This paper proposes "Generative Interfaces for Language Models," a novel interaction method leveraging large-scale language models (LLMs). To overcome the limitations of traditional linear question-and-answer methods, LLMs generate user interfaces (UIs) in response to user queries, enabling more adaptive and interactive engagement. User queries are transformed into task-specific UIs through structured interface-specific representations and iterative refinement. We introduce a multidimensional evaluation framework that compares generative interfaces with traditional chat-based interfaces across a variety of tasks, interaction patterns, and question types, capturing functional, interactional, and emotional aspects of user experience. Experimental results show that generative interfaces consistently outperform conversational interfaces, with over 70% of users preferring generative interfaces. These findings clarify when and why users prefer generative interfaces and pave the way for future advancements in human-AI interaction.
