Can LLMs Beat Humans on Discerning Human-written and LLM-generated Science News?
The paper introduces SANews, a manually annotated dataset of paired human-written and GPT-3.5-generated science news articles from ScienceAlert, and shows that a Guided Few-shot prompting template with a single example boosts LLMs (including open-weight LLaMA-3 70b) to match or exceed graduate-student performance in discerning their origin.Powered by Seed HypermediaOpen App