What is KuaiMod
KuaiMod is a short video quality assessment framework launched by Kuaishou, based on multimodal large models, capable of efficiently identifying and filtering harmful and low-quality content. The framework draws on the Common Law system and dynamically updates moderation policies in a case-driven manner, quickly adapting to the rapid changes in content on short video platforms. KuaiMod combines Vision-Language Models (VLM) and Chain-of-Thought (CoT) technology, and performs reinforcement learning based on user feedback to achieve precise content assessment. KuaiMod achieves an offline test accuracy of up to 92.4%, and in actual deploymentsignificantly reduces user report rates, improves user experience, and provides strong support for the healthy ecosystem of short video platforms.
Main features of KuaiMod
- Automatic moderation of harmful content: Identifies and filters violating videos such as violence, pornography, and rumors.
- Fine-grained content classification: Precisely classifies videos into different violation categories for ease of management.
- Dynamic policy updates: Adjusts moderation policies in real time based on user feedback to adapt to content changes.
- Reduced user report rates: Reduces users' exposure to harmful content and improves user experience.
- Supports personalized recommendations: Optimizes the recommendation system to push high-quality videos that match user interests.
Technical principles of KuaiMod
- Multimodal large model (VLM): Based on a visual encoder and a language model, it fuses the visual information of videos (such as video frames and covers) with textual information (such as titles, comments, and OCR/ASR text) to generate a comprehensive video representation. Based on large-scale pretraining, the VLM can understand the semantic content of videos and identify harmful information within them.
- Chain-of-Thought (CoT): Based on CoT technology, it generates a detailed reasoning process explaining why a certain video is judged as harmful or non-harmful. It decomposes the video moderation process into multiple states (such as content extraction, content analysis, and user feedback analysis) and reasons step by step to ultimately provide the moderation result.
- Offline adaptation and training: Uses large-scale annotated data to perform supervised fine-tuning of the VLM to adapt it to the video moderation task. Based on generated positive and negative sample pairs, it optimizes the model's preferences and improves the accuracy and consistency of moderation.
- Online updates and reinforcement learning: Reinforcement Learning from User Feedback (RLUF): Based on feedback such as user reports, likes, and comments, it constructs new training data and uses reinforcement learning to dynamically update moderation policies. The model is optimized based on real-time feedback to ensure rapid adaptation to changes in platform content.
- Dynamic classification system: Based on changes in platform content, it dynamically expands and adjusts the classification label system to ensure coverage of newly emerging types of harmful content.
KuaiMod project address
- Project official website：https://kuaimod.github.io/
- GitHub repository：https://github.com/KuaiMod/KuaiMod.github.io
- arXiv technical paper：https://arxiv.org/pdf/2504.14904v1
Application scenarios of KuaiMod
- Content moderation: Automatically identifies and filters harmful content, such as violence, pornography, and rumors.
- Recommendation optimization: Improves the quality of personalized recommendations and pushes higher-quality videos that better match user interests.
- User feedback processing: Dynamically adjusts moderation policies based on user reports and comments.
- Improve content quality: Filter low-quality content and optimize the overall content ecosystem of the platform.
- Maintain platform ecosystem: Adapt to content changes and ensure the healthy and sustainable development of the platform.