TidyBot is a real-world robot system that combines language-based planning and perception with the few-shot summarization capabilities of large language models to infer generalized user preferences. The system learns from a handful of examples how a particular person prefers objects to be organized, then applies those learned preferences to future interactions. The approach achieves 91.2% accuracy on unseen objects in benchmark testing and successfully puts away 85.0% of objects in real-world scenarios.