Literature reading is an indispensable process for understanding a research field, but it often involves a significant investment of time. To enable researchers to quickly understand the advancements in a particular research field, we leverage the writing capabilities of large language models (LLMs) and propose an end-to-end survey paper writing system that can complete a paper using only a few keywords. Specifically, we closely mimic the human writing process by (1) generating an outline from multiple perspectives, (2) selecting the best one from the generated candidate content, and (3) refining the selected content. Furthermore, to evaluate the generated content, we curate SurGen, a dataset of recent high-quality survey articles, and conduct tests. Experimental results demonstrate that our method significantly improves both automated metrics and human evaluations compared to direct generation. To the best of our knowledge, we are the first to attempt generating survey papers using large language models and to release the corresponding dataset.