The rise of large language models (LLMs) in education has sparked debate over how they should be used to foster learning. While prior research focuses on problem-solving domains, we investigate how students use LLMs in a seminar-style research methods course in which one of the main activities is critical analyses of research papers. Through assignment-level reports, we derive an emergent bottom-up categorization of LLMs usage by students, revealing diverse strategies from clarifying concepts to generating ideas and generating answers. Our findings highlight both LLMs productive uses and potential over-reliance, emphasizing the need for clear guidelines to handle with care the integration of LLMs in education (This work falls under and complies with University's policies 81 and LR11 regarding ownership and usage of course material for pedagogical quality assurance.)