Topic

Application Practices of Kuaishou’s Security Large Models within Its Content Governance System

In today’s era of explosive digital content growth, content security faces unprecedented challenges: massive volumes of information emerge in real time, violations evolve in unpredictable ways, and cross-scenario, cross-modal governance becomes increasingly complex. Traditional content moderation and rule-based engines can no longer meet the dual demands for “speed and accuracy,” leading to growing efficiency bottlenecks and risk blind spots. The advent of large language models has significantly advanced machines’ capabilities in multimodal understanding, active learning, and complex semantic reasoning, making them a key choice for breaking through these challenges. This talk will focus on content security for short video platforms and introduce Kuaishou’s in-house multimodal large-model solutions, showcasing how these technologies empower intelligent security governance across the full operational chain. Topics will include multimodal model training and optimization, security knowledge base construction with retrieval-augmented techniques, reasoning-based security moderation agents, and intelligent risk-control engines driven by multi-agent collaboration. The session aims to provide new insights into customizing vertical large models for specific business scenarios. Outline: Background and challenges of content security on short video platforms Core technologies and practical applications of Kuaishou’s security large models Multimodal large-model training and optimization Security knowledge base construction and retrieval augmentation Reasoning-based moderation agents for security review Intelligent risk-control engines with multi-agent collaboration Future trends in technology and business applications