Elamur Research⌕
☰
home / papers / 2026 / UNIVID: Unified Vision-Language Model for Video Moderation

UNIVID: Unified Vision-Language Model for Video Moderation

Paper · 2026 · Kejuan Yang, Yizhuo Zhang, Mingyuan Du et al.

Summary

Short-video platforms usually moderate content with a large collection of separate classifiers, one per policy. These systems are hard to maintain, give opaque scores that auditors cannot check, and cannot be reused for other business needs. Off-the-shelf VLMs are not a good substitute: they often refuse to describe harmful material, are not aligned with internal policies, and are too costly to run on all traffic.

The authors (Bytedance) built UNIVID, a captioning VLM based on the LLaVA-OneVision design with Mistral-v0.3-7B as the language model. It produces policy-aware captions that describe a video and name any policy it may violate. Training data came from GPT-4o captions and VQA pairs, which trained annotators then corrected for factual accuracy and linked to internal policy rules. UNIVID powers a three-stage pipeline:

  • a Risk Filter for early screening;
  • a Moderation Actor made of UNIVID-Lite and UNIVID-RAG;
  • a Trend Governance module that detects emerging risks.

In production simulations, the system substantially reduced both missed violations and false enforcement. A single model replaced more than a thousand others, which shows that caption-centric moderation can be interpretable and efficient at industrial scale.

Key points

  • Training: three stages (pretraining, fine-tuning, continued fine-tuning) using about 1.6M, 5.2M (3.2M captions plus 2.0M VQA) and 0.1M samples. Training took 120 hours on 32 H100 GPUs. A lighter UNIVID-1B variant uses a proprietary decoder.
  • UNIVID-Lite: built on UNIVID-1B and fine-tuned on 1M in-house videos at a 1:5 positive-to-negative ratio. It generates a structured text output (Approve or Violation, plus the violated policy) instead of using a classification head.
  • CapBench: a new human-verified benchmark with 17,210 items, 11,476 of them violative. It scores violation recall by splitting captions into atomic events.
  • Results: violation leakage fell by 42.7% and the overkill rate by 37.0% (both relative). Over 1,000 policy-specific models were replaced, and the system reached 81% accuracy in a Brand & Ads simulation.
  • Release and limitations: code, weights and datasets are not released. The system uses no RL (e.g., GRPO), and because it samples frames, violations confined to a single frame can be missed.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksVision & language · 811 works

Cite

Cite as Kejuan Yang, Yizhuo Zhang, Mingyuan Du, Yue Zhang, Dixin Zheng, Kaili Zhao, Yang Xiao, Hanzhong Liang and Kenan Xiao. 2026. UNIVID: Unified Vision-Language Model for Video Moderation. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.32

@inproceedings{yang2026univid,
  title = {{UNIVID: Unified Vision-Language Model for Video Moderation}},
  author = {Kejuan Yang, Yizhuo Zhang, Mingyuan Du, Yue Zhang, Dixin Zheng, Kaili Zhao, Yang Xiao, Hanzhong Liang and Kenan Xiao},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.32},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.32}
}

References

41 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Xiang An, Yin Xie, Kaicheng Yang, Wenkang Zhang, Xiuwei Zhao, Zheng Cheng, Yirui Wang, Songcen Xu, Changrui Chen, Chunsheng Wu, Huajie Tan, Chunyuan Li, Jing Yang, Jie Yu, Xiyao Wang, Bin Qin, Yumeng Wang, Zizhen Yan, Ziyong Feng, and 3 others. 2025. Llava-onevision-1.5: Fully open framework for dem… find
  2. Han Bao, Qinying Wang, Zhi Chen, Qingming Li, Xuhong Zhang, Changjiang Li, Zonghui Wang, Shouling Ji, and Wenzhi Chen. 2025. Vmoda: An effective framework for adaptive nsfw image moderation. arXiv preprint arXiv:2505.23386. — [arXiv:2505.23386](https://arxiv.org/abs/2505.23386) find
  3. Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lac… find
  4. Vivian Lai, Samuel Carton, Rajat Bhatnagar, Vera Liao, Yunfeng Zhang, Chenhao Tan, and Q. Liao. 2022. Human-ai collaboration via conditional delegation: A case study of content moderation. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems. find
  5. DongGeon Lee, Joonwon Jang, Jihae Jeong, and Hwanjo Yu. 2025. Are vision-language models safe in the wild? a meme-based benchmark study. arXiv preprint arXiv:2505.15389. find
  6. Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. 2015. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the ieee conference on computer vision and pattern recognition, pages 961–970. find
  7. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  8. Vaishali U. Gongane, M. Munot, and A. Anuse. 2022. Detection and moderation of detrimental content on social media platforms: current status and future directions. Social Network Analysis and Mining, 12. find
  9. Robert Gorwa, Reuben Binns, and Christian Katzenbach. 2020. Algorithmic content moderation: Technical and political challenges in the automation of platform governance. Big Data & Society, 7. find
  10. Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904–6913.
  11. Tal Hassner, Yossi Itcher, and Orit Kliper-Gross. 2012. Violent flows: Real-time detection of violent crowd behavior. In 2012 IEEE computer society conference on computer vision and pattern recognition workshops, pages 1–6. IEEE. find
  12. Adi Levi, Or Levi, Sardhendu Mishra, and Jonathan Morra. 2025. Ai vs. human moderators: A comparative evaluation of multimodal llms in content moderation for brand safety. 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 6024–6032. find
  13. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. 2024a. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326.
  14. Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR.
  15. Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, and 1 others. 2024b. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195–22206.
  16. Hanzhong Liang, Jinghao Shi, Xiang Shen, Zixuan Wang, Vera Wen, Ardalan Mehrani, Zhiqian Chen, Yifan Wu, and Zhixin Zhang. 2025. Embedding-based retrieval in multi-modal content moderation. In Proceedings of the 48th International ACM SI-GIR Conference on Research and Development in Information Retr… find
  17. Guolei Huang, Qinzhi Peng, Gan Xu, Yuxuan Lu, and Yongjun Shen. 2025. Llavashield: Safeguarding multimodal multi-turn dialogues in vision-language models. Preprint, arXiv:2509.25896. — [arXiv:2509.25896](https://arxiv.org/abs/2509.25896) find
  18. Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. 2023. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122.
  19. Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700–6709.
  20. Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024a. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26296–26306.
  21. Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024b. Llava-next: Improved reasoning, ocr, and world knowledge. find
  22. Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. 2023. Eva-clip: Improved training techniques for clip at scale. Preprint, arXiv:2303.15389.
  23. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. Advances in neural information processing systems, 36:34892– 34916.
  24. Sahana Udupa, Antonis Maronikolakis, and Axel Wisiorek. 2023. Ethical scaling for content moderation: Extreme speech and the (in)significance of artificial intelligence. Big Data & Society, 10. find
  25. 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. — [arXiv:2505.11842](https://arxiv.org/abs/2505.11842) find
  26. Hexu Wang, Wenlong Luo, Wei Wu, Fei Xie, Jindong Liu, Jing Li, and Shizhou Zhang. 2025a. Vips: Learning-view-invariant feature for person search. Sensors, 25(17):5362. find
  27. Xingyu Lu, Tianke Zhang, Chang Meng, Xiaobei Wang, Jinpeng Wang, Yi-Fan Zhang, Shisong Tang, Changyi Liu, Haojie Ding, Kaiyu Jiang, and 1 others. 2025. Vlm as policy: Common-law content moderation framework for short video platform. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Disco… find
  28. Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. 2024. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024).
  29. Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195–3204. find
  30. Amna Naseeb, Muhammad Zain, Nisar Hussain, Amna Qasim, Fiaz Ahmad, Grigori Sidorov, and A. Gelbukh. 2025. Machine learning- and deep learning-based multi-model system for hate speech detection on facebook. Algorithms. find
  31. OpenAI. 2023. GPT-4 technical report. abs/2303.08774.
  32. CoRR, Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, and 1 others. 2023. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems,… find
  33. Jinghao Shi, Xiang Shen, Kaili Zhao, Xuedong Wang, Vera Wen, Zixuan Wang, Yifan Wu, and Zhixin Zhang. 2024. Cpfd: Confidence-aware privileged feature distillation for short video classification. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 48… find
  34. Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. 2024. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634. — [arXiv:2407.00634](https://arxiv.org/abs/2407.00634) find
  35. Mo Wang, Kaixuan Ren, Pratik Jalan, Ahmed Ashraf, Tuong Vy Vu, Rahul Seetharaman, Shah Nawaz, and Usman Naseem. 2026. From native memes to global moderation: Cros-cultural evaluation of vision-language models for hateful meme detection. arXiv preprint arXiv:2602.07497. — [arXiv:2602.07497](https://arxiv.org/abs/2602.07497) find
  36. Wenxuan Wang, Jingyuan Huang, Chang Chen, Jiazhen Gu, Jianping Zhang, Weibin Wu, Pinjia He, and Michael R. Lyu. 2023. Validating multimedia content moderation software via semantic fusion. Proceedings of the 32nd ACM SIGSOFT International Symposium on Software Testing and Analysis. find
  37. Zixuan Wang, Jinghao Shi, Hanzhong Liang, Xiang Shen, Vera Wen, Zhiqian Chen, Yifan Wu, Zhixin Zhang, and Hongyu Xiong. 2025b. Filter-and-refine: A mllm based cascade system for industrial-scale video content moderation. In Proceedings of the 63rd Annual Meeting of the Association for Computational… find
  38. Zixuan Wang, Yu Sun, Hongwei Wang, Baoyu Jing, Xiang Shen, Xin Luna Dong, Zhuolin Hao, Hongyu Xiong, and Yang Song. 2025c. Reasoning-enhanced domain-adaptive pretraining of multimodal large language models for short video content governance. In Proceedings of the 2025 Conference on Empirical Methods… find
  39. Kanwal Yousaf and Tabassam Nawaz. 2022. A deep learning-based approach for inappropriate content detection and classification of youtube videos. IEEE Access, 10:16283–16298. find
  40. Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
  41. Waqas Sultani, Chen Chen, and Mubarak Shah. 2018. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479–6488. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-univid-unified-vision-language-model/