Large Language Models for Citation Function Classification

This study presents a comprehensive evaluation of state-of-the-art LLMs for citation function classification, achieving a new SOTA macro F1 score of 73.3% on the ACL-ARC dataset using a fine-tuned Falcon 7B. We systematically compare five models—Mistral, Orca 2, LLaMA 3.1, Falcon, and SciBERT—across zero-shot, few-shot, and fine-tuning scenarios. Additionally, we introduce AC^3, a novel dataset with a seven-category scheme that distinguishes between neutral acknowledgments and explicit evaluative stances. Our analysis evaluates four context extraction variants and provides experimental insights to address the current research gap in model comparison for citation analysis.