Structured persona extraction improves LLM digital twin accuracy by 1.91 percentage points in controlled tests

A new arXiv preprint introduces a structured approach to organizing persona information for LLM-based digital twins, arguing that the primary constraint is not how much information is provided but how it is structured.

The authors propose a hand-crafted schema called BDE—Background, Decision procedure, Evaluation—grounded in consumer-behavior theory. On a homogeneous benchmark (Twin-2K-500), this structured representation improved predictive accuracy by 1.91 percentage points over raw transcripts.

The fixed BDE schema did not generalize to more heterogeneous tasks, where performance was statistically indistinguishable from the raw transcript baseline. To address this, the researchers developed an automatic structure-discovery pipeline in which an LLM iteratively proposes and refines task-specific persona structures and extraction prompts.

On a benchmark of 13 diverse sub-studies, the automatic pipeline restored performance, improving mean accuracy by 1.91 percentage points over the raw transcript baseline and eliminating significant losses observed with the fixed schema.

The paper also reports similar gains on robustness checks using gpt-5.4-mini and Qwen3-8B, suggesting the findings are not limited to a single model family.