UNIVID: Unified Vision-Language Model for Video Moderation

2026-06-04

AI Summary

This work addresses the dual challenges of multimodal fine-grained reasoning and decision interpretability in global video moderation. Current systems rely on numerous fragmented black-box models, which are difficult to maintain and lack transparency. To overcome these limitations, we propose UNIVID, a unified vision-language model that introduces policy-aligned, interpretable captions as an intermediate representation. This approach enables a single model to replace thousands of specialized classifiers while supporting human verification and multi-task reuse. By integrating expert-annotated data with synthetic examples and employing end-to-end vision-language modeling coupled with a refusal control mechanism, UNIVID significantly enhances alignment with safety policies. Experiments demonstrate a 42.7% reduction in harmful content漏检率 and a 37.0% decrease in false positives, alongside substantial savings in computational resources and engineering maintenance costs.

Abstract

Global-scale video moderation faces a dual challenge: the need for fine-grained multi-modal reasoning and the demand for interpretable outputs to support downstream enforcement. Traditional moderation systems often rely on fragmented black-box classifiers that are difficult to maintain and lack transparency. In this paper, we present UNIVID, a UNIfied VIsion-language model for video moDeration. Unlike standard classification models, UNIVID generates policy-aware captions that serve as an interpretable intermediate representation, enabling human-verifiable decisions and multi-task reusability. While existing open-source and commercial VLMs often suffer from safety-guardrail refusals and lack fine-grained policy alignment, we develop a specialized training data recipe that combines expert human-refined labels with synthetic data to align the model with our safety guidelines. By integrating UNIVID as the core captioner, we design a novel end-to-end video moderation system that reduces violation leakage by 42.7% and overkill rate by 37.0% relatively. Meanwhile, by replacing over 1,000 policy-specific models with a single UNIVID backbone, we recycled extensive computation resources while reducing engineering maintenance overhead. To our knowledge, this is one of the first reports of a high-efficiency captioning VLM successfully supporting industrial-scale moderation and cross-functional business.

Problem

• video moderation
• fine-grained multi-modal reasoning
• interpretable outputs
• black-box classifiers
• policy alignment

Innovation

• vision-language model
• interpretable captioning
• policy-aware moderation
• synthetic data alignment
• end-to-end video moderation