From Shallow Humor to Metaphor: Towards Label-Free Harmful Meme Detection via LMM Agent Self-Improvement

This article describes research focused on detecting harmful memes without relying on labeled data. The approach uses a Large Multimodal Model (LMM) agent that improves its detection capabilities through self-improvement. The title suggests a progression from simple humor understanding to more complex metaphorical analysis, which is crucial for identifying subtle forms of harmful content. The research area is relevant to current challenges in AI safety and content moderation.

• •Focuses on label-free harmful meme detection.
• •Utilizes a Large Multimodal Model (LMM) agent.
• •Employs self-improvement for enhanced detection.
• •Addresses the challenge of identifying subtle harmful content.