Title: Enhancing Livestream Content Moderation: A Hybrid Approach Merging Supervised Classification with MLLM-Enhanced Similarity Detection

Abstract: For massive user-generated video platforms, particularly within livestreaming sectors, content moderation stands as a vital but complex challenge. Moderators must operate in real-time, handle multimodal data, and remain resilient against shifting tactics of harmful content. This paper introduces a hybrid moderation system implemented at a production scale, which integrates supervised classification for established policy breaches alongside reference-based similarity matching to identify new or nuanced infractions. This dual-strategy architecture ensures the robust identification of both clear-cut violations and novel edge cases that typically bypass standard classifiers. By processing multimodal inputs—including text, audio, and visual data—through both pipelines, the system leverages a multimodal large language model (MLLM) to distill knowledge, thereby enhancing accuracy without compromising inference speed. In live production environments, the classification pipeline records a recall rate of 67% at 80% precision, while the similarity pipeline reaches 76% recall at the same precision threshold. Extensive A/B testing indicates a 6-8% decrease in user exposure to undesirable livestreams. These findings highlight a scalable and flexible framework for multimodal content governance, effectively managing both explicit violations and emerging adversarial strategies.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC