al.: Everybodydance: Bipartite graph–based identity correspondence for multi-character animation. Advances in Neural Information Processing Systems pp. 14123–14151 (2025)

    2025

  31. [31]

    arXiv preprint arXiv:2502.14994 (2025)

    Liu, Q., et al.: Lavid: An agentic lvlm framework for diffusion-generated video detection. arXiv preprint arXiv:2502.14994 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    In: European Conference on Computer Vision

    Liu, S., et al.: Grounding dino: Marrying dino with grounded pre-training for open- set object detection. In: European Conference on Computer Vision. pp. 38–55 (2024)

    2024

  33. [33]

    Advances in Neural Information Processing Systems (2025)

    Liu,X.,etal.:Video-safetybench:Abenchmarkforsafetyevaluationofvideolvlms. Advances in Neural Information Processing Systems (2025)

    2025

  34. [34]

    In: 2025 IEEE International Conference on Multimedia and Expo (ICME)

    Ma, L., et al.: Detecting ai-generated video via frame consistency. In: 2025 IEEE International Conference on Multimedia and Expo (ICME). pp. 1–6 (2025)

    2025

  35. [35]

    Mellis, I.: Pika.https://pika.art/(2025), 2025-02-06

    2025

  36. [36]

    In: International Conference on Learning Representations

    Nan, K., et al.: Openvid-1m: A large-scale high-quality dataset for text-to-video generation. In: International Conference on Learning Representations. pp. 1045– 1064 (2025)

    2025

  37. [37]

    In: European Conference on Computer Vision

    Ni, B., et al.: Expanding language-image pretrained models for general video recog- nition. In: European Conference on Computer Vision. pp. 1–18 (2022)

    2022

  38. [38]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Ni, Z., et al.:Genvidbench: A6-millionbenchmark forai-generated videodetection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 15582–15590 (2026)

    2026

  39. [39]

    OpenAI: Sora system card.https://openai.com/index/sora- system- card/ (2024), accessed: 2025-08-14

    2024

  40. [40]

    Transactions on Machine Learning Research (2024)

    Oquab, M., et al.: Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024)

    2024

  41. [41]

    In: International Conference on Machine Learning

    Radford, A., et al.: Learning transferable visual models from natural language su- pervision. In: International Conference on Machine Learning. pp. 8748–8763 (2021)

    2021

  42. [42]

    In: International Conference on Learning Representations

    Ravi, N., et al.: Sam 2: Segment anything in images and videos. In: International Conference on Learning Representations. pp. 28085–28128 (2025) 18 W. Wu et al

    2025

  43. [43]

    Advances in Neural Information Processing Systems pp

    Song, X., et al.: On learning multi-modal forgery representation for diffusion gen- erated video detection. Advances in Neural Information Processing Systems pp. 122054–122077 (2024)

    2024

  44. [44]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Tan, C., et al.: Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 28130–28139 (2024)

    2024

  45. [45]

    In: European Conference on Computer Vision

    Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: European Conference on Computer Vision. pp. 402–419 (2020)

    2020

  46. [46]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Advances in Neural Information Processing Systems pp

    Wang, Q., et al.: Videorft: Incentivizing video reasoning capability in mllms via reinforced fine-tuning. Advances in Neural Information Processing Systems pp. 4350–4376 (2025)

    2025

  48. [48]

    5: Advancing open-source multimodal models in ver- satility, reasoning, and efficiency

    Wang, W., et al.: Internvl3. 5: Advancing open-source multimodal models in ver- satility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    Pith/arXiv arXiv 2025

  49. [49]

    Advances in Neural Information Processing Systems pp

    Wang, W., Yang, Y.: Vidprom: A million-scale real prompt-gallery dataset for text- to-video diffusion models. Advances in Neural Information Processing Systems pp. 65618–65642 (2024)

    2024

  50. [50]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, W., Yang, Y.: Tip-i2v: A million-scale real text and image prompt dataset for image-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14898–14908 (2025)

    2025

  51. [51]

    Advances in Neural Information Processing Systems (2025)

    Wang, W., Yang, Y.: Videoufo: A million-scale user-focused dataset for text-to- video generation. Advances in Neural Information Processing Systems (2025)

    2025

  52. [52]

    arXiv preprint arXiv:2510.16320 (2025)

    Wang, W., et al.: Scaling laws for deepfake detection. arXiv preprint arXiv:2510.16320 (2025)

    arXiv 2025

  53. [53]

    Advances in Neural Information Processing Systems pp

    Wang, X., et al.: Affordbot: 3d fine-grained embodied reasoning via multimodal large language models. Advances in Neural Information Processing Systems pp. 140367–140387 (2025)

    2025

  54. [54]

    arXiv preprint arXiv:2505.12620 (2025)

    Wen, H., et al.: Busterx: Mllm-powered ai-generated video forgery detection and explanation. arXiv preprint arXiv:2505.12620 (2025)

    Pith/arXiv arXiv 2025

  55. [55]

    arXiv preprint arXiv:2507.14632 (2025)

    Wen, H., et al.: Busterx++: Towards unified cross-modal ai-generated content detection and explanation with mllm. arXiv preprint arXiv:2507.14632 (2025)

    Pith/arXiv arXiv 2025

  56. [56]

    arXiv preprint arXiv:2508.02324 (2025)

    Wu, C., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    Pith/arXiv arXiv 2025

  57. [57]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xiao, J., et al.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9777–9786 (2021)

    2021

  58. [58]

    ResearchGate preprint (2026)

    Xu, Y., et al.: Self-evolving agents as dynamic graph transformation: A survey and new perspective. ResearchGate preprint (2026)

    2026

  59. [59]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Xu, Y., et al.: Tall: Thumbnail layout for deepfake video detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22658–22668 (2023)

    2023

  60. [60]

    In: Advances in Neural Information Processing Systems (2025)

    Xue, Z., et al.: Ultravideo: High-quality UHD video dataset with comprehensive captions. In: Advances in Neural Information Processing Systems (2025)

    2025

  61. [61]

    Advances in Neural Information Processing Systems pp

    Yang, L., et al.: Depth anything v2. Advances in Neural Information Processing Systems pp. 21875–21911 (2024)

    2024

  62. [62]

    In: Proceedings of the ACM Web Conference 2026

    Yang, S., et al.: Towards robust detection of chinese toxic variants via dynamic knowledge graph–llm reasoning. In: Proceedings of the ACM Web Conference 2026. pp. 3183–3194 (2026) SafeGuard 19

    2026

  63. [63]

    In: Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval

    Yang, X., et al.: Deconfounded video moment retrieval with causal intervention. In: Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. pp. 1–10 (2021)

    2021

  64. [64]

    IEEE Transactions on Image Processing pp

    Yang, X., et al.: Video moment retrieval with cross-modal neural architecture search. IEEE Transactions on Image Processing pp. 1204–1216 (2022)

    2022

  65. [65]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    Yang, X., et al.: Fine: Fine-grained neuron-level model editing for reliable and safe llms. IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    2026

  66. [66]

    Yang, Z., et al.: Cogvideox: Text-to-video diffusion models with an expert trans- former.In:InternationalConferenceonLearningRepresentations.pp.83048–83077 (2025)

    2025

  67. [67]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yao, X., et al.: A latent transformer for disentangled face editing in images and videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13789–13798 (2021)

    2021

  68. [68]

    International Conference on Learning Representations (2025)

    Ye, J., et al.: Loki: A comprehensive synthetic data detection benchmark using large multimodal models. International Conference on Learning Representations (2025)

    2025

  69. [69]

    YouTube LLC: Youtube.https://www.youtube.com/(2025), accessed: 2025-08-14

    2025

  70. [70]

    arXiv preprint arXiv:2501.13106 (2025)

    Zhang, B., et al.: Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106 (2025)

    Pith/arXiv arXiv 2025

  71. [71]

    Advances in Neural Information Processing Systems pp

    Zhang, S., et al.: Physics-driven spatiotemporal modeling for ai-generated video detection. Advances in Neural Information Processing Systems pp. 174683–174730 (2025)

    2025

  72. [72]

    Zhang, Y., et al.: Llava-next: A strong zero-shot video understanding model (April 2024),https://llava-vl.github.io/blog/2024-04-30-llava-next-video/

    2024

  73. [73]

    arXiv preprint arXiv:2510.14276 (2025)

    Zhao, H., et al.: Qwen3guard technical report. arXiv preprint arXiv:2510.14276 (2025)

    Pith/arXiv arXiv 2025

  74. [74]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zheng, C., et al.: D3: Training-free ai-generated video detection using second-order features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12852–12862 (2025)

    2025

  75. [75]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Zheng, Y., et al.: Exploring temporal coherence for more general video face forgery detection. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 15044–15054 (2021)

    2021

  76. [76]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhou, S., et al.: Egotextvqa: Towards egocentric scene-text aware video question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3363–3373 (2025)

    2025

  77. [77]

    IEEE Transactions on Multimedia pp

    Zhou, S., et al.: Scene-text grounding for text-based video question answering. IEEE Transactions on Multimedia pp. 1417–1430 (2025)

    2025