Reproduce “BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards” on OpenResearch

About this paper

Yiran Yang, Zhaowei Liu, Yuan Yuan, Yukun Song, Xiong Ma, Yinghao Song, Xiangji Zeng, Lu Sun and 5 more

Researchers from Kuaishou Technology, Beijing University of Posts and Telecommunications, and Shandong University developed BLM-Guard, an explainable multimodal framework for moderating short-video advertisements. The system integrates Chain-of-Thought reasoning with policy-aligned reinforcement learning, outperforming state-of-the-art baselines with a 0.914 strict accuracy and 0.845 reasoning consistency on a new benchmark while reducing annotation costs.

Key insights

• Explicitly integrating policy principles into Chain-of-Thought (CoT) reasoning, guided by a rule-anchored prior, is effective for achieving explainable and policy-compliant moderation decisions.
• Dynamic, principle-guided feedback through a Self-Consistency and Adaptive Reward (SCA-R) mechanism is crucial for refining reasoning quality and adapting to evolving moderation policies in an RL framework.
• Rule-driven Interleaved-modal Chain-of-Thought (ICoT) data synthesis can generate high-quality, structured multimodal moderation data efficiently, mitigating the high costs of manual annotation.

Summary by alphaXiv.