BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards

TL;DR

BLM-Guard is an explainable multimodal ad moderation framework that combines Chain-of-Thought reasoning, rule-based policy guidance, and reinforcement learning to improve detection accuracy and robustness in short-video ads.

Contribution

It introduces a novel rule-driven data synthesis pipeline and a reinforcement learning approach for policy-aligned, explainable multimodal ad moderation.

Findings

Outperforms strong baselines in accuracy and consistency

Enhances robustness to intra-modal manipulations and cross-modal mismatches

Demonstrates effective generalization on real-world short-video ads

Abstract

Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle-speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.