Too Long; Didn't Read

This study explores using pre-trained vision-language models (PVLMs) for hateful meme detection without fine-tuning, showcasing BERT's superiority and introducing PromptHate with probe-captioning. Limitations include heuristic probing question usage, suggesting future directions for optimization and deeper interpretation using gradient-based approaches.