Beyond a Joke: Multi-Angle Reasoning for Detecting and Explaining Harmful Humor in Memes

Explain Like I'm Normal

Researchers developed MAR-12, a new framework designed to help Vision Language Models (VLMs) understand the complex intersection of humor and hate speech in memes. By using multi-angle reasoning, the system doesn't just flag content but actually explains the cultural and textual context behind offensive jokes. This represents a significant step in multimodal reasoning where models must distinguish between irony, satire, and genuine harm.