PrismAgent: Zero-Shot Multi-Agent Framework for Harmful Meme Detection

A new research paper on arXiv (2605.02940) introduces PrismAgent, a zero-shot, multi-agent, interpretable framework designed to detect harmful content in memes. The framework conceptualizes the detection task as a criminal case investigation, employing four specialized agents: an analyst, an investigator, a prosecutor, and a judge. The analyst paraphrases memes under benevolent and malicious assumptions to probe underlying intent. The investigator retrieves supporting evidence from unannotated datasets and constructs contextual interpretations. The prosecutor then performs further analysis. This approach addresses limitations of existing methods that rely on high-volume annotated data, which incur substantial training costs and limited generalization. PrismAgent aims to curb the spread of misinformation by enabling effective identification of harmful memes without requiring annotated datasets.