PEST: Parameter Efficient Steering of Blackbox VLMs via Agentic Few-shot Alignment for Hateful Meme Moderation
A new framework uses task-specific generative AI agents to simultaneously classify, explain, and intervene in hateful memes. By generating 'silver training data' through three specialized agents—captioning, explanation generation, and intervention creation—the method overcomes data scarcity for end-to-end moderation. Results show large models like GPT-4o significantly outperform baselines across all three tasks.