Reasoning-Enhanced Domain-Adaptive Pretraining of Multimodal Large Language Models for Short Video Content Moderation

2025-09-25

AI Summary

Short-video platforms face significant challenges in content safety, including high annotation costs and poor cross-category generalization. To address these issues, we propose a unified multimodal large language model (MLLM) pretraining framework specifically designed for inappropriate content detection. Our method innovatively integrates three synergistic pretraining stages—caption generation, visual question answering, and chain-of-thought reasoning—to jointly enhance visual perception, semantic understanding, and logical reasoning capabilities. Leveraging domain-adaptive pretraining, vision-language alignment, instruction tuning, and chain-based reasoning, the framework enables end-to-end violation classification. Experimental results demonstrate substantial improvements over strong baselines under both zero-shot and supervised settings, with particularly robust generalization to unseen violation categories. This work provides an efficient, scalable, and unified solution for short-video content safety governance.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Safety and Robustness

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Short video platforms are evolving rapidly, making the identification of inappropriate content increasingly critical. Existing approaches typically train separate and small classification models for each type of issue, which requires extensive human-labeled data and lacks cross-issue generalization. We propose a reasoning-enhanced multimodal large language model (MLLM) pretraining paradigm for unified inappropriate content detection. To address the distribution gap between short video content and the original pretraining data of MLLMs, as well as the complex issue definitions, we introduce three targeted pretraining tasks: (1) extit{Caption}, to enhance the MLLM's perception of video details; (2) extit{Visual Question Answering (VQA)}, to deepen the MLLM's understanding of issue definitions and annotation guidelines; (3) extit{Chain-of-Thought (CoT)}, to enhance the MLLM's reasoning capability. Experimental results show that our pretraining approach significantly improves the MLLM's performance in both zero-shot and supervised fine-tuning (SFT) settings. In addition, our pretrained model demonstrates strong generalization capabilities to emergent, previously unseen issues.

Problem

• Detecting inappropriate content on rapidly evolving short video platforms
• Overcoming distribution gaps between short videos and pretraining data
• Addressing complex issue definitions requiring enhanced reasoning capabilities

Innovation

• Reasoning-enhanced MLLM pretraining for content moderation
• Domain-adaptive pretraining with three specialized tasks
• Caption, VQA, and Chain-of-Thought enhance video understanding