QwenSafe: Multimodal Content Rating Description Identification via Preference-Aligned VLMs

2026-05-19

AI Summary

Mobile applications are required to disclose Apple-defined Content Rating Descriptors (CRDs), yet their multimodal nature poses significant challenges for accurate identification. This work proposes QwenSafe, a novel approach built upon Qwen3-VL-8B that introduces a descriptor-aware multimodal alignment mechanism. By integrating supervised fine-tuning with Direct Preference Optimization (DPO) and leveraging a proprietary metadata2CRD pipeline to synthesize aligned question-answer pairs for training, QwenSafe achieves substantial performance gains. Evaluated across 12 CRD categories, it improves positive-class recall by 111.8%, 36.1%, and 2.1% over Qwen3-VL, LLaVA-1.6, and Gemini-2.5-Flash, respectively, demonstrating clear superiority over existing models.

Abstract

Mobile app marketplaces require developers to disclose standardized content rating descriptors (CRDs) to inform users about potentially sensitive or restricted content. Ensuring the accuracy and consistency of these disclosures remains challenging due to the multimodal nature of app content, which spans textual descriptions and visual interfaces. In this paper, we present QwenSafe, a Vision-Language Model (VLM) designed to automatically identify the presence of Apple-defined CRDs by jointly reasoning over app metadata and screenshots. To enable scalable training for this task, we introduce metadata2CRD, a data-construction pipeline that synthesizes descriptor-aligned question-answer pairs by combining app descriptions, screenshots, and formal descriptor definitions. We adapt Qwen3-VL-8B using supervised fine-tuning followed by Direct Preference Optimization (DPO) to align model predictions with descriptor-specific evidence and explanations across visual and textual modalities. We evaluate QwenSafe on 12 Apple-defined content rating descriptors and compare it against state-of-the-art vision-language models, including Qwen3-VL, LLaVA-1.6, and Gemini-2.5-Flash. QwenSafe consistently outperforms all baselines in binary CRD classification, achieving improvements in positive-class recall of 111.8%, 36.1%, and 2.1%, respectively. Our results demonstrate that descriptor-aware multimodal alignment substantially improves automated content classification and highlights the potential of vision-language models to support scalable and consistent content rating in mobile app marketplaces.

Problem

• content rating descriptors
• multimodal content
• mobile app marketplaces
• accuracy and consistency
• vision-language models

Innovation

• Vision-Language Model
• Content Rating Descriptor
• Direct Preference Optimization
• Multimodal Alignment
• metadata2CRD