The increasing complexity of cyber-physical systems, such as modern automotive E/E architectures, demands efficient, consistent, and scalable modeling processes. Traditional manual modeling of SysML diagrams from natural-language requirements is time-consuming and error-prone. We target to assist E/E architects in the process by utilizing Large Language Models (LLMs). Therefore, we designed a three-step approach for automated generation of SysML activity diagrams. First, the approach identifies diagram element candidates in the requirements and matches them to potential existing elements. Second, missing elements are generated, and finally, the elements are linked. Evaluated on real-world system-level requirements documents, comprising more than 720 requirements and 72 use cases, our approach generates system-level diagrams comparable to those created by experts and provides valuable architectural recommendations. A survey with nine industry architects shows that at the system-level LLM-generated diagrams were even rated higher than the diagrams modeled by experts. However, quality degrades at the subsystem-level, where human refinement remains essential for correctness and completeness.