Author-Informed Stance Detection: A Dataset and Findings

Reframing political stance detection as an author-informed task improves performance for both humans and LLMs, particularly on sarcastic content. Experiments show that fine-tuning LLMs to predict author political affiliation before stance detection yields better accuracy than standard fine-tuning methods using text alone.