Pith. sign in

REVIEW 4 major objections 6 minor 64 references

A question–answer probing framework with expert knowledge consistently outperforms single-pass models at spotting child-oriented risks in AI-generated videos.

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 13:07 UTC pith:5LXY6CAQ

load-bearing objection Useful benchmark, but the reported gains likely rest on test-set leakage in the experience memory; the evaluation needs to be redone before the numbers are credible. the 4 major comments →

arxiv 2607.22715 v1 pith:5LXY6CAQ submitted 2026-07-21 cs.CV cs.MM

Child-Oriented AIGC Video Risk Reviewing: A Benchmark and Knowledge-Supported Iterative Reasoning Framework

Lewen Mi , Manyi Li , Yuling Sun , Yufan Zhang , Yuxin Shi , Yulong Bian , Xiangxian Li , Juan Liu This is my paper
classification cs.CV cs.MM
keywords AIGC video reviewchild safetyvision-language modelsmulti-agent frameworkrisk taxonomybenchmarkiterative evidence constructioncontent moderation
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.

This paper argues that reviewing AI-generated videos for risks specific to children is a distinct task that cannot be solved by adult-oriented violation detection or single-pass classification. To make the case, it introduces CAVSR, a benchmark of 605 real-world AIGC videos annotated with a six-category, 26-label hierarchical risk taxonomy, and QVRS-E, a framework in which a Question Agent, Vision Agent, and Review Agent iteratively probe visual evidence under expert and experience knowledge before a Summary Agent issues a structured report. The paper reports that this iterative evidence construction substantially improves open-source vision-language models, e.g. raising Micro-F1 from 0.2414 to 0.4498 on the fine-grained setting, and that a 7–8B open-source backbone equipped with QVRS-E becomes competitive with or better than GPT-4o on this benchmark. If these results hold, they offer a path toward deployable, low-cost child-safety review for the growing flood of AIGC video content.

Core claim

The central discovery is that decomposing child-risk review into repeatable question–answer–verification loops, rather than asking a model to judge a video in one pass, yields large and consistent gains in fine-grained risk recognition. On the 26-label taxonomy, QVRS-E raises Micro-F1 from 0.2414 to 0.4498 with Qwen3-VL-8B, and from 0.1246 to 0.3208 with Qwen2.5-VL-7B; on the merged 6-class setting it reaches 0.6975 Micro-F1, exceeding GPT-4o's 0.5901. Ablations attribute the bulk of the improvement to the iterative Q-V-R probing itself, with expert knowledge contributing additional gains in fine-grained label ranking, and the paper shows the two components are complementary.

What carries the argument

The load-bearing mechanism is the Q-V-R probing loop: a Question Agent converts risk cues into targeted, visually verifiable questions; a Vision Agent answers them over 64 uniformly sampled frames; and a Review Agent decides whether accumulated evidence is sufficient, looping back if not. This loop is anchored by two knowledge stores—an expert rule base defining risk boundaries and typical manifestations, and an experience memory that saves past scene descriptions, effective inquiry paths, and risk conclusions for reuse. The framework's effect is to turn a single holistic judgment into a sequence of grounded verification steps, which the paper shows improves evidence coverage (probing) and f

Load-bearing premise

The evaluation assumes the experience-memory update in Eq. (7) does not leak information across the evaluated videos, but the paper never specifies a train/test split for memory, so later benchmark predictions may be partly conditioned on inquiry paths and conclusions drawn from earlier test videos.

What would settle it

Run the full CAVSR evaluation with the experience-memory write-back disabled (or with memory built only on a separate training set) and compare Micro-F1 against the reported 0.4498 and 0.6975; if the gap largely collapses, the gains are attributable to transductive leakage rather than to the probing framework itself.

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

If this is right

  • An open-source 7–8B model with QVRS-E matches or beats GPT-4o on the CAVSR benchmark, making child-oriented risk review feasible for local, low-cost deployment without proprietary APIs.
  • The benchmark's taxonomy foregrounds risks that are characteristic of AIGC failure modes (distorted biological appearance, body cognition errors, logical paradoxes), giving video-safety research a target aligned with how generated content actually breaks.
  • Iterative evidence construction reduces reliance on a single salient frame, enabling detection of implicit risks that are distributed across frames and depend on contextual interpretation.
  • The framework's gains are larger at the coarse 6-class level than the fine 26-class level, indicating that accumulated evidence translates more reliably into high-level risk judgments.

Where Pith is reading between the lines

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

  • If the experience memory is updated after every video in the 605-video benchmark without a train/test split, later predictions are conditioned on earlier test videos' inquiry paths; the reported gains may therefore partly reflect transductive leakage rather than the framework's standalone ability. A protocol that freezes memory on a held-out training set would provide a cleaner measurement.
  • The long-tail label distribution means Micro-F1 gains could be driven by a few frequent classes; per-label recall/precision breakdowns would clarify whether the framework improves rare-risk detection or mostly consolidates easy wins.
  • Since the framework errs toward over-interpretation in stylized content (flagged in the error analysis), deploying it in production would likely require a tunable risk-aversion threshold to avoid over-flagging benign cartoon content.
  • The reliance on uniform 64-frame sampling misses transient risk cues; an adaptive frame-selection mechanism keyed on motion or saliency is a natural, testable extension of the same Q-V-R structure.

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

4 major / 6 minor

Summary. The paper introduces CAVSR, a benchmark of 605 real-world AIGC videos annotated with a hierarchical taxonomy of 6 top-level and 26 fine-grained child-oriented risk labels, and proposes QVRS-E, a multi-agent framework that iteratively probes a VLM with question–answer–review cycles, augmented by expert knowledge rules and an experience memory that is updated after each assessed video. Experiments report substantial improvements over open- and closed-source baselines under both the fine-grained 26-class and coarse-grained 6-class settings, e.g., Micro-F1 rising from 0.2414 to 0.4498 for Qwen3-VL-8B in the 26-class setting.

Significance. The problem is timely and socially important, and the CAVSR benchmark could be a useful resource for the community if properly validated. The proposed framework is also methodologically interesting, combining multi-agent probing with knowledge injection. If the reported gains were measured under a sound protocol, the result would be significant: sub-10B open-source VLMs outperforming GPT-4o on fine-grained child-risk review. However, the current experimental design does not support the central claim because the experience-memory update leaks information across evaluated test videos, and because the evaluation is reported as single runs with no statistical significance or inter-annotator agreement. These issues undermine the validity of the headline comparisons.

major comments (4)
  1. [§4.1.3, Eq. (7); §5.1] The experience memory is updated after every assessed video via K_mem ← K_mem ∪ {m}, and Section 5.1 does not describe a train/test split or specify that K_mem is frozen/reset during evaluation. If all 605 benchmark videos are assessed sequentially, then each video after the first is reviewed with a memory base populated from earlier test videos' inquiry paths and risk conclusions. The baselines do not receive this memory, so the comparisons in Tables 1 and 2 are not apples-to-apples. This transductive leakage nullifies the claim of 'consistently outperforms' and likely inflates the reported gains. The ablation 'Ours(all)' vs 'Iterative Q-V-R Probing' in Tables 3–4 also includes the accumulating memory, so the isolated contribution of experience knowledge cannot be determined. The authors must either populate K_mem from a separate training corpus before evaluation, or freeze/reset memory
  2. [§3.2 and §4.1.2] There is a circularity concern: the risk taxonomy, the gold labels, and the expert-knowledge rules are all derived from the same child-education expert perspective. Injecting this expert knowledge into the model and then scoring against expert-provided labels means the method's improvement may reflect alignment with the annotators' conceptualization rather than better detection of an independent property. This does not by itself invalidate the method, but it is a load-bearing point for the claim that expert knowledge 'further improves fine-grained risk attribution.' A concrete mitigation would be to evaluate on labels from an independent annotator pool or to derive rules/labels from disjoint expert groups.
  3. [§5.2, Tables 1–4] All results are reported from single runs with no confidence intervals, no multiple seeds, and no significance tests. Given the stochasticity of LLM decoding and the order-dependent memory retrieval, the differences between the proposed method and baselines (and between ablation variants) could be within run-to-run noise. The paper repeatedly uses 'significantly enhances' and 'consistently outperforms,' but no statistical evidence is provided. The authors should run the evaluation multiple times (at least 3–5 seeds, or bootstrap over the test set) and report means/standard deviations along with paired significance tests.
  4. [§3.2] The benchmark annotation procedure reports that three annotators reached consensus after discussion, but no inter-annotator agreement metric (e.g., Cohen's kappa or Fleiss' kappa) is reported for the pre-consensus labels or for the final labels. Without this, the reliability of the gold standard used to evaluate all models is unquantified. The authors should report per-label and overall agreement statistics, at least on a subset.
minor comments (6)
  1. [§4.1.1] The heading '4.1.1 Risk-Guided Multi-Agent Design' is immediately followed by '4.1.2' with no actual text under 4.1.1. Either the subsection is empty or a paragraph is missing.
  2. [Figure 1 and §3 heading] The dataset is referred to as 'CA VSR' in Figure 1 and in the Section 3 heading ('The CA VSR Dataset'), while the acronym is elsewhere CAVSR. Please standardize.
  3. [Tables 1–4] The metric is labeled 'Hit Rate@3' in Tables 1–2 but 'Hit@3' in the text and Tables 3–4. Use one name consistently.
  4. [Eq. (7)] The initial state of K_mem is not specified. If it is empty at the start of evaluation, that should be stated; if it is pre-populated, the source should be described.
  5. [§5.1] The paper says the formal definition of Hit@K is provided in the supplementary material, but a standard ranking metric like Hit@K should be defined in the main text for readability.
  6. [Table 1] For Ours (Qwen3-VL-8B), Precision@1 = 0.6215 and Recall@1 = 0.1221. Given that these are rank-based metrics in a multi-label setting, the large discrepancy deserves a brief explanation to avoid appearing contradictory.

Circularity Check

1 steps flagged

Reported gains may be inflated by experience-memory leakage: Eq. (7) writes each assessed video into K_mem with no train/test split specified, so later test videos are reviewed using memory derived from earlier test videos.

specific steps
  1. fitted input called prediction [Section 4.1.3 (Experience Memory Update), Eq. (7); retrieval in Section 4.1.2; evaluation in Section 5.1]
    "After assessing a video, the system writes newly accumulated experience back to K mem. Specifically, it extracts the initial description d0, an effective inquiry path Π∗, and the final risk conclusion y ... K mem←K mem∪{m}. ... During probing, ΦQ uses d0 or the latest feedback to retrieve similar items from K mem as references."

    Eq. (7) appends each assessed video's final risk conclusion y to K_mem, and Section 4.1.2 makes this memory a retrieval source for later probing. Section 5.1 evaluates on the 605-video CAVSR benchmark with no stated training/evaluation split and no statement that K_mem is frozen or reset during benchmarking. Sequentially, later predictions are conditioned on question paths and risk conclusions extracted from earlier benchmark (test) videos; baselines receive no such memory. Thus the reported Micro-F1/Hit@3 gains are not an independent measurement: part of the 'prediction' is constructed from the evaluation set's own labels.

full rationale

The central derivation of QVRS-E is not circular in its architecture: question generation, visual verification, and evidence accumulation are defined independently of the gold labels. However, the evaluation claim that QVRS-E 'consistently outperforms' baselines rests on a contaminated protocol. The experience memory is updated with each assessed video's final risk conclusion (Eq. 7) and is retrieved during subsequent probing; the paper never specifies that the 605-video benchmark is split into training/validation sets or that K_mem is reset per video. Under sequential evaluation, later test predictions are partly built from earlier test videos' labels, a transductive leakage that baselines do not enjoy. This makes the headline performance comparison and the experience-knowledge ablation non-independent. The expert-knowledge/gold-label same-perspective issue is present but secondary; it is a benchmark-validity concern rather than a formal circular reduction. No self-citation chain or uniqueness import is load-bearing. Score reflects the partial circularity of the evaluation: the claimed prediction gains reduce substantially to test-set-conditioned memory.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its free parameters are design choices in frame sampling and retrieval. The key axioms are domain assumptions about label validity, VLM grounding, frame coverage, and memory isolation; the memory-isolation assumption is the most fragile because the paper's own equations update memory on every assessed video.

free parameters (2)
  • Uniform frame sample size = 64
    Section 4.1.2 fixes 64 uniformly sampled frames as the visual basis; this hand-chosen budget determines which transient risks are visible and is acknowledged in the error analysis as a cause of missed brief cues.
  • Retrieval top-k / similarity thresholds for Chroma memory
    The paper specifies Chroma with m3e-base embeddings but does not state retrieval counts or thresholds used for Z_mem and Z_rule; these design choices affect which expert rules and past cases influence questions and assessment.
axioms (5)
  • domain assumption Expert-consensus risk labels are valid ground truth for child developmental harm.
    The benchmark's gold labels are produced by three child-education experts through focus-group discussion, with no reported inter-annotator agreement or validation against external child-development standards; all evaluation scores inherit this assumption.
  • domain assumption The Vision Agent's answers are objective and grounded in visual facts.
    Section 4.1 describes Phi_V as returning 'objective answers grounded in visual facts,' but it is the same VLM backbone used elsewhere and is subject to hallucination; the framework's evidence accumulation depends on this reliability.
  • domain assumption Uniformly sampling 64 frames adequately represents each video's risk content.
    Section 4.1.2 samples 64 frames uniformly; Section 5.4 admits this causes failures when risk cues are transient or fine-grained, so evidence coverage is bounded by this sampling choice.
  • domain assumption Experience memory updates during evaluation do not contaminate the test set.
    Eq. (7) appends each assessed video's memory item to K_mem, and the experimental section does not state that memory is cleared or restricted to a training split. This assumption is likely violated, which would invalidate the reported gains.
  • domain assumption m3e-base text embeddings are adequate for retrieving relevant expert rules and historical cases.
    Section 5.1 selects m3e-base without evaluation against alternative embedding models; retrieval quality is load-bearing for the knowledge-support components.

pith-pipeline@v1.3.0-alltime-deepseek · 15272 in / 8463 out tokens · 95917 ms · 2026-08-01T13:07:37.610743+00:00 · methodology

0 comments
read the original abstract

The rapid growth of Artificial Intelligence-generated content (AIGC) is reshaping video production and circulation, exposing children to an increasing volume of AIGC videos. Unlike traditionally produced videos, AIGC videos often exhibit greater uncertainty in visual details, narrative coherence, and content expression, which may introduce developmentally inappropriate risks for children. However, existing video safety research is largely designed for general violation detection from an adult perspective and remains insufficient for identifying the fine-grained, implicit, and context-dependent risks that children may encounter when viewing AIGC videos. To address this gap, we study child-oriented AIGC video reviewing, making three contributions. First, we construct CAVSR, a benchmark of 605 real-world videos collected from multiple platforms, and develop a hierarchical risk taxonomy comprising 6 top-level categories and 26 fine-grained labels to support systematic evaluation of children's viewing risks. Second, we propose QVRS-E, a knowledge- and experience-augmented video reviewing framework that combines multi-agent collaboration with expert and experiential knowledge to support targeted evidence acquisition and fact-grounded reviewing decisions. Third, extensive experiments demonstrate that our method significantly enhances the reviewing of child-related risks integrated with vision-language models, and yields more robust review reports.

Figures

Figures reproduced from arXiv: 2607.22715 by Juan Liu, Lewen Mi, Manyi Li, Xiangxian Li, Yufan Zhang, Yuling Sun, Yulong Bian, Yuxin Shi.

Figure 1
Figure 1. Figure 1: Overview of our study on child-oriented AIGC video risk review. The figure illustrates the CAVSR benchmark, its risk [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the risk label taxonomy in the CAVSR benchmark and the distribution of collected data. The left part [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed knowledge-supported [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative cases in the error analysis. Cases (a)–(d) compare our method with the baseline on child-oriented risk [PITH_FULL_IMAGE:figures/full_fig_p008_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

64 extracted references · 15 linked inside Pith

  1. [1]

    Sajal Aggarwal and Dinesh Kumar Vishwakarma. 2023. Protecting our children from the dark corners of YouTube: A cutting-edge analysis. In2023 4th IEEE global conference for advancement in technology (GCAT). IEEE, 1–5

    2023

  2. [2]

    Syed Hammad Ahmed. 2024. A Multimodal Framework for Automated Content Moderation of Children’s Videos. (2024)

    2024

  3. [3]

    Syed Hammad Ahmed, Shengnan Hu, and Gita Sukthankar. 2023. The Potential of Vision-Language Models for Content Moderation of Children’s Videos. In2023 International Conference on Machine Learning and Applications (ICMLA). IEEE, 1237–1241

    2023

  4. [4]

    Syed Hammad Ahmed, Muhammad Junaid Khan, HM Qaisar, and Gita Sukthankar

  5. [5]

    Syed Hammad Ahmed, Muhammad Junaid Khan, and Gita Sukthankar. 2024. Enhanced multimodal content moderation of children’s videos using audiovisual fusion.arXiv preprint arXiv:2405.06128(2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Iftikhar Alam, Abdul Basit, and Riaz Ahmad Ziar. 2024. Utilizing Age-Adaptive Deep Learning Approaches for Detecting Inappropriate Video Content.Human Behavior and Emerging Technologies2024, 1 (2024), 7004031

    2024

  7. [7]

    R Ami. 2021. AI in automated content moderation on social media.International Journal of Artificial Intelligence and Machine Learning4, 3 (2021)

    2021

  8. [8]

    Rashid Azim, Naveed Abbas, Hend Khalid Alkahtani, and Ayman Qahmash. 2026. An explainable deep learning framework for video violence detection using unsupervised keyframe selection and attention-based CNN.Scientific Reports (2026)

    2026

  9. [9]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

    Pith/arXiv arXiv 2025

  10. [10]

    Mazen Balat, Mahmoud Gabr, Hend Bakr, and Ahmed B Zaky. 2024. Tikguard: A deep learning transformer-based solution for detecting unsuitable tiktok content for kids. In2024 6th Novel Intelligent and Leading Emerging Sciences Conference (NILES). IEEE, 337–340

    2024

  11. [11]

    Le Binh, Rajat Tandon, Chingis Oinar, Jeffrey Liu, Uma Durairaj, Jiani Guo, Spencer Zahabizadeh, Sanjana Ilango, Jeremy Tang, Fred Morstatter, et al. 2022. Samba: Identifying inappropriate videos for young children on YouTube. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 88–97

    2022

  12. [12]

    Jasmin Bogatinovski, Ljupčo Todorovski, Sašo Džeroski, and Dragi Kocev. 2022. Comprehensive comparative study of multi-label classification methods.Expert Systems with Applications203 (2022), 117215

    2022

  13. [13]

    J Alex Bonus, Rebecca A Dore, Brenna Hassinger-Das, Julia M Wilson, Elena O’Hara, and C Joseph Francemone. 2025. You do You [Tube]! The multifaceted roles of online video viewing in the lives of US children.Early Childhood Research Quarterly70 (2025), 167–177

    2025

  14. [14]

    Elena Bozzola, Giulia Spina, Rino Agostiniani, Sarah Barni, Rocco Russo, Elena Scarpato, Antonio Di Mauro, Antonella Vita Di Stefano, Cinthia Caruso, Giovanni Corsello, et al. 2022. The use of social media in children and adolescents: Scoping review on the potential risks.International journal of environmental research and public health19, 16 (2022), 9960

    2022

  15. [15]

    Yuxin Cao, Wei Song, Derui Wang, Jingling Xue, and Jin Song Dong. 2026. Failures to surface harmful contents in video large language models. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 35331–35339

    2026

  16. [16]

    Satish Prakash Chand. 2025. Methods of data collection in qualitative research: Interviews, focus groups, observations, and document analysis.Advances in Educational Research and Evaluation6, 1 (2025), 303–317

    2025

  17. [17]

    Hao Chen, Pingping Wang, and Shuaikang Hao. 2025. AI in the spotlight: The impact of artificial intelligence disclosure on user engagement in short-form videos.Computers in Human Behavior162 (2025), 108448

    2025

  18. [18]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al . 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Bianca Chu, Natansh D Modi, Bradley D Menz, Erik Cornelisse, Stephen Bac- chi, Norma Bulamu, Shahid Ullah, Ross A McKinnon, Kacper Gradon, Andrew Rowland, et al. 2026. Evaluation of Generative Artificial Intelligence Safeguards Against the Creation of Images and Videos Harmful to Public Health.Public Health Reports®(2026), 00333549261418596

    2026

  20. [20]

    Mohammad Yasser Chuttur and A Nazurally. 2022. A multi-modal approach to detect inappropriate cartoon video contents using deep learning networks. Multimedia Tools and Applications81, 12 (2022), 16881–16900

    2022

  21. [21]

    Johan Edstedt, Amanda Berg, Michael Felsberg, Johan Karlsson, Francisca Be- navente, Anette Novak, and Gustav Grund Pihlgren. 2022. Vidharm: A clip based dataset for harmful content detection. In2022 26th International Conference on Pattern Recognition (ICPR). IEEE, 1543–1549

    2022

  22. [22]

    Fatmaelzahraa Eltaher, Rahul Krishna Gajula, Luis Miralles-Pechuán, Patrick Crotty, Juan Martínez-Otero, Christina Thorpe, and Susan McKeever. 2025. Pro- tecting young users on social media: Evaluating the effectiveness of content moderation and legal safeguards on video sharing platforms.arXiv preprint arXiv:2505.11160(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. 2024. Videoagent: A memory-augmented multimodal agent for video understanding. InEuropean Conference on Computer Vision. Springer, 75–92

    2024

  24. [24]

    Laura Christina Anne Fearnley, Elly Cairns, Tom Stoneham, Philippa Mary Ryan, Jennifer Alison Chubb, Jo Iacovides, Cynthia Paola Iglesias Urrutia, Phillip David James Morgan, John Alexander McDermid, and Ibrahim Habli. 2025. Risk of What? Defining Harm in the Context of AI Safety. https://eprints.whiterose. ac.uk/id/eprint/223407/

    2025

  25. [25]

    Myrsini Gkolemi, Panagiotis Papadopoulos, Evangelos Markatos, and Nicolas Kourtellis. 2022. YouTubers Not MadeForKids: Detecting channels sharing inap- propriate videos targeting children. InProceedings of the 14th ACM Web Science Conference 2022. 370–381

    2022

  26. [26]

    Team Glm, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A fam- ily of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793(2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Google. 2025. Gemini 3 Flash: A faster, more efficient model for developers. https: //blog.google/products-and-platforms/products/gemini/gemini-3-flash/. Ac- cessed: 2025-05-22

    2025

  28. [28]

    Adamos Hadjipanayis, Elisavet Efstathiou, Peter Altorjai, Tom Stiris, Arunas Valiulis, Berthold Koletzko, and Helena Fonseca. 2019. Social media and children: what is the paediatrician’s role?European journal of pediatrics178, 10 (2019), 1605–1612

    2019

  29. [29]

    Junda He, Xin Zhou, Bowen Xu, Ting Zhang, Kisub Kim, Zhou Yang, Ferdian Thung, Ivana Clairine Irsan, and David Lo. 2024. Representation learning for stack overflow posts: How far are we?ACM Transactions on Software Engineering and Methodology33, 3 (2024), 1–24

    2024

  30. [30]

    YuFeng Huang, ShiJuan Lv, Kuo-Kun Tseng, Pin-Jen Tseng, Xin Xie, and Regina Fang-Ying Lin. 2023. Recent advances in artificial intelligence for video produc- tion system.Enterprise Information Systems17, 11 (2023), 2246188

    2023

  31. [31]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    Viktor Denes Huszar, Vamsi Kiran Adhikarla, Imre Negyesi, and Csaba Krasz- nay. 2023. Toward fast and accurate violence detection for automated video surveillance applications.IEEE Access11 (2023), 18772–18793. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Mi et al

    2023

  33. [33]

    Claire Wonjeong Jo, Magdalena Wojcieszak, et al. 2024. Harmful youtube video detection: A taxonomy of online harm and mllms as alternative annotators.arXiv preprint arXiv:2411.05854(2024)

    Pith/arXiv arXiv 2024

  34. [34]

    Hamza Khan, Xiaohong Yuan, Letu Qingge, and Kaushik Roy. 2025. Violence detection from industrial surveillance videos using deep learning.IEEE Access13 (2025), 15363–15375

    2025

  35. [35]

    Sena A Kojah, Kentaro Toyama, and Oliver L Haimson. 2025. Silencing the Voiceless: Social Media Content Moderation for At Risk Marginalized Populations in High-Stakes Violence and Human Rights Contexts in Nigeria.Proceedings of the ACM on Human-Computer Interaction9, 7 (2025), 1–25

    2025

  36. [36]

    Dimitrios Kollias, Damith C Senadeera, Jianian Zheng, Kaushal KK Yadav, Greg Slabaugh, Muhammad Awais, and Xiaoyun Yang. 2025. Dvd: A comprehensive dataset for advancing violence detection in real-world scenarios.arXiv preprint arXiv:2506.05372(2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Noriyuki Kugo, Xiang Li, Zixin Li, Ashish Gupta, Arpandeep Khatua, Nidhish Jain, Chaitanya Patel, Yuta Kyuragi, Yasunori Ishii, Masamoto Tanabiki, et al

  38. [38]

    Deepak Kumar, Yousef Anees AbuHashem, and Zakir Durumeric. 2024. Watch your language: Investigating content moderation with large language models. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 18. 865–878

    2024

  39. [39]

    Adi Levi, Or Levi, Sardhendu Mishra, and Jonathan Morra. 2025. AI vs. Human Moderators: A Comparative Evaluation of Multimodal LLMs in Content Modera- tion for Brand Safety. InProceedings of the IEEE/CVF International Conference on Computer Vision. 5965–5973

    2025

  40. [40]

    Xuannan Liu, Zekun Li, Zheqi He, Peipei Li, Shuhan Xia, Xing Cui, Huaibo Huang, Xi Yang, and Ran He. 2025. Video-SafetyBench: A benchmark for safety evaluation of video lvlms.arXiv preprint arXiv:2505.11842(2025)

    arXiv 2025

  41. [41]

    Rui Lu, Jinhe Bi, Yunpu Ma, Feng Xiao, Yuntao Du, and Yijun Tian. 2025. MV- Debate: Multi-view Agent Debate with Dynamic Reflection Gating for Multimodal Harmful Content Detection in Social Media.arXiv preprint arXiv:2508.05557 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    Yue Ma, Kunyu Feng, Zhongyuan Hu, Xinyu Wang, Yucheng Wang, Mingzhe Zheng, Bingyuan Wang, Qinghe Wang, Xuanhua He, Hongfa Wang, et al. 2025. Controllable video generation: A survey.arXiv preprint arXiv:2507.16869(2025)

    arXiv 2025

  43. [43]

    Batyrkhan Omarov, Sergazi Narynov, Zhandos Zhumanov, Aidana Gumar, and Mariyam Khassanova. 2022. State-of-the-art violence detection techniques in video surveillance security systems: a systematic review.PeerJ Computer Science 8 (2022), e920

    2022

  44. [44]

    Cihan Orak and Zeynep Turan. 2024. Using artificial intelligence in digital video production: A systematic review study.Journal of Educational Technology and Online Learning7, 3 (2024), 286–307

    2024

  45. [45]

    Xiaoye Ouyang, Aochuan Ba, Siyuan Zhu, Zhen Zhao, Xiaoling Zhu, Qingxin Sun, and Liu Yuan. 2025. Multimodal Perception and Agent-Based Decision- Making Integrated Approach for Short Video Content Compliance Regulation. InProceedings of the 2025 International Conference on Artificial Intelligence and Foundation Model. 146–151

    2025

  46. [46]

    Shraman Pramanick, Shivam Sharma, Dimitar Dimitrov, Md Shad Akhtar, Preslav Nakov, and Tanmoy Chakraborty. 2021. MOMENTA: A multimodal framework for detecting harmful memes and their targets. InFindings of the association for computational linguistics: EMNLP 2021. 4439–4455

    2021

  47. [47]

    Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. 2024. Sniffer: Multimodal large language model for explainable out-of-context misinformation detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 13052–13062

    2024

  48. [48]

    Bowen Qu, Xiaoyu Liang, Shangkun Sun, and Wei Gao. 2024. Exploring aigc video quality: A focus on visual harmony video-text consistency and domain distribution gap. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6652–6660

    2024

  49. [49]

    Krithika Ramesh, Ashiqur R KhudaBukhsh, and Sumeet Kumar. 2022. ‘beach’to ‘bitch’: Inadvertent unsafe transcription of kids’ content on youtube. InProceed- ings of the AAAI Conference on Artificial Intelligence, Vol. 36. 12108–12118

    2022

  50. [50]

    Shubham Singh, Rishabh Kaushal, Arun Balaji Buduru, and Ponnurangam Ku- maraguru. 2019. KidsGUARD: fine grained approach for child unsafe video representation and detection. InProceedings of the 34th ACM/SIGAPP symposium on applied computing. 2104–2111

    2019

  51. [51]

    Mohit Singhal, Chen Ling, Pujan Paudel, Poojitha Thota, Nihal Kumarswamy, Gianluca Stringhini, and Shirin Nilizadeh. 2023. SoK: Content moderation in social media, from guidelines to enforcement, and research to practice. In2023 IEEE 8th European Symposium on Security and Privacy (EuroS&P). IEEE, 868–895

    2023

  52. [52]

    Ella Steen, Kathryn Yurechko, and Daniel Klug. 2023. You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on TikTok.Social Media+ Society9, 3 (2023), 20563051231194586

    2023

  53. [53]

    Tan Tang, Yanhong Wu, Yingcai Wu, Lingyun Yu, and Yuhong Li. 2021. Videomod- erator: A risk-aware framework for multimodal video moderation in e-commerce. IEEE Transactions on Visualization and Computer Graphics28, 1 (2021), 846–856

    2021

  54. [54]

    Han Wang, Tan Rui Yang, Usman Naseem, and Roy Ka-Wei Lee. 2024. Multihate- clip: A multilingual benchmark dataset for hateful video detection on youtube and bilibili. InProceedings of the 32nd ACM International Conference on Multimedia. 7493–7502

    2024

  55. [55]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191(2024)

    Pith/arXiv arXiv 2024

  56. [56]

    Yixu Wang, Jiaxin Song, Yifeng Gao, Xin Wang, Yang Yao, Yan Teng, Xingjun Ma, Yingchun Wang, and Yu-Gang Jiang. 2025. Safevid: Toward safety aligned video large multimodal models.arXiv preprint arXiv:2505.11926(2025)

    Pith/arXiv arXiv 2025

  57. [57]

    Zixuan Wang, Jinghao Shi, Hanzhong Liang, Xiang Shen, Vera Wen, Zhiqian Chen, Yifan Wu, Zhixin Zhang, and Hongyu Xiong. 2025. Filter-And-Refine: A MLLM Based Cascade System for Industrial-Scale Video Content Moderation. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track). 873–880

    2025

  58. [58]

    Zheng Wei, Mingchen Li, Pu Zhang, Xinyu Liu, Huamin Qu, and Pan Hui. 2025. ContextAware: A Multi-Agent Framework for Detecting Harmful Image-Based Comments on Social Media. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI, Vol. 25

    2025

  59. [59]

    Umut Yildirim, Rohan Dutta, Burak Yildirim, and Atharva Vaidya. 2024. Experi- mentation in content moderation using RWKV.arXiv preprint arXiv:2409.03939 (2024)

    Pith/arXiv arXiv 2024

  60. [60]

    Kanwal Yousaf and Tabassam Nawaz. 2022. A deep learning-based approach for inappropriate content detection and classification of youtube videos.IEEE Access 10 (2022), 16283–16298

    2022

  61. [61]

    Jiaji Zhou, Jiabai Jin, Peng Zhuang, Dongdong Jia, and Chunli Wu. 2026. Semantic understanding and controllable generation methods of AIGC in virtual reality scene construction.Discover Computing29, 1 (2026), 94

    2026

  62. [62]

    Yiyang Zhou, Yangfan He, Yaofeng Su, Siwei Han, Joel Jang, Gedas Bertasius, Mohit Bansal, and Huaxiu Yao. 2025. Reagent-v: A reward-driven multi-agent framework for video understanding.arXiv preprint arXiv:2506.01300(2025). Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    Pith/arXiv arXiv 2025

  63. [2023]

    Malicious or Benign? Towards Effective Content Moderation for Children’s Videos.arXiv preprint arXiv:2305.15551(2023)

    Pith/arXiv arXiv 2023

  64. [2025]

    arXiv preprint arXiv:2504.20091(2025)

    Videomultiagents: A multi-agent framework for video question answering. arXiv preprint arXiv:2504.20091(2025)

    Pith/arXiv arXiv 2025