UniGuard-Cascade: A Unified LLM-Driven Safety Scoring and Multi-Stage Audit Framework for Cross-Platform User Comments

Abstract

User-generated comments across social media, e-commerce platforms, and online forums pose increasing challenges for automated content safety management. Existing moderation systems typically target a single platform, rely on inconsistent labeling standards, and struggle to balance accuracy with operational cost. To address these limitations, we propose UniGuard-Cascade, a unified safety scoring and cascading moderation framework that integrates large language models (LLMs), lightweight classifiers, and retrieval-augmented verification. UniGuard-Cascade introduces a platform-agnostic unified safety taxonomy covering toxicity, hate speech, spam, sexual/NSFW content, and misinformation, generated through LLM-based label alignment and normalization. The system operates in a three-stage cascade: (1) Fast Path Screening using a DistilBERT-based multi-label classifier for low-cost filtering; (2) Slow Path LLM Examination that provides refined labels and natural-language explanations; and (3) RAG-Enhanced Misinformation Verification, retrieving external evidence to validate factuality. Experiments conducted on four real-world datasets-Twitter, Amazon Reviews, Reddit, and YouTube Comments-show that UniGuard-Cascade consistently outperforms existing moderation baselines, achieving a Micro-F1 of 0.91 and a ROC-AUC of 0.97. The framework further reduces LLM usage by 87 percent compared with single-stage GPT-4 moderation, yielding a 5.7×reduction in overall inference cost while maintaining state-of-the-art multi-platform safety performance.