Filter-And-Refine: A MLLM Based Cascade System for Industrial-Scale Video Content Moderation

TL;DR

This paper introduces a cascade system combining multimodal large language models with a lightweight router to improve industrial-scale video content moderation, achieving high accuracy and efficiency with minimal data and computational resources.

Contribution

The authors propose a novel cascade system that transforms generative MLLMs into classifiers and integrates them with a lightweight router for scalable, cost-effective video moderation.

Findings

F1 score improved by 66.50% over traditional classifiers

System increases moderation volume by 41%

Reduces computational cost to 1.5% of full deployment

Abstract

Effective content moderation is essential for video platforms to safeguard user experience and uphold community standards. While traditional video classification models effectively handle well-defined moderation tasks, they struggle with complicated scenarios such as implicit harmful content and contextual ambiguity. Multimodal large language models (MLLMs) offer a promising solution to these limitations with their superior cross-modal reasoning and contextual understanding. However, two key challenges hinder their industrial adoption. First, the high computational cost of MLLMs makes full-scale deployment impractical. Second, adapting generative models for discriminative classification remains an open research problem. In this paper, we first introduce an efficient method to transform a generative MLLM into a multimodal classifier using minimal discriminative training data. To enable industry-scale deployment, we then propose a router-ranking cascade system that integrates MLLMs with a lightweight router model. Offline experiments demonstrate that our MLLM-based approach improves F1 score by 66.50% over traditional classifiers while requiring only 2% of the fine-tuning data. Online evaluations show that our system increases automatic content moderation volume by 41%, while the cascading deployment reduces computational cost to only 1.5% of direct full-scale deployment.