Yale NLP/LLM Interest Group Talks
BioCiteGPT: Recurrent Retrieval-Augmented LLMs for Faithful Biomedical Citation Recommendation
In the biomedical domain, finding relevant information in vast scientific literature is challenging, making citation recommendation crucial. Large language models (LLMs) like ChatGPT and GPT-4offer a significant advance by generating literature recommendations without labeled datasets, but they face issues with generating unfaithful citations. Additionally, current tools lack the ability to provide precise, sentence- or paragraph-specific recommendations, highlighting the need for more granular solutions in biomedical research. To address the above limitations, we introduce the first sentence-level biomedical citation recommendation method BioCiteGPT based on retrieval-augmented LLMs, providing reliable citations for input sentences. To support this, we created the first open-source data suite with over 255K articles relevant to Alzheimer's disease and 528K sentences from these articles, each paired with positive citations and 15contrastive negatives, providing a comprehensive resource for model developing and testing. Our experiments show that this approach significantly improves the reliability and quality of LLM-generated citation recommendations.