Automatically finding and synthesizing information from rich textual sources can support a wide range of use cases across work, education, and personal use. Artificial intelligence systems are already assisting users to fulfill their information needs, from providing encyclopedic facts to answering complex questions that require multiple steps of reasoning. Despite these triumphs, such systems often provide incorrect or outdated information while sounding plausible and authoritative. Furthermore, compared to conversing with domain experts who can answer our questions, interaction with current systems is limited. Instead of engaging in multi-turn interaction with users, asking clarifying questions or follow-up questions, systems mostly take a passive role, aiming to provide accurate information at once. This project envisions interactive systems that critically reason about textual sources to provide high-quality, up-to-date information. This research will advance how language systems interface with rich knowledge sources: parametric knowledge acquired during the language model (LM)’s massive pretraining, documents prepended at inference time, and users who can provide context for their initial input query. The devised systems will model the complexities of real-world scenarios, where users' questions are ambiguous, answers continuously change based on the context of the interaction, and heterogeneous knowledge sources contain imperfect and outdated information. It will