analiza aiqbench
This production-scale hybrid framework effectively combines supervised classification with MLLM-boosted similarity matching to address both known violations and novel edge cases in livestreaming. By distilling multimodal knowledge into lightweight inference pipelines, the system achieves a 6-8% reduction in unwanted content views while maintaining robust precision. The approach offers a scalable blueprint for content governance that adapts to evolving adversarial behaviors without sacrificing inference efficiency.
Dynamic Content Moderation in Livestreams: Combining Supervised Classification with MLLM-Boosted Similarity Matching
arXiv:2512.03553v3 Announce Type: replace-cross Abstract: Content moderation remains a critical yet challenging task for large-scale user-generated video platforms, especially in livestreaming environments where moderation must be timely, multimodal, and robust to evolving forms of unwanted content. We present a hybrid moderation framework deployed at production scale that combines supervised classification for known violations with reference-based similarity matching for novel or subtle cases. This hybrid design enables robust detection of both explicit violations and novel edge cases that evade traditional classifiers. Multimodal inputs (text, audio, visual) are processed through both pipelines, with a multimodal large language model (MLLM) distilling knowledge into each to boost accuracy while keeping inference lightweight. In production, the classification pipeline achieves 67% recall at 80% precision, and the similarity pipeline achieves 76% recall at 80% precision. Large-scale A/B tests show a 6-8% reduction in user views of unwanted livestreams}. These results demonstrate a scalable and adaptable approach to multimodal content governance, capable of addressing both explicit violations and emerging adversarial behaviors.
Abstract:Content moderation remains a critical yet challenging task for large-scale user-generated video platforms, especially in livestreaming environments where moderation must be timely, multimodal, and robust to evolving forms of unwanted content. We present a hybrid moderation framework deployed at production scale that combines supervised classification for known violations with reference-based similarity matching for novel or subtle cases. This hybrid design enables robust detection of both explicit violations and novel edge cases that evade traditional classifiers. Multimodal inputs (text, audio, visual) are processed through both pipelines, with a multimodal large language model (MLLM) distilling knowledge into each to boost accuracy while keeping inference lightweight. In production, the classification pipeline achieves 67% recall at 80% precision, and the similarity pipeline achieves 76% recall at 80% precision. Large-scale A/B tests show a 6-8% reduction in user views of unwanted livestreams}. These results demonstrate a scalable and adaptable approach to multimodal content governance, capable of addressing both explicit violations and emerging adversarial behaviors.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2512.03553 [cs.CV]
(or arXiv:2512.03553v3 [cs.CV] for this version)
https://doi.org/10.1145/33770854.3783936 DOI(s) linking to related resources
Text integral sub licența deschisă a publicației · 182 cuvinte
Subiecte conexe
Can Jev Be a Better Agent Evaluator?
LangChain Blog · 21 sept. 2026
Twitch will let parents stop their teens going live
The Verge · 21 iul. 2026
AI gets keys to unmapped data: theCUBE’s AI Cybersecurity and Resilience Summit day-two kickoff analysis
SiliconANGLE · 28 oct. 2026
Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week
Latent Space · 30 sept. 2026