Imagine sifting through mountains of medical research to find the exact evidence needed for a patient's unique cancer case. Molecular Tumor Boards face this challenge daily, as genomic sequencing reveals a growing array of patient-specific genetic alterations. This is where the CIViC knowledgebase comes in, connecting genomic variants, cancer types, and treatments with levels of clinical evidence. But labeling this data is a laborious, manual task. Could AI automate this process? Researchers explored this by using Large Language Models (LLMs) to automatically assign CIViC evidence levels to medical paper abstracts. They experimented with various LLMs, including BERT and RoBERTa, fine-tuning them on a dataset derived from CIViC. The results were promising. The LLMs, especially those pre-trained on biomedical texts (like BiomedBERT and BioLinkBERT), outperformed a traditional machine learning model based on tf-idf scores. The team even developed a specialized LLM, "Biomed-RoBERTa-Long," to handle the length of many medical abstracts. However, the research also highlighted some challenges. Not all abstracts contain the necessary information for accurate classification. Sometimes, the crucial details are buried within the full text of the article, not the abstract. The study found that readily available LLMs like OpenAI's GPT-4, while powerful, couldn't match the fine-tuned models without specialized prompting or further training. Additionally, performance varied across different evidence levels. The models struggled most with less common levels due to limited training data. This research reveals the exciting potential of LLMs in biomedical knowledge curation. Automating evidence labeling could save significant time and resources, accelerating the creation of comprehensive, up-to-date knowledgebases. Future research may focus on incorporating full-text information and developing smarter prompting strategies to boost LLM performance, ultimately aiding Molecular Tumor Boards and advancing precision oncology.