Pith. sign in

REVIEW 4 major objections 5 minor 80 references

Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Decomposing deepfake video detection into four specialist agent perspectives plus a judge produces better cross-generator accuracy than holistic single-model judgment.

desk verdict The benchmark is a real contribution, but the headline F1 gain is confounded with training: the multi-agent system is fine-tuned on 68.9K videos while closed-source baselines are zero-shot. read the letter →

arxiv 2608.06865 v1 pith:IDYFHHAZ submitted 2026-08-07 cs.CV cs.AIcs.MA

Xuechao Zou , Shun Zhang , Kai Li , Yi Zhou , Xinyu Sun , Yuhui Chen , Zhe Wu , Congyan Lang
show 1 more author
Junliang Xing
This is my paper · ORCID
classification cs.CVcs.AIcs.MA
keywords deepfakevideodetectionmulti-agentreasoningmultimodallargelanguagemodelsout-of-distributiongeneralizationforensicanalysisFaceVid-Forensics-100Kexplainableforgery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that single deepfake detectors fail on new AI video generators because they judge the whole video at once, so subtle evidence spread across different cues gets drowned out. It proposes splitting the examination into four independent expert perspectives—texture, lighting, motion, and physics—and letting a judge model reconcile the four reports into a verdict and an explanation. To train such agents it introduces FaceVid-Forensics-100K, a dataset of 100,000 face videos from 33 synthesis methods with fine-grained text annotations for each of the four perspectives. On videos from 20 unseen generators, the multi-agent system built from small open-source models reports higher accuracy, recall, and F1 than every single-model baseline, including closed-source GPT and Gemini systems. If the finding holds, decomposition into specialist perspectives is a practical route to cross-generator deepfake detection without depending on the largest proprietary models.

What carries the argument

The central mechanism is the decomposition of deepfake detection into four independent forensic perspectives—texture, lighting, motion, and physics—each handled by its own observation agent that reports cues without committing to a verdict, followed by a judge agent that reconciles the reports. The supporting machinery is FaceVid-Forensics-100K, whose automatically generated textual labels split observations into those same four dimensions, so each agent can be supervised on a single class of evidence. A separate aggregation step (five annotator MLLMs synthesized by a text-only aggregator, DeepSeek-V4 Pro) produces the observation and explanation labels, and GRPO refines the judge's decisions after supervised fine-tuning. The judge-with-video variant shows that the visual frames still add information on top of the text reports, while the text-only judge remains competitive.

What would settle it

Train the four observation agents on observation labels produced without telling the aggregator the ground-truth label, then evaluate on the same 20 unseen generators; if F1 drops to the level of the no-annotation baselines, the original labels were encoding the answer rather than transferable visual cues.

Watch

Extended reading notes

Core claim

On the paper's out-of-domain test set—5,716 real and 1,920 fake videos from 20 generation methods held out of training—the full framework reports 69.87% accuracy, 81.82% recall, and 53.28% F1, compared with 47.45% F1 for the strongest single-model baseline, Gemini-2.5-Pro. The system consists of four observation agents, each a small open-source multimodal large language model (Qwen2.5-VL-7B) directed to one forensic dimension, plus a judge agent trained with supervised fine-tuning and group relative policy optimization. The judge receives the four textual reports and optionally the sampled frames, and produces a binary verdict with a supporting explanation. The paper interprets the result as evidence that explicit multi-perspective collaborative reasoning generalizes across unseen generators better than holistic judgment by a single MLLM, and that the benchmark's dimension-specific textual supervision is what makes the specialized agents learn transferable cues.

Load-bearing premise

The load-bearing premise is that the automatically generated text labels describe real visual evidence in the video, and are not just paraphrases of the known real-or-fake answer that the aggregator was allowed to see when resolving conflicts.

Editorial extensions

If this is right

  • Small open-source multi-agent systems can outperform closed-source single models for out-of-domain deepfake video detection, lowering the capability threshold for deployment.
  • The four-perspective decomposition is robust across generators: the framework ranks first or second on 11 of the 20 held-out generators after GRPO with video input.
  • Adding direct video access to the judge improves F1 from 51.01% to 53.28%, so textual expert reports are useful but not a full substitute for visual evidence.
  • The framework is modular: mixing different MLLMs for observation and judge roles remains competitive, meaning components can be upgraded independently.
  • Training the judge with GRPO after SFT raises both accuracy and F1, indicating that decision-policy refinement adds generalization beyond learning to produce evidence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The headline result should be read with the supervision caveat: because the annotation pipeline conditions on ground-truth labels, the metric gap may partly reflect agents learning label-consistent language patterns rather than purely visual transferable cues; a blind-label variant would settle how much.
  • The same four-perspective decomposition could transfer to other forgery domains—audio deepfakes, image manipulation localization, or document forgery—where holistic judgment also tends to miss weak distributed evidence.
  • The OOD benchmark itself could be reused as a testbed for whether explanation quality tracks detection accuracy; the paper's own explanation scores show different judge models rank the system differently, so explanation gains are less settled than detection gains.
  • If label-conditioned supervision is indeed the key ingredient, then scaling the dataset's generator coverage may matter more than scaling model size, since the paper shows modest gains from 3B to 32B parameters compared with the jump from single-model to multi-agent reasoning.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces FaceVid-Forensics-100K, a 100,000-video deepfake dataset spanning 33 synthesis methods, with automatically generated textual observations in four forensic dimensions (texture, lighting, motion, physics) and verdict-consistent explanations. It also proposes a multi-agent forensic reasoning framework in which four specialized MLLM agents independently analyze these dimensions and a judge agent reconciles their outputs into a binary prediction and explanation. The main empirical claim is that, on a generator-disjoint out-of-distribution test set of 7,636 videos, the multi-agent system trained with SFT and GRPO on Qwen2.5-VL-7B achieves 69.87% accuracy, 81.82% recall, and 53.28% F1, outperforming all baselines including closed-source GPT and Gemini models evaluated zero-shot. The paper additionally reports ablations on the contribution of each agent, training stages, reasoning strategies, model combinations, and model scale.

Significance. If the results hold, the dataset is a substantial resource: 100K videos, 33 synthesis methods, generator-disjoint OOD evaluation, and per-dimension textual annotations with released prompts are valuable for the deepfake-detection community. The multi-agent decomposition is clearly described, and the ablation design is thoughtful, including training-free comparisons, model-scale experiments, and heterogeneous agent/judge combinations. The paper also ships detailed prompt templates and training dynamics in the appendix, which supports reproducibility. However, the headline comparison against closed-source MLLMs does not hold training status constant, the annotation pipeline is conditioned on ground-truth labels, and the OOD evaluation lacks a majority-class baseline and error bars. These issues leave the central attribution claim underdetermined and the reported margins potentially unquantified, so the paper needs substantial revision before its main claims can be accepted.

major comments (4)
  1. [Experiments, Tables 2 and 5] The central attribution claim is underdetermined because the comparison does not hold training status constant. In Table 2, GPT-4o, GPT-5-mini, and Gemini-2.5-Pro are evaluated zero-shot through their APIs, and the forensics-tuned MLLMs use released checkpoints, while 'Ours (w/ Video)' is Qwen2.5-VL-7B trained with SFT+GRPO on 68,906 videos of FaceVid-Forensics-100K. Table 5 shows that the training-free multi-agent system on the same backbone reaches only 33.53 F1 without video and 42.29 F1 with video, both below Gemini-2.5-Pro's 47.45 F1. The reported 5.83-point F1 gain may therefore come from supervised fine-tuning on the large labeled dataset rather than from the four-perspective decomposition and judge reconciliation. Please add an equally trained single-MLLM control (same backbone, same SFT+GRPO protocol, same data, but without the multi-agent decomposition) and an equally trained non-agent reasoning baseline, and report these controls in the main comparison table.
  2. [Label Generation, Listing 7] The textual supervision used to train the observation agents is generated with access to ground-truth labels, so the agents may learn language-label priors rather than transferable visual forensic cues. The aggregation prompt in Listing 7 instructs DeepSeek-V4 Pro to 'use the ground-truth label solely as internal guidance to resolve conflicts' between models, and the explanation-label section states that aggregation is 'strictly conditioned on the ground-truth video label' with reverse inference when all annotators are wrong. Although the output text is required not to mention the label, the resulting observations and explanations are still functions of the label. The OOD detection metric is not formally circular because the test labels are independent, but the stronger claim that the agents acquire visually grounded, transferable forensic knowledge is not established by the present experiments. Please add an ablation that trains agents on annotations produced without ground-truth access, and test whether observation text alone (without the judge) predicts the label on OOD data at above-chance rates while controlling for the base MLLM.
  3. [Appendix B.4, Listing 8] The explanation-quality evaluation is conditioned on the ground-truth answer and a reference rationale, which inflates the reported quality scores. In Listing 8, both evaluator judges receive the ground-truth answer, the reference rationale, the model answer, and the model rationale, and the rubric explicitly assigns low scores to rationales that contradict the ground-truth label. This makes the explanation scores a measure of agreement with label-conditioned references rather than an independent assessment of forensic quality. The claim that the framework produces high-quality explanations is therefore overstated. Please report a blind evaluation in which judges do not receive the ground-truth label or reference rationale, or provide a human study on a subset of the OOD explanations.
  4. [Experiments, Table 2 and Evaluation Protocols] The OOD evaluation is missing a majority-class baseline and error bars, which undermines the strength of the reported margins. The OOD test set contains 5,716 real and 1,920 fake videos, a 75.1% real base rate; predicting 'real' for every video yields 74.9% accuracy, which is higher than the reported 69.87% accuracy of the full system, and an undefined F1 (or 0 under the standard convention when no fakes are predicted). The paper should report majority-class accuracy and F1, per-class precision/recall, and a confusion matrix. In addition, all results are single runs without confidence intervals or significance tests, so the 5.83 F1 gain over Gemini-2.5-Pro and the 2.27 F1 gain from adding video frames to the judge cannot be distinguished from run-to-run variation. Please report multiple seeds or bootstrap confidence intervals for the main results.
minor comments (5)
  1. [Abstract and Section 4.2] The phrase 'ranks first across all reported metrics' is inaccurate on the OOD set because the reported accuracy of 69.87% is below the 74.9% always-real accuracy, and the in-domain results in Table 8 show small vision models outperforming the framework; please qualify this claim.
  2. [Figure 2 and Appendix A, Table 6] The OOD split contains only entire-face synthesis (EFS) generators, with no face-swap or face-reenactment examples, so the generalization claim should be scoped to entire-face synthesis rather than deepfake video detection in general.
  3. [Section 3.1, Collection and Processing] The collection counts are typeset without separators ('442,000,6,780,000,3,000, and5,000') and should be corrected for readability and verifiability.
  4. [Listing 7] The user prompt in Listing 7 passes 'This video is {ground_truth}' to the aggregator while the system prompt forbids mentioning the label in the output; this tension is the source of the label-conditioning issue and should be acknowledged explicitly in the dataset description.
  5. [Project page and reproducibility] The paper provides the project page URL but does not state whether the dataset and trained models will be released; please clarify data and code availability, since the dataset is a core contribution.

Circularity Check

2 steps flagged · score 4.0 of 10

Detection F1 is not formally circular because OOD labels are independent, but the agent supervision and explanation-quality evaluation are both conditioned on ground-truth labels, making the 'forensic evidence' and 'explanation quality' claims partially self-defined.

  1. self definitional [FaceVid-Forensics-100K / Label Generation; Training Pipeline Eqs. (1)-(2)]
    "Guided by the ground-truth video label to resolve cross-model contradictions and filter noise, the aggregator outputs a structured observation label ... This aggregation is strictly conditioned on the ground-truth video label: when correct predictions exist, the aggregator prioritizes and synthesizes explanations from correct annotators; if all annotators make incorrect predictions, it performs reverse inference by re-evaluating the merged observations against the ground-truth label."

    The textual 'observations' and 'verdict-consistent explanations' used to supervise the agents are manufactured with the ground-truth label as a filtering and tie-breaking input. Observation agents are then trained by Eq. (1) to reproduce these label-conditioned labels, and the judge by Eq. (2) to reproduce the label-conditioned explanation and verdict. The claimed forensic evidence is therefore defined in terms of the answer it is supposed to explain, so the 'evidence-driven' reasoning claim is partly self-definitional. This does not make the held-out F1 logically forced, because the model still maps frames to labels, but it means the transferable-visual-cue and explanation claims are not independently grounded.

  2. self definitional [Appendix B.4 Evaluation of Textual Explanation Quality; Listing 8]
    "Both judges receive the same ground-truth answer, reference rationale, model answer, and model rationale, and both apply an identical rubric covering evidence accuracy, reference alignment, specificity and grounding, clarity and conciseness, and consistency between the rationale and the verdict. ... A rationale that contradicts the ground-truth label or its own predicted label should not receive a high score."

    Explanation quality is scored by how well the model rationale aligns with a reference rationale that was itself aggregated under ground-truth-label guidance, and a rationale that contradicts the ground-truth label is penalized. The evaluator input includes the ground-truth answer, so the explanation-quality comparison measures agreement with a label-conditioned reference rather than independent forensic validity. This makes the reported explanation-quality ranking partially circular, although it does not affect the binary-detection F1 itself.

full rationale

The headline claim that the multi-agent framework outperforms closed-source MLLMs on an out-of-distribution benchmark is not formally circular: the OOD test labels are independent source labels, and the F1 values are computed against those labels rather than against the training annotations. The main circularity concern is upstream of the evaluation. The dataset's observation and explanation labels are generated by MLLMs and then aggregated by DeepSeek-V4 Pro with the ground-truth video label as explicit guidance: the aggregator uses the label to resolve cross-model contradictions, prioritizes explanations from correct annotators, and performs reverse inference when all annotators are wrong. The observation agents and judge are then fine-tuned to reproduce these label-conditioned texts. As a result, the 'forensic observations' that the framework is said to elicit are not independent of the answer they are used to predict; the model may learn label-consistent language priors rather than purely visual transferable cues. The explanation-quality evaluation in Appendix B.4 has the same structure: both evaluators receive the ground-truth answer and the ground-truth-conditioned reference rationale, and the rubric explicitly rewards reference alignment and penalizes contradiction with the ground-truth label, so the reported explanation scores partly measure agreement with the label-conditioned reference. These issues do not force the detection scores by construction, because the held-out set is generator-disjoint and the labels are external, so the paper does not rise to a 6 or above. There is no load-bearing self-citation chain or imported uniqueness theorem; the self-citations present are ordinary related-work citations. Overall, the detection claim has independent content, but the evidence and explanation claims are partially self-defined through ground-truth-conditioned annotation, warranting a moderate circularity score of 4.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central detection claim rests on no fitted numerical constants, but it does rest on several domain assumptions about the annotation pipeline, data curation, and OOD validity. The most consequential is that ground-truth-guided MLLM annotations constitute faithful visual supervision rather than label-encoded text.

free parameters (3)
  • Sampled frame count per video = not specified
    Section 'Multi-Agent System' says frames are uniformly sampled but does not state how many. This choice affects how much temporal evidence the observation agents and the judge receive.
  • Hard-sample retention threshold (AltFreezing score) = not specified
    Dataset collection: low-scoring fake videos are retained for four dominant models, reducing counts to 11,958, 15,163, 10,432, and 32,690, but the exact score cutoff is not reported. This selection may bias artifact diversity.
  • LoRA rank and GRPO group size = 8 responses per prompt for GRPO; LoRA rank not reported
    Training hyperparameters are fixed choices. The group size is reported, but LoRA rank is not, and these choices affect performance without being fitted to the test set.
assumptions (4)
  • domain assumption MLLM annotators (GPT-4o, Gemini 3.5-Flash, Qwen2.5-VL, Skyra, VideoVeritas) can produce reliable forensic observations from uniformly sampled frames.
    The label-generation pipeline presupposes that text reports from these models capture real visual artifacts; no human verification of the annotations is reported.
  • domain assumption Using the ground-truth video label to resolve conflicts in annotation does not leak the label into the generated observation or explanation text.
    The prompts forbid mentioning the label, but the aggregator is instructed to prioritize observations consistent with the label, which can bias the text toward the correct answer.
  • ad hoc to paper The AltFreezing detector score is a valid difficulty measure, and selecting low-scoring fake videos does not bias the artifact distribution.
    The curation step reduces 1,122,159 fake videos to 70,243 based on AltFreezing scores; no analysis shows that this selection is neutral for downstream generalization.
  • domain assumption The OOD test split (20 unseen generators, unseen identities) is free of overlap with training data and with the pretraining data of the base and closed-source models.
    The split is generator-disjoint by construction, but base models and closed-source APIs were pretrained on web data, so contamination cannot be ruled out.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection." pith.science (2026). https://pith.science/paper/IDYFHHAZ

@misc{pith2026260806865,
  author       = {Pith},
  title        = {Pith review of: Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDYFHHAZ}},
  note         = {Machine review of arXiv:2608.06865}
}
read the original abstract

The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective, often fail to capture subtle forgery artifacts, limiting their generalization to emerging AI-generated methods. To address these limitations, we introduce FaceVid-Forensics-100K, a large-scale deepfake video dataset comprising 100,000 videos and spanning 33 synthesis methods across face swapping, face reenactment, and entire-face synthesis, including recent generators such as Seedance 2.0. The dataset provides fine-grained textual annotations of visual observations and verdict-consistent forensic explanations, automatically synthesized through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs. Building on this benchmark, we propose a multi-agent forensic reasoning framework that employs four specialized domain-expert agents to independently analyze forgery cues from four perspectives: texture, lighting, motion, and physics. A judge agent then reconciles their reports to produce a final prediction together with an explanation. Extensive evaluations on out-of-domain test sets show that, despite being composed entirely of small open-source MLLMs, our framework outperforms all methods including closed-source GPT and Gemini models and ranks first across all reported metrics on this benchmark. The project page is available at https://xavierjiezou.github.io/ARGUS/.

Figures

Figures reproduced from arXiv: 2608.06865 by the authors.

Figure 1
Figure 1. Framework comparison. A single MLLM often [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of FaceVid-Forensics-100K. Training and Evaluation. We divide the dataset into train￾ing, in-domain test, and out-of-distribution (OOD) test sets. The training and in-domain test sets share the same families of synthesis methods. To rigorously evaluate cross-generator generalization, we assign videos from 20 completely un￾seen (held-out) EFS generators exclusively to the OOD test set, along with a portion o… view at source ↗
Figure 3
Figure 3. Pipeline of observation and explanation label generation. (a) DeepSeek-V4 Pro aggregates independent observations [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed multi-agent forensic reasoning framework. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on a challenging fake video. GPT-4o and MiMo-V2.5 incorrectly predict Real, treating the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the reasoning strategies evaluated [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of Single prediction, CoT, Multi-turn dialogue, and our multi-agent system (MAS) on a fake [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Training dynamics of the framework. (a–d) SFT loss and token accuracy for the four observation agents and for the [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 44 canonical work pages

  1. [1]

    2019 , month = sep, note =

    Dufour, Nicholas and Gully, Andrew and Karlsson, Per and Vorbyov, Alexey Victor and Leung, Thomas and Childs, Jeremiah and Bregler, Christoph , title =. 2019 , month = sep, note =

    work page 2019

  2. [2]

    arXiv preprint arXiv:2006.07397 , year=

    The deepfake detection challenge (dfdc) dataset , author=. arXiv preprint arXiv:2006.07397 , year=

    arXiv 2006

  3. [3]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Celeb-df: A large-scale challenging dataset for deepfake forensics , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Df40: Toward next-generation deepfake detection , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    arXiv preprint arXiv:2507.18015 , year=

    Celeb-df++: A large-scale challenging video deepfake benchmark for generalizable forensics , author=. arXiv preprint arXiv:2507.18015 , year=

    arXiv

  7. [7]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    A style-based generator architecture for generative adversarial networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  8. [8]

    International Conference on Learning Representations (ICLR) , year=

    Auto-Encoding Variational Bayes , author=. International Conference on Learning Representations (ICLR) , year=

Show all 80 references
  1. [9]

    2024 , howpublished =

    HeyGen , title =. 2024 , howpublished =

    2024

  2. [10]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Float: Generative motion latent flow matching for audio-driven talking portrait , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  3. [11]

    International Journal of Computer Vision , volume=

    Learning Spatiotemporal Inconsistency via Thumbnail Layout for Face Deepfake Detection , author=. International Journal of Computer Vision , volume=. 2024 , publisher=

    2024

  4. [12]

    for now , author=

    Shadows don't lie and lines can't bend! generative models don't know projective geometry... for now , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [13]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Face forgery video detection via temporal forgery cue unraveling , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  6. [14]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Towards more general video-based deepfake detection through facial component guided adaptation for foundation model , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [15]

    IEEE Transactions on Information Forensics and Security , volume=

    Where deepfakes gaze at? Spatial--temporal gaze inconsistency analysis for video face forgery detection , author=. IEEE Transactions on Information Forensics and Security , volume=. 2024 , publisher=

    2024

  8. [16]

    International Conference on Machine Learning , pages=

    Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  10. [18]

    2024 , month = aug, howpublished =

    2024

  11. [19]

    2507.06261 , archivePrefix =

    arXiv preprint arXiv:2507.06261 , year =. 2507.06261 , archivePrefix =

    arXiv

  12. [20]

    arXiv preprint arXiv:2502.13923 , year =

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu,...

    arXiv

  13. [21]

    2507.01006 , archivePrefix =

    arXiv preprint arXiv:2507.01006 , year =. 2507.01006 , archivePrefix =

    arXiv

  14. [22]

    Proceedings of the 43rd International Conference on Machine Learning , year =

    Tan, Hao and Lan, Jun and Shi, Senyuan and Tan, Zichang and Yu, Zijian and Zhu, Huijia and Wang, Weiqiang and Wan, Jun and Lei, Zhen , title =. Proceedings of the 43rd International Conference on Machine Learning , year =

  15. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GenVidBench: A 6-Million Benchmark for AI-Generated Video Detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2026 , doi=

    2026

  16. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Your One-Stop Solution for AI-Generated Video Detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=. doi:10.48550/arXiv.2601.11035 , url=

    doi

  17. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Faceforensics++: Learning to detect manipulated facial images , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    From Slow Bidirectional to Fast Autoregressive Video Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  19. [27]

    2023 , doi =

    Wang, Jiuniu and Yuan, Hangjie and Chen, Dayou and Zhang, Yingya and Wang, Xiang and Zhang, Shiwei , journal =. 2023 , doi =

    2023

  20. [28]

    Khachatryan, Levon and Movsisyan, Andranik and Tadevosyan, Vahram and Henschel, Roberto and Wang, Zhangyang and Navasardyan, Shant and Shi, Humphrey , booktitle =

  21. [29]

    Chen, Haoxin and Zhang, Yong and Cun, Xiaodong and Xia, Menghan and Wang, Xintao and Weng, Chao and Shan, Ying , booktitle =

  22. [30]

    2023 , url =

    Hong, Wenyi and Ding, Ming and Zheng, Wendi and Liu, Xinghan and Tang, Jie , booktitle =. 2023 , url =

    2023

  23. [31]

    2024 , doi =

    Kong, Weijie and Tian, Qi and Zhang, Zijian and Min, Rox and Dai, Zuozhuo and Zhou, Jin and Xiong, Jiangfeng and Li, Xin and Wu, Bo and Zhang, Jianwei and others , journal =. 2024 , doi =

    2024

  24. [32]

    2025 , doi =

    HaCohen, Yoav and Chiprut, Nisan and Brazowski, Benny and Shalem, Daniel and Moshe, Dudu and Richardson, Eitan and Levin, Eran and Shiran, Guy and Zabari, Nir and Gordon, Ori and Panet, Poriya and Weissbuch, Sapir and Kulikov, Victor and Bitterman, Yaki and Melumian, Zeev and ...

    2025

  25. [33]

    2025 , doi =

    Chen, Guibin and Lin, Dixuan and Yang, Jiangping and Lin, Chunze and Zhu, Juncheng and Fan, Mingyuan and Zhang, Hao and others , journal =. 2025 , doi =

    2025

  26. [34]

    doi:10.48550/arXiv.2503.20314 , url =

    arXiv preprint arXiv:2503.20314 , year =. doi:10.48550/arXiv.2503.20314 , url =

    doi arXiv

  27. [35]

    doi:10.48550/arXiv.2604.14148 , url =

    arXiv preprint arXiv:2604.14148 , year =. doi:10.48550/arXiv.2604.14148 , url =

    doi arXiv

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    DeepfakeBench: A Comprehensive Benchmark of Deepfake Detection , author=. Advances in Neural Information Processing Systems , volume=

  29. [37]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Altfreezing for more general video face forgery detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  30. [38]

    2026 , howpublished =

    2026

  31. [39]

    2026 , month = feb, howpublished =

    2026

  32. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unishield: An adaptive multi-agent framework for unified forgery image detection and localization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  33. [41]

    L ong V ideo A gent: Multi-Agent Reasoning with Long Videos

    Liu, Runtao and Liu, Ziyi and Tang, Jiaqi and Ma, Yue and Pi, Renjie and Zhang, Jipeng and Chen, Qifeng. L ong V ideo A gent: Multi-Agent Reasoning with Long Videos. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Pape...

    2026 doi

  34. [42]

    2025 , eprint=

    Single-agent or Multi-agent Systems? Why Not Both? , author=. 2025 , eprint=

    2025

  35. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    MARPO: A Reflective Policy Optimization for Multi-Agent Reinforcement Learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2026 , abstract=

    2026

  36. [44]

    2025 , eprint=

    Enhancing LLM Reasoning with Multi-Path Collaborative Reactive and Reflection agents , author=. 2025 , eprint=

    2025

  37. [45]

    Applied Intelligence , year =

    Lin, Enlu and Chen, Qiong and Qi, Xiaoming , title =. Applied Intelligence , year =. doi:10.1007/s10489-020-01637-z , abstract =

    doi

  38. [46]

    Proceedings of the Second Workshop of Identity Aware AI , month =

    Balancing the Scales: Reinforcement Learning for Fair Classification , author =. Proceedings of the Second Workshop of Identity Aware AI , month =. 2026 , pages =. doi:10.63317/4df2efew2ftw , abstract =

    2026 doi

  39. [47]

    Mixture of Global and Local Experts with Diffusion Transformer for Controllable Face Generation , year=

    Zou, Xuechao and Zhang, Shun and Fu, Xing and Li, Yue and Li, Kai and Cao, Yushe and Lang, Congyan and Tao, Pin and Xing, Junliang , journal=. Mixture of Global and Local Experts with Diffusion Transformer for Controllable Face Generation , year=

  40. [48]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Multivariate diffusion transformer with decoupled attention for high-fidelity mask-text collaborative facial generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  41. [49]

    ICCV , pages=

    Adding conditional control to text-to-image diffusion models , author=. ICCV , pages=

  42. [50]

    CVPR , pages=

    Collaborative diffusion for multi-modal face generation and editing , author=. CVPR , pages=

  43. [51]

    ICLR , year=

    Decoupled Weight Decay Regularization , author=. ICLR , year=

  44. [52]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Intern VL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks , author=. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2024 , organization=

    2024

  45. [53]

    Forty-first International Conference on Machine Learning , year=

    Improving Factuality and Reasoning in Language Models through Multiagent Debate , author=. Forty-first International Conference on Machine Learning , year=

  46. [54]

    AutoGen: Enabling Next-Gen

    Qingyun Wu and Gagan Bansal and Jieyu Zhang and Yiran Wu and Beibin Li and Erkang Zhu and Li Jiang and Xiaoyun Zhang and Shaokun Zhang and Jiale Liu and Ahmed Hassan Awadallah and Ryen W White and Doug Burger and Chi Wang , booktitle=. AutoGen: Enabling Next-Gen. 2024 , url=

    2024

  47. [55]

    Face X-Ray for More General Face Forgery Detection , year=

    Li, Lingzhi and Bao, Jianmin and Zhang, Ting and Yang, Hao and Chen, Dong and Wen, Fang and Guo, Baining , booktitle=. Face X-Ray for More General Face Forgery Detection , year=. doi:10.1109/CVPR42600.2020.00505 , ISSN=

    2020

  48. [56]

    Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection , year=

    Haliassos, Alexandros and Vougioukas, Konstantinos and Petridis, Stavros and Pantic, Maja , booktitle=. Lips Don't Lie: A Generalisable and Robust Approach to Face Forgery Detection , year=. doi:10.1109/CVPR46437.2021.00500 , ISSN=

    2021

  49. [57]

    Exploring Temporal Coherence for More General Video Face Forgery Detection , year=

    Zheng, Yinglin and Bao, Jianmin and Chen, Dong and Zeng, Ming and Wen, Fang , booktitle=. Exploring Temporal Coherence for More General Video Face Forgery Detection , year=. doi:10.1109/ICCV48922.2021.01477 , ISSN=

    2021

  50. [58]

    Proceedings of the 28th ACM International Conference on Multimedia , pages =

    Zi, Bojia and Chang, Minghao and Chen, Jingjing and Ma, Xingjun and Jiang, Yu-Gang , title =. Proceedings of the 28th ACM International Conference on Multimedia , pages =. 2020 , isbn =. doi:10.1145/3394171.3413769 , abstract =

    2020

  51. [59]

    Lang Feng and Longtao Zheng and Shuo He and Fuxiang Zhang and Bo An , booktitle=. Dr. 2026 , url=

    2026

  52. [60]

    Stronger-

    Yujie Zhao and Lanxiang Hu and Yang Wang and Minmin Hou and Hao Zhang and Ke Ding and Jishen Zhao , booktitle=. Stronger-. 2026 , url=

    2026

  53. [61]

    VidGuard-R1:

    Kyoungjun Park and Yifan Yang and Juheon Yi and Muhammad Muaz and Shicheng Zheng and Yifei Shen and Dongqi Han and Caihua Shan and Lili Qiu , booktitle=. VidGuard-R1:. 2026 , url=

    2026

  54. [62]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Face2face: Real-time face capture and reenactment of rgb videos , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  55. [63]

    Acm Transactions on Graphics (TOG) , volume=

    Deferred neural rendering: Image synthesis using neural textures , author=. Acm Transactions on Graphics (TOG) , volume=. 2019 , publisher=

    2019

  56. [64]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Advancing high fidelity identity swapping for forgery detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  57. [65]

    arXiv preprint arXiv:2510.16442 , year=

    EDVD-LLaMA: Explainable Deepfake Video Detection via Multimodal Large Language Model Reasoning , author=. arXiv preprint arXiv:2510.16442 , year=

  58. [66]

    2026 , url=

    Offline Multi-Agent Reinforcement Learning via Sequential Score Decomposition , author=. 2026 , url=

    2026

  59. [67]

    arXiv preprint arXiv:2606.19348 , year=

    Deepseek-v4: Towards highly efficient million-token context intelligence , author=. arXiv preprint arXiv:2606.19348 , year=

  60. [68]

    Gemini 3.5 Flash: Model Card , year =

  61. [69]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  62. [70]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Swift: a scalable lightweight infrastructure for fine-tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  63. [71]

    Edward J Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , pages=

    2022

  64. [72]

    2025 , month = aug, url =

    2025

  65. [73]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Cheng, Jikang and Yan, Renye and Yan, Zhiyuan and Gan, Yaozhong and Zhang, Xueyi and Wang, Zhongyuan and Peng, Wei and Liang, Ling , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

    2026

  66. [74]

    arXiv preprint arXiv:2511.18436 , year=

    When Generative Replay Meets Evolving Deepfakes: Domain-Aware Relative Weighting for Incremental Face Forgery Detection , author=. arXiv preprint arXiv:2511.18436 , year=

    arXiv

  67. [75]

    arXiv preprint arXiv:2607.16873 , year=

    InfoDense: Density-Aware Regional Decisive Replay for Memory-Efficient Incremental Face Forgery Detection , author=. arXiv preprint arXiv:2607.16873 , year=

    arXiv

  68. [76]

    Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , pages=

    Safeear: Content privacy-preserving audio deepfake detection , author=. Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , pages=

    2024

  69. [77]

    International Conference on Learning Representations , volume=

    Audiotrust: Benchmarking the multifaceted trustworthiness of audio large language models , author=. International Conference on Learning Representations , volume=

  70. [78]

    Advances in Neural Information Processing Systems , volume=

    Mmar: A challenging benchmark for deep reasoning in speech, audio, music, and their mix , author=. Advances in Neural Information Processing Systems , volume=

  71. [79]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Huang, Yihuan and Xue, Jun and Jiajun, Liu and Li, Daixian and Zhang, Tong and Yi, Zhuolin and Ren, Yanzhen and Li, Kai , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

    2026

  72. [80]

    2026 , eprint=

    BEAVER: A Training-Free Hierarchical Prompt Compression Method via Structure-Aware Page Selection , author=. 2026 , eprint=

    2026

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.