This study examines how seven cutting-edge LLMs respond to hate speech: LLaMA 2, Vicuna, LLaMA 3, Mistral, GPT-3.5, GPT-4, and Gemini Pro. The researchers want to demonstrate these models’ ability to process hate speech inputs by exposing the range of reactions these models generate through qualitative analysis. We also go over ways to reduce the production of hate speech by LLMs, especially through guardrailing guidelines and fine-tuning. Lastly, the researchers investigate how the models react to politically acceptable hate speech.
https://arxiv.org/abs/2410.00775