Researchers have developed a new approach to creating 'digital twins' - AI models that simulate human behavior. Instead of using raw transcripts or summaries of responses, they've created a structured way of organizing persona information. This involves categorizing background, decision-making processes, and evaluation criteria into a framework called BDE. The team tested this method on various tasks and found it improved accuracy by 1.91 percentage points compared to using raw transcripts. However, the fixed structure didn't generalize well across different tasks. To address this, they proposed an automatic pipeline that refines task-specific structures and extraction prompts. This approach restored performance, improving mean accuracy by 1.91 percentage points over the raw transcript baseline. --- Why it matters: This research matters to engineers working on AI models because it highlights the importance of structuring persona information in digital twins. The optimal structure depends on the specific task, and this study provides a framework for automatically discovering and refining these structures. Source: https://arxiv.org/abs/2608.20344