Given unprecedented growth in the scientific literature, researchers face an urgent challenge: How can we navigate, synthesize, and build upon vast and complex bodies of work? Large language models may accelerate discovery by improving our ability to retrieve, evaluate, and map scientific knowledge. In this talk, I will introduce a framework designed to address critical bottlenecks in understanding existing bodies of scholarly research. This includes new benchmarks for reasoning-based retrieval of scientific papers, evidence extraction against specific hypotheses, and quality assessment of study methodologies. I will also highlight how LLMs can be adapted to specialized domains like climate science and improved through domain-specific finetuning. Finally, I will discuss an ambitious effort to construct a ?map? of the biomedical literature -- a network of hypotheses, findings, and supporting evidence across the scientific record. Together, these approaches open new possibilities for scientists to more effectively understand and navigate the scientific literature.