People often approach complex documents (e.g., academic papers and professional reports) with diverse goals and varying levels of prior knowledge. Even when reading the same document, the paths to consuming relevant information can differ significantly depending on the reader's information needs. However, the linear format and dense content of these documents make users manually sift through content, often leading to inefficient and narrow information consumption. To address this, we explore design principles that guide users along customized reading paths via questing answering. Applying these principles to guide academic paper reading, we introduce DocVoyager, a novel document-reading interface that adapts to users' goals by suggesting tailored, goal-based questions. DocVoyager leverages Large Language Models (LLMs) to anticipate users' information needs and dynamically suggests questions based on prior interactions. Our study found that participants easily focused on information relevant to their goals and engaged effectively with the document content using DocVoyager.