KuaiMod - Kuaishou's automated short video quality assessment framework

What is KuaiMod?

KuaiMod is a short video quality assessment framework launched by Kuaishou, based on a multimodal large model. It efficiently identifies and filters harmful and low-quality content. The framework draws inspiration from the common law system, dynamically updating its review strategy based on a case-driven approach to quickly adapt to the rapid changes in content on short video platforms. KuaiMod combines Visual Language Modeling (VLM) and Chain-of-Thought (CoT) technologies, using reinforcement learning based on user feedback to achieve accurate content assessment. KuaiMod achieves an offline testing accuracy of up to 92.4%, and in actual deployment...It significantly reduces user complaint rates, improves user experience, and provides strong support for the healthy ecosystem of short video platforms.

KuaiMod's main functions

• Automatic review of harmful contentIt identifies and filters videos containing violence, pornography, rumors, and other illegal content.
• Fine-grained content categorizationThe video is accurately categorized into different violation categories for easier management.
• Dynamic strategy updateWe adjust our review strategy in real time based on user feedback to adapt to changes in content.
• Reduce user complaint rateReduce users' exposure to inappropriate content and improve user experience.
• Enabling personalized recommendationsOptimize the recommendation system to push high-quality videos that match users' interests.

KuaiMod's technical principles

• Multimodal Large Model (VLM)Based on a visual encoder and language model, VLM fuses visual information (such as video frames and cover images) and textual information (such as titles, comments, and OCR/ASR text) from a video to generate a comprehensive video representation. Based on large-scale pre-training, 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 to explain why a video is judged as harmful or not harmful. The video review process is broken down into multiple states (such as content extraction, content analysis, and user feedback analysis), and the final review result is given through step-by-step reasoning.
• Offline adaptation and trainingWe use large-scale labeled data to supervise and fine-tune the VLM (Video Management Model) to adapt it to video review tasks. Based on the generation of positive and negative sample pairs, we optimize the model's preferences to improve the accuracy and consistency of review.
• Online updates and reinforcement learningUser Feedback-Based Reinforcement Learning (RLUF): Based on user feedback such as reports, likes, and comments, new training data is constructed, and the moderation strategy is dynamically updated using reinforcement learning. The model is optimized based on real-time feedback to ensure rapid adaptation to changes in platform content.
• Dynamic classification systemBased on changes in platform content, dynamically expand and adjust the classification and tagging system to ensure coverage of newly emerging types of harmful content.

KuaiMod's 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 moderationAutomatically identifies and filters harmful content, such as violence, pornography, and rumors.
• Recommended optimizationImprove the quality of personalized recommendations and push high-quality videos that better match users' interests.
• User feedback handlingThe review strategy is dynamically adjusted based on user reports and comments.
• Improve content qualityFilter low-quality content and optimize the overall content ecosystem of the platform.
• Maintaining the platform ecosystemAdapt to content changes and ensure the platform's healthy and sustainable development.