Pith. sign in

REVIEW 2 major objections 4 minor 43 references

This paper claims that AI-generated video can be detected as it streams by reading the motion field the codec already wrote, and proves that one end-calibrated threshold—not per-prefix recalibration—controls false alarms at the data-depende

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 · deepseek-v4-flash

2026-08-01 12:40 UTC pith:2SQZAZJN

load-bearing objection The streaming reframing and the anytime-validity trick are real and correct; the headline paired gain in E3 is arithmetically inconsistent with its own flip counts and must be resolved from the released code before acceptance. the 2 major comments →

arxiv 2607.19476 v1 pith:2SQZAZJN submitted 2026-07-21 cs.CV

Detect Early, Escalate Rarely: Anytime Detection of AI-Generated Video from the Compressed Bitstream

Mert Onur Cakiroglu , Mehmet Dalkilic , Hasan Kurban This is my paper
classification cs.CV
keywords AI-generated video detectioncompressed-domain analysismotion vectorsstreaming perceptionanytime-valid inferencecascade deferralrunning maximumfalse-positive control
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.

The paper recasts AI-generated video detection as a streaming decision rather than a one-shot offline scoring problem. It argues that the motion field a video codec writes into the compressed bitstream is enough to drive an always-on cheap detector, and that because the detector's running score is a monotone maximum, a single threshold calibrated once on real clips controls the false-positive rate at whatever time the detector decides to stop—while re-calibrating the threshold separately at each prefix breaks that control. It then prices a cascade: clips whose final score lands in an uncertainty band are deferred to a more expensive pixel or vision-language model, with expected compute and accuracy given in closed form. Measured on a matched public benchmark, the cheap stage reaches an AUC of 0.64 at five orders of magnitude less compute than a pixel CNN, and deferring 15% of clips lifts accuracy from 0.75 to 0.78 at seven times less compute. If correct, detection cost becomes a codec parse instead of a large-model forward pass.

Core claim

The paper's central claim is Proposition 3: with M_t the running maximum of per-GOP scores, if a single threshold tau is calibrated so that P0(M_N >= tau) <= alpha under the real-clip distribution, then the alarm time sigma = inf{t <= N : M_t >= tau} satisfies P0(sigma <= N) = P0(M_N >= tau) <= alpha, and P0(M_t >= tau) <= alpha for every prefix t. Monotonicity collapses the union of crossing events into the single final event. The paper also shows the converse construction fails: choosing thresholds tau_t with P0(M_t >= tau_t) = alpha at each prefix gives a union probability at least alpha, with strict excess whenever an early crossing missed by the final test has positive probability. A co

What carries the argument

The running maximum M_t = max_{i <= t} s_i of per-chunk scores, whose monotonicity collapses the union of crossing events {exists t: M_t >= tau} into a single final event {M_N >= tau}; plus the deferral band W = [tau - w, tau) and the closed-form cascade accounting of Proposition 4, which expresses expected compute and expected error as linear functions of the deferral probability and the deferral gain.

Load-bearing premise

The practical guarantee holds only while deployed real clips have the same running-maximum distribution as the calibration sample; under the paper's own cross-dataset shift the stopping-time false-positive rate drifts from about 0.04 to 0.068 at a target of 0.05.

What would settle it

Take a held-out collection of real videos whose codec and motion statistics differ from the calibration set, run the gate with the end-calibrated threshold, and count the fraction of real clips that trigger 'generated' at their stopping time; a rate clearly above the declared alpha—the paper already reports 0.068 at alpha 0.05—would falsify the practical claim, even though Proposition 3 itself remains true under the calibrated null.

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

If this is right

  • An always-on detector can run on CPU inside the streaming path, adding roughly 10^5 MACs per chunk instead of a pixel-domain forward pass.
  • A single end-calibrated threshold gives a legitimate anytime-valid guarantee; re-calibrating at each prefix is not just redundant but inflates the stopping-time false-positive rate (measured 1.7–2.8x under distribution shift).
  • A compute budget maps directly to a deferral width, and the measured compute-accuracy frontier is monotone only where stage 2 beats stage 1 on the deferred band; beyond that knee the frontier bends back.
  • The result is not tied to the specific motion features: any monotone stage-1 score inherits both guarantees, so the same construction can carry a stronger cheap detector.
  • When real and generated clips are matched in length, the anytime curve is flat—the first GOP is as informative as the full clip—so streaming latency buys little there.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same monotone-score-plus-end-threshold construction should transfer to other streaming binary detection tasks (deepfake audio, fraud, content moderation), since the proof uses only monotonicity, not video-specific structure.
  • The five-orders-of-magnitude compute claim counts only the parse; a purpose-built bitstream parser is needed to turn that into wall-clock savings, because the reference implementation pays a frame decode to export motion vectors.
  • Calibration transfer, not the theorem, is the operational risk: re-calibrating on diverse real video samples, or adding a second monotone feature, could keep the stopping-time false-positive rate near target under distribution shift.
  • The paper's own ablation suggests a monotone transform of a cumulative average could recover some of the accuracy gap while preserving the anytime guarantee—an extension worth testing, since the running max is the weakest offline aggregate measured.

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

2 major / 4 minor

Summary. The paper recasts AI-generated video detection as a streaming perception problem. It scores a codec motion-vector field already present in the compressed bitstream, using a CPU parse of ~10^5 MACs per GOP, and aggregates per-GOP scores with a running maximum. A single end-calibrated threshold decides “generated” when the running maximum crosses it; clips ending below a deferral band are escalated to a pixel or VLM stage. The main formal contributions are Proposition 2 (monotonicity of the running maximum), Proposition 3 (one end-calibrated threshold controls the false-positive rate at the data-dependent stopping time under the calibration null, while per-prefix recalibration does not), and Proposition 4 / Corollary 5 (closed-form expected compute and cascade accuracy, with a frontier monotone exactly where the marginal deferral condition holds). Experiments on a matched ~27k-clip GenVidBench cell report codec AUC@N 0.64, sAUC(1) 0.80, and deferring 15% of clips lifting decision accuracy from 0.75 to 0.78 at ~2.6e9 expected MACs, claimed as 7x below pixel-everywhere; cross-dataset replication on AIGVDBench gives AUC 0.62. The paper is careful to state its limitations: Proposition 3 is valid under the calibration null, Proposition 4 is conditional on the deferral condition being tested, and motion-bias controls are reported.

Significance. The core theoretical observation is correct and clean: for the monotone running maximum, the union over prefixes collapses to the final-prefix event, so a threshold calibrated once at the end controls the false-positive rate at the stopping time. I verified the proof of Proposition 3; it is elementary but non-obvious and directly useful. The empirical apparatus is unusually careful for this area: an identity floor, leave-one-generator-out evaluation, per-generator spread, motion-bias control, re-run external baselines (ReStraV, VideoVeritas), and a released harness that asserts Propositions 2–4 and Corollary 5 on synthetic data. If the measured frontier survives verification, this is a valuable contribution: a cheap always-on front-end with honest guarantees and a priced cascade. However, the practical impact rests on two load-bearing empirical claims that need scrutiny: the paired accuracy gain at the 15% deferral point, and the wall-clock interpretation of the compute savings. The manuscript itself repeatedly and honestly flags the conditional nature of the guarantees, which is a strength, not a defect.

major comments (2)
  1. [E3 / Table 2 / Figure 3] The reported paired gain of +0.027 with “60 vs 130 flips” on ~27k matched clips is arithmetically inconsistent. For a paired binary outcome, the accuracy gain equals (b−c)/N where b and c are the discordant counts; (130−60)/27000 ≈ 0.0026, an order of magnitude below the reported +0.027. A +0.027 gain requires roughly 730 net flips. This is not a rounding issue: it is the central evidence for the headline operating point of 0.78 versus pixel-everywhere 0.76. If the true paired gain is ~0.003, the 15%-deferral decision accuracy is ~0.753, below the pixel baseline, and the claim “above pixel-everywhere at 7× less compute” fails. The released code must resolve whether the flip counts, the sample size, or the gain is mis-stated. I could not resolve this from the manuscript text.
  2. [§4 Cost and §J Compute accounting] The headline “5 orders of magnitude less compute” and “7× less compute” are based on MAC counts for a purpose-built bitstream parser, but the released reference implementation reads motion vectors through the decoder’s side-data export and pays a full frame decode (~0.4 s per GOP), compared with ~28 ms for the pixel forward pass. As the paper itself admits, this largely erases the wall-clock advantage. The MAC accounting is defensible for a deployed platform that already decodes the stream, but the abstract and Figure 3 present the MAC-based number as the operating-point benefit. Please either provide a measured parser benchmark or qualify the headline to “added compute” with the decode treated as a shared cost, so readers can see the wall-clock trade-off at the claimed operating point.
minor comments (4)
  1. [Figure 2 / Table 2] The caption states that Figure 2 endpoints (0.674/0.782) “sit consistently above” Table 2’s AUC@N values (0.64/0.76). Please clarify explicitly why this happens (e.g., different prefix truncation or population definition) so the reader does not perceive an inconsistency.
  2. [E3] The McNemar p-value of 4×10−7 is consistent with 60 vs 130 discordant counts, but not with the reported gain of +0.027. After the code is checked, please ensure the reported CI, p-value, flip counts, and gain all refer to the same paired analysis and same N.
  3. [Data and Code Availability] The statement “the released harness reproduces every number in this paper except the VLM MAC estimate” is strong; please add a commit hash or version identifier and, after resolving the E3 discrepancy, confirm that the reproduction statement still holds.
  4. [§7 Limitations] The limitation that Proposition 3 holds only under the calibration null is stated clearly, and the cross-dataset FPR drift to 0.068 is reported. Consider carrying this qualifier into the abstract, since the phrase “anytime-valid” could otherwise be read as unconditional.

Circularity Check

0 steps flagged

No load-bearing circularity: the anytime-validity and cascade theorems are self-contained, and empirical claims rest on public benchmarks, external baselines, and re-derived anchors; score reflects minor self-citation only.

full rationale

The paper's central derivation chain is not circular. Proposition 3 is proved in Appendix A from elementary probability; the substantive step is the union collapse {exists u<=t: M_u>=tau} = {M_t>=tau} from monotonicity, after which the stopping-time FPR equals the calibrated final-prefix tail. This is a theorem about the aggregation rule, independent of any fitted parameter, and the per-prefix recalibration comparison is also a valid quantile argument. Proposition 4 and Corollary 5 are accounting identities for the cascade; they state the error/compute decomposition of the defined deferral rule and are proved by substitution, so they are analytic rather than empirical predictions. The measured frontier (E3) is an honest measurement on the matched 27k cell, not a claim that the identity predicts external data. Empirically, the main numbers are anchored to public benchmarks (GenVidBench, AIGVDBench); external baselines are re-run under the audited protocol (ReStraV ported and re-trained, VideoVeritas run from the released checkpoint) rather than quoted. The only prominent self-citations are to the authors' VidAudit toolkit for the 13-d feature, base scorer, harness, splits, and the offline 0.832 anchor. These are not load-bearing for the new guarantees: the anytime/cascade theorems hold for any monotone stage-1 score ('any monotone stage 1 inherits both guarantees'), and E1 re-derives the prior anchor from the released feature table (0.830) as a positive control rather than importing it as an unverified premise. The paper also discloses the limits that would matter if they were hidden: the gate's stopping-time FPR drifts to 0.068 on AIGVDBench, the wall-clock advantage depends on a purpose-built parser, and the motion-bias control cannot exclude residual shortcuts on hard generators. The E3 paired-flip counts (60 vs 130 on ~27k clips versus a reported +0.027 gain) look internally inconsistent, but that is a correctness/consistency issue, not a circularity, and is outside this pass. Overall: no prediction reduces to a fitted input or to a self-citation chain; score 2 reflects only the repeated self-citation in the front-end/harness description.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claims rest on four fitted/design parameters (gate threshold tau calibrated to the held-out real null; deferral width w set at the measured frontier knee; base-scorer weights inherited from the authors' VidAudit toolkit [23]; chunk size 16 frames) and six premises: the calibration-null transfer assumption (Proposition 3's guarantee is stated against the null it was calibrated on — the paper's own E6/E7 show FPR drifting to 0.068 under distribution shift); the informativeness of the 13-d codec-motion feature (full-length AUC 0.64 vs the 0.63 identity floor; 3/7 generators at chance); bitstream/codec access with a canonical H.264 re-encode; the running-max monotonicity (standard math); the atomless-score-law feasibility condition for Corollary 5; and the confidence-deferral condition err2^W <= err1^W, which the paper tests rather than assumes. No invented entities: the paper explicitly introduces no new detector; the deferral band, identity floor, and anytime-AUC metrics are design choices and measurements.

free parameters (4)
  • Decision threshold tau at gate level alpha=0.05 = upper 5% quantile of final-prefix running max over held-out real clips (measured stopping-time FPR 0.039-0.041 on matche
    The one end-calibrated threshold Proposition 3 requires. Fitted to the calibration null by design; its validity is exactly the paper's theorem.
  • Deferral width w (band W=[tau-w, tau)) = chosen so Pr(s in W)=15% at the reported operating point
    The single knob setting the deferral rate; the operating point is selected at the measured frontier knee (disclosed in E3).
  • Base scorer f_R weights (13-d linear score on codec-motion features) = inherited from VidAudit [23]; E1 reproduces offline AUC 0.832 / re-extraction 0.830
    Features and scorer come from the authors' own prior toolkit; this paper contributes the streaming machinery and measurements on top of it.
  • Chunk size (frames per GOP window) = 16 (default); 8 and 32 ablated
    Design knob; ablations (E7/Table A1) show 8 starves spectral features and 16/32 tie on full-prefix AUC.
axioms (6)
  • domain assumption Deployed real clips share the calibration null's distribution of final-prefix running-max scores
    Proposition 3's control is stated against the calibration null; E6/E7 measure FPR drift to 0.068 at alpha=0.05 under distribution shift. This is the load-bearing premise for the practical anytime-validity claim (Sec. 5, E7-i).
  • domain assumption The 13-d codec-motion feature (4 global stats + 9 spectral) is informative for AIGV
    Borrowed from VidAudit [23]; E1 reproduces 0.83 offline but E2 shows full-length AUC 0.64 barely above the 0.63 identity floor and 3/7 generators at or near chance (Sec. D).
  • domain assumption A codec with an accessible, informative motion field; canonical H.264 closed-GOP-16 re-encode governs the benchmark
    Limitations: 'assumes bitstream access and a codec whose motion field is informative; transcoding is a calibration step, not free' (Sec. 7); harness described in Sec. 6.
  • standard math Running-max monotonicity: M_t = max(M_{t-1}, s_t) is pointwise non-decreasing
    Property of max; used by Propositions 2 and 3 and Lemma 6.
  • standard math Score law is atomless below tau for Corollary 5's budget-feasibility construction
    Stated in Corollary 5; needed so deferral probability is continuous in the width w.
  • domain assumption Stage 2 is at least as accurate as stage 1 on the deferred set (confidence-deferral condition err2^W <= err1^W)
    Proposition 4's gain is conditional on this; the paper tests rather than assumes it (E3/E4). Listed because the unconditional reading of the frontier's usefulness depends on it.

pith-pipeline@v1.3.0-alltime-deepseek · 21457 in / 37580 out tokens · 356016 ms · 2026-08-01T12:40:23.134964+00:00 · methodology

0 comments
read the original abstract

Detectors for AI-generated video are evaluated offline. A clip is decoded to pixels and scored once, increasingly by a large vision-language model. Detection, however, is deployed online. We recast the task as streaming perception and score the motion field the codec already wrote into the bitstream. Reading that field is a parse, not a pixel-domain forward pass. Because the running aggregate is monotone, one end-calibrated threshold is anytime-valid at the data-dependent decision time. Recalibrating at each prefix is not. Escalation is priced in closed form. A compute budget maps to a deferral window, on a frontier monotone exactly where the deferral condition holds. On matched GenVidBench the codec stage reaches full-length AUC 0.64 at five orders of magnitude less compute than a pixel CNN, on CPU. Its gate holds the stopping-time false-positive rate at target while the real data match its calibration, and drifts above it under distribution shift. Deferring 15% of clips lifts accuracy from 0.75 to 0.78 at $7\times$ less compute (paired: McNemar $p<10^{-6}$). The stage-1 ordering replicates on AIGVDBench. We introduce no new detector. The contribution is the reframing, two guarantees, and the measured frontiers. Code, configurations, and evaluation splits: https://github.com/KurbanIntelligenceLab/streamdet.

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

43 extracted references · 4 canonical work pages

  1. [1]

    VidGuard-R1: AI-generated video detection and explanation via reasoning MLLMs and RL

    Kyoungjun Park, Yifan Yang, Juheon Yi, Shicheng Zheng, Muhammad Muaz, Yifei Shen, Dongqi Han, Caihua Shan, and Lili Qiu. VidGuard-R1: AI-generated video detection and explanation via reasoning MLLMs and RL. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=gXjOsBcXIR

    2026

  2. [2]

    Skyra: AI-generated video detection via grounded artifact reasoning

    Yifei Li, Wenzhao Zheng, Yanran Zhang, Runze Sun, Yu Zheng, Lei Chen, Jie Zhou, and Jiwen Lu. Skyra: AI-generated video detection via grounded artifact reasoning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4482–4493, June 2026

    2026

  3. [3]

    Ivy-Fake: A unified explainable framework and benchmark for image and video AIGC detection

    Changjiang Jiang, Wenhui Dong, Zhonghao Zhang, Fengchang Yu, Wei Peng, Xinbin Yuan, Yifei Bi, Ming Zhao, Zian Zhou, Chenyang Si, and Caifeng Shan. Ivy-Fake: A unified explainable framework and benchmark for image and video AIGC detection. InProceedings of the 2026 International Conference on Multimedia Retrieval, ICMR ’26, page 2438–2447, New York, NY, US...

    arXiv 2026

  4. [4]

    BusterX: MLLM-powered AI-generated video forgery detection and explanation, 2026

    Haiquan Wen, Yiwei He, Zhenglin Huang, Tianxiao Li, Zihan Yu, Xingru Huang, Lu Qi, Baoyuan Wu, Xiangtai Li, and Guangliang Cheng. BusterX: MLLM-powered AI-generated video forgery detection and explanation, 2026. URLhttps://arxiv.org/abs/2505.12620

    Pith/arXiv arXiv 2026

  5. [5]

    VideoVeritas: AI-generated video detection via perception pretext reinforcement learning

    Hao Tan, Jun Lan, Senyuan Shi, Zichang Tan, Zijian Yu, Huijia Zhu, Weiqiang Wang, Jun Wan, and Zhen Lei. VideoVeritas: AI-generated video detection via perception pretext reinforcement learning. InForty-third International Conference on Machine Learning, 2026. URL https://openreview.net/for um?id=lbRi2aAWL7

    2026

  6. [6]

    Artifact-Bench: Evaluating MLLMs on detecting and assessing the artifacts of AI-generated videos, 2026

    Yuqi Tang, Yang Shi, Zhuoran Zhang, Qixun Wang, Xuehai Bai, Yue Ding, Ruizhe Chen, Bohan Zeng, Xinlong Chen, Xuanyu Zhu, Bozhou Li, Yuran Wang, Yifan Dai, Chengzhuo Tong, Xinyu Liu, Yiyan Ji, Yujie Wei, Yuhao Dong, Shilin Yan, Fengxiang Wang, Yi-Fan Zhang, Haotian Wang, Yuanxing Zhang, and Pengfei Wan. Artifact-Bench: Evaluating MLLMs on detecting and ass...

    Pith/arXiv arXiv 2026

  7. [7]

    GenVideoLens: Where LVLMs fall short in AI-generated video detection?, 2026

    Yueying Zou, Pei Pei Li, Zekun Li, Xinyu Guo, Xing Cui, Huaibo Huang, and Ran He. GenVideoLens: Where LVLMs fall short in AI-generated video detection?, 2026. URL https://arxiv.org/abs/2603 .18625

    2026

  8. [8]

    StreamYOLO: Real-time object detection for streaming perception, 2022

    Jinrong Yang, Songtao Liu, Zeming Li, Xiaoping Li, and Jian Sun. StreamYOLO: Real-time object detection for streaming perception, 2022. URLhttps://arxiv.org/abs/2207.10433

    Pith/arXiv arXiv 2022

  9. [9]

    EdgeOAR: Real-time online action recognition on edge devices.IEEE Transactions on Mobile Computing, 24(12):13426–13440, 2025

    Wei Luo, Deyu Zhang, Yin Tang, Fan Wu, and Yaoxue Zhang. EdgeOAR: Real-time online action recognition on edge devices.IEEE Transactions on Mobile Computing, 24(12):13426–13440, 2025. doi: 10.1109/TMC.2025.3591188

    arXiv 2025

  10. [10]

    Detection of real-time deepfakes in video conferencing with active probing and corneal reflection

    Hui Guo, Xin Wang, and Siwei Lyu. Detection of real-time deepfakes in video conferencing with active probing and corneal reflection. InICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023. doi: 10.1109/ICASSP49357.2023.10094720

    arXiv 2023

  11. [11]

    A spatial-frequency aware multi-scale fusion network for real-time deepfake detection

    Libo Lv, Tianyi Wang, Mengxiao Huang, Ruixia Liu, and Yinglong Wang. A spatial-frequency aware multi-scale fusion network for real-time deepfake detection. In Josef Kittler, Hongkai Xiong, Jian Yang, Xilin Chen, Jiwen Lu, Weiyao Lin, Jingyi Yu, and Weishi Zheng, editors,Pattern Recognition and Computer Vision, pages 536–550, Singapore, 2026. Springer Natu...

    2026

  12. [12]

    Deepfake video detection through optical flow based CNN

    Irene Amerini, Leonardo Galteri, Roberto Caldelli, and Alberto Del Bimbo. Deepfake video detection through optical flow based CNN. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019

    2019

  13. [13]

    Multi-domain awareness for compressed deepfake videos detection over social networks guided by common mechanisms between artifacts

    Yan Wang, Qindong Sun, Dongzhu Rong, and Rong Geng. Multi-domain awareness for compressed deepfake videos detection over social networks guided by common mechanisms between artifacts. 14 ANYTIME AIGV DETECTION FROM THE COMPRESSED BITSTREAMKurban Intelligence Lab Computer Vision and Image Understanding, 247:104072, 2024. ISSN 1077-3142. doi: https://doi.or...

    arXiv 2024

  14. [14]

    Efficient temporally- aware deepfake detection using H.264 motion vectors.Electronic Imaging, 36(4):335–1–335–1, 2024

    Peter Grönquist, Yufan Ren, Qingyi He, Alessio Verardo, and Sabine Süsstrunk. Efficient temporally- aware deepfake detection using H.264 motion vectors.Electronic Imaging, 36(4):335–1–335–1, 2024. doi: 10.2352/EI.2024.36.4.MWSF-335. URLhttps://library.imaging.org/ei/articles/36/4/MWSF-335

    work page doi:10.2352/ei.2024.36.4.mwsf-335 2024

  15. [15]

    Seeing what matters: Generalizable AI-generated video detection with forensic-oriented augmentation

    Riccardo Corvi, Davide Cozzolino, Ekta Prashnani, Shalini De Mello, Koki Nagano, and Luisa Verdoliva. Seeing what matters: Generalizable AI-generated video detection with forensic-oriented augmentation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https: //openreview.net/forum?id=dOGXKBL7IE

    2025

  16. [16]

    D3: Training-free AI-generated video detection using second-order features

    Chende Zheng, Ruiqi Suo, Chenhao Lin, Zhengyu Zhao, Le Yang, Shuai Liu, Minghui Yang, Cong Wang, and Chao Shen. D3: Training-free AI-generated video detection using second-order features. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12852–12862, October 2025

    2025

  17. [17]

    Preserving forgery artifacts: AI-generated video detection at native scale

    Zhengcen Li, Chenyang Jiang, Hang Zhao, Shiyang Zhou, Yunyang Mo, Feng Gao, Fan Yang, Qiben Shan, Shaocong Wu, and Jingyong Su. Preserving forgery artifacts: AI-generated video detection at native scale. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=XD43lfRCg6

    2026

  18. [18]

    Viola and M

    P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. InProceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I–I, 2001. doi: 10.1109/CVPR.2001.990517

    arXiv 2001

  19. [19]

    Surat Teerapittayanon, Bradley McDanel, and H.T. Kung. BranchyNet: Fast inference via early exiting from deep neural networks. In2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464–2469, 2016. doi: 10.1109/ICPR.2016.7900006

    arXiv 2016

  20. [20]

    Window-based early-exit cascades for uncertainty esti- mation: When deep ensembles are more efficient than single models

    Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty esti- mation: When deep ensembles are more efficient than single models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368–17380, October 2023

    2023

  21. [21]

    Wittawat Jitkrittum, Neha Gupta, Aditya Krishna Menon, Harikrishna Narasimhan, Ankit Singh Rawat, and Sanjiv Kumar. When does confidence-based cascade deferral suffice? In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors,Advances in Neural Information Processing Systems 36: Annual Conference on Neural Inform...

    2023

  22. [22]

    CoCoVideo: The high-quality commercial-model-based contrastive benchmark for AI-generated video detection

    Huidong Feng, Wentao Chen, Jie Chen, Xinqi Cai, Ruolong Ma, Yinglin Zheng, Yuxin Lin, and Ming Zeng. CoCoVideo: The high-quality commercial-model-based contrastive benchmark for AI-generated video detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11346–11356, June 2026

    2026

  23. [23]

    Auditing generalization in AI-generated video detection: A six-control protocol and the VidAudit toolkit, 2026

    Mert Onur Cakiroglu, Zhihe Lu, Mehmet Dalkilic, and Hasan Kurban. Auditing generalization in AI-generated video detection: A six-control protocol and the VidAudit toolkit, 2026. URL https: //arxiv.org/abs/2606.31004

    Pith/arXiv arXiv 2026

  24. [24]

    AI-generated video detection via perceptual straightening

    Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, and David Klindt. AI-generated video detection via perceptual straightening. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URLhttps://openreview.net/forum?id=LsmUgStXby

    2025

  25. [25]

    SpecSem-Net: Integrating spectral and semantic features for robust AI-generated video detection, 2026

    Zixi Wei, Huixuan Zhang, and Xiaojun Wan. SpecSem-Net: Integrating spectral and semantic features for robust AI-generated video detection, 2026. URLhttps://arxiv.org/abs/2605.17311. 15 ANYTIME AIGV DETECTION FROM THE COMPRESSED BITSTREAMKurban Intelligence Lab

    Pith/arXiv arXiv 2026

  26. [26]

    GenVidBench: A 6-million benchmark for AI-generated video detection.Proceedings of the AAAI Conference on Artificial Intelligence, 40(18):15582–15590, Mar

    Zhenliang Ni, Qiangyu Yan, Mouxiao Huang, Tianning Yuan, Yehui Tang, Hailin Hu, Xinghao Chen, and Yunhe Wang. GenVidBench: A 6-million benchmark for AI-generated video detection.Proceedings of the AAAI Conference on Artificial Intelligence, 40(18):15582–15590, Mar. 2026. doi: 10.1609/aaai.v40i1 8.38587. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/38587

    work page doi:10.1609/aaai.v40i1 2026

  27. [27]

    Your one-stop solution for AI-generated video detection

    Long Ma, Zihao Xue, Yan Wang, Zhiyuan Yan, Jin Xu, Xiaorui Jiang, Haiyang Yu, Yong Liao, and Zhen Bi. Your one-stop solution for AI-generated video detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4458–4470, June 2026

    2026

  28. [28]

    DeMamba: AI-generated video detection on million-scale GenVideo benchmark.Science China Information Sciences, 69(6):162103, May 2026

    Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, and Huaxiong Li. DeMamba: AI-generated video detection on million-scale GenVideo benchmark.Science China Information Sciences, 69(6):162103, May 2026. ISSN 1869-1919. doi: 10.1007/s11432-024-4894-0. URLhttps://doi.org/10.1007/s11432...

    work page doi:10.1007/s11432-024-4894-0 2026

  29. [29]

    Dataset biases and shortcut learning in motion-based AI-generated video detection, 2026

    Joren Michels, Lode Jorissen, and Nick Michiels. Dataset biases and shortcut learning in motion-based AI-generated video detection, 2026. URLhttps://arxiv.org/abs/2607.00948

    Pith/arXiv arXiv 2026

  30. [30]

    CodecSight: Leveraging video codec signals for efficient streaming VLM inference,

    Yulin Zou, Yan Chen, Wenyan Chen, JooYoung Park, Shivaraman Nitin, Luo Tao, Francisco Romero, and Dmitrii Ustiugov. CodecSight: Leveraging video codec signals for efficient streaming VLM inference,

  31. [31]

    Reasoning on a budget: A survey of adaptive and controllable test-time compute in llms, 2025

    Mohammad Ali Alomrani, Yingxue Zhang, Derek Li, Qianyi Sun, Soumyasundar Pal, Zhanguang Zhang, Yaochen Hu, Rohan Deepak Ajwani, Antonios Valkanas, Raika Karimi, Peng Cheng, Yunzhou Wang, Pengyi Liao, Hanrui Huang, Bin Wang, Jianye Hao, and Mark Coates. Reasoning on a budget: A survey of adaptive and controllable test-time compute in llms, 2025. URL https:...

    2025

  32. [32]

    Dujian Ding, Ankur Mallick, Shaokun Zhang, Chi Wang, Daniel Madrigal, Mirian Del Carmen Hipolito Garcia, Menglin Xia, Laks V. S. Lakshmanan, Qingyun Wu, and Victor Rühle. BEST- route: Adaptive LLM routing with test-time optimal compute. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerr...

    2025

  33. [33]

    See without decoding: Motion-vector-based tracking in compressed video, 2026

    Axel Duché, Clément Chatelain, and Gilles Gasso. See without decoding: Motion-vector-based tracking in compressed video, 2026. URLhttps://arxiv.org/abs/2602.00153

    arXiv 2026

  34. [34]

    Content credentials: C2PA technical specification, version 2.4

    Coalition for Content Provenance and Authenticity. Content credentials: C2PA technical specification, version 2.4. https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specifica tion.html, April 2026. Accessed July 20, 2026

    2026

  35. [35]

    Sven Gowal, Rudy Bunel, Florian Stimberg, David Stutz, Guillermo Ortiz-Jimenez, Christina Kouridi, Mel Vecerik, Jamie Hayes, Sylvestre-Alvise Rebuffi, Paul Bernard, Chris Gamble, Miklós Z. Horváth, Fabian Kaczmarczyck, Alex Kaskasoli, Aleksandar Petrov, Ilia Shumailov, Meghana Thotakuri, Olivia Wiles, Jessica Yung, Zahra Ahmed, Victor Martin, Simon Rosen,...

    arXiv 2025

  36. [36]

    European Parliament and Council of the European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU...

    2024

  37. [37]

    Wald.Sequential Tests of Statistical Hypotheses, pages 256–298

    A. Wald.Sequential Tests of Statistical Hypotheses, pages 256–298. Springer New York, New York, NY,

  38. [38]

    Game-Theoretic Statistics and Safe Anytime-Valid Inference.Statistical Science, 38(4):576 – 601, 2023

    Aaditya Ramdas, Peter Grünwald, Vladimir Vovk, and Glenn Shafer. Game-Theoretic Statistics and Safe Anytime-Valid Inference.Statistical Science, 38(4):576 – 601, 2023. doi: 10.1214/23-STS894. URL https://doi.org/10.1214/23-STS894

    doi:10.1214/23-sts894 2023

  39. [39]

    Hypothesis testing with e-values.Foundations and Trends in Statistics, 1(1-2):1–390, 07 2025

    Aaditya Ramdas and Ruodu Wang. Hypothesis testing with e-values.Foundations and Trends in Statistics, 1(1-2):1–390, 07 2025. ISSN 2978-4212. doi: 10.1561/3600000002. URL https://doi.org/10 .1561/3600000002

    doi:10.1561/3600000002 2025

  40. [40]

    Anytime validity is free: inducing sequential tests.Journal of the Royal Statistical Society Series B: Statistical Methodology, page qkag050, 02 2026

    Nick W Koning and Sam van Meer. Anytime validity is free: inducing sequential tests.Journal of the Royal Statistical Society Series B: Statistical Methodology, page qkag050, 02 2026. ISSN 1369-7412. doi: 10.1093/jrsssb/qkag050. URLhttps://doi.org/10.1093/jrsssb/qkag050

    doi:10.1093/jrsssb/qkag050 2026

  41. [41]

    generated

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machine ...

    2021

  42. [1992]

    doi: 10.1007/978-1-4612-0919-5_18

    ISBN 978-1-4612-0919-5. doi: 10.1007/978-1-4612-0919-5_18. URL https://doi.org/10.1007/ 978-1-4612-0919-5_18

    work page doi:10.1007/978-1-4612-0919-5_18

  43. [2026]

    URLhttps://arxiv.org/abs/2604.06036

    Pith/arXiv arXiv